ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-15 18:42:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

[Block Scale GEMM] Optimized block scale gemm (#1950)

Browse Source 
* Added two kernel for M=32 problem

* Comment the first one

* Enable multiply_multiply for Scale_Block_M = 1 for deepseek

* Modify the a_thread offset since the A data load is different from B.

* edit fp8 ab scale for Scale_Block_M=1

* edit GemmSpec to MNKPadding

* enable blockwise pipelie v1 and v2. v1 is work for small K.

* add instance for gemm_ab_scale

* fix cmakelist of ckProfiler

* optimize blockscale gemm. todo: reduce vgpr usage

* fix a correctness bug

* sanity checked

* revert ckprofiler cmake changes

* clang format

* revert unnecessary changes.

* remove commented codes.

* split weight preshuffle library targets

* bring back enable-post-misched=0

* fix build issues for gemm_multiply_multiply_fp8 instances

* fix clang format

* add verbose build flag when building for all targets

* reduce path names for new instances

* fix paths in cmake

* refactor gemm_multiply_multiply library target

* fix a bug in example

* fix example 65 cmake

* reduce the number of threads when building libs for all targets to 50

* use ninja to build for all targets

* reduce teh number of threads when building for all targets

* reduce the number of threads to 32 when building libs for all targets to 50

---------

Co-authored-by: mtgu0705 <mtgu@amd.com>
Co-authored-by: chenjun <junchen2@amd.com>
Co-authored-by: illsilin <Illia.Silin@amd.com>
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>

[ROCm/composable_kernel commit: cbd74c2d12]
This commit is contained in:
Haocong WANG
2025-03-12 01:11:21 +08:00
committed by GitHub
parent 6357ab039d
commit b31ac9d14d
75 changed files with 2997 additions and 1242 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
Jenkinsfile vendored
View File

@@ -1145,11 +1145,11 @@ pipeline {
}
agent{ label rocmnode("gfx90a") }
environment{
execute_args = """ cmake -D CMAKE_PREFIX_PATH=/opt/rocm \
execute_args = """ cmake -G Ninja -D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_CXX_COMPILER="${build_compiler()}" \
-D CMAKE_BUILD_TYPE=Release \
-D GPU_ARCHS="gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102" \
-D CMAKE_CXX_FLAGS=" -O3 " .. && make -j64 """
-D CMAKE_CXX_FLAGS=" -O3 " .. && ninja -j32 """
}
steps{
buildHipClangJobAndReboot(setup_cmd: "", build_cmd: "", no_reboot:true, build_type: 'Release', execute_cmd: execute_args)

72
example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp
View File

@@ -55,7 +55,7 @@ using CDEElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr ck::index_t Scale_Block_M = 128;
static constexpr ck::index_t Scale_Block_M = 1;
static constexpr ck::index_t Scale_Block_N = 128;
static constexpr ck::index_t Scale_Block_K = 128;
@@ -65,14 +65,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_ABScale_
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec,
256, Scale_Block_M, Scale_Block_N, Scale_Block_K,
128, 128,
128, 16, 16,
16, 128,
256, 16, 16,
16, 16,
4, 4,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
1, 2, S<1, 32, 1, 8>, S<8, 8, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>;
1, 2,
S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
1, 2, S<1, 16, 1, 16>, S<8>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>;
// clang-format on
int main(int argc, char* argv[])
@@ -80,11 +80,12 @@ int main(int argc, char* argv[])
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
bool flush_cache = true;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t M = 128;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = K;
ck::index_t StrideB = K;
@@ -100,7 +101,7 @@ int main(int argc, char* argv[])
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
else if(argc == 8)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
@@ -110,16 +111,19 @@ int main(int argc, char* argv[])
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideE = std::stoi(argv[9]);
flush_cache = std::stoi(argv[7]);
StrideA = K;
StrideB = K;
StrideE = N;
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n");
printf("arg4 to 6: M, N, K\n");
printf("arg7: flush both I$ and L2$ (0=no, 1=yes)\n");
exit(0);
}
@@ -182,9 +186,15 @@ int main(int argc, char* argv[])
b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 4:
a0_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
b0_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 5:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
break;
default:
@@ -194,6 +204,16 @@ int main(int argc, char* argv[])
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
}
#endif
#if 0
for(int im =0; im< (M + Scale_Block_M - 1) / Scale_Block_M; im++){
float row_sum = .0;
for(int ik =0; ik< (K + Scale_Block_K - 1) / Scale_Block_K; ik++){
printf("%lf ",a1_m_k(im, ik));
row_sum += a1_m_k(im, ik);
}
printf("sum: %lf\n", row_sum * 128);
}
#endif
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(A1DataType) * a1_m_k.mDesc.GetElementSpaceSize());
@@ -239,12 +259,24 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 20, 50});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
float ave_time = .0;
if(flush_cache)
{
int rotating_buf = (512 * 1024 * 1024 + num_btype - 1) / num_btype;
ave_time = invoker.Run(argument,
StreamConfig{nullptr, time_kernel, 0, 50, 100, true, rotating_buf});
}
else
{
ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 100});
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;

10
example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp
View File

@@ -140,14 +140,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShu
// clang-format off
< Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec, 256,
256, 256, 128,
128, 128, 128,
16, 16,
16, 16,
8, 8,
32, 32,
2, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
1, 2, S<1, 32, 1, 8>, S<8, 8, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>;
1, 1, S<1, 32, 1, 8>, S<8, 8, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>;
// clang-format on
int main(int argc, char* argv[])

2
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp
View File

@@ -453,7 +453,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1<BlockGemmPipelineScheduler::I
// latency
// __builtin_amdgcn_sched_barrier(0);
}
else
else if constexpr(TailNum == TailNumber::Odd)
{
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {

2
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp
View File

@@ -784,7 +784,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::I
// latency
// __builtin_amdgcn_sched_barrier(0);
}
else
else if constexpr(TailNum == TailNumber::Odd)
{
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {

615
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp
View File

@@ -7,10 +7,10 @@
namespace ck {
// Naive pipeline with lowest resource request per WGP
// GlobalPrefetchStages: 1
// Compute optimized pipeline
// GlobalPrefetchStages: 2
// LocalPreFillStages: 1
// LocalPreFetchStages: 0
// LocalPreFetchStages: 1
// LocalSharedMemoryBuffer: 1
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
@@ -96,7 +96,8 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
NPerXDL,
MRepeat,
NRepeat,
KPack>
KPack,
true>
{
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
@@ -117,10 +118,15 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
NPerXDL,
MRepeat,
NRepeat,
KPack>;
KPack,
true>;
using Base::A_K1;
using Base::B_K1;
using Base::I0;
using Base::I1;
using Base::KRepeat;
using Base::xdlops_gemm;
using typename Base::HotLoopInstList;
using Base::CalculateCThreadOriginDataIndex;
using Base::CalculateCThreadOriginDataIndex8D;
@@ -131,19 +137,43 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetWaveIdx;
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::a_block_desc_m0_m1_m2_k;
using Base::b_block_desc_n0_n1_n2_k;
using Base::AMmaKStride;
using Base::BMmaKStride;
static constexpr index_t AMmaKStride = xdlops_gemm.K0PerXdlops * KPack;
static constexpr index_t BMmaKStride = xdlops_gemm.K0PerXdlops * KPack;
static constexpr index_t PrefetchStages = 1;
static constexpr index_t PrefetchStages = 2;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = 1;
// Force mfma not cross the scaleblock
__device__ static auto CalculateAThreadOriginDataIndex()
{
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPack * xdlops_a_idx[I0]);
}
__device__ static auto CalculateBThreadOriginDataIndex()
{
const auto wave_idx = GetWaveIdx();
const auto waveId_n = wave_idx[I1];
const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPack * xdlops_b_idx[I0]);
}
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
@@ -151,11 +181,116 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
ignore = num_loop;
return TailNumber::Full;
return num_loop == 1 ? TailNumber::Odd : TailNumber::Full;
}
__device__ static constexpr auto HotLoopScheduler()
{
// A/B split schedule
// compiler is likely to use ds_read2 when instruction width smaller than 16bytes
constexpr auto num_ds_read_inst_a =
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16
? HotLoopInstList::A_LDS_Read_Inst_Num
: HotLoopInstList::A_LDS_Read_Inst_Num / 2;
constexpr auto num_ds_read_inst_b =
HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16
? HotLoopInstList::B_LDS_Read_Inst_Num
: HotLoopInstList::B_LDS_Read_Inst_Num / 2;
constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
constexpr auto num_ds_write_inst_b = HotLoopInstList::B_LDS_Write_Inst_Num;
constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
constexpr auto ds_read_a_issue_cycle =
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
constexpr auto ds_read_b_issue_cycle =
HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16 ? 8 : 4;
constexpr auto ds_read_a_mfma_rate =
(mfma_cycle - 4 + 2 * ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
constexpr auto ds_read_b_mfma_rate =
(mfma_cycle - 4 + 2 * ds_read_b_issue_cycle - 1) / (2 * ds_read_b_issue_cycle);
constexpr auto num_dsread_a_mfma =
(num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
constexpr auto num_dsread_b_mfma =
(num_ds_read_inst_b + ds_read_b_mfma_rate - 1) / ds_read_b_mfma_rate;
// stage 1
// Separate this part?
// constexpr auto num_mfma_per_ds_read = sizeof(ComputeDataType) / sizeof(ADataType) >
// sizeof(ComputeDataType) / sizeof(BDataType)
// ? sizeof(ComputeDataType) / sizeof(ADataType)
// : sizeof(ComputeDataType) / sizeof(BDataType);
constexpr auto num_mfma_stage1 = num_mfma_inst - (num_dsread_a_mfma + num_dsread_b_mfma);
constexpr auto num_mfma_per_issue =
num_mfma_stage1 / (num_buffer_load_inst_a + num_buffer_load_inst_b);
constexpr auto num_dswrite_per_issue_a = num_ds_write_inst_a / num_buffer_load_inst_a;
constexpr auto num_dswrite_per_issue_b = num_ds_write_inst_b / num_buffer_load_inst_b;
static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) {
ignore = i;
static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) {
ignore = idswrite;
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(
0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA
});
static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) {
ignore = i;
static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) {
ignore = idswrite;
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(
0x008, num_mfma_per_issue - num_dswrite_per_issue_b, 0); // MFMA
});
// stage 2
static_for<0, num_dsread_a_mfma, 1>{}([&](auto i) {
if constexpr((num_ds_read_inst_a - (i + 1) * ds_read_a_mfma_rate) >=
ds_read_a_mfma_rate)
{
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(0x100,
num_ds_read_inst_a - (num_dsread_a_mfma - 1) *
ds_read_a_mfma_rate,
0); // DS read
}
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
static_for<0, num_dsread_b_mfma, 1>{}([&](auto i) {
if constexpr((num_ds_read_inst_b - (i + 1) * ds_read_b_mfma_rate) >=
ds_read_b_mfma_rate)
{
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_b_mfma_rate, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(0x100,
num_ds_read_inst_b - (num_dsread_b_mfma - 1) *
ds_read_b_mfma_rate,
0); // DS read
}
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
}
template <bool HasMainLoop,
int NumKBlockPerScale,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
@@ -169,6 +304,7 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CScaleThreadDesc,
typename CThreadBuffer,
typename AScaleGridBuffer,
typename AScaleGridDesc,
@@ -196,6 +332,7 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
// CThread
const CScaleThreadDesc& c_scale_thread_desc,
CThreadBuffer& c_thread_buf,
// AScaleThreadCopy
const AScaleGridDesc& a_scale_grid_desc,
@@ -210,11 +347,10 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
const BScaleGridBuffer& b_scale_grid_buf,
const BScaleThreadTransferStep& b_scale_thread_copy_step,
// num_loop
index_t num_loop,
index_t num_loop_per_scale) const
index_t num_loop) const
{
__builtin_amdgcn_sched_barrier(0);
// assume kperblock = scaleblockk
ignore = num_loop_per_scale;
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
@@ -223,6 +359,8 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
a_scale_thread_desc.GetElementSpaceSize());
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
b_scale_thread_desc.GetElementSpaceSize());
auto c_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
c_scale_thread_desc.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
@@ -231,11 +369,26 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(I0, I0),
a_scale_thread_buf);
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(m0, I0),
a_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<0>{}));
});
if constexpr(NumKBlockPerScale == 1)
{
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<2>{}));
}
else
{
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<1>{}));
}
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
@@ -243,17 +396,101 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
make_tuple(I0, I0),
b_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
constexpr auto num_scale_k_block = CScaleThreadDesc{}.GetLength(Number<0>{});
constexpr auto num_scale_m_block = CScaleThreadDesc{}.GetLength(Number<1>{});
constexpr auto num_scale_n_block = CScaleThreadDesc{}.GetLength(Number<2>{});
static_for<0, num_scale_m_block, 1>{}([&](auto m0) {
static_for<0, num_scale_n_block, 1>{}([&](auto n0) {
static_for<0, num_scale_k_block, 1>{}([&](auto k0) {
constexpr index_t c_offset =
CScaleThreadDesc{}.CalculateOffset(make_tuple(k0, m0, n0));
constexpr index_t a_offset =
AScaleThreadDesc{}.CalculateOffset(make_tuple(m0, k0));
constexpr index_t b_offset =
BScaleThreadDesc{}.CalculateOffset(make_tuple(n0, k0));
c_scale_thread_buf(Number<c_offset>{}) =
a_scale_thread_buf[Number<a_offset>{}] *
b_scale_thread_buf[Number<b_offset>{}];
});
});
});
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
// Global prefetch 2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(m0, I0),
a_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<0>{}));
});
if constexpr(NumKBlockPerScale == 1)
{
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<2>{}));
}
else
{
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<1>{}));
}
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(I0, I0),
b_scale_thread_buf);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
// Initialize C
c_thread_buf.Clear();
auto c_thread_buf_per_scale = remove_cvref_t<decltype(c_thread_buf)>();
StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
AccDataType,
1,
xdlops_gemm.GetRegSizePerXdlops(),
true>
c_thread_buf_per_scale;
// Local prefetch 1
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
// main body
if constexpr(HasMainLoop)
@@ -261,13 +498,85 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
index_t i = 0;
do
{
// -------------------------------------------------------------------------------------------
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, num_scale_k_block, 1>{}([&](auto kscale0) {
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()(Number<t>{}) = 0;
});
static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0,
I0,
kscale0 * KRepeat / num_scale_k_block + k0,
ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0,
I0,
kscale0 * KRepeat / num_scale_k_block + k0,
ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
constexpr index_t cscale_offset =
CScaleThreadDesc{}.CalculateOffset(
make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()[Number<t>{}] *
type_convert<AccDataType>(
c_scale_thread_buf[Number<cscale_offset>{}]);
});
});
});
});
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, num_scale_n_block, 1>{}([&](auto n0) {
static_for<0, num_scale_k_block, 1>{}([&](auto k0) {
constexpr index_t c_offset =
CScaleThreadDesc{}.CalculateOffset(make_tuple(k0, m0, n0));
constexpr index_t a_offset =
AScaleThreadDesc{}.CalculateOffset(make_tuple(m0, k0));
constexpr index_t b_offset =
BScaleThreadDesc{}.CalculateOffset(make_tuple(n0, k0));
c_scale_thread_buf(Number<c_offset>{}) =
a_scale_thread_buf[Number<a_offset>{}] *
b_scale_thread_buf[Number<b_offset>{}];
});
});
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
@@ -289,19 +598,70 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
});
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
c_thread_buf_per_scale.Clear();
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(m0, I0),
a_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, a_scale_thread_copy_step.At(Number<0>{}));
});
if constexpr(NumKBlockPerScale == 1)
{
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{}));
}
else
{
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, a_scale_thread_copy_step.At(Number<1>{}));
}
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(I0, I0),
b_scale_thread_buf);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
i += 1;
} while(i < (num_loop - 2));
}
// tail
if constexpr(TailNum == TailNumber::Full)
{
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, num_scale_k_block, 1>{}([&](auto kscale0) {
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()(Number<t>{}) = 0;
});
static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
make_tuple(m0,
I0,
kscale0 * KRepeat / num_scale_k_block + k0,
ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
make_tuple(n0,
I0,
kscale0 * KRepeat / num_scale_k_block + k0,
ik))>{}];
});
using mfma_input_type =
@@ -311,46 +671,41 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset(
make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale[Number<t>{}] *
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
type_convert<AccDataType>(b_scale_thread_buf[I0]);
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()[Number<t>{}] *
type_convert<AccDataType>(
c_scale_thread_buf[Number<cscale_offset>{}]);
});
});
});
});
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(I0, I0),
a_scale_thread_buf);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, num_scale_n_block, 1>{}([&](auto n0) {
static_for<0, num_scale_k_block, 1>{}([&](auto k0) {
constexpr index_t c_offset =
CScaleThreadDesc{}.CalculateOffset(make_tuple(k0, m0, n0));
constexpr index_t a_offset =
AScaleThreadDesc{}.CalculateOffset(make_tuple(m0, k0));
constexpr index_t b_offset =
BScaleThreadDesc{}.CalculateOffset(make_tuple(n0, k0));
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(I0, I0),
b_scale_thread_buf);
c_scale_thread_buf(Number<c_offset>{}) =
a_scale_thread_buf[Number<a_offset>{}] *
b_scale_thread_buf[Number<b_offset>{}];
});
});
});
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
i += 1;
} while(i < (num_loop - 1));
}
// tail
if constexpr(TailNum == TailNumber::Full)
{
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
@@ -371,49 +726,143 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
c_thread_buf_per_scale.Clear();
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
static_for<0, num_scale_k_block, 1>{}([&](auto kscale0) {
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()(Number<t>{}) = 0;
});
static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0,
I0,
kscale0 * KRepeat / num_scale_k_block + k0,
ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0,
I0,
kscale0 * KRepeat / num_scale_k_block + k0,
ik))>{}];
});
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale[Number<t>{}] *
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
type_convert<AccDataType>(b_scale_thread_buf[I0]);
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset(
make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()[Number<t>{}] *
type_convert<AccDataType>(
c_scale_thread_buf[Number<cscale_offset>{}]);
});
});
});
});
__builtin_amdgcn_sched_barrier(0);
}
else if constexpr(TailNum == TailNumber::Odd)
{
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, num_scale_k_block, 1>{}([&](auto kscale0) {
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()(Number<t>{}) = 0;
});
static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0,
I0,
kscale0 * KRepeat / num_scale_k_block + k0,
ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0,
I0,
kscale0 * KRepeat / num_scale_k_block + k0,
ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset(
make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()[Number<t>{}] *
type_convert<AccDataType>(
c_scale_thread_buf[Number<cscale_offset>{}]);
});
});
});
});
__builtin_amdgcn_sched_barrier(0);
}
}
protected:
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
using Base::b_thread_desc_;
using Base::c_thread_desc_;
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType,
ComputeDataType,
decltype(a_block_desc_m0_m1_m2_k),
decltype(a_thread_desc_),
Sequence<1, 1, 1, KPack>,
Sequence<0, 1, 2, 3>,
3,
A_K1,
A_K1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<BDataType,
ComputeDataType,
decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_),
Sequence<1, 1, 1, KPack>,
Sequence<0, 1, 2, 3>,
3,
B_K1,
B_K1>;
AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
};
} // namespace ck

93
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp
View File

@@ -96,7 +96,8 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
NPerXDL,
MRepeat,
NRepeat,
KPack>
KPack,
true>
{
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
@@ -117,7 +118,8 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
NPerXDL,
MRepeat,
NRepeat,
KPack>;
KPack,
true>;
using Base::I0;
using Base::KRepeat;
using Base::xdlops_gemm;
@@ -270,11 +272,26 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(I0, I0),
a_scale_thread_buf);
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(m0, I0),
a_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<0>{}));
});
if(num_loop_per_scale == 1)
{
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<2>{}));
}
else
{
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<1>{}));
}
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
@@ -282,7 +299,6 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
make_tuple(I0, I0),
b_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
// Local prefill 1
@@ -360,17 +376,32 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale[Number<t>{}] *
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
type_convert<AccDataType>(a_scale_thread_buf[m0]) *
type_convert<AccDataType>(b_scale_thread_buf[I0]);
});
});
});
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(I0, I0),
a_scale_thread_buf);
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(m0, I0),
a_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, a_scale_thread_copy_step.At(Number<0>{}));
});
if(num_loop_per_scale == 1)
{
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{}));
}
else
{
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, a_scale_thread_copy_step.At(Number<1>{}));
}
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
@@ -378,8 +409,6 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
make_tuple(I0, I0),
b_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step);
@@ -453,17 +482,32 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale[Number<t>{}] *
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
type_convert<AccDataType>(a_scale_thread_buf[m0]) *
type_convert<AccDataType>(b_scale_thread_buf[I0]);
});
});
});
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(I0, I0),
a_scale_thread_buf);
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(m0, I0),
a_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, a_scale_thread_copy_step.At(Number<0>{}));
});
if(num_loop_per_scale == 1)
{
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{}));
}
else
{
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, a_scale_thread_copy_step.At(Number<1>{}));
}
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
@@ -471,7 +515,6 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
make_tuple(I0, I0),
b_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
block_sync_lds();
@@ -528,7 +571,7 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale[Number<t>{}] *
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
type_convert<AccDataType>(a_scale_thread_buf[m0]) *
type_convert<AccDataType>(b_scale_thread_buf[I0]);
});
});
@@ -586,7 +629,7 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale[Number<t>{}] *
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
type_convert<AccDataType>(a_scale_thread_buf[m0]) *
type_convert<AccDataType>(b_scale_thread_buf[I0]);
});
});

153
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp
View File

@@ -96,7 +96,8 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
NPerXDL,
MRepeat,
NRepeat,
KPack>
KPack,
true>
{
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
@@ -117,7 +118,8 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
NPerXDL,
MRepeat,
NRepeat,
KPack>;
KPack,
true>;
using Base::I0;
using Base::KRepeat;
using Base::xdlops_gemm;
@@ -177,11 +179,11 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
constexpr auto ds_read_a_issue_cycle = 4;
// HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
constexpr auto ds_read_b_issue_cycle = 4;
// HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16 ? 8 : 4;
constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
constexpr auto ds_read_a_issue_cycle =
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
constexpr auto ds_read_b_issue_cycle =
HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16 ? 8 : 4;
constexpr auto ds_read_a_mfma_rate =
(mfma_cycle - 4 + 2 * ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
constexpr auto ds_read_b_mfma_rate =
@@ -262,6 +264,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
}
template <bool HasMainLoop,
int NumKBlockPerScale,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
@@ -275,6 +278,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CScaleThreadDesc,
typename CThreadBuffer,
typename AScaleGridBuffer,
typename AScaleGridDesc,
@@ -302,6 +306,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
// CThread
const CScaleThreadDesc& c_scale_thread_desc,
CThreadBuffer& c_thread_buf,
// AScaleThreadCopy
const AScaleGridDesc& a_scale_grid_desc,
@@ -316,12 +321,14 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
const BScaleGridBuffer& b_scale_grid_buf,
const BScaleThreadTransferStep& b_scale_thread_copy_step,
// num_loop
index_t num_loop,
index_t num_loop_per_scale) const
index_t num_loop) const
{
__builtin_amdgcn_sched_barrier(0);
static_assert(CScaleThreadDesc{}.GetLength(Number<0>{}) == 1,
"Pipeline v3 only support scaleblocksliceK=1");
static_assert(CScaleThreadDesc{}.GetLength(Number<2>{}) == 1,
"Pipeline v3 only support scaleblocksliceN=1");
// assume kperblock = scaleblockk
ignore = num_loop_per_scale;
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
@@ -330,6 +337,8 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
a_scale_thread_desc.GetElementSpaceSize());
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
b_scale_thread_desc.GetElementSpaceSize());
auto c_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
c_scale_thread_desc.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
@@ -338,11 +347,26 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(I0, I0),
a_scale_thread_buf);
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(m0, I0),
a_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<0>{}));
});
if constexpr(NumKBlockPerScale == 1)
{
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<2>{}));
}
else
{
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<1>{}));
}
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
@@ -350,8 +374,12 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
make_tuple(I0, I0),
b_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
static_for<0, MRepeat, 1>{}([&](auto m0) {
c_scale_thread_buf(m0) = a_scale_thread_buf[m0] * b_scale_thread_buf[I0];
});
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
@@ -363,10 +391,44 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(m0, I0),
a_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<0>{}));
});
if constexpr(NumKBlockPerScale == 1)
{
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<2>{}));
}
else
{
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<1>{}));
}
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(I0, I0),
b_scale_thread_buf);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
// Initialize C
c_thread_buf.Clear();
auto c_thread_buf_per_scale = remove_cvref_t<decltype(c_thread_buf)>();
StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
AccDataType,
1,
xdlops_gemm.GetRegSizePerXdlops(),
true>
c_thread_buf_per_scale;
// Local prefetch 1
block_sync_lds();
@@ -409,7 +471,10 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
c_thread_buf_per_scale.Clear();
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()(Number<t>{}) = 0;
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
@@ -430,19 +495,23 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale[Number<t>{}] *
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
type_convert<AccDataType>(b_scale_thread_buf[I0]);
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()[Number<t>{}] *
type_convert<AccDataType>(c_scale_thread_buf[m0]);
});
});
});
static_for<0, MRepeat, 1>{}([&](auto m0) {
c_scale_thread_buf(m0) = a_scale_thread_buf[m0] * b_scale_thread_buf[I0];
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
@@ -462,11 +531,27 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
b_thread_buf);
});
});
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(I0, I0),
a_scale_thread_buf);
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(m0, I0),
a_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, a_scale_thread_copy_step.At(Number<0>{}));
});
if constexpr(NumKBlockPerScale == 1)
{
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{}));
}
else
{
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, a_scale_thread_copy_step.At(Number<1>{}));
}
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
@@ -474,7 +559,6 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
make_tuple(I0, I0),
b_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
@@ -487,7 +571,10 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
{
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
c_thread_buf_per_scale.Clear();
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()(Number<t>{}) = 0;
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
@@ -507,15 +594,15 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale[Number<t>{}] *
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
type_convert<AccDataType>(b_scale_thread_buf[I0]);
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()[Number<t>{}] *
type_convert<AccDataType>(c_scale_thread_buf[m0]);
});
});
});

195
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp
View File

@@ -15,6 +15,7 @@
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/host_utility/flush_cache.hpp"
namespace ck {
namespace tensor_operation {
@@ -177,14 +178,57 @@ struct DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3
const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
const auto Run = [&](const auto& kernel) {
if(arg.KBatch > 1)
hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
0,
arg.M * arg.N * sizeof(CDataType),
stream_config.stream_id_));
if(stream_config.flush_cache)
{
Argument arg_ = arg;
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(
arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0);
const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(
arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0);
auto size_a_buffer =
a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType);
auto size_b_buffer =
b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType);
ck::utility::RotatingMemWrapper<Argument> rotating_mem(
arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer);
rotating_mem.Print();
auto run_flush_cache = [&]() {
// flush icache
ck::utility::flush_icache();
// rotating mem
rotating_mem.Next();
// clear c mem
if(arg_.KBatch > 1)
hipGetErrorString(hipMemsetAsync(arg_.p_c_grid,
0,
arg_.M * arg_.N * sizeof(CDataType),
stream_config.stream_id_));
};
ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
stream_config,
run_flush_cache,
kernel,
dim3(gdx, gdy, gdz),
dim3(BlockSize),
0,
arg_);
}
else
{
if(arg.KBatch > 1)
hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
0,
arg.M * arg.N * sizeof(CDataType),
stream_config.stream_id_));
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
}
};
constexpr index_t minimum_occupancy =
@@ -195,7 +239,7 @@ struct DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3
if(has_main_k_block_loop)
{
// Tail number always 1
// Tail number always full
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1 ||
BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
{
@@ -208,127 +252,13 @@ struct DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3
Run(kernel);
}
}
// Tail number could be One to Seven
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
{
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::One>;
Run(kernel);
}
else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Full)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Full>;
Run(kernel);
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Two>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Three)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Three>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Four)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Four>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Five)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Five>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Six>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Seven)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Seven>;
Run(kernel);
}
}
}
}
}
else
{
// Tail number always 1
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Full)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
@@ -337,6 +267,16 @@ struct DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3
minimum_occupancy>;
Run(kernel);
}
else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
}
}
return ave_time;
@@ -363,10 +303,11 @@ struct DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3
return false;
}
if(ScaleBlockM % MPerBlock != 0 || ScaleBlockN % NPerBlock != 0 || ScaleBlockK != KPerBlock)
{
return false;
}
// if(ScaleBlockM % MPerBlock != 0 || ScaleBlockN % NPerBlock != 0 || ScaleBlockK !=
// KPerBlock)
// {
// return false;
// }
if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::NKPadding ||

234
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp
View File

@@ -234,7 +234,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
}
__device__ static auto MakeAGridDescriptor_AK0_M_AK1(
__host__ __device__ static auto MakeAGridDescriptor_AK0_M_AK1(
index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0)
{
const auto a_grid_desc_mraw_kraw = [&]() {
@@ -316,7 +316,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
}
}
__device__ static auto MakeBGridDescriptor_BK0_N_BK1(
__host__ __device__ static auto MakeBGridDescriptor_BK0_N_BK1(
index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
{
const auto b_grid_desc_nraw_kraw = [&]() {
@@ -431,6 +431,13 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
}
}();
// pad M and N
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
#if 0
using GemmSpecialization = tensor_operation::device::GemmSpecialization;
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
@@ -468,6 +475,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
// not pad M or N
return c_grid_desc_mraw_nraw;
}
#endif
}
__host__ __device__ static auto MakeDsGridDescriptor_M_N(
@@ -665,40 +673,19 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
// in some cases.
else if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
constexpr auto MLdsLayer = 32 * 4 / KPerBlock / sizeof(LDSTypeA) < 1
? 1
: 32 * 4 / KPerBlock / sizeof(LDSTypeA);
constexpr auto a_lds_block_desc = make_naive_tensor_descriptor(
make_tuple(
AK0Number * Number<MLdsLayer>{}, Number<MPerBlock / MLdsLayer>{}, AK1Number),
make_tuple(AK1Number, Number<KPerBlock * MLdsLayer>{}, I1));
constexpr auto a_lds_block_desc =
make_naive_tensor_descriptor(make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
make_tuple(AK1Number, Number<KPerBlock>{}, I1));
constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
a_lds_block_desc,
make_tuple(make_xor_with_modulo_transform(make_tuple(
Number<MPerBlock / MLdsLayer>{}, Number<AK0Number * MLdsLayer>{})),
make_tuple(make_xor_with_modulo_transform(
make_tuple(Number<MPerBlock>{}, Number<AK0Number>{})),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
constexpr auto a_lds_block_desc_ak0_mldslayer_m_ak1 = transform_tensor_descriptor(
a_lds_block_desc_permuted,
make_tuple(make_unmerge_transform(make_tuple(AK0Number, Number<MLdsLayer>{})),
make_pass_through_transform(Number<MPerBlock / MLdsLayer>{}),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_lds_block_desc_ak0_mldslayer_m_ak1,
make_tuple(make_pass_through_transform(AK0Number),
make_merge_transform_v3_division_mod(
make_tuple(Number<MPerBlock / MLdsLayer>{}, Number<MLdsLayer>{})),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return a_lds_block_desc_ak0_m_ak1;
return a_lds_block_desc_permuted;
}
else // ColumnMajor A
{
@@ -800,42 +787,19 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
// NLdsLayer * K0 as logical Bank
constexpr auto NLdsLayer = 32 * 4 / KPerBlock / sizeof(LDSTypeB) < 1
? 1
: 32 * 4 / KPerBlock / sizeof(LDSTypeB);
;
constexpr auto b_lds_block_desc = make_naive_tensor_descriptor(
make_tuple(
BK0Number * Number<NLdsLayer>{}, Number<NPerBlock / NLdsLayer>{}, BK1Number),
make_tuple(BK1Number, Number<KPerBlock * NLdsLayer>{}, I1));
constexpr auto b_lds_block_desc =
make_naive_tensor_descriptor(make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
make_tuple(BK1Number, Number<KPerBlock>{}, I1));
constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
b_lds_block_desc,
make_tuple(make_xor_with_modulo_transform(make_tuple(
Number<NPerBlock / NLdsLayer>{}, Number<BK0Number * NLdsLayer>{})),
make_tuple(make_xor_with_modulo_transform(
make_tuple(Number<NPerBlock>{}, Number<BK0Number>{})),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
constexpr auto b_lds_block_desc_bk0_nldslayer_n_bk1 = transform_tensor_descriptor(
b_lds_block_desc_permuted,
make_tuple(make_unmerge_transform(make_tuple(BK0Number, Number<NLdsLayer>{})),
make_pass_through_transform(Number<NPerBlock / NLdsLayer>{}),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_lds_block_desc_bk0_nldslayer_n_bk1,
make_tuple(make_pass_through_transform(BK0Number),
make_merge_transform_v3_division_mod(
make_tuple(Number<NPerBlock / NLdsLayer>{}, Number<NLdsLayer>{})),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return b_lds_block_desc_bk0_n_bk1;
return b_lds_block_desc_permuted;
}
else // RowMajor B
{
@@ -1001,7 +965,8 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
!(is_same<tensor_layout::gemm::RowMajor, ALayout>::value))
{
if(!(karg.M % MPerBlock == 0))
{
@@ -1018,7 +983,8 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
(is_same<tensor_layout::gemm::RowMajor, BLayout>::value))
{
if(!(karg.N % NPerBlock == 0))
{
@@ -1366,28 +1332,39 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
const index_t ScaleSliceSizeM = 1;
const index_t ScaleSliceSizeN = 1;
const index_t ScaleSliceSizeK = 1;
constexpr index_t ScaleSliceSizeM = MXdlPerWave;
constexpr index_t ScaleSliceSizeN = math::integer_divide_ceil(NPerBlock, ScaleBlockN);
constexpr index_t ScaleSliceSizeK = math::integer_divide_ceil(KPerBlock, ScaleBlockK);
// ScaleSliceSizeK is last dimension in A/B scale for vector memory access
// ScaleSliceSizeK is first dimension in C scale for packed math
constexpr auto a_scale_thread_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<ScaleSliceSizeM>{}, Number<ScaleSliceSizeK>{}));
constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
auto a_thread_offset =
get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 64) / NWaves * MPerXdl;
constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<ScaleSliceSizeM>{}, Number<ScaleSliceSizeK>{}));
make_tuple(Number<ScaleSliceSizeN>{}, Number<ScaleSliceSizeK>{}));
constexpr auto c_scale_thread_desc = make_naive_tensor_descriptor_packed(make_tuple(
Number<ScaleSliceSizeK>{}, Number<ScaleSliceSizeM>{}, Number<ScaleSliceSizeN>{}));
auto a_scale_thread_copy =
ThreadwiseTensorSliceTransfer_v2<AScaleType,
AScaleType,
decltype(a_scale_grid_desc_am_ak),
decltype(a_scale_thread_desc),
Sequence<ScaleSliceSizeM, ScaleSliceSizeK>,
Sequence<1, ScaleSliceSizeK>,
Sequence<0, 1>,
1,
1,
ScaleSliceSizeK,
1,
false>(
a_scale_grid_desc_am_ak, make_multi_index(block_m_id * MPerBlock / ScaleBlockM, 0));
a_scale_grid_desc_am_ak,
make_multi_index(block_m_id * MPerBlock / ScaleBlockM + a_thread_offset, 0));
auto b_scale_thread_copy =
ThreadwiseTensorSliceTransfer_v2<BScaleType,
@@ -1397,17 +1374,21 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
Sequence<ScaleSliceSizeN, ScaleSliceSizeK>,
Sequence<0, 1>,
1,
1,
ScaleSliceSizeK,
1,
false>(
b_scale_grid_desc_bn_ak, make_multi_index(block_n_id * NPerBlock / ScaleBlockN, 0));
constexpr auto a_scale_thread_slice_copy_step = make_multi_index(0, 1);
constexpr auto b_scale_thread_slice_copy_step = make_multi_index(0, 1);
// constexpr auto a_scale_thread_slice_copy_step = make_multi_index(0, 1);
constexpr auto a_scale_thread_slice_copy_step =
make_tuple(make_multi_index(MWaves * MPerXdl, 0),
make_multi_index(-MPerBlock, 0),
make_multi_index(-MPerBlock, ScaleSliceSizeK));
constexpr auto b_scale_thread_slice_copy_step = make_multi_index(0, ScaleSliceSizeK);
const index_t num_k_block_per_scale = ScaleBlockK / KPerBlock;
constexpr auto NumKBlockPerScale = math::integer_divide_ceil(ScaleBlockK, KPerBlock);
blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(
blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, NumKBlockPerScale, TailNum>(
a_grid_desc_ak0_m_ak1,
a_block_desc_ak0_m_ak1,
a_blockwise_copy,
@@ -1420,6 +1401,8 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
c_scale_thread_desc,
c_thread_buf,
a_scale_grid_desc_am_ak,
@@ -1434,8 +1417,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
b_scale_grid_buf,
b_scale_thread_slice_copy_step,
num_k_block_main_loop,
num_k_block_per_scale);
num_k_block_main_loop);
// shuffle C and write out
{
@@ -1446,23 +1428,24 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// transposed XDL
// // TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// // TODO: hacky, fix it!
// only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp =
blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5);
constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6);
constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
@@ -1471,24 +1454,24 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
static_cast<CShuffleDataType*>(p_shared),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
M2)), // M2 = MPerXdl
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
N2, // N2 * N3 * N4 = NPerXdl
N3,
N4))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
@@ -1498,57 +1481,57 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
const auto m_thread_data_on_block_to_m0_m1_m2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
make_tuple(make_merge_transform(make_tuple(M0, M1, M2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
CShuffleDataType,
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
decltype(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
decltype(c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
M2,
I1,
M4,
I1>,
N2,
I1,
N4>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3],
m_thread_data_on_block_idx[I4],
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
n_thread_data_on_block_idx[I2],
n_thread_data_on_block_idx[I3],
n_thread_data_on_block_idx[I4]),
tensor_operation::element_wise::PassThrough{}};
using EDataType = CDataType;
@@ -1630,18 +1613,17 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
make_tuple(make_multi_index(block_m_id, 0, block_n_id, 0)),
c_element_op};
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, 1, N2, 1, N4>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
1,
1,
M2,
1,
M4,
1>>{};
N2,
1,
N4>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
@@ -1661,10 +1643,10 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3
block_sync_lds();
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
c_shuffle_block_buf);
// make sure it's safe to read from LDS

88
library/include/ck/library/tensor_operation_instance/gpu/gemm_ab_scale.hpp
View File

@@ -17,7 +17,7 @@ namespace tensor_operation {
namespace device {
namespace instance {
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instances(
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
@@ -28,14 +28,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_i
F32,
Tuple<>,
BF16,
128,
1,
128,
128,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instances(
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
@@ -46,14 +46,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_
F32,
Tuple<>,
BF16,
128,
1,
128,
128,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instances(
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
@@ -64,14 +64,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding
F32,
Tuple<>,
BF16,
128,
1,
128,
128,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instances(
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
@@ -82,61 +82,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpaddin
F32,
Tuple<>,
BF16,
128,
128,
128,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
Row,
F8,
F32,
F8,
F32,
Tuple<>,
BF16,
128,
128,
128,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
Row,
F8,
F32,
F8,
F32,
Tuple<>,
BF16,
128,
128,
128,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
Row,
F8,
F32,
F8,
F32,
Tuple<>,
BF16,
128,
1,
128,
128,
PassThrough,
@@ -163,7 +109,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
B1DataType,
Tuple<>,
CDataType,
128,
1,
128,
128,
ck::tensor_operation::element_wise::PassThrough,
@@ -180,7 +126,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
B1DataType,
Tuple<>,
CDataType,
128,
1,
128,
128,
ck::tensor_operation::element_wise::PassThrough,
@@ -198,20 +144,14 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instances(
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances(
op_ptrs);
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instances(
op_ptrs);
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instances(
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_instances(
op_ptrs);
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instances(
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances(
op_ptrs);
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instances(
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_instances(
op_ptrs);
}
}

60
library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply.hpp
View File

@@ -122,6 +122,32 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kp
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instances_part3(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part3(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
@@ -360,6 +386,32 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpa
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instances_part3(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part3(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
@@ -556,6 +608,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part2(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instances_part3(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part3(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances(
op_ptrs);
@@ -593,6 +649,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part2(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instances_part3(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part3(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v1_default_instances(
op_ptrs);

317
library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle.hpp
View File

@@ -1,317 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#if(defined(CK_ENABLE_F16) || defined(CK_ENABLE_FP8))
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
#endif
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
#endif
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmMultipleDSplitKBPreShuffle<
ALayout,
BLayout,
Tuple<Row, Col>,
CLayout,
ADataType,
BDataType,
Tuple<F32, F32>,
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::MultiplyMultiply>>
{
using DeviceOp =
DeviceGemmMultipleDSplitKBPreShuffle<ALayout,
BLayout,
Tuple<Row, Col>,
CLayout,
ADataType,
BDataType,
Tuple<F32, F32>,
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::MultiplyMultiply>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
// TODO: Add MFMA layout into tensor layout
#if(defined(CK_ENABLE_F16) || defined(CK_ENABLE_FP8))
if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances(
op_ptrs);
}
}
#endif
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instances(
op_ptrs);
}
}
#endif
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

664
library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply_wp.hpp Normal file
View File

@@ -0,0 +1,664 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#if(defined(CK_ENABLE_F16) || defined(CK_ENABLE_FP8))
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances_p1(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances_p2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p1(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p3(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p4(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p5(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p6(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
#endif
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instances_p1(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instances_p2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p1(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p3(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p4(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p5(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p6(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>&
instances);
#endif
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmMultipleDSplitKBPreShuffle<
ALayout,
BLayout,
Tuple<Row, Col>,
CLayout,
ADataType,
BDataType,
Tuple<F32, F32>,
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::MultiplyMultiply>>
{
using DeviceOp =
DeviceGemmMultipleDSplitKBPreShuffle<ALayout,
BLayout,
Tuple<Row, Col>,
CLayout,
ADataType,
BDataType,
Tuple<F32, F32>,
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::MultiplyMultiply>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
// TODO: Add MFMA layout into tensor layout
#if(defined(CK_ENABLE_F16) || defined(CK_ENABLE_FP8))
if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances_p1(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances_p2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p1(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p3(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p4(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p5(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p6(
op_ptrs);
}
}
#endif
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instances_v2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instances_p1(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instances_p2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p1(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p2(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p3(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p4(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p5(
op_ptrs);
add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p6(
op_ptrs);
}
}
#endif
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

36
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View File

@@ -77,7 +77,7 @@ function(add_instance_library INSTANCE_NAME)
# Do not build gemm_universal_f8 or gemm_multiply_multiply_f8 for any targets except gfx94
if(NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH)
foreach(source IN LISTS ARGN)
if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx95" AND source MATCHES "gemm_multiply_multiply_xdl_f8")
if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx95" AND source MATCHES "gemm_multiply_multiply" AND source MATCHES "_f8_")
message("removing gemm_multiply_multiply_f8 instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
@@ -88,18 +88,6 @@ function(add_instance_library INSTANCE_NAME)
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
foreach(source IN LISTS ARGN)
if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx95" AND source MATCHES "batched_gemm_xdl_universal" AND source MATCHES "_f8_")
message("removing batched_gemm_universal_f8 instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
foreach(source IN LISTS ARGN)
if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx95" AND source MATCHES "gemm_xdl_universal_streamk" AND source MATCHES "_f8_")
message("removing gemm_universal_streamk_f8 instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
endif()
#only continue if there are some source files left on the list
if(ARGN)
@@ -113,17 +101,21 @@ function(add_instance_library INSTANCE_NAME)
elseif(source MATCHES "mha")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif()
#only build the fp8 gemm instances for gfx908/90a if the build argument is set
#only build the fp8 gemm instances for gfx90a if the build argument is set, otherwise only build for gfx942/gfx950
if(NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH)
if(source MATCHES "gemm_xdl_universal" AND source MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif()
if(source MATCHES "gemm_multiply_multiply_f8")
if(source MATCHES "gemm_multiply_multiply" AND source MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif()
if(source MATCHES "bached_gemm_multiply_multiply_f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
else()
if(source MATCHES "gemm_xdl_universal" AND source MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif()
if(source MATCHES "gemm_multiply_multiply" AND source MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif()
endif()
set(offload_targets)
foreach(target IN LISTS INST_TARGETS)
@@ -243,7 +235,7 @@ FOREACH(subdir_path ${dir_list})
message("Found only xdl instances, but gfx9 is not on the targets list. Skipping.")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "ONLY WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12"))
if(("${cmake_instance}" MATCHES "ONLY WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12"))
message("Found only wmma instances, but gfx11 is not on the targets list. Skipping.")
set(add_inst 0)
endif()
@@ -251,14 +243,18 @@ FOREACH(subdir_path ${dir_list})
message("Found only xdl and dl instances, but gfx9 is not on the targets listand DL_KERNELS is not set. Skipping.")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "ONLY XDL_AND_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12") AND (NOT INST_TARGETS MATCHES "gfx9"))
if(("${cmake_instance}" MATCHES "ONLY XDL_AND_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12") AND (NOT INST_TARGETS MATCHES "gfx9"))
message("Found only xdl and wmma instances, but gfx11 and gfx9 are not on the targets list. Skipping.")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "XDL_DL_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12") AND (NOT INST_TARGETS MATCHES "gfx9") AND (NOT DEFINED DL_KERNELS))
if(("${cmake_instance}" MATCHES "XDL_DL_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12") AND (NOT INST_TARGETS MATCHES "gfx9") AND (NOT DEFINED DL_KERNELS))
message("Found xdl, dl, and wmma instances, but none of those meet the target list. Skipping.")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "gemm_multiply_multiply" AND "${cmake_instance}" MATCHES "_f8_" ) AND (NOT INST_TARGETS MATCHES "gfx94") AND (NOT INST_TARGETS MATCHES "gfx95") AND (NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH))
message("Found gemm_multiply_multiply_f8 instances, but gfx94/gfx95 not on the target list. Skipping.")
set(add_inst 0)
endif()
if((add_inst EQUAL 1))
get_filename_component(target_dir ${subdir_path} NAME)
add_subdirectory(${target_dir})

7
library/src/tensor_operation_instance/gpu/gemm_ab_scale/CMakeLists.txt
View File

@@ -4,16 +4,13 @@ set(GEMM_AB_SCALE_INSTANCES)
list(APPEND GEMM_AB_SCALE_INSTANCES
device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp
device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp
device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instance.cpp
device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instance.cpp
device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp
device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp
device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp
)
set_source_files_properties(device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
add_instance_library(device_gemm_ab_scale_instance ${GEMM_AB_SCALE_INSTANCES})

69
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp
View File

@@ -34,49 +34,50 @@ static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
template <GemmSpecialization GemmSpec>
using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances = std::tuple<
using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances = std::tuple<
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | M| N| K| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
//################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | M| N| K| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Compute friendly
// Spill in current compiler
// DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
// DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 224, 128, 16, 16, 16, 16, 8, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 128, 64, 128, 16, 16, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 64, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 64, 64, 128, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 64, 128, 16, 16, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 64, 128, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances = std::tuple<
using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances = std::tuple<
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | M| N| K| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
//################################| ALayout| BLayout| DsLayout| ELayout|AData | BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | M| N| K| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Latency friendly
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 32, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 128, 128, 128, 16, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 16, 32, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
// Memory friendly
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 128, 32, 128, 16, 16, 32, 32, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 128, 16, 128, 16, 16, 16, 16, 4, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 64, 32, 128, 16, 16, 32, 32, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 64, 16, 128, 16, 16, 16, 16, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 32, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 128, 128, 128, 16, 16, 64, 16, 16, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 128, 128, 128, 16, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 16, 32, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 16, 64, 128, 16, 16, 16, 16, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 32, 64, 128, 16, 16, 32, 32, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 16, 128, 128, 16, 16, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>
// Memory friendly
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 256, 128, 8, 16, 16, 16, 1, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 128, 128, 8, 16, 16, 16, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 64, 128, 8, 16, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 128, 256, 16, 16, 16, 16, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 64, 256, 16, 16, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 256, 128, 16, 16, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 128, 128, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 64, 128, 16, 16, 16, 16, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 64, 256, 16, 16, 16, 16, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 256, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 128, 128, 16, 16, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 64, 128, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 128, 256, 16, 16, 32, 32, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 64, 256, 16, 16, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>
// clang-format on
>;
} // namespace instance

6
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp
View File

@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instances(
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_i
F32,
Tuple<>,
BF16,
128,
1,
128,
128,
PassThrough,
@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_i
{
add_device_operation_instances(
instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances<GemmDefault>{});
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances<GemmDefault>{});
}
} // namespace instance

6
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp
View File

@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instances(
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_
F32,
Tuple<>,
BF16,
128,
1,
128,
128,
PassThrough,
@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_
{
add_device_operation_instances(
instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances<GemmKPadding>{});
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances<GemmKPadding>{});
}
} // namespace instance

37
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instance.cpp
View File

@@ -1,37 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
Row,
F8,
F32,
F8,
F32,
Tuple<>,
BF16,
128,
128,
128,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances<GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

37
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instance.cpp
View File

@@ -1,37 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
Row,
F8,
F32,
F8,
F32,
Tuple<>,
BF16,
128,
128,
128,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances<GemmMNPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

8
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp
View File

@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instances(
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default
F32,
Tuple<>,
BF16,
128,
1,
128,
128,
PassThrough,
@@ -28,8 +28,8 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default
{
add_device_operation_instances(
instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances<Intrawave,
GemmDefault>{});
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances<Intrawave,
GemmDefault>{});
}
} // namespace instance

8
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp
View File

@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instances(
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpaddin
F32,
Tuple<>,
BF16,
128,
1,
128,
128,
PassThrough,
@@ -28,8 +28,8 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpaddin
{
add_device_operation_instances(
instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances<Intrawave,
GemmKPadding>{});
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances<Intrawave,
GemmKPadding>{});
}
} // namespace instance

38
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp
View File

@@ -1,38 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col,
Tuple<>,
Row,
F8,
F32,
F8,
F32,
Tuple<>,
BF16,
128,
128,
128,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances<Intrawave,
GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

8
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/CMakeLists.txt
View File

@@ -10,6 +10,8 @@ list(APPEND GEMM_MULTIPLY_MULTIPLY_INSTANCES
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part3.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part3.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_default_instance.cpp
@@ -23,6 +25,8 @@ list(APPEND GEMM_MULTIPLY_MULTIPLY_INSTANCES
device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp
device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp
device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp
device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part3.cpp
device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part3.cpp
device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v1_default_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v2_default_instance.cpp
@@ -44,6 +48,8 @@ set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_
set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part3.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part3.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance_part1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance_part1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
@@ -53,6 +59,8 @@ set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_g
set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part3.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part3.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")

39
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp
View File

@@ -36,12 +36,11 @@ static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part1 = std::tuple<
// clang-format off
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// Compute friendly
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -60,18 +59,16 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part1
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 32, 32, 1, 3, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 256, 16, 16, 32, 32, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
#endif
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part2 = std::tuple<
// clang-format off
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// Compute friendly
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 128, 128, 16, 16, 32, 32, 3, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -100,19 +97,17 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part2
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 192, 256, 16, 16, 16, 16, 1, 3, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 128, 256, 16, 16, 16, 16, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
#endif
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_instances_part1 =
std::tuple<
// clang-format off
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// Compute friendly
// 256x[64, 256, 32]x128
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -125,7 +120,20 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_insta
// 224x[64, 256, 32]x128
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 224, 128, 16, 16, 16, 16, 7, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 192, 128, 16, 16, 16, 16, 7, 6, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 192, 128, 16, 16, 16, 16, 7, 6, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_instances_part2 =
std::tuple<
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Compute friendly
// 224x[64, 256, 32]x128
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 160, 128, 16, 16, 16, 16, 7, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 128, 128, 16, 16, 16, 16, 7, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 96, 128, 16, 16, 16, 16, 7, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -139,19 +147,17 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_insta
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 96, 128, 16, 16, 16, 16, 6, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 128, 16, 16, 16, 16, 6, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 256, 16, 16, 16, 16, 6, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16,16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
#endif
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_instances_part2 =
using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_instances_part3 =
std::tuple<
// clang-format off
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// Compute friendly
// 160x[64, 256, 32]x128, 160x[64, 96, 32]x256
@@ -180,18 +186,16 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_insta
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 16, 16, 4, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 16, 16, 4, 3, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 256, 16, 16, 16, 16, 4, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
#endif
// clang-format on
>;
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_instances = std::tuple<
// clang-format off
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// Latency friendly
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 128, 32, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 64, 16, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
@@ -227,8 +231,7 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_instances = std:
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 128, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
#endif
// clang-format on
>;
} // namespace instance
} // namespace device

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part3.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instances_part3(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_instances_part3<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part3.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part3(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_instances_part3<
GemmKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

39
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp
View File

@@ -36,12 +36,11 @@ static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances_part1 = std::tuple<
// clang-format off
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// Compute friendly
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -59,18 +58,16 @@ using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances_part1
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 32, 32, 1, 3, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 256, 16, 16, 32, 32, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
#endif
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances_part2 = std::tuple<
// clang-format off
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// Compute friendly
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 128, 16, 16, 32, 32, 1, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -92,19 +89,17 @@ using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances_part2
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 192, 256, 16, 16, 16, 16, 1, 3, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 128, 256, 16, 16, 16, 16, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
#endif
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instances_part1 =
std::tuple<
// clang-format off
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// Compute friendly
// 256x[64, 256, 32]x128
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -117,7 +112,20 @@ using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instan
// 224x[64, 256, 32]x128
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 224, 128, 16, 16, 16, 16, 7, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 192, 128, 16, 16, 16, 16, 7, 6, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 192, 128, 16, 16, 16, 16, 7, 6, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instances_part2 =
std::tuple<
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Compute friendly
// 224x[64, 256, 32]x128
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 160, 128, 16, 16, 16, 16, 7, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 128, 128, 16, 16, 16, 16, 7, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 96, 128, 16, 16, 16, 16, 7, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -131,19 +139,17 @@ using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instan
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 96, 128, 16, 16, 16, 16, 6, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 128, 16, 16, 16, 16, 6, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 256, 16, 16, 16, 16, 6, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16,16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
#endif
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instances_part2 =
using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instances_part3 =
std::tuple<
// clang-format off
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// Compute friendly
// 160x[64, 256, 32]x128, 160x[64, 96, 32]x256
@@ -167,18 +173,16 @@ using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instan
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 16, 16, 4, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 16, 16, 4, 3, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 256, 16, 16, 16, 16, 4, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
#endif
// clang-format on
>;
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_instances = std::tuple<
// clang-format off
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// Latency friendly
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 128, 32, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 64, 16, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
@@ -214,8 +218,7 @@ using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_instances = std::
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 128, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
#endif
// clang-format on
>;
} // namespace instance
} // namespace device

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part3.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instances_part3(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instances_part3<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part3.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part3(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instances_part3<
GemmKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

42
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/CMakeLists.txt
View File

@@ -1,42 +0,0 @@
# ONLY XDL_KERNELS
set(GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES)
list(APPEND GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance.cpp
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance.cpp
)
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
add_instance_library(device_gemm_multiply_multiply_weight_preshuffle_instance ${GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES})

82
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/CMakeLists.txt Normal file
View File

@@ -0,0 +1,82 @@
# ONLY XDL_KERNELS
set(GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES)
list(APPEND GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instance.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instance.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instance_v2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instance_v2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance_p1.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance_p2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p1.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p3.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p4.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p5.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p6.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instance.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instance.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instance_v2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instance_v2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance_p1.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance_p2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p1.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p3.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p4.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p5.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p6.cpp
)
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance_p1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance_p2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p3.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p4.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p5.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p6.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance_p1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance_p2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p3.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p4.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p5.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p6.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
add_instance_library(device_gemm_multiply_multiply_wp_instance ${GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES})

6
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p1.cpp
View File

@@ -1,14 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instances(
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p1(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
@@ -23,7 +23,7 @@ void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances<
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances_p1<
GemmDefault>{});
}

6
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p2.cpp
View File

@@ -1,14 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances(
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
@@ -23,7 +23,7 @@ void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma1
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_instances<
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances_p2<
GemmDefault>{});
}

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p3.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p3(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances_p3<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p4.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p4(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances_p4<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p5.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p5(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances_p5<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p6.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances_p6(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances_p6<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

202
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp
View File

@@ -40,107 +40,125 @@ static constexpr auto v2 = BlockGemmPipelineVersion::v2;
template <BlockGemmPipelineVersion BlkGemmPipeVer, GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_instances =
std::tuple<
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 32, 32, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 32, 32, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 128, 128, 16, 16, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 128, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// N 256
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 8, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 32, 32, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 128, 16, 16, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// N 512
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 512, 128, 16, 16, 32, 32, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 512, 128, 16, 16, 32, 32, 1, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
#endif
// clang-format on
>;
template <BlockGemmPipelineVersion BlkGemmPipeVer, GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_instances =
std::tuple<
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 32, 32, 4, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 128, 256, 16, 16, 32, 32, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 128, 256, 16, 16, 32, 32, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 128, 512, 16, 16, 32, 32, 2, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 512, 16, 16, 32, 32, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// N 256
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 256, 16, 16, 32, 32, 4, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 256, 16, 16, 32, 32, 2, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 256, 16, 16, 32, 32, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 512, 16, 16, 32, 32, 2, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 512, 16, 16, 32, 32, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// N 512
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 512, 256, 16, 16, 32, 32, 2, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 512, 256, 16, 16, 32, 32, 1, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
#endif
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 512, 16, 16, 32, 32, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
// clang-format on
>;
template <BlockGemmPipelineVersion BlkGemmPipeVer, GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_instances =
std::tuple<
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 256, 16, 16, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 128, 256, 16, 16, 16, 16, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 256, 16, 16, 16, 16, 1, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 512, 256, 16, 16, 16, 16, 1, 8, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// N 256
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 256, 16, 16, 32, 32, 2, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 256, 16, 16, 32, 32, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 512, 16, 16, 32, 32, 2, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 512, 16, 16, 32, 32, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// N 512
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 512, 256, 16, 16, 32, 32, 2, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 512, 256, 16, 16, 32, 32, 1, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
// clang-format on
>;
template <BlockGemmPipelineVersion BlkGemmPipeVer, GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p4_instances =
std::tuple<
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 128, 512, 16, 16, 16, 16, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 512, 16, 16, 16, 16, 1, 4, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
#endif
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 128, 512, 16, 16, 16, 16, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 512, 16, 16, 16, 16, 1, 4, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
// clang-format on
>;
template <BlockGemmPipelineVersion BlkGemmPipeVer, GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p5_instances =
std::tuple<
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 256, 16, 16, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 128, 256, 16, 16, 16, 16, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 256, 16, 16, 16, 16, 1, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 512, 256, 16, 16, 16, 16, 1, 8, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_instances =
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_instances_p1 =
std::tuple<
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 256, 128, 16, 16, 32, 32, 7, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 256, 128, 16, 16, 32, 32, 6, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 256, 128, 16, 16, 32, 32, 5, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 32, 32, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 32, 32, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_instances_p2 =
std::tuple<
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 32, 32, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 128, 128, 16, 16, 32, 32, 7, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 128, 128, 16, 16, 32, 32, 6, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 128, 128, 16, 16, 32, 32, 5, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 32, 32, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
#endif
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances =
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances_p1 =
std::tuple<
// clang-format off
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// Compute friendly
// 256x[64, 256, 32]x128
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -149,7 +167,18 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x1
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 160, 128, 16, 16, 16, 16, 8, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 16, 16, 8, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 96, 128, 16, 16, 16, 16, 8, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 64, 128, 16, 16, 16, 16, 8, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 64, 128, 16, 16, 16, 16, 8, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances_p2 =
std::tuple<
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// 224x[64, 256, 32]x128
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 224, 128, 16, 16, 16, 16, 7, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -157,7 +186,17 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x1
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 160, 128, 16, 16, 16, 16, 7, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 128, 128, 16, 16, 16, 16, 7, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 96, 128, 16, 16, 16, 16, 7, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 64, 128, 16, 16, 16, 16, 7, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 64, 128, 16, 16, 16, 16, 7, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances_p3 =
std::tuple<
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// 192x[64, 256, 32]x128, 192x[64]x256
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 256, 128, 16, 16, 16, 16, 6, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 224, 128, 16, 16, 16, 16, 6, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -165,7 +204,17 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x1
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 160, 128, 16, 16, 16, 16, 6, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 128, 128, 16, 16, 16, 16, 6, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 96, 128, 16, 16, 16, 16, 6, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 128, 16, 16, 16, 16, 6, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 128, 16, 16, 16, 16, 6, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances_p4 =
std::tuple<
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// 160x[64, 256, 32]x128, 160x[64, 96, 32]x256
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 256, 128, 16, 16, 16, 16, 5, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 224, 128, 16, 16, 16, 16, 5, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -173,19 +222,38 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x1
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 160, 128, 16, 16, 16, 16, 5, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 128, 128, 16, 16, 16, 16, 5, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 96, 128, 16, 16, 16, 16, 5, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 64, 128, 16, 16, 16, 16, 5, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
// 128x[64, 256, 32]x128, 128x[64, 128, 32]x256
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 16, 16, 4, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 224, 128, 16, 16, 16, 16, 4, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 192, 128, 16, 16, 16, 16, 4, 6, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 160, 128, 16, 16, 16, 16, 4, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 16, 16, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 64, 128, 16, 16, 16, 16, 5, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances_p5 =
std::tuple<
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 128, 16, 16, 16, 16, 4, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 128, 16, 16, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 16, 16, 4, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 16, 16, 4, 3, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 256, 16, 16, 16, 16, 4, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
#endif
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances_p6 =
std::tuple<
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 16, 16, 4, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 224, 128, 16, 16, 16, 16, 4, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 192, 128, 16, 16, 16, 16, 4, 6, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 160, 128, 16, 16, 16, 16, 4, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 16, 16, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance_p1.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instances_p1(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_instances_p1<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance_p2.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instances_p2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_instances_p2<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

2
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp
View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {

2
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp
View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {

2
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp
View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {

2
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp
View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {

2
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp
View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {

2
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp
View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {

34
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instance.cpp Normal file
View File

@@ -0,0 +1,34 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p4_instances<
v1,
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

34
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instance_v2.cpp Normal file
View File

@@ -0,0 +1,34 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p4_instances<
v2,
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

34
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instance.cpp Normal file
View File

@@ -0,0 +1,34 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p5_instances<
v1,
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

34
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instance_v2.cpp Normal file
View File

@@ -0,0 +1,34 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p5_instances<
v2,
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

6
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p1.cpp
View File

@@ -1,14 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances(
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p1(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
@@ -23,7 +23,7 @@ void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances<
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances_p1<
GemmDefault>{});
}

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p2.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances_p2<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p3.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p3(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances_p3<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p4.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p4(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances_p4<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p5.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p5(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances_p5<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p6.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances_p6(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances_p6<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

202
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp
View File

@@ -40,107 +40,125 @@ static constexpr auto v2 = BlockGemmPipelineVersion::v2;
template <BlockGemmPipelineVersion BlkGemmPipeVer, GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_instances =
std::tuple<
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 32, 32, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 32, 32, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 128, 128, 16, 16, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 128, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// N 256
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 8, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 32, 32, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 128, 16, 16, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// N 512
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 512, 128, 16, 16, 32, 32, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 512, 128, 16, 16, 32, 32, 1, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
#endif
// clang-format on
>;
template <BlockGemmPipelineVersion BlkGemmPipeVer, GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_instances =
std::tuple<
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 32, 32, 4, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 128, 256, 16, 16, 32, 32, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 128, 256, 16, 16, 32, 32, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 128, 512, 16, 16, 32, 32, 2, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 512, 16, 16, 32, 32, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// N 256
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 256, 16, 16, 32, 32, 4, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 256, 16, 16, 32, 32, 2, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 256, 16, 16, 32, 32, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 512, 16, 16, 32, 32, 2, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 512, 16, 16, 32, 32, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// N 512
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 512, 256, 16, 16, 32, 32, 2, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 512, 256, 16, 16, 32, 32, 1, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
#endif
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 512, 16, 16, 32, 32, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
// clang-format on
>;
template <BlockGemmPipelineVersion BlkGemmPipeVer, GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_instances =
std::tuple<
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 256, 16, 16, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 128, 256, 16, 16, 16, 16, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 256, 16, 16, 16, 16, 1, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 512, 256, 16, 16, 16, 16, 1, 8, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// N 256
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 256, 16, 16, 32, 32, 2, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 256, 16, 16, 32, 32, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 512, 16, 16, 32, 32, 2, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 512, 16, 16, 32, 32, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
// N 512
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 512, 256, 16, 16, 32, 32, 2, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 512, 256, 16, 16, 32, 32, 1, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
// clang-format on
>;
template <BlockGemmPipelineVersion BlkGemmPipeVer, GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p4_instances =
std::tuple<
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 128, 512, 16, 16, 16, 16, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 512, 16, 16, 16, 16, 1, 4, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
#endif
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 128, 512, 16, 16, 16, 16, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 512, 16, 16, 16, 16, 1, 4, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
// clang-format on
>;
template <BlockGemmPipelineVersion BlkGemmPipeVer, GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p5_instances =
std::tuple<
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 256, 16, 16, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 128, 256, 16, 16, 16, 16, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 256, 16, 16, 16, 16, 1, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 512, 256, 16, 16, 16, 16, 1, 8, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_instances =
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_instances_p1 =
std::tuple<
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 256, 128, 16, 16, 32, 32, 7, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 256, 128, 16, 16, 32, 32, 6, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 256, 128, 16, 16, 32, 32, 5, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 32, 32, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 32, 32, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_instances_p2 =
std::tuple<
// clang-format off
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 32, 32, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 128, 128, 16, 16, 32, 32, 7, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 128, 128, 16, 16, 32, 32, 6, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 128, 128, 16, 16, 32, 32, 5, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 32, 32, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
#endif
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances =
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances_p1 =
std::tuple<
// clang-format off
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
// Compute friendly
// 256x[64, 256, 32]x128
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -149,7 +167,18 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 160, 128, 16, 16, 16, 16, 8, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 16, 16, 8, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 96, 128, 16, 16, 16, 16, 8, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 64, 128, 16, 16, 16, 16, 8, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 64, 128, 16, 16, 16, 16, 8, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances_p2 =
std::tuple<
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// 224x[64, 256, 32]x128
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 224, 128, 16, 16, 16, 16, 7, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -157,7 +186,17 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 160, 128, 16, 16, 16, 16, 7, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 128, 128, 16, 16, 16, 16, 7, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 96, 128, 16, 16, 16, 16, 7, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 64, 128, 16, 16, 16, 16, 7, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 64, 128, 16, 16, 16, 16, 7, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances_p3 =
std::tuple<
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// 192x[64, 256, 32]x128, 192x[64]x256
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 256, 128, 16, 16, 16, 16, 6, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 224, 128, 16, 16, 16, 16, 6, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -165,7 +204,17 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 160, 128, 16, 16, 16, 16, 6, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 128, 128, 16, 16, 16, 16, 6, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 96, 128, 16, 16, 16, 16, 6, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 128, 16, 16, 16, 16, 6, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 128, 16, 16, 16, 16, 6, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances_p4 =
std::tuple<
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// 160x[64, 256, 32]x128, 160x[64, 96, 32]x256
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 256, 128, 16, 16, 16, 16, 5, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 224, 128, 16, 16, 16, 16, 5, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
@@ -173,19 +222,38 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 160, 128, 16, 16, 16, 16, 5, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 128, 128, 16, 16, 16, 16, 5, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 96, 128, 16, 16, 16, 16, 5, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 64, 128, 16, 16, 16, 16, 5, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
// 128x[64, 256, 32]x128, 128x[64, 128, 32]x256
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 16, 16, 4, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 224, 128, 16, 16, 16, 16, 4, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 192, 128, 16, 16, 16, 16, 4, 6, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 160, 128, 16, 16, 16, 16, 4, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 16, 16, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 64, 128, 16, 16, 16, 16, 5, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances_p5 =
std::tuple<
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 128, 16, 16, 16, 16, 4, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 128, 16, 16, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 16, 16, 4, 4, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 16, 16, 4, 3, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 256, 16, 16, 16, 16, 4, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
#endif
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances_p6 =
std::tuple<
// clang-format off
//############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 16, 16, 4, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 224, 128, 16, 16, 16, 16, 4, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 192, 128, 16, 16, 16, 16, 4, 6, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 160, 128, 16, 16, 16, 16, 4, 5, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 16, 16, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;

6
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance_p1.cpp
View File

@@ -1,14 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances(
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances_p1(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
@@ -23,7 +23,7 @@ void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_m
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_instances<
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_instances_p1<
GemmDefault>{});
}

33
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance_p2.cpp Normal file
View File

@@ -0,0 +1,33 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances_p2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_instances_p2<
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

2
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp
View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {

2
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp
View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {

2
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp
View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {

2
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp
View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {

2
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp
View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {

2
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp → library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp
View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {

34
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instance.cpp Normal file
View File

@@ -0,0 +1,34 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p4_instances<
v1,
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

34
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instance_v2.cpp Normal file
View File

@@ -0,0 +1,34 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p4_instances<
v2,
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

34
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instance.cpp Normal file
View File

@@ -0,0 +1,34 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p5_instances<
v1,
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

34
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instance_v2.cpp Normal file
View File

@@ -0,0 +1,34 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instances_v2(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
F16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p5_instances<
v2,
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

2
profiler/include/profiler/profile_gemm_multiply_multiply_weight_preshuffle_impl.hpp → profiler/include/profiler/profile_gemm_multiply_multiply_wp_impl.hpp
View File

@@ -12,7 +12,7 @@
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply_wp.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"

4
profiler/src/CMakeLists.txt
View File

@@ -50,7 +50,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply_weight_preshuffle.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply_wp.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
endif()
list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
@@ -140,7 +140,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_weight_preshuffle_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_wp_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)

8
profiler/src/profile_gemm_ab_scale.cpp
View File

@@ -32,6 +32,7 @@ enum struct GemmDataType
enum struct ScaleBlockTile
{
Tile_128_128_128, // 0
Tile_1_128_128, // 1
};
#define OP_NAME "gemm_ab_scale"
@@ -49,7 +50,8 @@ int profile_gemm_ab_scale(int argc, char* argv[])
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: scale block tile (0: ScaleBlockM/N/K = [128, 128, 128];\n");
printf("arg4: scale block tile (0: ScaleBlockM/N/K = [128, 128, 128]; 1: ScaleBlockM/N/K = "
"[1, 128, 128];\n");
printf("arg5: verification (0: no; 1: yes)\n");
printf("arg6: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg7: print tensor value (0: no; 1: yes)\n");
@@ -155,7 +157,7 @@ int profile_gemm_ab_scale(int argc, char* argv[])
};
if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN &&
scale_block_tile == ScaleBlockTile::Tile_128_128_128)
scale_block_tile == ScaleBlockTile::Tile_1_128_128)
{
return profile(F8{},
F32{},
@@ -164,7 +166,7 @@ int profile_gemm_ab_scale(int argc, char* argv[])
F8{},
F32{},
BF16{},
ck::Number<128>{},
ck::Number<1>{},
ck::Number<128>{},
ck::Number<128>{},
Row{},

2
profiler/src/profile_gemm_multiply_multiply_weight_preshuffle.cpp → profiler/src/profile_gemm_multiply_multiply_wp.cpp
View File

@@ -6,7 +6,7 @@
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_multiply_multiply_weight_preshuffle_impl.hpp"
#include "profiler/profile_gemm_multiply_multiply_wp_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout