ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-15 19:44:39 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/wjx/moe_v3_aiter' into moe_merge_v3_bs_for_aiter

Browse Source
This commit is contained in:
OscarXu
2025-05-14 15:05:08 +08:00
parent 0a845c0b7d eaebefb278
commit 9ecbcce2e9
14 changed files with 581 additions and 420 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
example/65_gemm_multiply_multiply/CMakeLists.txt
View File

@@ -41,8 +41,16 @@ set(GEMM_OPTIONS)
list(APPEND GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32")
list(APPEND GEMM_OPTIONS -v --save-temps -Wno-gnu-line-marker)
set(BLOCKSCALE_GEMM_OPTIONS)
list(APPEND BLOCKSCALE_GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32 -mllvm --schedmodel=0 -mllvm --amdgpu-sched-strategy=gcn-iterative-max-occupancy-experimental")
list(APPEND BLOCKSCALE_GEMM_OPTIONS -v --save-temps -Wno-gnu-line-marker)
target_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE ${GEMM_OPTIONS})
target_compile_options(example_moe_gemm1_xdl_fp8 PRIVATE ${GEMM_OPTIONS})
target_compile_options(example_moe_gemm2_xdl_fp8 PRIVATE ${GEMM_OPTIONS})
target_compile_options(example_gemm_multiply_multiply_xdl_fp8_ab_scale PRIVATE ${BLOCKSCALE_GEMM_OPTIONS})
target_compile_options(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle PRIVATE ${BLOCKSCALE_GEMM_OPTIONS})
# list(APPEND BLOCKSCALE_GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32 -mllvm --disable-schedmodel-in-sched-mi=1 -mllvm --amdgpu-sched-strategy=gcn-iterative-max-occupancy-experimental -mllvm --misched-bottomup=1")
list(APPEND BLOCKSCALE_GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32")
list(APPEND BLOCKSCALE_GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32 -mllvm --schedmodel=0 -mllvm --amdgpu-sched-strategy=gcn-iterative-max-occupancy-experimental")
list(APPEND BLOCKSCALE_GEMM_OPTIONS -v --save-temps -Wno-gnu-line-marker)
target_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE ${GEMM_OPTIONS})
target_compile_options(example_moe_gemm1_xdl_fp8 PRIVATE ${GEMM_OPTIONS})
@@ -53,4 +61,4 @@ target_compile_options(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshu
target_compile_options(example_moe_gemm2_xdl_fp8_blockscale PRIVATE ${BLOCKSCALE_GEMM_OPTIONS})
#hacky fix for bs_moe_stage2 with rocm < 6.4
target_compile_definitions(example_moe_gemm2_xdl_fp8_blockscale PRIVATE MOE_STAGE2_ASM_DIR="${CMAKE_CURRENT_SOURCE_DIR}/hsa/")
target_compile_definitions(example_moe_gemm2_xdl_fp8_blockscale PRIVATE MOE_STAGE2_ASM_DIR="${CMAKE_CURRENT_SOURCE_DIR}/hsa/")

4
example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp
View File

@@ -139,10 +139,10 @@ 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,
128, 128, 128,
256, 256, 128,
16, 16,
16, 16,
8, 2,
16, 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,
2, 1, S<1, 32, 1, 8>, S<8, 8, 1>,

19
example/65_gemm_multiply_multiply/moe_gemm2_xdl_fp8.cpp
View File

@@ -168,8 +168,8 @@ using DeviceOpInstance = ck::tensor_operation::device::Devic
// CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
4, 2, S<1, CShuffleMLane, 1, CShuffleNLane>, S<EVec, D0Vec, D1Vec, D2Vec>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, 0, false, false, MulRoutedWeight, false, int32_t, A0DataType>;
2, 1, S<1, CShuffleMLane, 1, CShuffleNLane>, S<EVec, D0Vec, D1Vec, D2Vec>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, 0, false, false, MulRoutedWeight, false, int32_t, A0DataType>;
// kernel 2: 128->32x128x128
// < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 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>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, EDataType>;
@@ -186,11 +186,11 @@ int main(int argc, char* argv[])
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t experts = 8;
ck::index_t sorted_tile_num = 16;
ck::index_t valid_tile_num = 13;
ck::index_t sorted_tile_num = 133;
ck::index_t valid_tile_num = 128;
ck::index_t sorted_size = sorted_tile_num * MPerBlock;
ck::index_t valid_size = valid_tile_num * MPerBlock;
ck::index_t tokens = 128;
ck::index_t tokens = 16384;
ck::index_t topk = 2;
if(argc == 1)
@@ -245,10 +245,11 @@ int main(int argc, char* argv[])
Tensor<ck::index_t> expert_ids(HostTensorDescriptor({sorted_tile_num}, {1}));
Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1}));
max_token_id.mData = {valid_size, 0, 2, 3, 4, 6, 8, 10, 12, 13};
int eids[] = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 7, 7, 3, 3, 3};
// max_token_id.mData[0] = valid_size;
// max_token_id.mData = {valid_size, 0, 2, 3, 4, 6, 8, 10, 12, 13};
// int eids[] = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 7, 7, 3, 3, 3};
max_token_id.mData = {valid_size, 0, 1, 2, 3, 4, 5, 6, 7, 8};
// int eids[] = {0, 1, 2, 3, 4, 5, 6, 7, 3, 3, 3}; // {2, 1, 1, 2, 2, 2, 1, 2}
for(int i = 0; i < sorted_tile_num; i++)
{
expert_ids.mData[i] = i / ((valid_tile_num + experts - 1) / experts);

50
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_v1.hpp
View File

@@ -122,6 +122,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_v1<BlockGemmPipelineSch
using Base::B_K1;
using Base::I0;
using Base::I1;
using Base::KGroup;
using Base::KRepeat;
using Base::xdlops_gemm;
using typename Base::HotLoopInstList;
@@ -154,9 +155,9 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_v1<BlockGemmPipelineSch
constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{});
constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{});
constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{});
constexpr index_t K2 = KPack;
constexpr index_t K2 = KPack / KGroup;
constexpr index_t K1 = 64 / NPerXDL;
constexpr index_t K0 = KRepeat;
constexpr index_t K0 = KRepeat * KGroup;
return transform_tensor_descriptor(
TileDesc_M0_M1_M2_K{},
@@ -298,12 +299,14 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_v1<BlockGemmPipelineSch
block_sync_lds();
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, k0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, I0),
a_thread_buf);
static_for<0, KGroup, 1>{}([&](auto kg0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
a_thread_buf);
});
});
});
@@ -382,12 +385,15 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_v1<BlockGemmPipelineSch
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, k0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, I0),
a_thread_buf);
static_for<0, KGroup, 1>{}([&](auto kg0) {
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
a_thread_buf);
});
});
});
@@ -458,12 +464,14 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_v1<BlockGemmPipelineSch
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, k0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, I0),
a_thread_buf);
static_for<0, KGroup, 1>{}([&](auto kg0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
a_thread_buf);
});
});
});
@@ -556,7 +564,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_v1<BlockGemmPipelineSch
ComputeDataType,
decltype(a_block_desc_m0_m1_m2_k0_k1_k2),
decltype(a_thread_desc_),
Sequence<1, 1, 1, 1, 1, KPack>,
Sequence<1, 1, 1, 1, 1, KPack / KGroup>,
Sequence<0, 1, 2, 3, 4, 5>,
5,
A_K1,

6
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp
View File

@@ -283,7 +283,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1<BlockGemmPipelineScheduler::I
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KGroup, 1>{}([&](auto kg0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, Number<k0 * 2 + kg0>{}, I0, I0),
make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
@@ -354,7 +354,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1<BlockGemmPipelineScheduler::I
static_for<0, KGroup, 1>{}([&](auto kg0) {
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, Number<k0 * 2 + kg0>{}, I0, I0),
make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
@@ -419,7 +419,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1<BlockGemmPipelineScheduler::I
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KGroup, 1>{}([&](auto kg0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, Number<k0 * 2 + kg0>{}, I0, I0),
make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),

24
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp
View File

@@ -336,9 +336,11 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::I
constexpr auto buffer_load_issue_point_b = 0;
constexpr auto buffer_load_issue_point_interval_more =
num_mfma_perstage / buffer_load_perstage_more;
math::integer_divide_ceil(num_mfma_perstage, buffer_load_perstage_more);
constexpr auto buffer_load_issue_point_interval_less =
num_mfma_perstage / buffer_load_perstage_less;
buffer_load_perstage_less == 0
? INT32_MAX
: math::integer_divide_ceil(num_mfma_perstage, buffer_load_perstage_less);
constexpr auto ds_write_issue_point = 0;
constexpr auto buffer_load_issue_point_a = num_mfma_perstage >= 3 ? 1 : 0;
@@ -470,7 +472,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::I
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KGroup, 1>{}([&](auto kg0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
make_tuple(m0, I0, I0, Number<k0 * KGroup+ kg0>{}, I0, I0),
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
@@ -549,7 +551,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::I
make_tuple(Number<(m0 + 2) % MRepeat>{},
I0,
I0,
Number<k0 * KGroup + kg0>{},
Number<k0 * KGroup+ kg0>{},
I0,
I0),
a_block_buf.At(local_read_buf),
@@ -575,7 +577,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::I
make_tuple(Number<(m0 + 2) % MRepeat>{},
I0,
I0,
Number<k0 * KGroup + kg0>{},
Number<k0 * KGroup+ kg0>{},
I0,
I0),
a_block_buf.At(local_read_buf),
@@ -601,7 +603,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::I
make_tuple(Number<(m0 + 2) % MRepeat>{},
I0,
I0,
Number<k0 * KGroup + kg0>{},
Number<k0 * KGroup+ kg0>{},
I0,
I0),
a_block_buf.At(mfma_reg_buf),
@@ -676,7 +678,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::I
make_tuple(Number<(m0 + 2) % MRepeat>{},
I0,
I0,
Number<k0 * KGroup + kg0>{},
Number<k0 * KGroup+ kg0>{},
I0,
I0),
a_block_buf.At(I1),
@@ -696,7 +698,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::I
make_tuple(Number<(m0 + 2) % MRepeat>{},
I0,
I0,
Number<k0 * KGroup + kg0>{},
Number<k0 * KGroup+ kg0>{},
I0,
I0),
a_block_buf.At(I1),
@@ -716,7 +718,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::I
make_tuple(Number<(m0 + 2) % MRepeat>{},
I0,
I0,
Number<k0 * KGroup + kg0>{},
Number<k0 * KGroup+ kg0>{},
I0,
I0),
a_block_buf.At(I0),
@@ -765,7 +767,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::I
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(
Number<m0 + 2>{}, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
Number<m0 + 2>{}, I0, I0, Number<k0 * KGroup+ kg0>{}, I0, I0),
a_block_buf.At(I1),
a_thread_desc_,
make_tuple(Number<(m0 + 2 + HotloopLocalBufSwitch) % 2>{},
@@ -820,7 +822,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::I
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(
Number<m0 + 2>{}, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
Number<m0 + 2>{}, I0, I0, Number<k0 * KGroup+ kg0>{}, I0, I0),
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(

333
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp
View File

@@ -194,6 +194,183 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
__device__ static constexpr auto HotLoopScheduler()
{
#if 0
constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num;
constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
constexpr auto num_buffer_load_inst_b = MWaves * HotLoopInstList::B_Buffer_Load_Inst_Num;
constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat;
constexpr auto staged_num_mfma = num_mfma / MRepeat;
constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a;
if constexpr(stage.value == 0)
{
// B VMEM access.
constexpr auto staged_num_buffer_load_b_per_ds_read_a =
num_buffer_load_inst_b / staged_num_ds_read_inst_a;
constexpr auto staged_num_mfma_per_buffer_load_b =
staged_num_mfma / num_buffer_load_inst_b;
// B global
static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
ignore = i_inst;
static_for<0, staged_num_buffer_load_b_per_ds_read_a - 1, 1>{}([&](auto ibuf_inst) {
ignore = ibuf_inst;
static_for<0, staged_num_mfma_per_buffer_load_b, 1>{}([&](auto i_mfma) {
ignore = i_mfma;
__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, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
static_for<0, staged_num_mfma_per_buffer_load_b - 1, 1>{}([&](auto i_mfma) {
ignore = i_mfma;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
});
__builtin_amdgcn_sched_barrier(0);
}
else if constexpr(stage.value == 1)
{
// A LDS write access.
constexpr auto staged_num_mfma_per_ds_write_a =
math::integer_divide_ceil(staged_num_mfma, num_ds_write_inst_a);
constexpr auto stage_more_mfma =
staged_num_mfma - (staged_num_mfma_per_ds_write_a - 1) * num_ds_write_inst_a;
// A local write
static_for<0, num_ds_write_inst_a, 1>{}([&](auto i_inst) {
if constexpr(i_inst.value < stage_more_mfma)
{
if(i_inst.value < staged_num_ds_read_inst_a)
{
static_for<0, staged_num_mfma_per_ds_write_a - 1, 1>{}([&](auto i_mfma) {
ignore = i_mfma;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
}
else
{
static_for<0, staged_num_mfma_per_ds_write_a, 1>{}([&](auto i_mfma) {
ignore = i_mfma;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
}
}
else
{
if(i_inst.value < staged_num_ds_read_inst_a)
{
static_for<0, staged_num_mfma_per_ds_write_a - 2, 1>{}([&](auto i_mfma) {
ignore = i_mfma;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
}
else
{
static_for<0, staged_num_mfma_per_ds_write_a - 1, 1>{}([&](auto i_mfma) {
ignore = i_mfma;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
}
}
});
__builtin_amdgcn_sched_barrier(0);
}
else if constexpr(stage.value == 2)
{
// A VMEM access.
constexpr auto staged_num_mfma_per_buffer_load_a =
math::integer_divide_ceil(staged_num_mfma, num_buffer_load_inst_a);
constexpr auto stage_more_mfma =
staged_num_mfma - (staged_num_mfma_per_buffer_load_a - 1) * num_buffer_load_inst_a;
// A global
static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i_inst) {
if constexpr(i_inst.value < stage_more_mfma)
{
if(i_inst.value < staged_num_ds_read_inst_a)
{
static_for<0, staged_num_mfma_per_buffer_load_a - 1, 1>{}([&](auto i_mfma) {
ignore = i_mfma;
__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, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
}
else
{
static_for<0, staged_num_mfma_per_buffer_load_a, 1>{}([&](auto i_mfma) {
ignore = i_mfma;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
}
}
else
{
if(i_inst.value < staged_num_ds_read_inst_a)
{
static_for<0, staged_num_mfma_per_buffer_load_a - 2, 1>{}([&](auto i_mfma) {
ignore = i_mfma;
__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, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
}
else
{
static_for<0, staged_num_mfma_per_buffer_load_a - 1, 1>{}([&](auto i_mfma) {
ignore = i_mfma;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
}
}
});
__builtin_amdgcn_sched_barrier(0);
}
else
{
// A local Read
static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
ignore = i_inst;
static_for<0, staged_num_mfma_per_ds_read_a, 1>{}([&](auto i_mfma) {
ignore = i_mfma;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
});
__builtin_amdgcn_sched_barrier(0);
}
#elif 1
// A/B split schedule
// compiler is likely to use ds_read2 when instruction width smaller than 16bytes
constexpr auto num_ds_read_inst_a =
@@ -262,13 +439,6 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
// B global read
static_for<0, buffer_load_b_stages, 1>{}([&](auto i) {
// Scale load, 1B
if constexpr (i.value==0){
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
}
// Scale load, 1A
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
static_for<0, num_mfma_perstage, 1>{}([&](auto imfma) {
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
@@ -286,15 +456,11 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
{
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
}
__builtin_amdgcn_sched_group_barrier(0x800, 2, 0); // v_pk_fma
});
// __builtin_amdgcn_sched_barrier(0);
});
// A global read + A local write
static_for<0, buffer_load_a_stages, 1>{}([&](auto i) {
// Scale load, 1A
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
static_for<0, num_mfma_perstage, 1>{}([&](auto imfma) {
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
if constexpr((((i + buffer_load_b_stages) < buffer_load_stages_more) &&
@@ -319,9 +485,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
{
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
}
__builtin_amdgcn_sched_group_barrier(0x800, 2, 0); // v_pk_fma
});
// __builtin_amdgcn_sched_barrier(0);
});
// lds synchronization, prefetch next loop local A
@@ -329,19 +493,13 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
ignore = i;
static_for<0, num_mfma_perstage, 1>{}([&](auto imfma) {
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
// Scale load, 1A
if constexpr(imfma == 0){
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
}
if constexpr(imfma >= (num_mfma_perstage - num_ds_read_a_mfma_perstage))
{
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
}
__builtin_amdgcn_sched_group_barrier(0x800, 2, 0); // v_pk_fma
});
// __builtin_amdgcn_sched_barrier(0);
});
#endif
}
template <bool HasMainLoop,
@@ -419,7 +577,6 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
auto a_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
a_scale_thread_desc.GetElementSpaceSize());
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
@@ -427,10 +584,6 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
auto c_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
c_scale_thread_desc.GetElementSpaceSize());
StaticallyIndexedArray<decltype(a_scale_thread_buf), Number<2>{}> a_scale_thread_bufs;
StaticallyIndexedArray<decltype(b_scale_thread_buf), Number<2>{}> b_scale_thread_bufs;
// StaticallyIndexedArray<decltype(c_scale_thread_buf), Number<2>{}> c_scale_thread_bufs;
// Global prefetch A1 B1, AScale1 BScale1
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
@@ -448,7 +601,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
a_scale_grid_buf,
a_scale_thread_desc,
make_tuple(m0, I0),
a_scale_thread_bufs(I0));
a_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<0>{}));
});
@@ -468,12 +621,12 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(I0, I0),
b_scale_thread_bufs(I0));
b_scale_thread_buf);
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_bufs[I0][m0] * b_scale_thread_bufs[I0][I0];
c_scale_thread_buf(m0) = __builtin_elementwise_fma(a_scale_thread_buf[m0], b_scale_thread_buf[I0], .0f);
});
// Local prefill A1
@@ -483,13 +636,12 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
#if 1
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_bufs(I0));
a_scale_thread_buf);
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
a_scale_thread_copy_step.At(Number<0>{}));
});
@@ -509,16 +661,13 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(I0, I0),
b_scale_thread_bufs(I0));
b_scale_thread_buf);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
#endif
// Initialize C
c_thread_buf.Clear();
// Double register buffer for non-scaled gemm computation
// 1. Reduce register pressure
// 2. Decouple the dependency between mfma instruction and scale-fma instruction following.
StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
AccDataType,
1,
@@ -532,7 +681,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KGroup, 1>{}([&](auto kg0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
make_tuple(m0, I0, I0, Number<k0 * 2 + kg0>{}, I0, I0),
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
@@ -541,32 +690,6 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
});
});
#if 0
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()(Number<t>{}) = 0;
});
// Fill first mfma buffer
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(I0, I0, I0, k0, I0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) = b_thread_bufs
[I0][Number<b_thread_desc_.CalculateOffset(make_tuple(I0, I0, 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>{}));
});
#endif
__builtin_amdgcn_sched_barrier(0);
// main body
@@ -586,36 +709,6 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(local_read_buf));
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_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_bufs(local_read_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_bufs(local_read_buf));
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step);
static_for<0, MRepeat, 1>{}([&](auto m0) {
vector_type<AccDataType, 2> c_scale_thread_vec;
@@ -686,7 +779,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
make_tuple(Number<(m0 + 2) % MRepeat>{},
I0,
I0,
Number<k0 * KGroup + kg0>{},
Number<k0 * 2 + kg0>{},
I0,
I0),
a_block_buf.At(local_read_buf),
@@ -712,7 +805,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
make_tuple(Number<(m0 + 2) % MRepeat>{},
I0,
I0,
Number<k0 * KGroup + kg0>{},
Number<k0 * 2 + kg0>{},
I0,
I0),
a_block_buf.At(local_read_buf),
@@ -738,7 +831,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
make_tuple(Number<(m0 + 2) % MRepeat>{},
I0,
I0,
Number<k0 * KGroup + kg0>{},
Number<k0 * 2 + kg0>{},
I0,
I0),
a_block_buf.At(mfma_reg_buf),
@@ -756,12 +849,44 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
});
}
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
static_for<0, MRepeat, 1>{}([&](auto m0) {
c_scale_thread_buf(m0) = a_scale_thread_bufs[mfma_reg_buf][m0] * b_scale_thread_bufs[mfma_reg_buf][I0];
c_scale_thread_buf(m0) = __builtin_elementwise_fma(a_scale_thread_buf[m0], b_scale_thread_buf[I0], .0f);
});
HotLoopScheduler();
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);
// __builtin_amdgcn_sched_group_barrier(0x020, MRepeat + 1, 0); // VMEM read
__builtin_amdgcn_sched_barrier(0);
};
@@ -782,7 +907,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
b_thread_bufs(I1));
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
vector_type<AccDataType, 2> c_scale_thread_vec;
c_scale_thread_vec.template AsType<AccDataType>()(Number<0>{}) =
c_scale_thread_buf[m0];
@@ -843,7 +968,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
make_tuple(Number<(m0 + 2) % MRepeat>{},
I0,
I0,
Number<k0 * KGroup + kg0>{},
Number<k0 * 2 + kg0>{},
I0,
I0),
a_block_buf.At(I1),
@@ -863,7 +988,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
make_tuple(Number<(m0 + 2) % MRepeat>{},
I0,
I0,
Number<k0 * KGroup + kg0>{},
Number<k0 * 2 + kg0>{},
I0,
I0),
a_block_buf.At(I1),
@@ -883,7 +1008,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
make_tuple(Number<(m0 + 2) % MRepeat>{},
I0,
I0,
Number<k0 * KGroup + kg0>{},
Number<k0 * 2 + kg0>{},
I0,
I0),
a_block_buf.At(I0),
@@ -899,7 +1024,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
HotLoopScheduler();
static_for<0, MRepeat, 1>{}([&](auto m0) {
c_scale_thread_buf(m0) = a_scale_thread_bufs[I0][m0] * b_scale_thread_bufs[I0][I0];
c_scale_thread_buf(m0) = __builtin_elementwise_fma(a_scale_thread_buf[m0], b_scale_thread_buf[I0], .0f);
});
static_for<0, MRepeat, 1>{}([&](auto m0) {
@@ -958,7 +1083,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(
Number<m0 + 2>{}, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
Number<m0 + 2>{}, I0, I0, Number<k0 * 2 + kg0>{}, I0, I0),
a_block_buf.At(I1),
a_thread_desc_,
make_tuple(Number<(m0 + 2 + HotloopLocalBufSwitch) % 2>{},
@@ -1034,7 +1159,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3<BlockGemmPipelineS
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(
Number<m0 + 2>{}, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
Number<m0 + 2>{}, I0, I0, Number<k0 * 2 + kg0>{}, I0, I0),
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(

10
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp
View File

@@ -266,7 +266,6 @@ struct DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle
}
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
{
#if 0
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
@@ -289,15 +288,6 @@ struct DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle
TailNumber::Even>;
Run(kernel);
}
#endif
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle_2lds<
GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
else

3
include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp
View File

@@ -189,8 +189,7 @@ struct GridwiseMoeGemm
static constexpr index_t KLane =
mfma_selector::GetKPerXdlops() / mfma_selector::GetK1PerXdlops();
static constexpr index_t KGroup = mfma_selector::selected_mfma.k_per_blk == 32 ? 2 : 1;
// static_assert(KGroup == 2, "");
static constexpr index_t KGroup = mfma_selector::selected_mfma.k_per_blk == 32 ? 2 : 1;
static constexpr index_t KRepeat = KPerBlock / KLane / (KPack / KGroup);
static constexpr index_t NLane = NPerXdl;
static constexpr index_t NWave = NPerBlock / NPerXdl / NXdlPerWave;

102
library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply_wp.hpp
View File

@@ -18,6 +18,7 @@ namespace device {
namespace instance {
#if(defined(CK_ENABLE_F16) || defined(CK_ENABLE_FP8))
#if 0
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,
@@ -185,7 +186,7 @@ void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_m
PassThrough,
MultiplyMultiply>>>&
instances);
#endif
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,
@@ -268,6 +269,7 @@ void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16
#endif
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
#if 0
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,
@@ -435,7 +437,7 @@ void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_
PassThrough,
MultiplyMultiply>>>&
instances);
#endif
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,
@@ -562,32 +564,32 @@ struct DeviceOperationInstanceFactory<
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(
// 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_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_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);
@@ -612,32 +614,32 @@ struct DeviceOperationInstanceFactory<
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(
// 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_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_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);

96
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/CMakeLists.txt
View File

@@ -2,18 +2,18 @@
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_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
@@ -21,18 +21,18 @@ list(APPEND GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES
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_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
@@ -41,18 +41,18 @@ list(APPEND GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES
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_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")
@@ -60,18 +60,18 @@ set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f
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_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")

44
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

@@ -171,13 +171,13 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x1
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// 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>,
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, 16, 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, 2, 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, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 192, 128, 16, 16, 16, 16, 8, 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, 256, 192, 128, 16, 16, 16, 16, 16, 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, 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, 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, 128, 128, 16, 16, 16, 16, 16, 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, 2, 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, 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, 16, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
@@ -190,13 +190,13 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x1
//############################################| | | | | | | | | | | 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, 256, 128, 16, 16, 16, 16, 14, 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, 2, 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, 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_BPreshuffle< 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_BPreshuffle< 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, 14, 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, 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, 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, 128, 128, 16, 16, 16, 16, 14, 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, 2, 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, 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, 14, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
@@ -208,13 +208,13 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x1
//############################################| | | | | | | | | | | 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, 256, 128, 16, 16, 16, 16, 12, 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, 2, 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, 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>,
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, 192, 128, 16, 16, 16, 16, 6, 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, 192, 192, 128, 16, 16, 16, 16, 12, 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, 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, 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, 128, 128, 16, 16, 16, 16, 12, 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, 2, 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, 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, 12, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
@@ -226,13 +226,13 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x1
//############################################| | | | | | | | | | | 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, 256, 128, 16, 16, 16, 16, 10, 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, 2, 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, 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>,
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, 192, 128, 16, 16, 16, 16, 5, 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, 160, 192, 128, 16, 16, 16, 16, 10, 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, 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, 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, 128, 128, 16, 16, 16, 16, 10, 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, 2, 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, 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>
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, 10, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
@@ -244,10 +244,10 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x1
//############################################| | | | | | | | | | | 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, 64, 128, 16, 16, 16, 16, 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, 2, 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, 256, 16, 16, 16, 16, 8, 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, 2, 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, 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>
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, 8, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
@@ -259,11 +259,11 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x1
//############################################| | | | | 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, 256, 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, 2, 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, 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, 192, 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, 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, 128, 128, 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, 2, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;

272
profiler/src/CMakeLists.txt
View File

@@ -1,89 +1,89 @@
# ckProfiler
set(PROFILER_SOURCES
profiler.cpp
profile_gemm.cpp
profile_reduce.cpp
profile_groupnorm_bwd_data.cpp
profile_groupnorm_fwd.cpp
profile_layernorm_bwd_data.cpp
profile_layernorm_bwd_gamma_beta.cpp
profile_groupnorm_bwd_gamma_beta.cpp
profile_layernorm_fwd.cpp
profile_max_pool2d_fwd.cpp
profile_pool3d_fwd.cpp
profile_avg_pool3d_bwd.cpp
profile_max_pool3d_bwd.cpp
profile_avg_pool2d_bwd.cpp
profile_max_pool2d_bwd.cpp
profile_softmax.cpp
profile_batchnorm_fwd.cpp
profile_batchnorm_bwd.cpp
profile_batchnorm_infer.cpp
profile_conv_tensor_rearrange.cpp
profile_transpose.cpp
profile_permute_scale.cpp
# profile_gemm.cpp
# profile_reduce.cpp
# profile_groupnorm_bwd_data.cpp
# profile_groupnorm_fwd.cpp
# profile_layernorm_bwd_data.cpp
# profile_layernorm_bwd_gamma_beta.cpp
# profile_groupnorm_bwd_gamma_beta.cpp
# profile_layernorm_fwd.cpp
# profile_max_pool2d_fwd.cpp
# profile_pool3d_fwd.cpp
# profile_avg_pool3d_bwd.cpp
# profile_max_pool3d_bwd.cpp
# profile_avg_pool2d_bwd.cpp
# profile_max_pool2d_bwd.cpp
# profile_softmax.cpp
# profile_batchnorm_fwd.cpp
# profile_batchnorm_bwd.cpp
# profile_batchnorm_infer.cpp
# profile_conv_tensor_rearrange.cpp
# profile_transpose.cpp
# profile_permute_scale.cpp
)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
# list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
# list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp)
# list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
# list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_add.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
# list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
# list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
# list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
# list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp)
# list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
endif()
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_add.cpp)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9[45]")
# list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply_wp.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_blockscale_wp.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_blockscale_wp.cpp)
endif()
list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_b_scale.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_b_scale.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp)
list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp)
list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp)
# list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
# list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_b_scale.cpp)
# list(APPEND PROFILER_SOURCES profile_batched_gemm_b_scale.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
# list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
# list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp)
# list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp)
# list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp)
# list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp)
endif()
if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
endif()
list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
# list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp)
# list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp)
# list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
endif()
if(DL_KERNELS)
list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
# list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp)
# list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
endif()
set(PROFILER_EXECUTABLE ckProfiler)
@@ -97,91 +97,91 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance)
endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
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(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9[45]")
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_wp_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_blockscale_wp_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_blockscale_wp_instance)
endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_b_scale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_b_scale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_b_scale_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_b_scale_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance)
endif()
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
endif()
if(DL_KERNELS)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
endif()
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)

26
wip.sh Normal file
View File

@@ -0,0 +1,26 @@
bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 8192 16384 5120 -1 -1 0 0 -1 1 20 50 512
echo
echo
echo
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 8 131072 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 16 131072 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 32 131072 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 64 131072 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 128 131072 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 256 131072 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 512 131072 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
echo
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 1024 1024 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 2048 1024 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 4096 1024 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 16384 1024 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 32768 1024 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
echo
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 32 16384 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 64 16384 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 128 16384 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 4096 16384 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"
bash -xc "bin/ckProfiler gemm_multiply_multiply_weight_preshuffle 1 0 0 2 0 1 8192 16384 5120 -1 -1 0 0 -1 1 20 50 512 | grep Best"