mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-14 02:02:46 +00:00
Allow distinct K0/K1 values for A/B block descriptor (#98)
* add gitignore
* host tensor: allow generating sequentially increasing value in a given dimension
* gridwise gemm v3r1: allow distinct K0/K1 values for A/B block descriptor
- remove dangling header include
- modify example gemm_xdl accordingly
- infer KPack value from M/NPerXdl
- device conv2d fwd: update parameters accordingly for the underlying gridwise gemm v3r1
(API for conv2d fwd stays the same for now until we decide to expose individual K0s for activation and weight)
* add LDS data dump utility
* profiler: reflect API change for distinct K0/K1 for A/B matrices
* profiler: add conflict-free LDS write FP16 kernel instances
* fix accidental perf regression
* address feedback; cosmetic changes
* clang-format for new files
* format
Co-authored-by: Chao Liu <chao.liu2@amd.com>
[ROCm/composable_kernel commit: 6d4450ef15]
This commit is contained in:
Anthony Chang
48
.gitignore
vendored
Normal file
48
.gitignore
vendored
Normal file
@@ -0,0 +1,48 @@
|
||||
# Compiled Object files
|
||||
*.slo
|
||||
*.lo
|
||||
*.o
|
||||
*.obj
|
||||
|
||||
# Precompiled Headers
|
||||
*.gch
|
||||
*.pch
|
||||
*.ipch
|
||||
|
||||
# Compiled Dynamic libraries
|
||||
*.so
|
||||
*.dylib
|
||||
*.dll
|
||||
|
||||
# Fortran module files
|
||||
*.mod
|
||||
|
||||
# Compiled Static libraries
|
||||
*.lai
|
||||
*.la
|
||||
*.a
|
||||
*.lib
|
||||
|
||||
# Executables
|
||||
*.exe
|
||||
*.out
|
||||
*.app
|
||||
|
||||
# vim tags
|
||||
tags
|
||||
.tags
|
||||
.*.swp
|
||||
|
||||
# Editors
|
||||
.vscode
|
||||
|
||||
# build-in-source directory
|
||||
build*
|
||||
|
||||
# emacs temporary/backup files
|
||||
.\#*
|
||||
\#*\#
|
||||
*~
|
||||
|
||||
# GDB temporary files
|
||||
.gdb_history
|
||||
@@ -17,7 +17,7 @@ template <index_t BlockSize,
|
||||
index_t NPerXDL,
|
||||
index_t MRepeat,
|
||||
index_t NRepeat,
|
||||
index_t K1>
|
||||
index_t KPack>
|
||||
struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
|
||||
{
|
||||
static constexpr auto I0 = Number<0>{};
|
||||
@@ -29,10 +29,15 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
|
||||
|
||||
static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1);
|
||||
static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1);
|
||||
static constexpr index_t KPerBlock =
|
||||
BK0NK1BlockDesc{}.GetLength(I0) * BK0NK1BlockDesc{}.GetLength(I2);
|
||||
|
||||
static constexpr index_t K0 = BK0NK1BlockDesc{}.GetLength(I0);
|
||||
static constexpr index_t A_K0 = AK0MK1BlockDesc{}.GetLength(I0);
|
||||
static constexpr index_t B_K0 = BK0NK1BlockDesc{}.GetLength(I0);
|
||||
static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
|
||||
static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
|
||||
|
||||
static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, K1>{};
|
||||
static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack>{};
|
||||
|
||||
static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
|
||||
static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
|
||||
@@ -66,7 +71,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
|
||||
|
||||
const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
|
||||
|
||||
return make_tuple(xdlops_a_idx[I0], 0, waveId_m, xdlops_a_idx[I1], 0);
|
||||
return make_tuple(0, waveId_m, xdlops_a_idx[I1], Number<KPack>{} * xdlops_a_idx[I0]);
|
||||
}
|
||||
|
||||
__device__ static auto CalculateBThreadOriginDataIndex()
|
||||
@@ -77,7 +82,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
|
||||
|
||||
const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
|
||||
|
||||
return make_tuple(xdlops_b_idx[I0], 0, waveId_n, xdlops_b_idx[I1], 0);
|
||||
return make_tuple(0, waveId_n, xdlops_b_idx[I1], Number<KPack>{} * xdlops_b_idx[I0]);
|
||||
}
|
||||
|
||||
template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
|
||||
@@ -115,12 +120,6 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
|
||||
BK0NK1BlockDesc::IsKnownAtCompileTime(),
|
||||
"wrong! Desc should be known at compile-time");
|
||||
|
||||
static_assert(AK0MK1BlockDesc{}.GetLength(I0) == BK0NK1BlockDesc{}.GetLength(I0),
|
||||
"wrong! K0 dimension not consistent");
|
||||
|
||||
static_assert(AK0MK1BlockDesc{}.GetLength(I2) == BK0NK1BlockDesc{}.GetLength(I2),
|
||||
"wrong! K1 dimension not consistent");
|
||||
|
||||
static_assert(BlockSize == MWaves * NWaves * WaveSize,
|
||||
"BlockSize != MWaves * NWaves * WaveSize\n");
|
||||
|
||||
@@ -219,32 +218,32 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
|
||||
c_grid_desc_g_m0_n0_m1_n1_m2_n2);
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr auto MakeABlockDescriptor_K0_M0_M1_M2_K1()
|
||||
__host__ __device__ static constexpr auto MakeABlockDescriptor_M0_M1_M2_K()
|
||||
{
|
||||
return transform_tensor_descriptor(
|
||||
AK0MK1BlockDesc{},
|
||||
make_tuple(make_pass_through_transform(Number<K0>{}),
|
||||
make_unmerge_transform(
|
||||
make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{})),
|
||||
make_pass_through_transform(Number<K1>{})),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}));
|
||||
make_tuple(
|
||||
make_merge_transform_v3_division_mod(make_tuple(Number<A_K0>{}, Number<A_K1>{})),
|
||||
make_unmerge_transform(
|
||||
make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{}))),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr auto MakeBBlockDescriptor_K0_N0_N1_N2_K1()
|
||||
__host__ __device__ static constexpr auto MakeBBlockDescriptor_N0_N1_N2_K()
|
||||
{
|
||||
return transform_tensor_descriptor(
|
||||
BK0NK1BlockDesc{},
|
||||
make_tuple(make_pass_through_transform(Number<K0>{}),
|
||||
make_unmerge_transform(
|
||||
make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerXDL>{})),
|
||||
make_pass_through_transform(Number<K1>{})),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}));
|
||||
make_tuple(
|
||||
make_merge_transform_v3_division_mod(make_tuple(Number<B_K0>{}, Number<B_K1>{})),
|
||||
make_unmerge_transform(
|
||||
make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerXDL>{}))),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
|
||||
}
|
||||
|
||||
static constexpr auto a_block_desc_k0_m0_m1_m2_k1 = MakeABlockDescriptor_K0_M0_M1_M2_K1();
|
||||
static constexpr auto b_block_desc_k0_n0_n1_n2_k1 = MakeBBlockDescriptor_K0_N0_N1_N2_K1();
|
||||
static constexpr auto a_block_desc_m0_m1_m2_k = MakeABlockDescriptor_M0_M1_M2_K();
|
||||
static constexpr auto b_block_desc_n0_n1_n2_k = MakeBBlockDescriptor_N0_N1_N2_K();
|
||||
|
||||
template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
|
||||
__device__ void Run(const ABlockBuffer& a_block_buf,
|
||||
@@ -258,31 +257,31 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
|
||||
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
// read A
|
||||
a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
|
||||
make_tuple(I0, m0, I0, I0, I0),
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, I0),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(I0, I0, I0, I0, I0),
|
||||
make_tuple(I0, I0, I0, I0),
|
||||
a_thread_buf);
|
||||
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
// read B
|
||||
b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
|
||||
make_tuple(I0, n0, I0, I0, I0),
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, I0),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(I0, I0, I0, I0, I0),
|
||||
make_tuple(I0, I0, I0, I0),
|
||||
b_thread_buf);
|
||||
|
||||
static_for<0, K0, xdlops_gemm.K0PerXdlops>{}([&](auto k0) {
|
||||
vector_type<FloatAB, K1> a_thread_vec;
|
||||
vector_type<FloatAB, K1> b_thread_vec;
|
||||
static_for<0, KPerBlock, KPack * xdlops_gemm.K0PerXdlops>{}([&](auto k) {
|
||||
vector_type<FloatAB, KPack> a_thread_vec;
|
||||
vector_type<FloatAB, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, K1, 1>{}([&](auto i) {
|
||||
static_for<0, KPack, 1>{}([&](auto i) {
|
||||
a_thread_vec.template AsType<FloatAB>()(i) = a_thread_buf
|
||||
[Number<a_thread_desc_.CalculateOffset(make_tuple(k0, 0, 0, 0, i))>{}];
|
||||
[Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
|
||||
b_thread_vec.template AsType<FloatAB>()(i) = b_thread_buf
|
||||
[Number<b_thread_desc_.CalculateOffset(make_tuple(k0, 0, 0, 0, i))>{}];
|
||||
[Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
@@ -301,13 +300,13 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
|
||||
}
|
||||
|
||||
private:
|
||||
// A[K0, M0, M1, M2, K1]
|
||||
// A[M0, M1, M2, KPerBlock]
|
||||
static constexpr auto a_thread_desc_ =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(Number<K0>{}, I1, I1, I1, Number<K1>{}));
|
||||
make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerBlock>{}));
|
||||
|
||||
// B[K0, N0, N1, N2, K1]
|
||||
// B[N0, N1, N2, KPerBlock]
|
||||
static constexpr auto b_thread_desc_ =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(Number<K0>{}, I1, I1, I1, Number<K1>{}));
|
||||
make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerBlock>{}));
|
||||
|
||||
// C[M, N, NumRegXdlops]
|
||||
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
|
||||
@@ -315,23 +314,23 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
|
||||
|
||||
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
|
||||
FloatAB,
|
||||
decltype(a_block_desc_k0_m0_m1_m2_k1),
|
||||
decltype(a_block_desc_m0_m1_m2_k),
|
||||
decltype(a_thread_desc_),
|
||||
Sequence<K0, 1, 1, 1, K1>,
|
||||
Sequence<0, 1, 2, 3, 4>,
|
||||
4,
|
||||
K1,
|
||||
K1>;
|
||||
Sequence<1, 1, 1, KPerBlock>,
|
||||
Sequence<0, 1, 2, 3>,
|
||||
3,
|
||||
A_K1,
|
||||
A_K1>;
|
||||
|
||||
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
|
||||
FloatAB,
|
||||
decltype(b_block_desc_k0_n0_n1_n2_k1),
|
||||
decltype(b_block_desc_n0_n1_n2_k),
|
||||
decltype(b_thread_desc_),
|
||||
Sequence<K0, 1, 1, 1, K1>,
|
||||
Sequence<0, 1, 2, 3, 4>,
|
||||
4,
|
||||
K1,
|
||||
K1>;
|
||||
Sequence<1, 1, 1, KPerBlock>,
|
||||
Sequence<0, 1, 2, 3>,
|
||||
3,
|
||||
B_K1,
|
||||
B_K1>;
|
||||
|
||||
AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
|
||||
BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
|
||||
|
||||
@@ -16,8 +16,8 @@ namespace ck {
|
||||
template <typename GridwiseGemm,
|
||||
typename FloatAB,
|
||||
typename FloatC,
|
||||
typename AGridDesc_K0_M_K1,
|
||||
typename BGridDesc_K0_N_K1,
|
||||
typename AGridDesc_AK0_M_AK1,
|
||||
typename BGridDesc_BK0_N_BK1,
|
||||
typename CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
@@ -32,8 +32,8 @@ __global__ void
|
||||
const FloatAB* __restrict__ p_a_grid,
|
||||
const FloatAB* __restrict__ p_b_grid,
|
||||
FloatC* __restrict__ p_c_grid,
|
||||
const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
|
||||
const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
|
||||
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
|
||||
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
|
||||
const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
|
||||
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
|
||||
const AElementwiseOperation a_element_op,
|
||||
@@ -48,8 +48,8 @@ __global__ void
|
||||
p_b_grid,
|
||||
p_c_grid,
|
||||
p_shared,
|
||||
a_grid_desc_k0_m_k1,
|
||||
b_grid_desc_k0_n_k1,
|
||||
a_grid_desc_ak0_m_ak1,
|
||||
b_grid_desc_bk0_n_bk1,
|
||||
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
@@ -63,21 +63,22 @@ template <
|
||||
typename FloatAcc,
|
||||
typename FloatC,
|
||||
InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
|
||||
typename AGridDesc_K0_M_K1,
|
||||
typename BGridDesc_K0_N_K1,
|
||||
typename AGridDesc_AK0_M_AK1,
|
||||
typename BGridDesc_BK0_N_BK1,
|
||||
typename CGridDesc_M_N,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CElementwiseOperation,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t K0PerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t AK1Value,
|
||||
index_t BK1Value,
|
||||
index_t MPerXdl,
|
||||
index_t NPerXdl,
|
||||
index_t K1Value,
|
||||
index_t MXdlPerWave,
|
||||
index_t NXdlPerWave,
|
||||
typename ABlockTransferThreadClusterLengths_K0_M_K1,
|
||||
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
typename ABlockTransferThreadClusterArrangeOrder,
|
||||
typename ABlockTransferSrcAccessOrder,
|
||||
index_t ABlockTransferSrcVectorDim,
|
||||
@@ -85,7 +86,7 @@ template <
|
||||
index_t ABlockTransferDstScalarPerVector_K1,
|
||||
bool AThreadTransferSrcResetCoordinateAfterRun,
|
||||
bool ABlockLdsExtraM,
|
||||
typename BBlockTransferThreadClusterLengths_K0_N_K1,
|
||||
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
|
||||
typename BBlockTransferThreadClusterArrangeOrder,
|
||||
typename BBlockTransferSrcAccessOrder,
|
||||
index_t BBlockTransferSrcVectorDim,
|
||||
@@ -110,50 +111,53 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
static constexpr auto I7 = Number<7>{};
|
||||
|
||||
// K1 should be Number<...>
|
||||
static constexpr auto K1 = Number<K1Value>{};
|
||||
static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
|
||||
static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
|
||||
static constexpr auto AK1 = Number<AK1Value>{};
|
||||
static constexpr auto BK1 = Number<BK1Value>{};
|
||||
|
||||
__host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
|
||||
__host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
|
||||
{
|
||||
constexpr auto max_lds_align = K1;
|
||||
constexpr auto max_lds_align = AK1;
|
||||
|
||||
// A matrix in LDS memory, dst of blockwise copy
|
||||
constexpr auto a_block_desc_k0_m_k1 = [&]() {
|
||||
constexpr auto a_block_desc_ak0_m_ak1 = [&]() {
|
||||
if constexpr(ABlockLdsExtraM)
|
||||
{
|
||||
return make_naive_tensor_descriptor(
|
||||
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
|
||||
make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
|
||||
make_tuple(AK0, Number<MPerBlock>{}, AK1),
|
||||
make_tuple(Number<MPerBlock + 1>{} * AK1, AK1, I1));
|
||||
}
|
||||
else
|
||||
{
|
||||
return make_naive_tensor_descriptor_aligned(
|
||||
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
|
||||
make_tuple(AK0, Number<MPerBlock>{}, AK1), max_lds_align);
|
||||
}
|
||||
}();
|
||||
|
||||
return a_block_desc_k0_m_k1;
|
||||
return a_block_desc_ak0_m_ak1;
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
|
||||
__host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
|
||||
{
|
||||
constexpr auto max_lds_align = K1;
|
||||
constexpr auto max_lds_align = BK1;
|
||||
|
||||
// B matrix in LDS memory, dst of blockwise copy
|
||||
constexpr auto b_block_desc_k0_n_k1 = [&]() {
|
||||
constexpr auto b_block_desc_bk0_n_bk1 = [&]() {
|
||||
if constexpr(BBlockLdsExtraN)
|
||||
{
|
||||
return make_naive_tensor_descriptor(
|
||||
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
|
||||
make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
|
||||
make_tuple(BK0, Number<NPerBlock>{}, BK1),
|
||||
make_tuple(Number<NPerBlock + 1>{} * BK1, BK1, I1));
|
||||
}
|
||||
else
|
||||
{
|
||||
return make_naive_tensor_descriptor_aligned(
|
||||
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
|
||||
make_tuple(BK0, Number<NPerBlock>{}, BK1), max_lds_align);
|
||||
}
|
||||
}();
|
||||
|
||||
return b_block_desc_k0_n_k1;
|
||||
return b_block_desc_bk0_n_bk1;
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr auto
|
||||
@@ -178,17 +182,15 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
|
||||
{
|
||||
// LDS allocation for A and B: be careful of alignment
|
||||
constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
|
||||
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
|
||||
|
||||
constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
|
||||
|
||||
constexpr auto max_lds_align = K1;
|
||||
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
|
||||
|
||||
constexpr auto a_block_space_size_aligned =
|
||||
math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
|
||||
math::integer_least_multiple(a_block_desc_ak0_m_ak1.GetElementSpaceSize(), AK1);
|
||||
|
||||
constexpr auto b_block_space_size_aligned =
|
||||
math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
|
||||
math::integer_least_multiple(b_block_desc_bk0_n_bk1.GetElementSpaceSize(), BK1);
|
||||
|
||||
// LDS allocation for C shuffle in LDS
|
||||
constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
|
||||
@@ -205,29 +207,28 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
|
||||
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
|
||||
__host__ __device__ static constexpr bool
|
||||
CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
|
||||
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
|
||||
CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
|
||||
const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
|
||||
const CGridDesc_M_N& c_grid_desc_m_n,
|
||||
index_t M01,
|
||||
index_t N01)
|
||||
{
|
||||
static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
|
||||
"wrong! K1 need to be known at compile-time");
|
||||
// static_assert(is_known_at_compile_time<remove_cv_t<decltype(AK1)>>::value &&
|
||||
// is_known_at_compile_time<remove_cv_t<decltype(BK1)>>::value,
|
||||
// "wrong! K1 need to be known at compile-time");
|
||||
|
||||
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
|
||||
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
|
||||
"Invalid tuning param!");
|
||||
|
||||
const auto M = a_grid_desc_k0_m_k1.GetLength(I1);
|
||||
const auto N = b_grid_desc_k0_n_k1.GetLength(I1);
|
||||
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
|
||||
const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
|
||||
const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
|
||||
const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
|
||||
|
||||
if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
|
||||
K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
|
||||
K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
|
||||
if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
|
||||
return false;
|
||||
|
||||
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
|
||||
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
|
||||
return false;
|
||||
|
||||
// check NumPrefetch
|
||||
@@ -239,7 +240,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
{
|
||||
// 2-stage prefetch currently only support even number of K0 loop
|
||||
// TODO: add support for odd number of K0 loop
|
||||
if(!((K0 / K0PerBlock) % 2 == 0))
|
||||
if(!((K / KPerBlock) % 2 == 0))
|
||||
{
|
||||
return false;
|
||||
}
|
||||
@@ -277,7 +278,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
// TODO move this function into GEMM-pipeline class
|
||||
__host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
|
||||
{
|
||||
const bool has_main_k0_block_loop = (K0 / (NumPrefetch * K0PerBlock)) > 1;
|
||||
const bool has_main_k0_block_loop = ((K0 * AK1) / (NumPrefetch * KPerBlock)) > 1;
|
||||
|
||||
return has_main_k0_block_loop;
|
||||
}
|
||||
@@ -357,8 +358,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
const FloatAB* __restrict__ p_b_grid,
|
||||
FloatC* __restrict__ p_c_grid,
|
||||
void* __restrict__ p_shared,
|
||||
const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
|
||||
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
|
||||
const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
|
||||
const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
|
||||
const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
|
||||
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
|
||||
const AElementwiseOperation& a_element_op,
|
||||
@@ -367,16 +368,14 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
const Block2CTileMap& block_2_ctile_map)
|
||||
{
|
||||
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
|
||||
p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
|
||||
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
|
||||
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
|
||||
p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
|
||||
p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
|
||||
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
|
||||
p_c_grid,
|
||||
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
|
||||
.GetElementSpaceSize());
|
||||
|
||||
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
|
||||
|
||||
// divide block work by [M, N]
|
||||
const auto block_work_idx =
|
||||
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
|
||||
@@ -389,13 +388,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
|
||||
|
||||
// lds max alignment
|
||||
constexpr auto max_lds_align = K1;
|
||||
constexpr auto max_lds_align = math::lcm(AK1, BK1);
|
||||
|
||||
// A matrix in LDS memory, dst of blockwise copy
|
||||
constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
|
||||
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
|
||||
|
||||
// B matrix in LDS memory, dst of blockwise copy
|
||||
constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
|
||||
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
|
||||
|
||||
// A matrix blockwise copy
|
||||
auto a_blockwise_copy =
|
||||
@@ -403,13 +402,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
AElementwiseOperation,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
InMemoryDataOperationEnum_t::Set,
|
||||
Sequence<K0PerBlock, MPerBlock, K1>,
|
||||
ABlockTransferThreadClusterLengths_K0_M_K1,
|
||||
Sequence<AK0, MPerBlock, AK1>,
|
||||
ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
ABlockTransferThreadClusterArrangeOrder,
|
||||
FloatAB,
|
||||
FloatAB,
|
||||
decltype(a_grid_desc_k0_m_k1),
|
||||
decltype(a_block_desc_k0_m_k1),
|
||||
decltype(a_grid_desc_ak0_m_ak1),
|
||||
decltype(a_block_desc_ak0_m_ak1),
|
||||
ABlockTransferSrcAccessOrder,
|
||||
Sequence<1, 0, 2>,
|
||||
ABlockTransferSrcVectorDim,
|
||||
@@ -421,10 +420,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
AThreadTransferSrcResetCoordinateAfterRun,
|
||||
true,
|
||||
NumPrefetch>(
|
||||
a_grid_desc_k0_m_k1,
|
||||
a_grid_desc_ak0_m_ak1,
|
||||
make_multi_index(0, m_block_data_idx_on_grid, 0),
|
||||
a_element_op,
|
||||
a_block_desc_k0_m_k1,
|
||||
a_block_desc_ak0_m_ak1,
|
||||
make_multi_index(0, 0, 0),
|
||||
ck::tensor_operation::element_wise::PassThrough{});
|
||||
|
||||
@@ -434,13 +433,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
BElementwiseOperation,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
InMemoryDataOperationEnum_t::Set,
|
||||
Sequence<K0PerBlock, NPerBlock, K1>,
|
||||
BBlockTransferThreadClusterLengths_K0_N_K1,
|
||||
Sequence<BK0, NPerBlock, BK1>,
|
||||
BBlockTransferThreadClusterLengths_BK0_N_BK1,
|
||||
BBlockTransferThreadClusterArrangeOrder,
|
||||
FloatAB,
|
||||
FloatAB,
|
||||
decltype(b_grid_desc_k0_n_k1),
|
||||
decltype(b_block_desc_k0_n_k1),
|
||||
decltype(b_grid_desc_bk0_n_bk1),
|
||||
decltype(b_block_desc_bk0_n_bk1),
|
||||
BBlockTransferSrcAccessOrder,
|
||||
Sequence<1, 0, 2>,
|
||||
BBlockTransferSrcVectorDim,
|
||||
@@ -452,10 +451,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
BThreadTransferSrcResetCoordinateAfterRun,
|
||||
true,
|
||||
NumPrefetch>(
|
||||
b_grid_desc_k0_n_k1,
|
||||
b_grid_desc_bk0_n_bk1,
|
||||
make_multi_index(0, n_block_data_idx_on_grid, 0),
|
||||
b_element_op,
|
||||
b_block_desc_k0_n_k1,
|
||||
b_block_desc_bk0_n_bk1,
|
||||
make_multi_index(0, 0, 0),
|
||||
ck::tensor_operation::element_wise::PassThrough{});
|
||||
|
||||
@@ -466,45 +465,47 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
|
||||
// register
|
||||
// sanity check
|
||||
constexpr index_t k_pack = math::max(
|
||||
math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
|
||||
|
||||
auto blockwise_gemm =
|
||||
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
|
||||
FloatAB,
|
||||
FloatAcc,
|
||||
decltype(a_block_desc_k0_m_k1),
|
||||
decltype(b_block_desc_k0_n_k1),
|
||||
decltype(a_block_desc_ak0_m_ak1),
|
||||
decltype(b_block_desc_bk0_n_bk1),
|
||||
MPerXdl,
|
||||
NPerXdl,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
K1>{};
|
||||
k_pack>{};
|
||||
|
||||
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
|
||||
|
||||
// LDS allocation for A and B: be careful of alignment
|
||||
constexpr auto a_block_space_size_aligned =
|
||||
math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
|
||||
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
|
||||
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
|
||||
|
||||
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
|
||||
static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
|
||||
static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
|
||||
|
||||
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
|
||||
static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
|
||||
b_block_desc_k0_n_k1.GetElementSpaceSize());
|
||||
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
|
||||
|
||||
constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
|
||||
constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
|
||||
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
|
||||
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
|
||||
|
||||
// gridwise GEMM pipeline
|
||||
const auto gridwise_gemm_pipeline =
|
||||
GridwiseGemmPipeline_v1<remove_cvref_t<decltype(a_grid_desc_k0_m_k1)>,
|
||||
remove_cvref_t<decltype(a_block_desc_k0_m_k1)>,
|
||||
GridwiseGemmPipeline_v1<remove_cvref_t<decltype(a_grid_desc_ak0_m_ak1)>,
|
||||
remove_cvref_t<decltype(a_block_desc_ak0_m_ak1)>,
|
||||
remove_cvref_t<decltype(a_blockwise_copy)>,
|
||||
remove_cvref_t<decltype(a_grid_buf)>,
|
||||
remove_cvref_t<decltype(a_block_buf)>,
|
||||
remove_cvref_t<decltype(a_block_slice_copy_step)>,
|
||||
remove_cvref_t<decltype(b_grid_desc_k0_n_k1)>,
|
||||
remove_cvref_t<decltype(b_block_desc_k0_n_k1)>,
|
||||
remove_cvref_t<decltype(b_grid_desc_bk0_n_bk1)>,
|
||||
remove_cvref_t<decltype(b_block_desc_bk0_n_bk1)>,
|
||||
remove_cvref_t<decltype(b_blockwise_copy)>,
|
||||
remove_cvref_t<decltype(b_grid_buf)>,
|
||||
remove_cvref_t<decltype(b_block_buf)>,
|
||||
@@ -514,23 +515,25 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
|
||||
NumPrefetch,
|
||||
HasMainK0BlockLoop>{};
|
||||
|
||||
const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
|
||||
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
|
||||
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
|
||||
KPerBlock);
|
||||
|
||||
gridwise_gemm_pipeline.Run(a_grid_desc_k0_m_k1,
|
||||
a_block_desc_k0_m_k1,
|
||||
gridwise_gemm_pipeline.Run(a_grid_desc_ak0_m_ak1,
|
||||
a_block_desc_ak0_m_ak1,
|
||||
a_blockwise_copy,
|
||||
a_grid_buf,
|
||||
a_block_buf,
|
||||
a_block_slice_copy_step,
|
||||
b_grid_desc_k0_n_k1,
|
||||
b_block_desc_k0_n_k1,
|
||||
b_grid_desc_bk0_n_bk1,
|
||||
b_block_desc_bk0_n_bk1,
|
||||
b_blockwise_copy,
|
||||
b_grid_buf,
|
||||
b_block_buf,
|
||||
b_block_slice_copy_step,
|
||||
blockwise_gemm,
|
||||
c_thread_buf,
|
||||
K0BlockMainLoop);
|
||||
num_k_block_main_loop);
|
||||
|
||||
// shuffle C and write out
|
||||
{
|
||||
|
||||
@@ -35,6 +35,7 @@
|
||||
#include "transpose_vectors.hpp"
|
||||
#include "inner_product.hpp"
|
||||
#include "element_wise_operation.hpp"
|
||||
#include "debug.hpp"
|
||||
|
||||
// TODO: remove this
|
||||
#if CK_USE_AMD_INLINE_ASM
|
||||
|
||||
77
composable_kernel/include/utility/debug.hpp
Normal file
77
composable_kernel/include/utility/debug.hpp
Normal file
@@ -0,0 +1,77 @@
|
||||
#ifndef UTILITY_DEBUG_HPP
|
||||
#define UTILITY_DEBUG_HPP
|
||||
|
||||
namespace ck {
|
||||
namespace debug {
|
||||
|
||||
namespace detail {
|
||||
template <typename T, typename Enable = void>
|
||||
struct PrintAsType;
|
||||
|
||||
template <typename T>
|
||||
struct PrintAsType<T, typename std::enable_if<std::is_floating_point<T>::value>::value>
|
||||
{
|
||||
using type = float;
|
||||
};
|
||||
|
||||
template <>
|
||||
struct PrintAsType<ck::half_t, void>
|
||||
{
|
||||
using type = float;
|
||||
};
|
||||
|
||||
template <typename T>
|
||||
struct PrintAsType<T, typename std::enable_if<std::is_integral<T>::value>::value>
|
||||
{
|
||||
using type = int;
|
||||
};
|
||||
} // namespace detail
|
||||
|
||||
// Print at runtime the data in shared memory in 128 bytes per row format given shared mem pointer
|
||||
// and the number of elements. Can optionally specify strides between elements and how many bytes'
|
||||
// worth of data per row.
|
||||
//
|
||||
// Usage example:
|
||||
//
|
||||
// debug::print_shared(a_block_buf.p_data_, index_t(a_block_desc_k0_m_k1.GetElementSpaceSize()));
|
||||
//
|
||||
template <typename T, index_t element_stride = 1, index_t row_bytes = 128>
|
||||
__device__ void print_shared(T const* p_shared, index_t num_elements)
|
||||
{
|
||||
using PrintType = typename detail::PrintAsType<T>::type;
|
||||
constexpr index_t row_elements = row_bytes / sizeof(T);
|
||||
static_assert((element_stride >= 1 && element_stride <= row_elements),
|
||||
"element_stride should between [1, row_elements]");
|
||||
|
||||
index_t wgid = blockIdx.x + blockIdx.y * gridDim.x + gridDim.x * gridDim.y * blockIdx.z;
|
||||
index_t tid =
|
||||
(threadIdx.z * (blockDim.x * blockDim.y)) + (threadIdx.y * blockDim.x) + threadIdx.x;
|
||||
|
||||
__syncthreads();
|
||||
|
||||
if(tid == 0)
|
||||
{
|
||||
printf("\nWorkgroup id %d, bytes per row %d, element stride %d\n\n",
|
||||
wgid,
|
||||
row_bytes,
|
||||
element_stride);
|
||||
for(index_t i = 0; i < num_elements; i += row_elements)
|
||||
{
|
||||
printf("elem %5d: ", i);
|
||||
for(index_t j = 0; j < row_elements; j += element_stride)
|
||||
{
|
||||
printf("%.0f ", static_cast<PrintType>(p_shared[i + j]));
|
||||
}
|
||||
|
||||
printf("\n");
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
} // namespace debug
|
||||
} // namespace ck
|
||||
|
||||
#endif // UTILITY_DEBUG_HPP
|
||||
@@ -432,10 +432,11 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
|
||||
OutElementwiseOperation,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
K0PerBlock,
|
||||
K0PerBlock * K1,
|
||||
K1, // AK1
|
||||
K1, // BK1
|
||||
MPerXdl,
|
||||
NPerXdl,
|
||||
K1,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
ABlockTransferThreadClusterLengths_K0_M_K1,
|
||||
|
||||
@@ -29,8 +29,9 @@ template <
|
||||
ck::index_t BlockSize,
|
||||
ck::index_t MPerBlock,
|
||||
ck::index_t NPerBlock,
|
||||
ck::index_t K0PerBlock,
|
||||
ck::index_t K1,
|
||||
ck::index_t KPerBlock,
|
||||
ck::index_t AK1,
|
||||
ck::index_t BK1,
|
||||
ck::index_t MPerXDL,
|
||||
ck::index_t NPerXDL,
|
||||
ck::index_t MXdlPerWave,
|
||||
@@ -61,13 +62,11 @@ struct DeviceGemmXdl_C_Shuffle
|
||||
static constexpr auto I1 = Number<1>{};
|
||||
static constexpr auto I2 = Number<2>{};
|
||||
|
||||
static constexpr auto K1Number = Number<K1>{};
|
||||
|
||||
static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
|
||||
{
|
||||
assert(K % K1 == 0);
|
||||
assert(K % AK1 == 0);
|
||||
|
||||
const index_t K0 = K / K1;
|
||||
const index_t K0 = K / AK1;
|
||||
|
||||
const auto a_grid_desc_m_k = [&]() {
|
||||
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
|
||||
@@ -80,21 +79,20 @@ struct DeviceGemmXdl_C_Shuffle
|
||||
}
|
||||
}();
|
||||
|
||||
const auto a_grid_desc_k0_m_k1 =
|
||||
transform_tensor_descriptor(a_grid_desc_m_k,
|
||||
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
|
||||
make_pass_through_transform(M)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
const auto a_grid_desc_k0_m_k1 = transform_tensor_descriptor(
|
||||
a_grid_desc_m_k,
|
||||
make_tuple(make_unmerge_transform(make_tuple(K0, AK1)), make_pass_through_transform(M)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
return a_grid_desc_k0_m_k1;
|
||||
}
|
||||
|
||||
static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
|
||||
{
|
||||
assert(K % K1 == 0);
|
||||
assert(K % BK1 == 0);
|
||||
|
||||
const index_t K0 = K / K1;
|
||||
const index_t K0 = K / BK1;
|
||||
|
||||
const auto b_grid_desc_k_n = [&]() {
|
||||
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
|
||||
@@ -107,12 +105,11 @@ struct DeviceGemmXdl_C_Shuffle
|
||||
}
|
||||
}();
|
||||
|
||||
const auto b_grid_desc_k0_n_k1 =
|
||||
transform_tensor_descriptor(b_grid_desc_k_n,
|
||||
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
|
||||
make_pass_through_transform(N)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
const auto b_grid_desc_k0_n_k1 = transform_tensor_descriptor(
|
||||
b_grid_desc_k_n,
|
||||
make_tuple(make_unmerge_transform(make_tuple(K0, BK1)), make_pass_through_transform(N)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
return b_grid_desc_k0_n_k1;
|
||||
}
|
||||
@@ -148,10 +145,11 @@ struct DeviceGemmXdl_C_Shuffle
|
||||
CElementwiseOperation,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
K0PerBlock,
|
||||
KPerBlock,
|
||||
AK1,
|
||||
BK1,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
K1,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
ABlockTransferThreadClusterLengths_K0_M_K1,
|
||||
@@ -461,7 +459,9 @@ struct DeviceGemmXdl_C_Shuffle
|
||||
<< BlockSize << ", "
|
||||
<< MPerBlock << ", "
|
||||
<< NPerBlock << ", "
|
||||
<< K0PerBlock
|
||||
<< KPerBlock << ", "
|
||||
<< AK1 << ", "
|
||||
<< BK1
|
||||
<< ">";
|
||||
// clang-format on
|
||||
|
||||
|
||||
@@ -23,23 +23,23 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
|
||||
using device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//#####################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Num|
|
||||
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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| Prefetch|
|
||||
//#####################| | | | | | | | 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| |
|
||||
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>
|
||||
//#####################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Num|
|
||||
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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| Prefetch|
|
||||
//#####################| | | | | | | | 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| |
|
||||
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
|
||||
@@ -21,23 +21,30 @@ using S = ck::Sequence<Is...>;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
|
||||
using device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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|
|
||||
//#####################| | | | | | | | 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|
|
||||
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
using device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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|
|
||||
//#####################| | | | | | | | 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|
|
||||
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 4, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
|
||||
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
|
||||
|
||||
@@ -21,23 +21,30 @@ using S = ck::Sequence<Is...>;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
|
||||
using device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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|
|
||||
//#####################| | | | | | | | 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|
|
||||
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
using device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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|
|
||||
//#####################| | | | | | | | 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|
|
||||
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
|
||||
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
|
||||
|
||||
@@ -21,23 +21,30 @@ using S = ck::Sequence<Is...>;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
|
||||
using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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|
|
||||
//#####################| | | | | | | | 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|
|
||||
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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|
|
||||
//#####################| | | | | | | | 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|
|
||||
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
|
||||
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
|
||||
|
||||
@@ -21,28 +21,27 @@ using S = ck::Sequence<Is...>;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
|
||||
using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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|
|
||||
//#####################| | | | | | | | 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|
|
||||
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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|
|
||||
//#####################| | | | | | | | 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|
|
||||
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
|
||||
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
|
||||
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
|
||||
|
||||
@@ -45,7 +45,7 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpeciali
|
||||
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization_t::MNPadding;
|
||||
|
||||
// clang-format off
|
||||
#if 1
|
||||
#if 0
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
|
||||
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| Num|
|
||||
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch|
|
||||
@@ -53,6 +53,13 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
// [256, 128, 4, 8], 1 stage, 2 occupancy
|
||||
< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1>;
|
||||
#elif 1
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle
|
||||
//######|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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|
|
||||
//######| | | | | | | | 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|
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>;
|
||||
#elif 0
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
|
||||
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| Num|
|
||||
@@ -82,14 +89,6 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
|
||||
// [ 64, 144, 4, 8], 2 stage, 2 occupancy
|
||||
// 85 TFlops
|
||||
// < F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 144, 4, 8, 16, 16, 1, 9, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 2>;
|
||||
#elif 1
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle
|
||||
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Num|
|
||||
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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| Prefetch|
|
||||
//######| | | | | | | | 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| |
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
// [128, 144, 8, 8], 1 stage, 1 occupancy, bounded by LDS size
|
||||
< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 144, 8, 8, 16, 16, 2, 9, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 8, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 1, 9, S<1, 1, 8, 1, 9, 2>, 8, 1>;
|
||||
#endif
|
||||
// clang-format on
|
||||
|
||||
@@ -156,8 +155,8 @@ int main(int argc, char* argv[])
|
||||
|
||||
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
|
||||
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
||||
Tensor<BDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
Tensor<BDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
|
||||
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
|
||||
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
|
||||
@@ -170,9 +169,13 @@ int main(int argc, char* argv[])
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
|
||||
break;
|
||||
default:
|
||||
case 2:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
break;
|
||||
default:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
|
||||
}
|
||||
|
||||
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
|
||||
|
||||
@@ -154,4 +154,15 @@ struct GeneratorTensor_Checkboard
|
||||
}
|
||||
};
|
||||
|
||||
template <ck::index_t Dim>
|
||||
struct GeneratorTensor_Sequential
|
||||
{
|
||||
template <typename... Ts>
|
||||
float operator()(Ts... Xs) const
|
||||
{
|
||||
std::array<ck::index_t, sizeof...(Ts)> dims = {{static_cast<ck::index_t>(Xs)...}};
|
||||
return dims[Dim];
|
||||
}
|
||||
};
|
||||
|
||||
#endif
|
||||
|
||||
Reference in New Issue
Block a user