mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-04 13:41:24 +00:00
Manual control of MAC cluster for improved interwave performance (#184)
* manual control of MAC cluster for improved 2-wave performance ensure setprio's order; ensure inner loop size >= local read size synchronize when single mac cluster * format * use value field from ck::integral_constant * roll out inter-wave loop scheduler to c-shuffle gemm variants will gradually roll out to other applicable device ops when occasional reg spill is resolved * additional comments * format * fix mismatch between inter-wave pipeline and interwave blockwise gemm * address review feedback * amend
This commit is contained in:
Anthony Chang
@@ -1,5 +1,6 @@
|
||||
#pragma once
|
||||
#include "common_header.hpp"
|
||||
#include "tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
@@ -248,4 +249,116 @@ struct GridwiseGemmPipeline_v1<2>
|
||||
}
|
||||
};
|
||||
|
||||
template <index_t NumPrefetch>
|
||||
struct GridwiseGemmPipelineInterwave_v1;
|
||||
|
||||
template <>
|
||||
struct GridwiseGemmPipelineInterwave_v1<1>
|
||||
{
|
||||
__host__ __device__ static constexpr bool IsSupported(index_t /* num_loop */) { return true; }
|
||||
|
||||
__host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
|
||||
{
|
||||
return num_loop > 1;
|
||||
}
|
||||
|
||||
template <bool HasMainLoop,
|
||||
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 BlockwiseGemm,
|
||||
typename CThreadBuffer>
|
||||
static __device__ void Run(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,
|
||||
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,
|
||||
const BlockwiseGemm& blockwise_gemm,
|
||||
CThreadBuffer& c_thread_buf,
|
||||
index_t num_loop)
|
||||
{
|
||||
// preload data into LDS
|
||||
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();
|
||||
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
|
||||
|
||||
// main body
|
||||
if constexpr(HasMainLoop)
|
||||
{
|
||||
index_t i = 0;
|
||||
|
||||
do
|
||||
{
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
|
||||
block_sync_lds();
|
||||
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
|
||||
|
||||
// block_sync_lds(); // moved into blockwise_gemm
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
|
||||
|
||||
++i;
|
||||
} while(i < (num_loop - 1));
|
||||
}
|
||||
|
||||
// tail
|
||||
{
|
||||
block_sync_lds();
|
||||
|
||||
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
// Note: 2 stage prefetch not optimized for inter-wave loop scheduler
|
||||
template <>
|
||||
struct GridwiseGemmPipelineInterwave_v1<2> : public GridwiseGemmPipeline_v1<2>
|
||||
{
|
||||
};
|
||||
|
||||
template <index_t NumPrefetch, LoopScheduler LoopSched>
|
||||
constexpr auto GridwiseGemmPipeline_v1_Selector()
|
||||
{
|
||||
if constexpr(LoopSched == LoopScheduler::Default)
|
||||
{
|
||||
return GridwiseGemmPipeline_v1<NumPrefetch>{};
|
||||
}
|
||||
else if constexpr(LoopSched == LoopScheduler::Interwave)
|
||||
{
|
||||
return GridwiseGemmPipelineInterwave_v1<NumPrefetch>{};
|
||||
}
|
||||
}
|
||||
|
||||
} // namespace ck
|
||||
|
||||
@@ -134,7 +134,8 @@ template <typename FloatAB,
|
||||
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
|
||||
typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
|
||||
index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
|
||||
index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock>
|
||||
index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
|
||||
LoopScheduler LoopSched>
|
||||
struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
|
||||
{
|
||||
static constexpr auto I0 = Number<0>{};
|
||||
@@ -473,17 +474,18 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
|
||||
constexpr index_t KPack = 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,
|
||||
FloatGemmAcc,
|
||||
decltype(a_block_desc_ak0_m_ak1),
|
||||
decltype(b_block_desc_bk0_n_bk1),
|
||||
MPerXdl,
|
||||
NPerXdl,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
KPack>{};
|
||||
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
|
||||
BlockSize,
|
||||
FloatAB,
|
||||
FloatGemmAcc,
|
||||
decltype(a_block_desc_ak0_m_ak1),
|
||||
decltype(b_block_desc_bk0_n_bk1),
|
||||
MPerXdl,
|
||||
NPerXdl,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
KPack,
|
||||
LoopSched>();
|
||||
|
||||
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
|
||||
|
||||
@@ -502,25 +504,28 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
|
||||
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
|
||||
|
||||
// gridwise GEMM pipeline
|
||||
const auto gridwise_gemm_pipeline =
|
||||
GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
|
||||
|
||||
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);
|
||||
|
||||
GridwiseGemmPipe::template Run<HasMainKBlockLoop>(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_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,
|
||||
num_k_block_main_loop);
|
||||
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(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_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,
|
||||
num_k_block_main_loop);
|
||||
|
||||
// shuffle C and write out
|
||||
{
|
||||
|
||||
@@ -107,7 +107,8 @@ template <typename FloatAB,
|
||||
index_t CShuffleMXdlPerWavePerShuffle,
|
||||
index_t CShuffleNXdlPerWavePerShuffle,
|
||||
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
|
||||
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
|
||||
LoopScheduler LoopSched>
|
||||
struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
|
||||
{
|
||||
static constexpr auto I0 = Number<0>{};
|
||||
@@ -416,17 +417,18 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
|
||||
constexpr index_t KPack = 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,
|
||||
FloatGemmAcc,
|
||||
decltype(a_block_desc_ak0_m_ak1),
|
||||
decltype(b_block_desc_bk0_n_bk1),
|
||||
MPerXdl,
|
||||
NPerXdl,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
KPack>{};
|
||||
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
|
||||
BlockSize,
|
||||
FloatAB,
|
||||
FloatGemmAcc,
|
||||
decltype(a_block_desc_ak0_m_ak1),
|
||||
decltype(b_block_desc_bk0_n_bk1),
|
||||
MPerXdl,
|
||||
NPerXdl,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
KPack,
|
||||
LoopSched>();
|
||||
|
||||
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
|
||||
|
||||
@@ -445,25 +447,28 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
|
||||
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
|
||||
|
||||
// gridwise GEMM pipeline
|
||||
const auto gridwise_gemm_pipeline =
|
||||
GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
|
||||
|
||||
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);
|
||||
|
||||
GridwiseGemmPipe::template Run<HasMainKBlockLoop>(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_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,
|
||||
num_k_block_main_loop);
|
||||
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(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_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,
|
||||
num_k_block_main_loop);
|
||||
|
||||
// shuffle C and write out
|
||||
{
|
||||
|
||||
Reference in New Issue
Block a user