mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-04 13:41:24 +00:00
[GEMM] Gemm universal device operation (#1154)
* Optimize GEMM on MI200/300: 1. Add new blockwise gemm pipeline 2. Add irregular splitk intances * clang format + typo fix * Fix a bug * initial commit * Add more instances to irregular splitk * blkgemm pipeline v1~4 prototype * Sanity Checked. Known issue: 1. Poor performance of splitk 2. Register spill on blkgemmpipeline v3 * Sanity and Performance fix: 1. fix a bug related to sanity in grouped b2c mapping 2. fix a bug related to sanity and performance in splitk offset * Sanity and API update: 1. Remove prefetch stage 2. Fix valid check bug 3, Add first gemm_universal instance into ckProfiler * Add NN instances for gemm universal * 1. Add NT instances for gemm_universal 2. Fix a bug about Kpadding in gemm_universal * Fix a bug regarding padding Odd K number * remove kernel print * Fix KPadding bug... * Update safety check * another try to fix kpadding.. * Sanity checked * new instances.. * clang format+typo fix * remove clang format script's change * Add non-hotloop compile option * 1. Add fp16xfp8 example 2. pull packed convert f8 from pr1150 * Some miscs.. opt and fix * Add pipeline description docs * Split universal gemm instance library to cut profiler compiling time * uncomment cmakefile * Fix a bug caused by blockwise_gemm_pipe_v2 * reduce default splitk to 1 * Add 224x256x64 tile size * update, including: 1. Experiment pipeline 5~7 2. Optimization for pipeline 4 3. Organized instance library * temp save * temp save * Permuted lds layout, sanity and function checked * clang format * Move OOB check from RunRead to RunWrite, for better software pipeline. TODO: agpr spill when NN layout * clangformat * A/B splitpipe scheduler for v3 * Fix two bugs * bug fix * fix a bug in oob check * Example for mixed fp16_fp8 gemm * Clean experimental code blocks * Add mixed precision gemm into profiler * tempsave * optimize m/n major lds layout * Add RRR GEMM mixed precision instances * Optimize f8 matrix transpose * Add test_gemm_universal * A/B spilt schedule for blkpip v5 * Take ds_read2 into iglp scheduling scheme * format * fixed cmake * Add llvm-option into CI cmake flag --------- Co-authored-by: Jing Zhang <jizhan@amd.com>
This commit is contained in:
Haocong WANG
@@ -960,13 +960,13 @@ struct BlockwiseGemmXdlops_pipeline_v4
|
||||
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor(
|
||||
make_tuple(Number<MRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}),
|
||||
make_tuple(
|
||||
Number<KPack>{}, Number<KPack * MRepeat * KPack>{}, Number<MRepeat * KPack>{}, I1));
|
||||
Number<KPack>{}, Number<KRepeat * MRepeat * KPack>{}, Number<MRepeat * KPack>{}, I1));
|
||||
|
||||
// B[N0, N1, N2, KPack]
|
||||
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor(
|
||||
make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}),
|
||||
make_tuple(
|
||||
Number<KPack>{}, Number<KPack * MRepeat * KPack>{}, Number<MRepeat * KPack>{}, I1));
|
||||
Number<KPack>{}, Number<KRepeat * NRepeat * KPack>{}, Number<NRepeat * KPack>{}, I1));
|
||||
|
||||
// C[M, N, NumRegXdlops]
|
||||
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
|
||||
|
||||
@@ -0,0 +1,354 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/utility/common_header.hpp"
|
||||
#include "ck/utility/blkgemmpipe_scheduler.hpp"
|
||||
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
|
||||
#include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
|
||||
#include "ck/tensor_description/tensor_adaptor.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
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 = false>
|
||||
struct BlockwiseGemmXdlops_pipeline_base
|
||||
{
|
||||
static constexpr auto I0 = Number<0>{};
|
||||
static constexpr auto I1 = Number<1>{};
|
||||
static constexpr auto I2 = Number<2>{};
|
||||
static constexpr auto I3 = Number<3>{};
|
||||
|
||||
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
|
||||
|
||||
// Hardcode to 64, as HIP-provided "warpSize" would return 32 on RDNA GPUs.
|
||||
static constexpr index_t WaveSize = 64;
|
||||
|
||||
static constexpr index_t A_K0 = ATileDesc{}.GetLength(I0);
|
||||
static constexpr index_t B_K0 = BTileDesc{}.GetLength(I0);
|
||||
static constexpr index_t A_K1 = ATileDesc{}.GetLength(I2);
|
||||
static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2);
|
||||
|
||||
static constexpr auto xdlops_gemm =
|
||||
XdlopsGemm<ComputeDataType, MPerXDL, NPerXDL, KPack, ComputeDataType, TransposeC>{};
|
||||
|
||||
static constexpr index_t AMmaKStride = KPack;
|
||||
static constexpr index_t BMmaKStride = KPack;
|
||||
|
||||
static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
|
||||
static constexpr index_t KRepeat = KPerThread / KPack;
|
||||
|
||||
static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
|
||||
static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
|
||||
|
||||
using HotLoopInstList =
|
||||
ck::BlockwiseGemmXdlops_pipeline_hotloop_inst<BlockSize,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
A_K1,
|
||||
B_K1,
|
||||
A_K1,
|
||||
B_K1,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
xdlops_gemm.KPerXdlops>;
|
||||
|
||||
static_assert(KPerThread % KPack == 0,
|
||||
"Wrong KPack setting; try increasing KPerThread or decreasing KPack");
|
||||
|
||||
StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
|
||||
AccDataType,
|
||||
MRepeat * NRepeat,
|
||||
xdlops_gemm.GetRegSizePerXdlops(),
|
||||
true>
|
||||
c_thread_buf_;
|
||||
|
||||
__host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
|
||||
|
||||
__device__ static auto GetWaveIdx()
|
||||
{
|
||||
const index_t thread_id = ThisThreadBlock::GetThreadId();
|
||||
|
||||
constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
|
||||
make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
|
||||
make_tuple(Sequence<0, 1, 2>{}),
|
||||
make_tuple(Sequence<0>{}));
|
||||
|
||||
return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
|
||||
}
|
||||
|
||||
__device__ static auto CalculateAThreadOriginDataIndex()
|
||||
{
|
||||
const auto wave_idx = GetWaveIdx();
|
||||
|
||||
const auto waveId_m = wave_idx[I0];
|
||||
|
||||
const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
|
||||
|
||||
return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPerThread * xdlops_a_idx[I0]);
|
||||
}
|
||||
|
||||
__device__ static auto CalculateBThreadOriginDataIndex()
|
||||
{
|
||||
const auto wave_idx = GetWaveIdx();
|
||||
|
||||
const auto waveId_n = wave_idx[I1];
|
||||
|
||||
const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
|
||||
|
||||
return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPerThread * xdlops_b_idx[I0]);
|
||||
}
|
||||
|
||||
template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
|
||||
__device__ static auto
|
||||
CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
|
||||
{
|
||||
const auto wave_idx = GetWaveIdx();
|
||||
|
||||
const auto waveId_m = wave_idx[I0];
|
||||
const auto waveId_n = wave_idx[I1];
|
||||
|
||||
const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
|
||||
|
||||
constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
|
||||
make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
|
||||
make_tuple(Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 1, 2>{}));
|
||||
|
||||
constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
|
||||
make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
|
||||
make_tuple(Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 1, 2>{}));
|
||||
|
||||
const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
|
||||
make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
|
||||
const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
|
||||
make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
|
||||
|
||||
return make_tuple(c_thread_m, c_thread_n);
|
||||
}
|
||||
|
||||
template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
|
||||
__device__ static auto
|
||||
CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
|
||||
{
|
||||
const auto wave_idx = GetWaveIdx();
|
||||
|
||||
const auto waveId_m = wave_idx[I0];
|
||||
const auto waveId_n = wave_idx[I1];
|
||||
|
||||
const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
|
||||
|
||||
return make_tuple(
|
||||
m0, n0, waveId_m, waveId_n, blk_idx[I0], blk_idx[I1], blk_idx[I2], blk_idx[I3]);
|
||||
}
|
||||
|
||||
using Tuple4 = decltype(CalculateAThreadOriginDataIndex());
|
||||
|
||||
__host__ __device__
|
||||
BlockwiseGemmXdlops_pipeline_base(Tuple4 a_origin = CalculateAThreadOriginDataIndex(),
|
||||
Tuple4 b_origin = CalculateBThreadOriginDataIndex())
|
||||
: a_thread_copy_(a_origin), b_thread_copy_(b_origin)
|
||||
{
|
||||
static_assert(AMmaTileDesc::IsKnownAtCompileTime() && BMmaTileDesc::IsKnownAtCompileTime(),
|
||||
"wrong! Desc should be known at compile-time");
|
||||
|
||||
static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
|
||||
"ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
|
||||
|
||||
static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
|
||||
"wrong!");
|
||||
}
|
||||
|
||||
// transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
|
||||
__host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
|
||||
{
|
||||
constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
|
||||
|
||||
constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
|
||||
constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
|
||||
constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
|
||||
constexpr auto N = c_m0_m1_m2_n_tblk_lens[I3];
|
||||
|
||||
return make_naive_tensor_descriptor_packed(
|
||||
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, N, M0, M1, M2));
|
||||
}
|
||||
|
||||
// XDL output supporting C_xdl = A_xdl * B_xdl
|
||||
__host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
|
||||
{
|
||||
constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
|
||||
|
||||
constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
|
||||
constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
|
||||
constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
|
||||
constexpr auto N = c_m0_m1_m2_n_tblk_lens[I3];
|
||||
|
||||
return make_naive_tensor_descriptor_packed(
|
||||
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
|
||||
{
|
||||
constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
|
||||
|
||||
constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
|
||||
constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
|
||||
constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
|
||||
constexpr auto N = c_m0_m1_m2_n_tblk_lens[I3];
|
||||
|
||||
return make_naive_tensor_descriptor_packed(
|
||||
make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
|
||||
}
|
||||
|
||||
// transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
|
||||
__host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
|
||||
{
|
||||
constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
|
||||
Number<NRepeat>{},
|
||||
Number<MWaves>{},
|
||||
Number<NWaves>{},
|
||||
Number<MPerXDL>{},
|
||||
Number<NPerXDL>{}));
|
||||
|
||||
return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(c_block_desc_m0_n0_m1_n1_m2_n2);
|
||||
}
|
||||
|
||||
// XDL output supporting C_xdl = A_xdl * B_xdl
|
||||
__host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
|
||||
{
|
||||
constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
|
||||
Number<NRepeat>{},
|
||||
Number<MWaves>{},
|
||||
Number<NWaves>{},
|
||||
Number<MPerXDL>{},
|
||||
Number<NPerXDL>{}));
|
||||
|
||||
return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
|
||||
{
|
||||
constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(I1,
|
||||
Number<MRepeat>{},
|
||||
Number<NRepeat>{},
|
||||
Number<MWaves>{},
|
||||
Number<NWaves>{},
|
||||
Number<MPerXDL>{},
|
||||
Number<NPerXDL>{}));
|
||||
|
||||
return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
|
||||
c_block_desc_g_m0_n0_m1_n1_m2_n2);
|
||||
}
|
||||
|
||||
template <typename CGridDesc_M_N>
|
||||
__host__ __device__ static constexpr auto
|
||||
MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
|
||||
{
|
||||
const auto M = c_grid_desc_m_n.GetLength(I0);
|
||||
const auto N = c_grid_desc_m_n.GetLength(I1);
|
||||
|
||||
const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
|
||||
c_grid_desc_m_n,
|
||||
make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
|
||||
make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
|
||||
|
||||
return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
|
||||
}
|
||||
|
||||
template <typename CGridDesc_G_M_N>
|
||||
__host__ __device__ static constexpr auto
|
||||
MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
|
||||
{
|
||||
const auto G = c_grid_desc_g_m_n.GetLength(I0);
|
||||
const auto M = c_grid_desc_g_m_n.GetLength(I1);
|
||||
const auto N = c_grid_desc_g_m_n.GetLength(I2);
|
||||
|
||||
const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
|
||||
c_grid_desc_g_m_n,
|
||||
make_tuple(make_pass_through_transform(G),
|
||||
make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
|
||||
make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
|
||||
|
||||
return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
|
||||
c_grid_desc_g_m0_n0_m1_n1_m2_n2);
|
||||
}
|
||||
|
||||
static constexpr AMmaTileDesc a_block_desc_m0_m1_m2_k;
|
||||
static constexpr BMmaTileDesc b_block_desc_n0_n1_n2_k;
|
||||
|
||||
protected:
|
||||
// M1, N1 as double buffer index
|
||||
// Read buffer + Compute buffer
|
||||
// A[M0, M1, M2, KPack]
|
||||
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor(
|
||||
make_tuple(Number<MRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}),
|
||||
make_tuple(
|
||||
Number<KPack>{}, Number<KRepeat * MRepeat * KPack>{}, Number<MRepeat * KPack>{}, I1));
|
||||
|
||||
// B[N0, N1, N2, KPack]
|
||||
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor(
|
||||
make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}),
|
||||
make_tuple(
|
||||
Number<KPack>{}, Number<KRepeat * NRepeat * KPack>{}, Number<NRepeat * KPack>{}, I1));
|
||||
|
||||
// C[M, N, NumRegXdlops]
|
||||
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
|
||||
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
|
||||
|
||||
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType,
|
||||
ComputeDataType,
|
||||
decltype(a_block_desc_m0_m1_m2_k),
|
||||
decltype(a_thread_desc_),
|
||||
Sequence<1, 1, 1, KPack>,
|
||||
Sequence<0, 1, 2, 3>,
|
||||
3,
|
||||
A_K1,
|
||||
A_K1>;
|
||||
|
||||
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<BDataType,
|
||||
ComputeDataType,
|
||||
decltype(b_block_desc_n0_n1_n2_k),
|
||||
decltype(b_thread_desc_),
|
||||
Sequence<1, 1, 1, KPack>,
|
||||
Sequence<0, 1, 2, 3>,
|
||||
3,
|
||||
B_K1,
|
||||
B_K1>;
|
||||
|
||||
AThreadCopy a_thread_copy_;
|
||||
BThreadCopy b_thread_copy_;
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,167 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v5.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
enum struct BlockGemmPipelineVersion
|
||||
{
|
||||
v1, // Naive
|
||||
v2, // Mem
|
||||
v3, // Comp
|
||||
v4, // Comp, double lds buffer
|
||||
v5, // Comp, double global prefetch register buffer
|
||||
};
|
||||
|
||||
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 BlockGemmPipeline_Selector()
|
||||
{
|
||||
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
|
||||
{
|
||||
return BlockwiseGemmXdlops_pipeline_v1<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<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<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::v4)
|
||||
{
|
||||
return BlockwiseGemmXdlops_pipeline_v4<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::v5)
|
||||
{
|
||||
return BlockwiseGemmXdlops_pipeline_v5<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 +1,732 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, 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
|
||||
{
|
||||
};
|
||||
|
||||
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<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>
|
||||
__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,
|
||||
CThreadBuffer& c_thread_buf,
|
||||
index_t num_loop) const
|
||||
{
|
||||
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());
|
||||
|
||||
// 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);
|
||||
|
||||
// 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();
|
||||
|
||||
// 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, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
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, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
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_;
|
||||
};
|
||||
|
||||
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<BlockGemmPipelineScheduler::Interwave,
|
||||
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::A_K1;
|
||||
using Base::B_K1;
|
||||
using Base::I0;
|
||||
using Base::I1;
|
||||
using Base::KPerThread;
|
||||
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;
|
||||
|
||||
static constexpr index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS;
|
||||
static constexpr index_t KPerInnerLoop = math::max(KPerThread / NumMacClusters, KPack);
|
||||
static constexpr index_t KRepeat = KPerThread / KPerInnerLoop;
|
||||
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>
|
||||
__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,
|
||||
CThreadBuffer& c_thread_buf,
|
||||
index_t num_loop) const
|
||||
{
|
||||
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());
|
||||
|
||||
// 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);
|
||||
|
||||
// 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();
|
||||
|
||||
// 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 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 * KPerInnerLoop>{}),
|
||||
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 * KPerInnerLoop>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k0, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
// NOTE: Synchronize threads in a workgroup at the start of each MAC cluster,
|
||||
// but except the first, as we can shorten non-MAC cluster a bit and there's no
|
||||
// observable negative impact. The desired effect is waves in a workgroup
|
||||
// executing MAC in sync. This avoids some out-of-sync waves hijacking MAC
|
||||
// resource from other workgroups and reducing the chance of latency hiding by
|
||||
// waiting for the rest of the workgroup at the eventual sync point.
|
||||
if constexpr(k0.value != 0 || KRepeat == 1)
|
||||
{
|
||||
__builtin_amdgcn_s_barrier();
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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, k_ + ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, k_ + ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType,
|
||||
xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
// The block_sync_lds() here performs double duty:
|
||||
// A) safeguard against data hazard because barrier from
|
||||
// blockwise_gemm is moved here B) reduce VMEM FIFO congestion by
|
||||
// applying small delays to different wavefronts It is performed
|
||||
// near the end of MAC cluster to minimize lgkmcnt penalty
|
||||
if constexpr(k0.value == KRepeat - 1 &&
|
||||
k_.value == KPerInnerLoop - KPack &&
|
||||
m0.value == MRepeat - 1 && n0.value == NRepeat - 1)
|
||||
{
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
block_sync_lds();
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
|
||||
{
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
__builtin_amdgcn_s_setprio(1);
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
});
|
||||
});
|
||||
});
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
__builtin_amdgcn_s_setprio(0);
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
});
|
||||
|
||||
// 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 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 * KPerInnerLoop>{}),
|
||||
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 * KPerInnerLoop>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k0, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
if constexpr(k0.value != 0 || KRepeat == 1)
|
||||
{
|
||||
__builtin_amdgcn_s_barrier();
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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, k_ + ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, k_ + ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType,
|
||||
xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
if constexpr(k0.value == KRepeat - 1 &&
|
||||
k_.value == KPerInnerLoop - KPack &&
|
||||
m0.value == MRepeat - 1 && n0.value == NRepeat - 1)
|
||||
{
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
block_sync_lds();
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
|
||||
{
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
__builtin_amdgcn_s_setprio(1);
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
});
|
||||
});
|
||||
});
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
__builtin_amdgcn_s_setprio(0);
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
protected:
|
||||
// K->M loopover
|
||||
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor(
|
||||
make_tuple(Number<MRepeat>{}, I1, Number<KRepeat>{}, Number<KPerInnerLoop>{}),
|
||||
make_tuple(Number<KPerInnerLoop>{},
|
||||
Number<KRepeat * MRepeat * KPerInnerLoop>{},
|
||||
Number<MRepeat * KPerInnerLoop>{},
|
||||
I1));
|
||||
|
||||
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor(
|
||||
make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPerInnerLoop>{}),
|
||||
make_tuple(Number<KPerInnerLoop>{},
|
||||
Number<KRepeat * NRepeat * KPerInnerLoop>{},
|
||||
Number<NRepeat * KPerInnerLoop>{},
|
||||
I1));
|
||||
|
||||
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType,
|
||||
ComputeDataType,
|
||||
decltype(a_block_desc_m0_m1_m2_k),
|
||||
decltype(a_thread_desc_),
|
||||
Sequence<1, 1, 1, KPerInnerLoop>,
|
||||
Sequence<0, 1, 2, 3>,
|
||||
3,
|
||||
A_K1,
|
||||
A_K1>;
|
||||
|
||||
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<BDataType,
|
||||
ComputeDataType,
|
||||
decltype(b_block_desc_n0_n1_n2_k),
|
||||
decltype(b_thread_desc_),
|
||||
Sequence<1, 1, 1, KPerInnerLoop>,
|
||||
Sequence<0, 1, 2, 3>,
|
||||
3,
|
||||
B_K1,
|
||||
B_K1>;
|
||||
|
||||
AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex()};
|
||||
BThreadCopy b_thread_copy_{Base::CalculateBThreadOriginDataIndex()};
|
||||
using Base::c_thread_desc_;
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,439 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, 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
|
||||
{
|
||||
};
|
||||
|
||||
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<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::I1;
|
||||
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 =
|
||||
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
|
||||
constexpr auto ds_read_b_issue_cycle =
|
||||
HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16 ? 8 : 4;
|
||||
constexpr auto ds_read_a_mfma_rate =
|
||||
(mfma_cycle - 8 + ds_read_a_issue_cycle - 1) / ds_read_a_issue_cycle;
|
||||
constexpr auto ds_read_b_mfma_rate =
|
||||
(mfma_cycle - 8 + ds_read_b_issue_cycle - 1) / ds_read_b_issue_cycle;
|
||||
|
||||
// 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_mfma_per_ds_read * (num_ds_read_inst_a / ds_read_a_mfma_rate +
|
||||
num_ds_read_inst_b / ds_read_b_mfma_rate);
|
||||
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_ds_read_inst_a / ds_read_a_mfma_rate, 1>{}([&](auto i) {
|
||||
ignore = i;
|
||||
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, num_mfma_per_ds_read, 0); // MFMA
|
||||
});
|
||||
|
||||
static_for<0, num_ds_read_inst_b / ds_read_b_mfma_rate, 1>{}([&](auto i) {
|
||||
ignore = i;
|
||||
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_b_mfma_rate, 0); // DS read
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, num_mfma_per_ds_read, 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>
|
||||
__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,
|
||||
CThreadBuffer& c_thread_buf,
|
||||
index_t num_loop) const
|
||||
{
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
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());
|
||||
|
||||
// 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);
|
||||
|
||||
// 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();
|
||||
|
||||
// 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, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
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);
|
||||
});
|
||||
});
|
||||
|
||||
HotLoopScheduler();
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
|
||||
i += 1;
|
||||
} while(i < (num_loop - 1));
|
||||
}
|
||||
// tail
|
||||
if constexpr(TailNum == TailNumber::Full)
|
||||
{
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
}
|
||||
|
||||
protected:
|
||||
using Base::a_thread_copy_;
|
||||
using Base::a_thread_desc_;
|
||||
using Base::b_thread_copy_;
|
||||
using Base::b_thread_desc_;
|
||||
using Base::c_thread_desc_;
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,597 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
// Compute optimimal pipeline with highest resource request
|
||||
// GlobalPrefetchStages: 4
|
||||
// LocalPreFillStages: 2
|
||||
// LocalPreFetchStages: 1
|
||||
// LocalSharedMemoryBuffer: 2
|
||||
|
||||
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_v4
|
||||
{
|
||||
};
|
||||
|
||||
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_v4<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::I1;
|
||||
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 = 4;
|
||||
static constexpr index_t PrefillStages = 2;
|
||||
static constexpr index_t GlobalBufferNum = 2;
|
||||
static constexpr index_t HotloopUnroll = 2;
|
||||
|
||||
__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 % HotloopUnroll == 1)
|
||||
{
|
||||
return TailNumber::Odd;
|
||||
}
|
||||
else
|
||||
{
|
||||
return TailNumber::Even;
|
||||
}
|
||||
}
|
||||
|
||||
template <typename ScheduleGroup>
|
||||
__device__ static constexpr void HotLoopScheduler(ScheduleGroup schedule_group)
|
||||
{
|
||||
// TODO: Take data type into consideration as pipe ver 3
|
||||
// A-B splited schedule
|
||||
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_issue_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
|
||||
constexpr auto num_dswrite_per_issue_a =
|
||||
(HotLoopInstList::A_LDS_Write_Inst_Num + num_issue_a - 1) / num_issue_a;
|
||||
constexpr auto num_dsread_per_issue_a = num_ds_read_inst_a / num_issue_a;
|
||||
|
||||
constexpr auto num_issue_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
|
||||
constexpr auto num_dswrite_per_issue_b =
|
||||
(HotLoopInstList::B_LDS_Write_Inst_Num + num_issue_b - 1) / num_issue_b;
|
||||
constexpr auto num_dsread_per_issue_b = num_ds_read_inst_b / num_issue_b;
|
||||
|
||||
constexpr auto num_mfma_per_issue =
|
||||
HotLoopInstList::C_MFMA_Inst_Num / (num_issue_a + num_issue_b);
|
||||
|
||||
static_for<0, num_issue_a, 1>{}([&](auto i) {
|
||||
ignore = i;
|
||||
static_for<0, num_dsread_per_issue_a, 1>{}([&](auto idsread) {
|
||||
ignore = idsread;
|
||||
__builtin_amdgcn_sched_group_barrier(0x100, 1, schedule_group); // DS read
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, 1, schedule_group); // MFMA
|
||||
});
|
||||
|
||||
static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) {
|
||||
ignore = idswrite;
|
||||
__builtin_amdgcn_sched_group_barrier(0x200, 1, schedule_group); // DS write
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, 1, schedule_group); // MFMA
|
||||
});
|
||||
|
||||
__builtin_amdgcn_sched_group_barrier(0x020, 1, schedule_group); // VMEM read
|
||||
__builtin_amdgcn_sched_group_barrier(0x008,
|
||||
num_mfma_per_issue - num_dsread_per_issue_a -
|
||||
num_dswrite_per_issue_a,
|
||||
schedule_group); // MFMA
|
||||
});
|
||||
|
||||
static_for<0, num_issue_b, 1>{}([&](auto i) {
|
||||
ignore = i;
|
||||
static_for<0, num_dsread_per_issue_b, 1>{}([&](auto idsread) {
|
||||
ignore = idsread;
|
||||
__builtin_amdgcn_sched_group_barrier(0x100, 1, schedule_group); // DS read
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, 1, schedule_group); // MFMA
|
||||
});
|
||||
|
||||
static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) {
|
||||
ignore = idswrite;
|
||||
__builtin_amdgcn_sched_group_barrier(0x200, 1, schedule_group); // DS write
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, 1, schedule_group); // MFMA
|
||||
});
|
||||
|
||||
__builtin_amdgcn_sched_group_barrier(0x020, 1, schedule_group); // VMEM read
|
||||
__builtin_amdgcn_sched_group_barrier(0x008,
|
||||
num_mfma_per_issue - num_dsread_per_issue_a -
|
||||
num_dswrite_per_issue_b,
|
||||
schedule_group); // MFMA
|
||||
});
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
|
||||
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>
|
||||
__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,
|
||||
CThreadBuffer& c_thread_buf,
|
||||
index_t num_loop) const
|
||||
{
|
||||
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());
|
||||
|
||||
StaticallyIndexedArray<decltype(a_thread_buf), Number<2>{}> a_thread_bufs;
|
||||
StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
|
||||
|
||||
// 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);
|
||||
|
||||
// Global prefetch 2
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
// Local prefill 1
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0), I0);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I0), I0);
|
||||
|
||||
// Local prefill 2
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1), I1);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I1), I1);
|
||||
|
||||
// Local prefetch 1
|
||||
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.At(I0),
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_bufs(I0));
|
||||
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.At(I0),
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_bufs(I0));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
// Global prefetch 3
|
||||
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);
|
||||
|
||||
// Global prefetch 4
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1);
|
||||
|
||||
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();
|
||||
|
||||
// main body
|
||||
if constexpr(HasMainLoop)
|
||||
{
|
||||
index_t i = 0;
|
||||
// This hot loop has two legacy loopover, to implement the double local buffer strategy
|
||||
do
|
||||
{
|
||||
auto LoopFunc = [&](auto lds_read_buf,
|
||||
auto lds_read_reg_buf,
|
||||
auto lds_write_buf,
|
||||
auto vmem_buf,
|
||||
auto mfma_reg_buf,
|
||||
auto schedule_group) {
|
||||
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.At(lds_read_buf),
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_bufs(lds_read_reg_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.At(lds_read_buf),
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_bufs(lds_read_reg_buf));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
a_blockwise_copy.RunWrite(
|
||||
a_block_desc, a_block_buf.At(lds_write_buf), vmem_buf);
|
||||
b_blockwise_copy.RunWrite(
|
||||
b_block_desc, b_block_buf.At(lds_write_buf), vmem_buf);
|
||||
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, vmem_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, vmem_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, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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_bufs[mfma_reg_buf]
|
||||
[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_bufs[mfma_reg_buf]
|
||||
[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType,
|
||||
xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
HotLoopScheduler(schedule_group);
|
||||
};
|
||||
|
||||
LoopFunc(I1, I1, I0, I0, I0, I0);
|
||||
LoopFunc(I0, I0, I1, I1, I1, I0);
|
||||
|
||||
i += HotloopUnroll;
|
||||
} while(i < (num_loop - PrefetchStages));
|
||||
}
|
||||
|
||||
auto ReadWriteCompFunc = [&](auto lds_read_buf,
|
||||
auto lds_read_reg_buf,
|
||||
auto lds_write_buf,
|
||||
auto vmem_buf,
|
||||
auto mfma_reg_buf,
|
||||
auto schedule_group) {
|
||||
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.At(lds_read_buf),
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_bufs(lds_read_reg_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.At(lds_read_buf),
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_bufs(lds_read_reg_buf));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(lds_write_buf), vmem_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(lds_write_buf), vmem_buf);
|
||||
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
HotLoopScheduler(schedule_group);
|
||||
};
|
||||
|
||||
auto ReadCompFunc = [&](auto lds_read_buf,
|
||||
auto lds_read_reg_buf,
|
||||
auto mfma_reg_buf,
|
||||
auto schedule_group) {
|
||||
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.At(lds_read_buf),
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_bufs(lds_read_reg_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.At(lds_read_buf),
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_bufs(lds_read_reg_buf));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
HotLoopScheduler(schedule_group);
|
||||
};
|
||||
|
||||
auto CompFunc = [&](auto mfma_reg_buf) {
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
};
|
||||
// tail
|
||||
if constexpr(TailNum == TailNumber::Odd)
|
||||
{
|
||||
ReadWriteCompFunc(I1, I1, I0, I0, I0, I1);
|
||||
ReadCompFunc(I0, I0, I1, I1);
|
||||
CompFunc(I0);
|
||||
}
|
||||
else if constexpr(TailNum == TailNumber::Even)
|
||||
{
|
||||
ReadWriteCompFunc(I1, I1, I0, I0, I0, I1);
|
||||
ReadWriteCompFunc(I0, I0, I1, I1, I1, I1);
|
||||
ReadCompFunc(I1, I1, I0, I1);
|
||||
CompFunc(I1);
|
||||
}
|
||||
}
|
||||
|
||||
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 +1,667 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, 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: 3
|
||||
// LocalPreFillStages: 1
|
||||
// LocalPreFetchStages: 1
|
||||
// LocalSharedMemoryBuffer: 2
|
||||
|
||||
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_v5
|
||||
{
|
||||
};
|
||||
|
||||
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_v5<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::A_K1;
|
||||
using Base::B_K1;
|
||||
using Base::I0;
|
||||
using Base::I1;
|
||||
using Base::KRepeat;
|
||||
using Base::xdlops_gemm;
|
||||
using typename Base::HotLoopInstList;
|
||||
|
||||
using Base::CalculateCThreadOriginDataIndex;
|
||||
using Base::CalculateCThreadOriginDataIndex8D;
|
||||
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 = 3;
|
||||
static constexpr index_t PrefillStages = 1;
|
||||
static constexpr index_t GlobalBufferNum = 2;
|
||||
static constexpr index_t HotloopUnroll = 2;
|
||||
|
||||
__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 % HotloopUnroll == 1)
|
||||
{
|
||||
return TailNumber::Odd;
|
||||
}
|
||||
else
|
||||
{
|
||||
return TailNumber::Even;
|
||||
}
|
||||
}
|
||||
|
||||
__device__ static constexpr auto HotLoopScheduler()
|
||||
{
|
||||
// TODO: Take data type into consideration as pipe ver 3
|
||||
// A/B split schedule
|
||||
// compiler is likely to use ds_read2 when instruction width smaller than 16bytes
|
||||
constexpr auto num_ds_read_inst_a =
|
||||
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16
|
||||
? HotLoopInstList::A_LDS_Read_Inst_Num
|
||||
: HotLoopInstList::A_LDS_Read_Inst_Num / 2;
|
||||
constexpr auto num_ds_read_inst_b =
|
||||
HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16
|
||||
? HotLoopInstList::B_LDS_Read_Inst_Num
|
||||
: HotLoopInstList::B_LDS_Read_Inst_Num / 2;
|
||||
|
||||
constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
|
||||
constexpr auto num_ds_write_inst_b = HotLoopInstList::B_LDS_Write_Inst_Num;
|
||||
|
||||
constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
|
||||
constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
|
||||
|
||||
constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
|
||||
|
||||
constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
|
||||
constexpr auto ds_read_a_issue_cycle =
|
||||
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
|
||||
constexpr auto ds_read_b_issue_cycle =
|
||||
HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16 ? 8 : 4;
|
||||
constexpr auto ds_read_a_mfma_rate =
|
||||
(mfma_cycle - 8 + ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
|
||||
constexpr auto ds_read_b_mfma_rate =
|
||||
(mfma_cycle - 8 + ds_read_b_issue_cycle - 1) / (2 * ds_read_b_issue_cycle);
|
||||
|
||||
constexpr auto num_dsread_stage1_a = num_ds_read_inst_a / KRepeat * (KRepeat - 1);
|
||||
constexpr auto num_dsread_stage1_b = num_ds_read_inst_b / KRepeat * (KRepeat - 1);
|
||||
constexpr auto num_dsread_stage3_a = num_ds_read_inst_a / KRepeat;
|
||||
constexpr auto num_dsread_stage3_b = num_ds_read_inst_b / KRepeat;
|
||||
|
||||
constexpr auto num_dsread_stage1_a_mfma =
|
||||
(num_dsread_stage1_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
|
||||
constexpr auto num_dsread_stage1_b_mfma =
|
||||
(num_dsread_stage1_b + ds_read_b_mfma_rate - 1) / ds_read_b_mfma_rate;
|
||||
constexpr auto num_dsread_stage3_a_mfma =
|
||||
(num_dsread_stage3_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
|
||||
constexpr auto num_dsread_stage3_b_mfma =
|
||||
(num_dsread_stage3_b + ds_read_b_mfma_rate - 1) / ds_read_b_mfma_rate;
|
||||
|
||||
constexpr auto num_mfma_stage2 = num_mfma_inst - num_ds_read_inst_a / ds_read_a_mfma_rate -
|
||||
num_ds_read_inst_b / ds_read_b_mfma_rate;
|
||||
constexpr auto num_mfma_per_issue =
|
||||
num_mfma_stage2 / (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;
|
||||
|
||||
// stage 1
|
||||
static_for<0, num_dsread_stage1_a_mfma, 1>{}([&](auto i) {
|
||||
ignore = i;
|
||||
if constexpr((num_dsread_stage1_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_dsread_stage1_a - (num_dsread_stage1_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_stage1_b_mfma, 1>{}([&](auto i) {
|
||||
ignore = i;
|
||||
if constexpr((num_dsread_stage1_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_dsread_stage1_b - (num_dsread_stage1_b_mfma - 1) * ds_read_b_mfma_rate,
|
||||
0); // DS read
|
||||
}
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
|
||||
});
|
||||
|
||||
// stage 2
|
||||
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 3
|
||||
static_for<0, num_dsread_stage3_a_mfma, 1>{}([&](auto i) {
|
||||
ignore = i;
|
||||
if constexpr((num_dsread_stage3_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_dsread_stage3_a - (num_dsread_stage3_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_stage3_b_mfma, 1>{}([&](auto i) {
|
||||
ignore = i;
|
||||
if constexpr((num_dsread_stage3_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_dsread_stage3_b - (num_dsread_stage3_b_mfma - 1) * ds_read_b_mfma_rate,
|
||||
0); // DS read
|
||||
}
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
|
||||
});
|
||||
|
||||
// IGLP COMPILER BUG:
|
||||
// If comment out following scheduler barrier would cause sanity fail.
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
|
||||
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>
|
||||
__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,
|
||||
CThreadBuffer& c_thread_buf,
|
||||
index_t num_loop) const
|
||||
{
|
||||
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());
|
||||
|
||||
// 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);
|
||||
|
||||
// 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);
|
||||
|
||||
// Global prefetch 2
|
||||
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);
|
||||
|
||||
// Global prefetch 3
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1);
|
||||
|
||||
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();
|
||||
|
||||
// Local prefetch 1
|
||||
block_sync_lds();
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, I0),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, I0, 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, I0),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, I0, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
|
||||
// main body
|
||||
if constexpr(HasMainLoop)
|
||||
{
|
||||
index_t i = 0;
|
||||
do
|
||||
{
|
||||
auto LoopFunc = [&](auto vmem_buf) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
if constexpr(k0 == (KRepeat - 1))
|
||||
{
|
||||
block_sync_lds();
|
||||
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, vmem_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, vmem_buf);
|
||||
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, vmem_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, vmem_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, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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, I0, ik))>{}];
|
||||
});
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, I0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType,
|
||||
xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
|
||||
a_thread_copy_.Run(
|
||||
a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<(k0 + 1) % KRepeat * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, I0, 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 + 1) % KRepeat * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, I0, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
|
||||
HotLoopScheduler();
|
||||
};
|
||||
|
||||
LoopFunc(I0);
|
||||
LoopFunc(I1);
|
||||
|
||||
i += HotloopUnroll;
|
||||
} while(i < (num_loop - PrefetchStages));
|
||||
}
|
||||
// tail
|
||||
auto ReadWriteCompFunc = [&](auto vmem_buf) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
if constexpr(k0 == (KRepeat - 1))
|
||||
{
|
||||
block_sync_lds();
|
||||
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, vmem_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, vmem_buf);
|
||||
|
||||
block_sync_lds();
|
||||
}
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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, I0, ik))>{}];
|
||||
});
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, I0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
a_thread_copy_.Run(
|
||||
a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<(k0 + 1) % KRepeat * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, I0, 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 + 1) % KRepeat * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, I0, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
|
||||
HotLoopScheduler();
|
||||
};
|
||||
auto ReadCompFunc = [&]() {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KRepeat - 1, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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, I0, ik))>{}];
|
||||
});
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, I0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
|
||||
a_thread_copy_.Run(
|
||||
a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<(k0 + 1) % KRepeat * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, I0, 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 + 1) % KRepeat * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, I0, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
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, I0, ik))>{}];
|
||||
});
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) = b_thread_buf
|
||||
[Number<b_thread_desc_.CalculateOffset(make_tuple(n0, I0, I0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.template Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
|
||||
HotLoopScheduler();
|
||||
};
|
||||
|
||||
if constexpr(TailNum == TailNumber::Odd)
|
||||
{
|
||||
ReadWriteCompFunc(I0);
|
||||
ReadWriteCompFunc(I1);
|
||||
ReadCompFunc();
|
||||
}
|
||||
else if constexpr(TailNum == TailNumber::Even)
|
||||
{
|
||||
ReadWriteCompFunc(I0);
|
||||
ReadCompFunc();
|
||||
}
|
||||
}
|
||||
|
||||
protected:
|
||||
// A[MRepeat, I1, I1, KPack]
|
||||
static constexpr auto a_thread_desc_ =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, I1, I1, Number<KPack>{}));
|
||||
|
||||
// B[NRepeat, N1, N2, KPack]
|
||||
static constexpr auto b_thread_desc_ =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(Number<NRepeat>{}, I1, I1, Number<KPack>{}));
|
||||
|
||||
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType,
|
||||
ComputeDataType,
|
||||
decltype(a_block_desc_m0_m1_m2_k),
|
||||
decltype(a_thread_desc_),
|
||||
Sequence<1, 1, 1, KPack>,
|
||||
Sequence<0, 1, 2, 3>,
|
||||
3,
|
||||
A_K1,
|
||||
A_K1>;
|
||||
|
||||
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<BDataType,
|
||||
ComputeDataType,
|
||||
decltype(b_block_desc_n0_n1_n2_k),
|
||||
decltype(b_thread_desc_),
|
||||
Sequence<1, 1, 1, KPack>,
|
||||
Sequence<0, 1, 2, 3>,
|
||||
3,
|
||||
B_K1,
|
||||
B_K1>;
|
||||
|
||||
AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex()};
|
||||
BThreadCopy b_thread_copy_{Base::CalculateBThreadOriginDataIndex()};
|
||||
using Base::c_thread_desc_;
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user