mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-18 03:49:41 +00:00
[GEMM] F8 GEMM, performance optimized. (#1384)
* add ab_scale init support * enabled interwave * add scale type; update isSupport * adjust example * clean * enable f8 pure gemm rcr ckprofiler * Add gemm_multiply_multiply instances * clang format * Optimize for ScaleBlockMNK=128 * enable abscale f8 gemm ck profiler * Add pure f8 gemm test suite * Reverting to the state of project atf60fd77* update copyright * clang format * update copyright --------- Co-authored-by: root <jizhan@amd.com> [ROCm/composable_kernel commit:8c90f25be3]
This commit is contained in:
Haocong WANG
@@ -0,0 +1,117 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
enum struct BlockGemmPipelineVersion
|
||||
{
|
||||
v1, // Naive
|
||||
v2, // Mem
|
||||
v3, // Comp
|
||||
};
|
||||
|
||||
template <BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSche,
|
||||
index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
typename AccDataType,
|
||||
typename ATileDesc,
|
||||
typename BTileDesc,
|
||||
typename AMmaTileDesc,
|
||||
typename BMmaTileDesc,
|
||||
index_t ABlockTransferSrcScalarPerVector,
|
||||
index_t BBlockTransferSrcScalarPerVector,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t MPerXDL,
|
||||
index_t NPerXDL,
|
||||
index_t MRepeat,
|
||||
index_t NRepeat,
|
||||
index_t KPack>
|
||||
constexpr auto BlockGemmABScalePipeline_Selector()
|
||||
{
|
||||
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
|
||||
{
|
||||
return BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlkGemmPipeSche,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>{};
|
||||
}
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
|
||||
{
|
||||
return BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlkGemmPipeSche,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>{};
|
||||
}
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
|
||||
{
|
||||
return BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlkGemmPipeSche,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>{};
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cerr << "BlockGemmPipeline configuration is not available" << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,418 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
// Naive pipeline with lowest resource request per WGP
|
||||
// GlobalPrefetchStages: 1
|
||||
// LocalPreFillStages: 1
|
||||
// LocalPreFetchStages: 0
|
||||
// LocalSharedMemoryBuffer: 1
|
||||
|
||||
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
|
||||
index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
typename AccDataType,
|
||||
typename ATileDesc,
|
||||
typename BTileDesc,
|
||||
typename AMmaTileDesc,
|
||||
typename BMmaTileDesc,
|
||||
index_t ABlockTransferSrcScalarPerVector,
|
||||
index_t BBlockTransferSrcScalarPerVector,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t MPerXDL,
|
||||
index_t NPerXDL,
|
||||
index_t MRepeat,
|
||||
index_t NRepeat,
|
||||
index_t KPacks>
|
||||
struct BlockwiseGemmXdlops_pipeline_v1_ab_scale
|
||||
{
|
||||
};
|
||||
|
||||
template <index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
typename AccDataType,
|
||||
typename ATileDesc,
|
||||
typename BTileDesc,
|
||||
typename AMmaTileDesc,
|
||||
typename BMmaTileDesc,
|
||||
index_t ABlockTransferSrcScalarPerVector,
|
||||
index_t BBlockTransferSrcScalarPerVector,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t MPerXDL,
|
||||
index_t NPerXDL,
|
||||
index_t MRepeat,
|
||||
index_t NRepeat,
|
||||
index_t KPack
|
||||
// ,bool TransposeC //disable transposec right now...
|
||||
>
|
||||
struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intrawave,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>
|
||||
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>
|
||||
|
||||
{
|
||||
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>;
|
||||
using Base::I0;
|
||||
using Base::KRepeat;
|
||||
using Base::xdlops_gemm;
|
||||
|
||||
using Base::CalculateCThreadOriginDataIndex;
|
||||
using Base::CalculateCThreadOriginDataIndex8D;
|
||||
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
|
||||
using Base::GetCThreadBuffer;
|
||||
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
|
||||
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
|
||||
using Base::a_block_desc_m0_m1_m2_k;
|
||||
using Base::b_block_desc_n0_n1_n2_k;
|
||||
|
||||
using Base::AMmaKStride;
|
||||
using Base::BMmaKStride;
|
||||
|
||||
static constexpr index_t PrefetchStages = 1;
|
||||
static constexpr index_t PrefillStages = 1;
|
||||
static constexpr index_t GlobalBufferNum = 1;
|
||||
|
||||
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
|
||||
{
|
||||
return num_loop > PrefetchStages;
|
||||
}
|
||||
|
||||
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
|
||||
{
|
||||
ignore = num_loop;
|
||||
return TailNumber::Full;
|
||||
}
|
||||
|
||||
template <bool HasMainLoop,
|
||||
TailNumber TailNum,
|
||||
typename AGridDesc,
|
||||
typename ABlockDesc,
|
||||
typename ABlockTransfer,
|
||||
typename AGridBuffer,
|
||||
typename ABlockBuffer,
|
||||
typename ABlockTransferStep,
|
||||
typename BGridDesc,
|
||||
typename BBlockDesc,
|
||||
typename BBlockTransfer,
|
||||
typename BGridBuffer,
|
||||
typename BBlockBuffer,
|
||||
typename BBlockTransferStep,
|
||||
typename CThreadBuffer,
|
||||
typename AScaleGridBuffer,
|
||||
typename AScaleGridDesc,
|
||||
typename AScaleThreadDesc,
|
||||
typename AScaleThreadTransfer,
|
||||
typename AScaleThreadTransferStep,
|
||||
typename BScaleGridBuffer,
|
||||
typename BScaleGridDesc,
|
||||
typename BScaleThreadDesc,
|
||||
typename BScaleThreadTransfer,
|
||||
typename BScaleThreadTransferStep>
|
||||
__device__ void Run(
|
||||
// ABlockCopy
|
||||
const AGridDesc& a_grid_desc,
|
||||
const ABlockDesc& a_block_desc,
|
||||
ABlockTransfer& a_blockwise_copy,
|
||||
const AGridBuffer& a_grid_buf,
|
||||
ABlockBuffer& a_block_buf,
|
||||
const ABlockTransferStep& a_block_copy_step,
|
||||
// BBlockCopy
|
||||
const BGridDesc& b_grid_desc,
|
||||
const BBlockDesc& b_block_desc,
|
||||
BBlockTransfer& b_blockwise_copy,
|
||||
const BGridBuffer& b_grid_buf,
|
||||
BBlockBuffer& b_block_buf,
|
||||
const BBlockTransferStep& b_block_copy_step,
|
||||
// CThread
|
||||
CThreadBuffer& c_thread_buf,
|
||||
// AScaleThreadCopy
|
||||
const AScaleGridDesc& a_scale_grid_desc,
|
||||
const AScaleThreadDesc& a_scale_thread_desc,
|
||||
AScaleThreadTransfer& a_scale_thread_copy,
|
||||
const AScaleGridBuffer& a_scale_grid_buf,
|
||||
const AScaleThreadTransferStep& a_scale_thread_copy_step,
|
||||
// BScaleThreadCopy
|
||||
const BScaleGridDesc& b_scale_grid_desc,
|
||||
const BScaleThreadDesc& b_scale_thread_desc,
|
||||
BScaleThreadTransfer& b_scale_thread_copy,
|
||||
const BScaleGridBuffer& b_scale_grid_buf,
|
||||
const BScaleThreadTransferStep& b_scale_thread_copy_step,
|
||||
// num_loop
|
||||
index_t num_loop,
|
||||
index_t num_loop_per_scale) const
|
||||
{
|
||||
// assume kperblock = scaleblockk
|
||||
ignore = num_loop_per_scale;
|
||||
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
a_thread_desc_.GetElementSpaceSize());
|
||||
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
b_thread_desc_.GetElementSpaceSize());
|
||||
auto a_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
|
||||
a_scale_thread_desc.GetElementSpaceSize());
|
||||
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
|
||||
b_scale_thread_desc.GetElementSpaceSize());
|
||||
|
||||
// Global prefetch 1
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
a_scale_thread_copy.Run(a_scale_grid_desc,
|
||||
a_scale_grid_buf,
|
||||
a_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
a_scale_thread_buf);
|
||||
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
b_scale_thread_buf);
|
||||
|
||||
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
|
||||
|
||||
// Local prefill 1
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
|
||||
|
||||
// Initialize C
|
||||
c_thread_buf.Clear();
|
||||
|
||||
auto c_thread_buf_per_scale = remove_cvref_t<decltype(c_thread_buf)>();
|
||||
|
||||
// main body
|
||||
if constexpr(HasMainLoop)
|
||||
{
|
||||
index_t i = 0;
|
||||
do
|
||||
{
|
||||
// -------------------------------------------------------------------------------------------
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
block_sync_lds();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_buf);
|
||||
});
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
c_thread_buf_per_scale.Clear();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType,
|
||||
xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf_per_scale.GetVectorTypeReference(I0));
|
||||
});
|
||||
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
|
||||
c_thread_buf(Number<c_offset>{}) +=
|
||||
c_thread_buf_per_scale[Number<t>{}] *
|
||||
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
|
||||
type_convert<AccDataType>(b_scale_thread_buf[I0]);
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
a_scale_thread_copy.Run(a_scale_grid_desc,
|
||||
a_scale_grid_buf,
|
||||
a_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
a_scale_thread_buf);
|
||||
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
b_scale_thread_buf);
|
||||
|
||||
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
|
||||
|
||||
block_sync_lds();
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
|
||||
|
||||
i += 1;
|
||||
|
||||
} while(i < (num_loop - 1));
|
||||
}
|
||||
|
||||
// tail
|
||||
if constexpr(TailNum == TailNumber::Full)
|
||||
{
|
||||
block_sync_lds();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_buf);
|
||||
});
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
c_thread_buf_per_scale.Clear();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf_per_scale.GetVectorTypeReference(I0));
|
||||
});
|
||||
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
|
||||
c_thread_buf(Number<c_offset>{}) +=
|
||||
c_thread_buf_per_scale[Number<t>{}] *
|
||||
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
|
||||
type_convert<AccDataType>(b_scale_thread_buf[I0]);
|
||||
});
|
||||
});
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
protected:
|
||||
using Base::a_thread_copy_;
|
||||
using Base::a_thread_desc_;
|
||||
using Base::b_thread_copy_;
|
||||
using Base::b_thread_desc_;
|
||||
using Base::c_thread_desc_;
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,631 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
// Maximum Global Memory throughput pipeline with >=32KB data in fly
|
||||
// GlobalPrefetchStages: >=2
|
||||
// LocalPreFillStages: 1
|
||||
// LocalPreFetchStages: 0
|
||||
// LocalSharedMemoryBuffer: 1
|
||||
|
||||
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
|
||||
index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
typename AccDataType,
|
||||
typename ATileDesc,
|
||||
typename BTileDesc,
|
||||
typename AMmaTileDesc,
|
||||
typename BMmaTileDesc,
|
||||
index_t ABlockTransferSrcScalarPerVector,
|
||||
index_t BBlockTransferSrcScalarPerVector,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t MPerXDL,
|
||||
index_t NPerXDL,
|
||||
index_t MRepeat,
|
||||
index_t NRepeat,
|
||||
index_t KPacks>
|
||||
struct BlockwiseGemmXdlops_pipeline_v2_ab_scale
|
||||
{
|
||||
};
|
||||
|
||||
template <index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
typename AccDataType,
|
||||
typename ATileDesc,
|
||||
typename BTileDesc,
|
||||
typename AMmaTileDesc,
|
||||
typename BMmaTileDesc,
|
||||
index_t ABlockTransferSrcScalarPerVector,
|
||||
index_t BBlockTransferSrcScalarPerVector,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t MPerXDL,
|
||||
index_t NPerXDL,
|
||||
index_t MRepeat,
|
||||
index_t NRepeat,
|
||||
index_t KPack
|
||||
// ,bool TransposeC //disable transposec right now...
|
||||
>
|
||||
struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intrawave,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>
|
||||
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>
|
||||
|
||||
{
|
||||
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>;
|
||||
using Base::I0;
|
||||
using Base::KRepeat;
|
||||
using Base::xdlops_gemm;
|
||||
|
||||
using Base::CalculateCThreadOriginDataIndex;
|
||||
using Base::CalculateCThreadOriginDataIndex8D;
|
||||
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
|
||||
using Base::GetCThreadBuffer;
|
||||
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
|
||||
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
|
||||
using Base::a_block_desc_m0_m1_m2_k;
|
||||
using Base::b_block_desc_n0_n1_n2_k;
|
||||
|
||||
using Base::AMmaKStride;
|
||||
using Base::BMmaKStride;
|
||||
|
||||
static constexpr index_t WgpPerCU =
|
||||
(4 * warpSize / BlockSize) >= 1 ? 4 * warpSize / BlockSize : 1;
|
||||
static constexpr index_t FullMemBandPrefetchStages = math::integer_divide_ceil(
|
||||
32768 / WgpPerCU,
|
||||
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);
|
||||
static constexpr index_t PrefetchStages =
|
||||
FullMemBandPrefetchStages >= 2
|
||||
? FullMemBandPrefetchStages <= 8 ? FullMemBandPrefetchStages : 8
|
||||
: 2;
|
||||
|
||||
static constexpr index_t PrefillStages = 1;
|
||||
static constexpr index_t GlobalBufferNum = PrefetchStages;
|
||||
|
||||
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
|
||||
{
|
||||
return num_loop > PrefetchStages;
|
||||
}
|
||||
|
||||
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
|
||||
{
|
||||
if(num_loop % PrefetchStages == 1)
|
||||
{
|
||||
return TailNumber::One;
|
||||
}
|
||||
else if(num_loop % PrefetchStages == 2)
|
||||
{
|
||||
return TailNumber::Two;
|
||||
}
|
||||
else if(num_loop % PrefetchStages == 3)
|
||||
{
|
||||
return TailNumber::Three;
|
||||
}
|
||||
else if(num_loop % PrefetchStages == 4)
|
||||
{
|
||||
return TailNumber::Four;
|
||||
}
|
||||
else if(num_loop % PrefetchStages == 5)
|
||||
{
|
||||
return TailNumber::Five;
|
||||
}
|
||||
else if(num_loop % PrefetchStages == 6)
|
||||
{
|
||||
return TailNumber::Six;
|
||||
}
|
||||
else if(num_loop % PrefetchStages == 7)
|
||||
{
|
||||
return TailNumber::Seven;
|
||||
}
|
||||
else
|
||||
{
|
||||
return TailNumber::Full;
|
||||
}
|
||||
}
|
||||
|
||||
template <bool HasMainLoop,
|
||||
TailNumber TailNum,
|
||||
typename AGridDesc,
|
||||
typename ABlockDesc,
|
||||
typename ABlockTransfer,
|
||||
typename AGridBuffer,
|
||||
typename ABlockBuffer,
|
||||
typename ABlockTransferStep,
|
||||
typename BGridDesc,
|
||||
typename BBlockDesc,
|
||||
typename BBlockTransfer,
|
||||
typename BGridBuffer,
|
||||
typename BBlockBuffer,
|
||||
typename BBlockTransferStep,
|
||||
typename CThreadBuffer,
|
||||
typename AScaleGridBuffer,
|
||||
typename AScaleGridDesc,
|
||||
typename AScaleThreadDesc,
|
||||
typename AScaleThreadTransfer,
|
||||
typename AScaleThreadTransferStep,
|
||||
typename BScaleGridBuffer,
|
||||
typename BScaleGridDesc,
|
||||
typename BScaleThreadDesc,
|
||||
typename BScaleThreadTransfer,
|
||||
typename BScaleThreadTransferStep>
|
||||
__device__ void Run(
|
||||
// ABlockCopy
|
||||
const AGridDesc& a_grid_desc,
|
||||
const ABlockDesc& a_block_desc,
|
||||
ABlockTransfer& a_blockwise_copy,
|
||||
const AGridBuffer& a_grid_buf,
|
||||
ABlockBuffer& a_block_buf,
|
||||
const ABlockTransferStep& a_block_copy_step,
|
||||
// BBlockCopy
|
||||
const BGridDesc& b_grid_desc,
|
||||
const BBlockDesc& b_block_desc,
|
||||
BBlockTransfer& b_blockwise_copy,
|
||||
const BGridBuffer& b_grid_buf,
|
||||
BBlockBuffer& b_block_buf,
|
||||
const BBlockTransferStep& b_block_copy_step,
|
||||
// CThread
|
||||
CThreadBuffer& c_thread_buf,
|
||||
// AScaleThreadCopy
|
||||
const AScaleGridDesc& a_scale_grid_desc,
|
||||
const AScaleThreadDesc& a_scale_thread_desc,
|
||||
AScaleThreadTransfer& a_scale_thread_copy,
|
||||
const AScaleGridBuffer& a_scale_grid_buf,
|
||||
const AScaleThreadTransferStep& a_scale_thread_copy_step,
|
||||
// BScaleThreadCopy
|
||||
const BScaleGridDesc& b_scale_grid_desc,
|
||||
const BScaleThreadDesc& b_scale_thread_desc,
|
||||
BScaleThreadTransfer& b_scale_thread_copy,
|
||||
const BScaleGridBuffer& b_scale_grid_buf,
|
||||
const BScaleThreadTransferStep& b_scale_thread_copy_step,
|
||||
// num_loop
|
||||
index_t num_loop,
|
||||
index_t num_loop_per_scale) const
|
||||
{
|
||||
// assume kperblock = scaleblockk
|
||||
ignore = num_loop_per_scale;
|
||||
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
a_thread_desc_.GetElementSpaceSize());
|
||||
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
b_thread_desc_.GetElementSpaceSize());
|
||||
auto a_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
|
||||
a_scale_thread_desc.GetElementSpaceSize());
|
||||
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
|
||||
b_scale_thread_desc.GetElementSpaceSize());
|
||||
|
||||
// Global prefetch 1
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
a_scale_thread_copy.Run(a_scale_grid_desc,
|
||||
a_scale_grid_buf,
|
||||
a_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
a_scale_thread_buf);
|
||||
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
b_scale_thread_buf);
|
||||
|
||||
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
|
||||
|
||||
// Local prefill 1
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I0);
|
||||
|
||||
// Initialize C
|
||||
c_thread_buf.Clear();
|
||||
|
||||
// Global prefetch [2, PrefetchStages]
|
||||
static_for<1, PrefetchStages, 1>{}([&](auto iprefetch) {
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, iprefetch);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, iprefetch);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
});
|
||||
|
||||
auto c_thread_buf_per_scale = remove_cvref_t<decltype(c_thread_buf)>();
|
||||
|
||||
// main body
|
||||
if constexpr(HasMainLoop)
|
||||
{
|
||||
index_t i = 0;
|
||||
do
|
||||
{
|
||||
static_for<0, PrefetchStages, 1>{}([&](auto iprefetch) {
|
||||
block_sync_lds();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_buf);
|
||||
});
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
c_thread_buf_per_scale.Clear();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType,
|
||||
xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf_per_scale.GetVectorTypeReference(I0));
|
||||
});
|
||||
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
|
||||
c_thread_buf(Number<c_offset>{}) +=
|
||||
c_thread_buf_per_scale[Number<t>{}] *
|
||||
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
|
||||
type_convert<AccDataType>(b_scale_thread_buf[I0]);
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
a_scale_thread_copy.Run(a_scale_grid_desc,
|
||||
a_scale_grid_buf,
|
||||
a_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
a_scale_thread_buf);
|
||||
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
b_scale_thread_buf);
|
||||
|
||||
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
|
||||
a_scale_thread_copy_step);
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step);
|
||||
|
||||
block_sync_lds();
|
||||
a_blockwise_copy.RunWrite(
|
||||
a_block_desc, a_block_buf, Number<(iprefetch + 1) % PrefetchStages>{});
|
||||
b_blockwise_copy.RunWrite(
|
||||
b_block_desc, b_block_buf, Number<(iprefetch + 1) % PrefetchStages>{});
|
||||
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, iprefetch);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, iprefetch);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
});
|
||||
|
||||
i += PrefetchStages;
|
||||
} while(i < (num_loop - PrefetchStages));
|
||||
}
|
||||
|
||||
// tail
|
||||
auto LoopTailFunc = [&](auto tail_num) {
|
||||
static_for<1, tail_num, 1>{}([&](auto iprefetch) {
|
||||
block_sync_lds();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_buf);
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
c_thread_buf_per_scale.Clear();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType,
|
||||
xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf_per_scale.GetVectorTypeReference(I0));
|
||||
});
|
||||
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
|
||||
c_thread_buf(Number<c_offset>{}) +=
|
||||
c_thread_buf_per_scale[Number<t>{}] *
|
||||
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
|
||||
type_convert<AccDataType>(b_scale_thread_buf[I0]);
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
a_scale_thread_copy.Run(a_scale_grid_desc,
|
||||
a_scale_grid_buf,
|
||||
a_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
a_scale_thread_buf);
|
||||
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
b_scale_thread_buf);
|
||||
|
||||
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
|
||||
|
||||
block_sync_lds();
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, iprefetch);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, iprefetch);
|
||||
});
|
||||
|
||||
block_sync_lds();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_buf);
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
c_thread_buf_per_scale.Clear();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf_per_scale.GetVectorTypeReference(I0));
|
||||
});
|
||||
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
|
||||
c_thread_buf(Number<c_offset>{}) +=
|
||||
c_thread_buf_per_scale[Number<t>{}] *
|
||||
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
|
||||
type_convert<AccDataType>(b_scale_thread_buf[I0]);
|
||||
});
|
||||
});
|
||||
});
|
||||
};
|
||||
|
||||
if constexpr(TailNum == TailNumber::One)
|
||||
{
|
||||
block_sync_lds();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_buf);
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
c_thread_buf_per_scale.Clear();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf_per_scale.GetVectorTypeReference(I0));
|
||||
});
|
||||
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
|
||||
c_thread_buf(Number<c_offset>{}) +=
|
||||
c_thread_buf_per_scale[Number<t>{}] *
|
||||
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
|
||||
type_convert<AccDataType>(b_scale_thread_buf[I0]);
|
||||
});
|
||||
});
|
||||
});
|
||||
}
|
||||
else if constexpr(TailNum == TailNumber::Two)
|
||||
{
|
||||
LoopTailFunc(Number<2>{});
|
||||
}
|
||||
else if constexpr(TailNum == TailNumber::Three)
|
||||
{
|
||||
LoopTailFunc(Number<3>{});
|
||||
}
|
||||
else if constexpr(TailNum == TailNumber::Four)
|
||||
{
|
||||
LoopTailFunc(Number<4>{});
|
||||
}
|
||||
else if constexpr(TailNum == TailNumber::Five)
|
||||
{
|
||||
LoopTailFunc(Number<5>{});
|
||||
}
|
||||
else if constexpr(TailNum == TailNumber::Six)
|
||||
{
|
||||
LoopTailFunc(Number<6>{});
|
||||
}
|
||||
else if constexpr(TailNum == TailNumber::Seven)
|
||||
{
|
||||
LoopTailFunc(Number<7>{});
|
||||
}
|
||||
else if constexpr(TailNum == TailNumber::Full)
|
||||
{
|
||||
LoopTailFunc(Number<PrefetchStages>{});
|
||||
}
|
||||
}
|
||||
|
||||
protected:
|
||||
using Base::a_thread_copy_;
|
||||
using Base::a_thread_desc_;
|
||||
using Base::b_thread_copy_;
|
||||
using Base::b_thread_desc_;
|
||||
using Base::c_thread_desc_;
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,533 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
// Compute optimized pipeline
|
||||
// GlobalPrefetchStages: 2
|
||||
// LocalPreFillStages: 1
|
||||
// LocalPreFetchStages: 1
|
||||
// LocalSharedMemoryBuffer: 1
|
||||
|
||||
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
|
||||
index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
typename AccDataType,
|
||||
typename ATileDesc,
|
||||
typename BTileDesc,
|
||||
typename AMmaTileDesc,
|
||||
typename BMmaTileDesc,
|
||||
index_t ABlockTransferSrcScalarPerVector,
|
||||
index_t BBlockTransferSrcScalarPerVector,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t MPerXDL,
|
||||
index_t NPerXDL,
|
||||
index_t MRepeat,
|
||||
index_t NRepeat,
|
||||
index_t KPacks>
|
||||
struct BlockwiseGemmXdlops_pipeline_v3_ab_scale
|
||||
{
|
||||
};
|
||||
|
||||
template <index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
typename AccDataType,
|
||||
typename ATileDesc,
|
||||
typename BTileDesc,
|
||||
typename AMmaTileDesc,
|
||||
typename BMmaTileDesc,
|
||||
index_t ABlockTransferSrcScalarPerVector,
|
||||
index_t BBlockTransferSrcScalarPerVector,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t MPerXDL,
|
||||
index_t NPerXDL,
|
||||
index_t MRepeat,
|
||||
index_t NRepeat,
|
||||
index_t KPack
|
||||
// ,bool TransposeC //disable transposec right now...
|
||||
>
|
||||
struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intrawave,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>
|
||||
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>
|
||||
|
||||
{
|
||||
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>;
|
||||
using Base::I0;
|
||||
using Base::KRepeat;
|
||||
using Base::xdlops_gemm;
|
||||
using typename Base::HotLoopInstList;
|
||||
|
||||
using Base::CalculateCThreadOriginDataIndex;
|
||||
using Base::CalculateCThreadOriginDataIndex8D;
|
||||
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
|
||||
using Base::GetCThreadBuffer;
|
||||
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
|
||||
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
|
||||
|
||||
using Base::a_block_desc_m0_m1_m2_k;
|
||||
using Base::b_block_desc_n0_n1_n2_k;
|
||||
|
||||
using Base::AMmaKStride;
|
||||
using Base::BMmaKStride;
|
||||
|
||||
static constexpr index_t PrefetchStages = 2;
|
||||
static constexpr index_t PrefillStages = 1;
|
||||
static constexpr index_t GlobalBufferNum = 1;
|
||||
|
||||
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
|
||||
{
|
||||
return num_loop > PrefetchStages;
|
||||
}
|
||||
|
||||
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
|
||||
{
|
||||
ignore = num_loop;
|
||||
return TailNumber::Full;
|
||||
}
|
||||
|
||||
__device__ static constexpr auto HotLoopScheduler()
|
||||
{
|
||||
// A/B split schedule
|
||||
// compiler is likely to use ds_read2 when instruction width smaller than 16bytes
|
||||
constexpr auto num_ds_read_inst_a =
|
||||
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16
|
||||
? HotLoopInstList::A_LDS_Read_Inst_Num
|
||||
: HotLoopInstList::A_LDS_Read_Inst_Num / 2;
|
||||
constexpr auto num_ds_read_inst_b =
|
||||
HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16
|
||||
? HotLoopInstList::B_LDS_Read_Inst_Num
|
||||
: HotLoopInstList::B_LDS_Read_Inst_Num / 2;
|
||||
|
||||
constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
|
||||
constexpr auto num_ds_write_inst_b = HotLoopInstList::B_LDS_Write_Inst_Num;
|
||||
|
||||
constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
|
||||
constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
|
||||
|
||||
constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
|
||||
|
||||
constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
|
||||
constexpr auto ds_read_a_issue_cycle = 4;
|
||||
// HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
|
||||
constexpr auto ds_read_b_issue_cycle = 4;
|
||||
// HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16 ? 8 : 4;
|
||||
constexpr auto ds_read_a_mfma_rate =
|
||||
(mfma_cycle - 4 + 2 * ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
|
||||
constexpr auto ds_read_b_mfma_rate =
|
||||
(mfma_cycle - 4 + 2 * ds_read_b_issue_cycle - 1) / (2 * ds_read_b_issue_cycle);
|
||||
|
||||
constexpr auto num_dsread_a_mfma =
|
||||
(num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
|
||||
constexpr auto num_dsread_b_mfma =
|
||||
(num_ds_read_inst_b + ds_read_b_mfma_rate - 1) / ds_read_b_mfma_rate;
|
||||
|
||||
// stage 1
|
||||
// Separate this part?
|
||||
// constexpr auto num_mfma_per_ds_read = sizeof(ComputeDataType) / sizeof(ADataType) >
|
||||
// sizeof(ComputeDataType) / sizeof(BDataType)
|
||||
// ? sizeof(ComputeDataType) / sizeof(ADataType)
|
||||
// : sizeof(ComputeDataType) / sizeof(BDataType);
|
||||
constexpr auto num_mfma_stage1 = num_mfma_inst - (num_dsread_a_mfma + num_dsread_b_mfma);
|
||||
constexpr auto num_mfma_per_issue =
|
||||
num_mfma_stage1 / (num_buffer_load_inst_a + num_buffer_load_inst_b);
|
||||
constexpr auto num_dswrite_per_issue_a = num_ds_write_inst_a / num_buffer_load_inst_a;
|
||||
constexpr auto num_dswrite_per_issue_b = num_ds_write_inst_b / num_buffer_load_inst_b;
|
||||
|
||||
static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) {
|
||||
ignore = i;
|
||||
static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) {
|
||||
ignore = idswrite;
|
||||
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
|
||||
});
|
||||
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
|
||||
__builtin_amdgcn_sched_group_barrier(
|
||||
0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA
|
||||
});
|
||||
static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) {
|
||||
ignore = i;
|
||||
static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) {
|
||||
ignore = idswrite;
|
||||
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
|
||||
});
|
||||
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
|
||||
__builtin_amdgcn_sched_group_barrier(
|
||||
0x008, num_mfma_per_issue - num_dswrite_per_issue_b, 0); // MFMA
|
||||
});
|
||||
|
||||
// stage 2
|
||||
static_for<0, num_dsread_a_mfma, 1>{}([&](auto i) {
|
||||
if constexpr((num_ds_read_inst_a - (i + 1) * ds_read_a_mfma_rate) >=
|
||||
ds_read_a_mfma_rate)
|
||||
{
|
||||
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
|
||||
}
|
||||
else
|
||||
{
|
||||
__builtin_amdgcn_sched_group_barrier(0x100,
|
||||
num_ds_read_inst_a - (num_dsread_a_mfma - 1) *
|
||||
ds_read_a_mfma_rate,
|
||||
0); // DS read
|
||||
}
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
|
||||
});
|
||||
|
||||
static_for<0, num_dsread_b_mfma, 1>{}([&](auto i) {
|
||||
if constexpr((num_ds_read_inst_b - (i + 1) * ds_read_b_mfma_rate) >=
|
||||
ds_read_b_mfma_rate)
|
||||
{
|
||||
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_b_mfma_rate, 0); // DS read
|
||||
}
|
||||
else
|
||||
{
|
||||
__builtin_amdgcn_sched_group_barrier(0x100,
|
||||
num_ds_read_inst_b - (num_dsread_b_mfma - 1) *
|
||||
ds_read_b_mfma_rate,
|
||||
0); // DS read
|
||||
}
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
|
||||
});
|
||||
}
|
||||
|
||||
template <bool HasMainLoop,
|
||||
TailNumber TailNum,
|
||||
typename AGridDesc,
|
||||
typename ABlockDesc,
|
||||
typename ABlockTransfer,
|
||||
typename AGridBuffer,
|
||||
typename ABlockBuffer,
|
||||
typename ABlockTransferStep,
|
||||
typename BGridDesc,
|
||||
typename BBlockDesc,
|
||||
typename BBlockTransfer,
|
||||
typename BGridBuffer,
|
||||
typename BBlockBuffer,
|
||||
typename BBlockTransferStep,
|
||||
typename CThreadBuffer,
|
||||
typename AScaleGridBuffer,
|
||||
typename AScaleGridDesc,
|
||||
typename AScaleThreadDesc,
|
||||
typename AScaleThreadTransfer,
|
||||
typename AScaleThreadTransferStep,
|
||||
typename BScaleGridBuffer,
|
||||
typename BScaleGridDesc,
|
||||
typename BScaleThreadDesc,
|
||||
typename BScaleThreadTransfer,
|
||||
typename BScaleThreadTransferStep>
|
||||
__device__ void Run(
|
||||
// ABlockCopy
|
||||
const AGridDesc& a_grid_desc,
|
||||
const ABlockDesc& a_block_desc,
|
||||
ABlockTransfer& a_blockwise_copy,
|
||||
const AGridBuffer& a_grid_buf,
|
||||
ABlockBuffer& a_block_buf,
|
||||
const ABlockTransferStep& a_block_copy_step,
|
||||
// BBlockCopy
|
||||
const BGridDesc& b_grid_desc,
|
||||
const BBlockDesc& b_block_desc,
|
||||
BBlockTransfer& b_blockwise_copy,
|
||||
const BGridBuffer& b_grid_buf,
|
||||
BBlockBuffer& b_block_buf,
|
||||
const BBlockTransferStep& b_block_copy_step,
|
||||
// CThread
|
||||
CThreadBuffer& c_thread_buf,
|
||||
// AScaleThreadCopy
|
||||
const AScaleGridDesc& a_scale_grid_desc,
|
||||
const AScaleThreadDesc& a_scale_thread_desc,
|
||||
AScaleThreadTransfer& a_scale_thread_copy,
|
||||
const AScaleGridBuffer& a_scale_grid_buf,
|
||||
const AScaleThreadTransferStep& a_scale_thread_copy_step,
|
||||
// BScaleThreadCopy
|
||||
const BScaleGridDesc& b_scale_grid_desc,
|
||||
const BScaleThreadDesc& b_scale_thread_desc,
|
||||
BScaleThreadTransfer& b_scale_thread_copy,
|
||||
const BScaleGridBuffer& b_scale_grid_buf,
|
||||
const BScaleThreadTransferStep& b_scale_thread_copy_step,
|
||||
// num_loop
|
||||
index_t num_loop,
|
||||
index_t num_loop_per_scale) const
|
||||
{
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
// assume kperblock = scaleblockk
|
||||
ignore = num_loop_per_scale;
|
||||
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
a_thread_desc_.GetElementSpaceSize());
|
||||
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
b_thread_desc_.GetElementSpaceSize());
|
||||
auto a_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
|
||||
a_scale_thread_desc.GetElementSpaceSize());
|
||||
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
|
||||
b_scale_thread_desc.GetElementSpaceSize());
|
||||
|
||||
// Global prefetch 1
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
a_scale_thread_copy.Run(a_scale_grid_desc,
|
||||
a_scale_grid_buf,
|
||||
a_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
a_scale_thread_buf);
|
||||
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
b_scale_thread_buf);
|
||||
|
||||
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
|
||||
// Local prefill 1
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
|
||||
|
||||
// Global prefetch 2
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
// Initialize C
|
||||
c_thread_buf.Clear();
|
||||
|
||||
auto c_thread_buf_per_scale = remove_cvref_t<decltype(c_thread_buf)>();
|
||||
|
||||
// Local prefetch 1
|
||||
block_sync_lds();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k0, I0),
|
||||
a_thread_buf);
|
||||
});
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k0 * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k0, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
|
||||
// main body
|
||||
if constexpr(HasMainLoop)
|
||||
{
|
||||
index_t i = 0;
|
||||
do
|
||||
{
|
||||
block_sync_lds();
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
|
||||
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
c_thread_buf_per_scale.Clear();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType,
|
||||
xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf_per_scale.GetVectorTypeReference(I0));
|
||||
});
|
||||
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
|
||||
c_thread_buf(Number<c_offset>{}) +=
|
||||
c_thread_buf_per_scale[Number<t>{}] *
|
||||
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
|
||||
type_convert<AccDataType>(b_scale_thread_buf[I0]);
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
block_sync_lds();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_buf);
|
||||
});
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
a_scale_thread_copy.Run(a_scale_grid_desc,
|
||||
a_scale_grid_buf,
|
||||
a_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
a_scale_thread_buf);
|
||||
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(I0, I0),
|
||||
b_scale_thread_buf);
|
||||
|
||||
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
|
||||
HotLoopScheduler();
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
i += 1;
|
||||
} while(i < (num_loop - 1));
|
||||
}
|
||||
|
||||
// tail
|
||||
if constexpr(TailNum == TailNumber::Full)
|
||||
{
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
c_thread_buf_per_scale.Clear();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf_per_scale.GetVectorTypeReference(I0));
|
||||
});
|
||||
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
|
||||
c_thread_buf(Number<c_offset>{}) +=
|
||||
c_thread_buf_per_scale[Number<t>{}] *
|
||||
type_convert<AccDataType>(a_scale_thread_buf[I0]) *
|
||||
type_convert<AccDataType>(b_scale_thread_buf[I0]);
|
||||
});
|
||||
});
|
||||
});
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
}
|
||||
|
||||
protected:
|
||||
using Base::a_thread_copy_;
|
||||
using Base::a_thread_desc_;
|
||||
using Base::b_thread_copy_;
|
||||
using Base::b_thread_desc_;
|
||||
using Base::c_thread_desc_;
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,65 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <array>
|
||||
|
||||
#include "ck/tensor_operation/gpu/device/device_base.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
|
||||
// GEMM:
|
||||
// input : A[M, K], B[K, N],
|
||||
// input : D0[M, N], D1[M, N], ...
|
||||
// output : E[M, N]
|
||||
// C = a_op(A) * b_op(B)
|
||||
// E = cde_op(C, D0, D1, ...)
|
||||
// Assume:
|
||||
// D0, D1, ... and E have the same layout
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename ADataType,
|
||||
typename AScaleType,
|
||||
typename BDataType,
|
||||
typename BScaleType,
|
||||
typename DsDataType,
|
||||
typename EDataType,
|
||||
index_t ScaleBlockM,
|
||||
index_t ScaleBlockN,
|
||||
index_t ScaleBlockK,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CDEElementwiseOperation>
|
||||
struct DeviceGemmMultipleD_ABScale : public BaseOperator
|
||||
{
|
||||
static constexpr index_t NumDTensor = DsDataType::Size();
|
||||
|
||||
virtual std::unique_ptr<BaseArgument>
|
||||
MakeArgumentPointer(const void* p_a,
|
||||
const void* p_b,
|
||||
std::array<const void*, NumDTensor> p_ds,
|
||||
void* p_e,
|
||||
const ck::index_t M,
|
||||
const ck::index_t N,
|
||||
const ck::index_t K,
|
||||
const ck::index_t StrideA,
|
||||
const ck::index_t StrideB,
|
||||
const std::array<ck::index_t, NumDTensor> StrideDs,
|
||||
const ck::index_t StrideE,
|
||||
const void* p_a_scale,
|
||||
const void* p_b_scale,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation cde_element_op) = 0;
|
||||
|
||||
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -1,5 +1,5 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
@@ -182,18 +182,6 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleD<ALayout,
|
||||
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1 ||
|
||||
BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
|
||||
{
|
||||
#if 0
|
||||
if(arg.KBatch > 1)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy>;
|
||||
Run(kernel);
|
||||
}
|
||||
else
|
||||
#endif
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
@@ -206,121 +194,6 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleD<ALayout,
|
||||
// Tail number could be One to Seven
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
|
||||
{
|
||||
#if 0
|
||||
if(arg.KBatch > 1)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::One>;
|
||||
Run(kernel);
|
||||
}
|
||||
else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Full)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Full>;
|
||||
Run(kernel);
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Two>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Three)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Three>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Four)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Four>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Five)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Five>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Six>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Seven)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Seven>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
}
|
||||
else
|
||||
#endif
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
|
||||
{
|
||||
@@ -436,32 +309,7 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleD<ALayout,
|
||||
// Tail number could be Odd or Even
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
|
||||
{
|
||||
#if 0
|
||||
if(arg.KBatch > 1)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Odd>;
|
||||
Run(kernel);
|
||||
}
|
||||
else
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Even>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
else
|
||||
#endif
|
||||
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
|
||||
{
|
||||
@@ -487,32 +335,6 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleD<ALayout,
|
||||
}
|
||||
else
|
||||
{
|
||||
#if 0
|
||||
if(arg.KBatch > 1)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Odd>;
|
||||
Run(kernel);
|
||||
}
|
||||
else
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Even>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
else
|
||||
#endif
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
|
||||
{
|
||||
@@ -542,18 +364,6 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleD<ALayout,
|
||||
// Tail number always 1
|
||||
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
|
||||
{
|
||||
#if 0
|
||||
if(arg.KBatch > 1)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
false,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy>;
|
||||
Run(kernel);
|
||||
}
|
||||
else
|
||||
#endif
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
|
||||
@@ -0,0 +1,516 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <iostream>
|
||||
#include <sstream>
|
||||
|
||||
#include "ck/utility/common_header.hpp"
|
||||
#include "ck/tensor_description/tensor_descriptor.hpp"
|
||||
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_ab_scale.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp"
|
||||
#include "ck/host_utility/device_prop.hpp"
|
||||
#include "ck/host_utility/kernel_launch.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename CLayout,
|
||||
typename ADataType,
|
||||
typename AScaleDataType,
|
||||
typename BDataType,
|
||||
typename BScaleDataType,
|
||||
typename DsDataType,
|
||||
typename CDataType,
|
||||
typename GemmAccDataType,
|
||||
typename CShuffleDataType,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CElementwiseOperation,
|
||||
GemmSpecialization GemmSpec,
|
||||
index_t BlockSize,
|
||||
index_t ScaleBlockM,
|
||||
index_t ScaleBlockN,
|
||||
index_t ScaleBlockK,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t AK1,
|
||||
index_t BK1,
|
||||
index_t MPerXDL,
|
||||
index_t NPerXDL,
|
||||
index_t MXdlPerWave,
|
||||
index_t NXdlPerWave,
|
||||
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
typename ABlockTransferThreadClusterArrangeOrder,
|
||||
typename ABlockTransferSrcAccessOrder,
|
||||
index_t ABlockTransferSrcVectorDim,
|
||||
index_t ABlockTransferSrcScalarPerVector,
|
||||
index_t ABlockTransferDstScalarPerVector_AK1,
|
||||
bool ABlockLdsExtraM,
|
||||
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
|
||||
typename BBlockTransferThreadClusterArrangeOrder,
|
||||
typename BBlockTransferSrcAccessOrder,
|
||||
index_t BBlockTransferSrcVectorDim,
|
||||
index_t BBlockTransferSrcScalarPerVector,
|
||||
index_t BBlockTransferDstScalarPerVector_BK1,
|
||||
bool BBlockLdsExtraN,
|
||||
index_t CShuffleMXdlPerWavePerShuffle,
|
||||
index_t CShuffleNXdlPerWavePerShuffle,
|
||||
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
typename CDEShuffleBlockTransferScalarPerVectors,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
|
||||
typename ComputeTypeA = CDataType,
|
||||
typename ComputeTypeB = ComputeTypeA,
|
||||
typename LDSTypeA = ComputeTypeA,
|
||||
typename LDSTypeB = ComputeTypeB>
|
||||
struct DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3
|
||||
: public DeviceGemmMultipleD_ABScale<ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
CLayout,
|
||||
ADataType,
|
||||
AScaleDataType,
|
||||
BDataType,
|
||||
BScaleDataType,
|
||||
DsDataType,
|
||||
CDataType,
|
||||
ScaleBlockM,
|
||||
ScaleBlockN,
|
||||
ScaleBlockK,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CElementwiseOperation>
|
||||
{
|
||||
static constexpr index_t NumDTensor = DsDataType::Size();
|
||||
|
||||
// GridwiseGemm
|
||||
using GridwiseGemm = GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3<
|
||||
ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
CLayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
GemmAccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
CDataType,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CElementwiseOperation,
|
||||
GemmSpec,
|
||||
BlockSize,
|
||||
ScaleBlockM,
|
||||
ScaleBlockN,
|
||||
ScaleBlockK,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
AK1,
|
||||
BK1,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
ABlockTransferThreadClusterArrangeOrder,
|
||||
ABlockTransferSrcAccessOrder,
|
||||
ABlockTransferSrcVectorDim,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
ABlockTransferDstScalarPerVector_AK1,
|
||||
false,
|
||||
ABlockLdsExtraM,
|
||||
BBlockTransferThreadClusterLengths_BK0_N_BK1,
|
||||
BBlockTransferThreadClusterArrangeOrder,
|
||||
BBlockTransferSrcAccessOrder,
|
||||
BBlockTransferSrcVectorDim,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
BBlockTransferDstScalarPerVector_BK1,
|
||||
false,
|
||||
BBlockLdsExtraN,
|
||||
CShuffleMXdlPerWavePerShuffle,
|
||||
CShuffleNXdlPerWavePerShuffle,
|
||||
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
CDEShuffleBlockTransferScalarPerVectors,
|
||||
BlkGemmPipeSched,
|
||||
BlkGemmPipelineVer,
|
||||
ComputeTypeA,
|
||||
ComputeTypeB,
|
||||
LDSTypeA,
|
||||
LDSTypeB>;
|
||||
|
||||
using Argument = typename GridwiseGemm::Argument;
|
||||
|
||||
// Invoker
|
||||
struct Invoker : public BaseInvoker
|
||||
{
|
||||
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
|
||||
{
|
||||
if(stream_config.log_level_ > 0)
|
||||
{
|
||||
arg.Print();
|
||||
}
|
||||
|
||||
if(!GridwiseGemm::CheckValidity(arg))
|
||||
{
|
||||
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
|
||||
}
|
||||
|
||||
index_t gdx, gdy, gdz;
|
||||
std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.KBatch);
|
||||
|
||||
float ave_time = 0;
|
||||
|
||||
index_t k_grain = arg.KBatch * KPerBlock;
|
||||
index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock;
|
||||
|
||||
const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
|
||||
|
||||
const auto Run = [&](const auto& kernel) {
|
||||
if(arg.KBatch > 1)
|
||||
hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
|
||||
0,
|
||||
arg.M * arg.N * sizeof(CDataType),
|
||||
stream_config.stream_id_));
|
||||
|
||||
ave_time = launch_and_time_kernel(
|
||||
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
|
||||
};
|
||||
|
||||
constexpr index_t minimum_occupancy =
|
||||
(BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave &&
|
||||
MPerBlock * NPerBlock / BlockSize > 64)
|
||||
? 1
|
||||
: 2;
|
||||
|
||||
if(has_main_k_block_loop)
|
||||
{
|
||||
// Tail number always 1
|
||||
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1 ||
|
||||
BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
|
||||
{
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
// Tail number could be One to Seven
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
|
||||
{
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::One>;
|
||||
Run(kernel);
|
||||
}
|
||||
else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Full)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Full>;
|
||||
Run(kernel);
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Two>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Three)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Three>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Four)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Four>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Five)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Five>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Six>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Seven)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Seven>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
// Tail number always 1
|
||||
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
|
||||
{
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
false,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
}
|
||||
return ave_time;
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
float Run(const BaseArgument* p_arg,
|
||||
const StreamConfig& stream_config = StreamConfig{}) override
|
||||
{
|
||||
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
|
||||
}
|
||||
};
|
||||
|
||||
static constexpr bool IsValidCompilationParameter()
|
||||
{
|
||||
// TODO: properly implement this check
|
||||
return true;
|
||||
}
|
||||
|
||||
static bool IsSupportedArgument(const Argument& arg)
|
||||
{
|
||||
if(!ck::is_xdl_supported())
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
if(ScaleBlockM % MPerBlock != 0 || ScaleBlockN % NPerBlock != 0 || ScaleBlockK != KPerBlock)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
|
||||
GemmSpec == GemmSpecialization::NKPadding ||
|
||||
GemmSpec == GemmSpecialization::MNKPadding ||
|
||||
GemmSpec == GemmSpecialization::KPadding))
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
return GridwiseGemm::CheckValidity(arg);
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
bool IsSupportedArgument(const BaseArgument* p_arg) override
|
||||
{
|
||||
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
|
||||
}
|
||||
|
||||
static auto MakeArgument(const void* p_a,
|
||||
const void* p_b,
|
||||
std::array<const void*, NumDTensor> p_ds,
|
||||
void* p_c,
|
||||
const index_t M,
|
||||
const index_t N,
|
||||
const index_t K,
|
||||
const index_t StrideA,
|
||||
const index_t StrideB,
|
||||
const std::array<index_t, NumDTensor> StrideDs,
|
||||
const index_t StrideC,
|
||||
const void* p_a_scale,
|
||||
const void* p_b_scale,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CElementwiseOperation c_element_op)
|
||||
{
|
||||
return Argument{static_cast<const ADataType*>(p_a),
|
||||
static_cast<const BDataType*>(p_b),
|
||||
p_ds,
|
||||
static_cast<CDataType*>(p_c),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideDs,
|
||||
StrideC,
|
||||
static_cast<const AScaleDataType*>(p_a_scale),
|
||||
static_cast<const BScaleDataType*>(p_b_scale),
|
||||
1,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op};
|
||||
}
|
||||
|
||||
static auto MakeInvoker() { return Invoker{}; }
|
||||
|
||||
// polymorphic
|
||||
std::unique_ptr<BaseArgument>
|
||||
MakeArgumentPointer(const void* p_a,
|
||||
const void* p_b,
|
||||
std::array<const void*, NumDTensor> p_ds,
|
||||
void* p_c,
|
||||
const index_t M,
|
||||
const index_t N,
|
||||
const index_t K,
|
||||
const index_t StrideA,
|
||||
const index_t StrideB,
|
||||
const std::array<ck::index_t, NumDTensor> StrideDs,
|
||||
const index_t StrideC,
|
||||
const void* p_a_scale,
|
||||
const void* p_b_scale,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CElementwiseOperation c_element_op) override
|
||||
{
|
||||
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
|
||||
static_cast<const BDataType*>(p_b),
|
||||
p_ds,
|
||||
static_cast<CDataType*>(p_c),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideDs,
|
||||
StrideC,
|
||||
static_cast<const AScaleDataType*>(p_a_scale),
|
||||
static_cast<const BScaleDataType*>(p_b_scale),
|
||||
1,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op);
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
|
||||
{
|
||||
return std::make_unique<Invoker>(Invoker{});
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
std::string GetTypeString() const override
|
||||
{
|
||||
auto str = std::stringstream();
|
||||
|
||||
std::map<BlockGemmPipelineScheduler, std::string> BlkGemmPipelineSchedulerToString{
|
||||
{BlockGemmPipelineScheduler::Intrawave, "Intrawave"},
|
||||
{BlockGemmPipelineScheduler::Interwave, "Interwave"}};
|
||||
|
||||
std::map<BlockGemmPipelineVersion, std::string> BlkGemmPipelineVersionToString{
|
||||
{BlockGemmPipelineVersion::v1, "v1"},
|
||||
{BlockGemmPipelineVersion::v2, "v2"},
|
||||
{BlockGemmPipelineVersion::v3, "v3"}};
|
||||
|
||||
// clang-format off
|
||||
str << "DeviceGemmXdlUniversal"
|
||||
<< "<"
|
||||
<< getGemmSpecializationString(GemmSpec) << ", "
|
||||
<< std::string(ALayout::name)[0]
|
||||
<< std::string(BLayout::name)[0]
|
||||
<< std::string(CLayout::name)[0]
|
||||
<< ">"
|
||||
<< " BlkSize: "
|
||||
<< BlockSize << ", "
|
||||
<< "BlkTile: "
|
||||
<< MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
|
||||
<< "WaveTile: "
|
||||
<< MPerXDL<<"x"<<NPerXDL << ", "
|
||||
<< "WaveMap: "
|
||||
<< MXdlPerWave<<"x" << NXdlPerWave<<", "
|
||||
<< "VmemReadVec: "
|
||||
<< ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
|
||||
<< "BlkGemmPipelineScheduler: "
|
||||
<< BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
|
||||
<< "BlkGemmPipelineVersion: "
|
||||
<< BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
|
||||
<< "BlkGemmPipelinePrefetchStages: "
|
||||
<< GridwiseGemm::BlockwiseGemmPipe::PrefetchStages;
|
||||
// clang-format on
|
||||
|
||||
return str.str();
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -1,5 +1,5 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
@@ -249,6 +249,31 @@ struct MultiplyAdd
|
||||
}
|
||||
};
|
||||
|
||||
struct MultiplyMultiply
|
||||
{
|
||||
template <typename E, typename C, typename D0, typename D1>
|
||||
__host__ __device__ constexpr void
|
||||
operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
|
||||
|
||||
template <>
|
||||
__host__ __device__ constexpr void operator()<ck::half_t, float, float, float>(
|
||||
ck::half_t& e, const float& c, const float& d0, const float& d1) const
|
||||
{
|
||||
const float x0_f = c * d0 * d1;
|
||||
|
||||
e = ck::type_convert<ck::half_t>(x0_f);
|
||||
}
|
||||
|
||||
template <>
|
||||
__host__ __device__ constexpr void operator()<ck::bhalf_t, float, float, float>(
|
||||
ck::bhalf_t& e, const float& c, const float& d0, const float& d1) const
|
||||
{
|
||||
const float x0_f = c * d0 * d1;
|
||||
|
||||
e = ck::type_convert<ck::bhalf_t>(x0_f);
|
||||
}
|
||||
};
|
||||
|
||||
struct MultiplyAddFastGelu
|
||||
{
|
||||
template <typename E, typename C, typename D0, typename D1>
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
Reference in New Issue
Block a user