mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-04 13:41:24 +00:00
GEMM with Multiple Source, GEMM+Bias+Add+FastGeLU example and ckProfiler (#241)
* ad gelu and fast_gelu * added GeLU and fast GeLU * clean up * add gemm+fastgelu example * add gemm+gelu instances * update profiler * clean up * clean up * adding gemm+bias+activation * clean * adding bias * clean * adding gemm multiple d * debugging * add gemm bias add fastgelu * rename, clean * refactoring; add readme * refactor * refactor * refactor * refactor * refactor * refactor * fix * fix * update example * update example * rename * update example * add ckProfiler * clean * clean * clean * clean * add comment * use type_convert * clean * clean element wise op
This commit is contained in:
Chao Liu
@@ -0,0 +1,668 @@
|
||||
#pragma once
|
||||
|
||||
#include "common_header.hpp"
|
||||
#include "multi_index_transform_helper.hpp"
|
||||
#include "tensor_descriptor.hpp"
|
||||
#include "tensor_descriptor_helper.hpp"
|
||||
#include "tensor_operation/gpu/grid/block_to_ctile_map.hpp"
|
||||
#include "blockwise_gemm_xdlops.hpp"
|
||||
#include "thread_group_tensor_slice_transfer_v4r1.hpp"
|
||||
#include "thread_group_tensor_slice_transfer_v7.hpp"
|
||||
#include "threadwise_tensor_slice_transfer.hpp"
|
||||
#include "gridwise_gemm_pipeline_v1.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
// input : A[AK0, M, AK1]
|
||||
// input : B[AK0, N, AK1]
|
||||
// input : D0[M, N], D1[M, N], ...
|
||||
// output : E[M, N]
|
||||
// C = a_op(A) * b_op(B)
|
||||
// E = cde_op(C, D0, D1, ...)
|
||||
template <typename FloatAB,
|
||||
typename FloatGemmAcc,
|
||||
typename FloatCShuffle,
|
||||
typename DsDataType,
|
||||
typename FloatE,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CDEElementwiseOperation,
|
||||
InMemoryDataOperationEnum EGlobalMemoryDataOperation,
|
||||
typename AGridDesc_AK0_M_AK1,
|
||||
typename BGridDesc_BK0_N_BK1,
|
||||
typename EGridDesc_M_N,
|
||||
index_t NumGemmKPrefetchStage,
|
||||
index_t BlockSize,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t AK1Value,
|
||||
index_t BK1Value,
|
||||
index_t MPerXdl,
|
||||
index_t NPerXdl,
|
||||
index_t MXdlPerWave,
|
||||
index_t NXdlPerWave,
|
||||
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
typename ABlockTransferThreadClusterArrangeOrder,
|
||||
typename ABlockTransferSrcAccessOrder,
|
||||
index_t ABlockTransferSrcVectorDim,
|
||||
index_t ABlockTransferSrcScalarPerVector,
|
||||
index_t ABlockTransferDstScalarPerVector_AK1,
|
||||
bool AThreadTransferSrcResetCoordinateAfterRun,
|
||||
index_t ABlockLdsExtraM,
|
||||
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
|
||||
typename BBlockTransferThreadClusterArrangeOrder,
|
||||
typename BBlockTransferSrcAccessOrder,
|
||||
index_t BBlockTransferSrcVectorDim,
|
||||
index_t BBlockTransferSrcScalarPerVector,
|
||||
index_t BBlockTransferDstScalarPerVector_BK1,
|
||||
bool BThreadTransferSrcResetCoordinateAfterRun,
|
||||
index_t BBlockLdsExtraN,
|
||||
index_t CShuffleMXdlPerWavePerShuffle,
|
||||
index_t CShuffleNXdlPerWavePerShuffle,
|
||||
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
|
||||
LoopScheduler LoopSched>
|
||||
struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
|
||||
{
|
||||
static constexpr index_t NumDTensor = DsDataType::Size();
|
||||
|
||||
static constexpr auto I0 = Number<0>{};
|
||||
static constexpr auto I1 = Number<1>{};
|
||||
static constexpr auto I2 = Number<2>{};
|
||||
static constexpr auto I3 = Number<3>{};
|
||||
static constexpr auto I4 = Number<4>{};
|
||||
static constexpr auto I5 = Number<5>{};
|
||||
static constexpr auto I6 = Number<6>{};
|
||||
static constexpr auto I7 = Number<7>{};
|
||||
|
||||
// K1 should be Number<...>
|
||||
static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
|
||||
static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
|
||||
static constexpr auto AK1 = Number<AK1Value>{};
|
||||
static constexpr auto BK1 = Number<BK1Value>{};
|
||||
|
||||
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
|
||||
|
||||
using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemmKPrefetchStage>;
|
||||
|
||||
__host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
|
||||
{
|
||||
// A matrix in LDS memory, dst of blockwise copy
|
||||
return make_naive_tensor_descriptor(
|
||||
make_tuple(AK0, Number<MPerBlock>{}, AK1),
|
||||
make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
|
||||
{
|
||||
// B matrix in LDS memory, dst of blockwise copy
|
||||
return make_naive_tensor_descriptor(
|
||||
make_tuple(BK0, Number<NPerBlock>{}, BK1),
|
||||
make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr auto
|
||||
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
|
||||
{
|
||||
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
|
||||
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
|
||||
|
||||
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
|
||||
make_naive_tensor_descriptor_packed(
|
||||
make_tuple(I1,
|
||||
Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
|
||||
I1,
|
||||
Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
|
||||
|
||||
return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
|
||||
}
|
||||
|
||||
// ck::Tuple<const D0DataType*, const D1DataType*, ...>
|
||||
static constexpr auto MakeDsGridPointer()
|
||||
{
|
||||
return generate_tuple(
|
||||
[&](auto i) {
|
||||
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
|
||||
|
||||
return static_cast<const DDataType*>(nullptr);
|
||||
},
|
||||
Number<NumDTensor>{});
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
|
||||
{
|
||||
// LDS allocation for A and B: be careful of alignment
|
||||
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
|
||||
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
|
||||
|
||||
// lds max alignment
|
||||
constexpr auto max_lds_align = math::lcm(AK1, BK1);
|
||||
|
||||
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
|
||||
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
|
||||
|
||||
constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
|
||||
b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
|
||||
|
||||
// LDS allocation for C shuffle in LDS
|
||||
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
|
||||
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
|
||||
|
||||
constexpr auto c_block_size =
|
||||
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
|
||||
|
||||
return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
|
||||
sizeof(FloatAB),
|
||||
c_block_size * sizeof(FloatCShuffle));
|
||||
}
|
||||
|
||||
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
|
||||
template <typename Block2ETileMap>
|
||||
__host__ __device__ static constexpr bool
|
||||
CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
|
||||
const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
|
||||
const EGridDesc_M_N& e_grid_desc_m_n,
|
||||
const Block2ETileMap& block_2_etile_map)
|
||||
{
|
||||
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
|
||||
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
|
||||
"Invalid tuning param!");
|
||||
|
||||
const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
|
||||
const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
|
||||
const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
|
||||
|
||||
if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
|
||||
return false;
|
||||
|
||||
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
|
||||
return false;
|
||||
|
||||
// check gridwise gemm pipeline
|
||||
const auto num_k_loop = K / KPerBlock;
|
||||
|
||||
if(!GridwiseGemmPipe::IsSupported(num_k_loop))
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
|
||||
return true;
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
|
||||
{
|
||||
const index_t num_loop = K / KPerBlock;
|
||||
|
||||
return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr auto
|
||||
MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
|
||||
{
|
||||
const auto M = e_grid_desc_m_n.GetLength(I0);
|
||||
const auto N = e_grid_desc_m_n.GetLength(I1);
|
||||
|
||||
const auto MBlock = M / MPerBlock;
|
||||
const auto NBlock = N / NPerBlock;
|
||||
|
||||
const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
|
||||
e_grid_desc_m_n,
|
||||
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
|
||||
make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
|
||||
|
||||
return e_grid_desc_mblock_mperblock_nblock_nperblock;
|
||||
}
|
||||
|
||||
// return block_id to E matrix tile idx (m0, n0) mapping
|
||||
__host__ __device__ static constexpr auto
|
||||
MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
|
||||
{
|
||||
return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
|
||||
e_grid_desc_m_n);
|
||||
}
|
||||
|
||||
using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
|
||||
MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
|
||||
|
||||
using DefaultBlock2ETileMap =
|
||||
remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
|
||||
|
||||
using DsGridPointer = decltype(MakeDsGridPointer());
|
||||
|
||||
template <bool HasMainKBlockLoop, typename Block2ETileMap>
|
||||
__device__ static void
|
||||
Run(const FloatAB* __restrict__ p_a_grid,
|
||||
const FloatAB* __restrict__ p_b_grid,
|
||||
DsGridPointer p_ds_grid,
|
||||
FloatE* __restrict__ p_e_grid,
|
||||
void* __restrict__ p_shared,
|
||||
const AElementwiseOperation& a_element_op,
|
||||
const BElementwiseOperation& b_element_op,
|
||||
const CDEElementwiseOperation& cde_element_op,
|
||||
const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
|
||||
const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
|
||||
const StaticallyIndexedArray<EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
NumDTensor>&
|
||||
ds_grid_desc_mblock_mperblock_nblock_nperblock, // FIXME: Ds desc may be of different
|
||||
// type from E
|
||||
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
|
||||
e_grid_desc_mblock_mperblock_nblock_nperblock,
|
||||
const Block2ETileMap& block_2_etile_map)
|
||||
{
|
||||
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
|
||||
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
|
||||
|
||||
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
|
||||
p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
|
||||
|
||||
const auto ds_grid_buf = generate_tuple(
|
||||
[&](auto i) {
|
||||
return make_dynamic_buffer<AddressSpaceEnum::Global>(
|
||||
p_ds_grid[i],
|
||||
ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
|
||||
},
|
||||
Number<NumDTensor>{});
|
||||
|
||||
auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
|
||||
p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
|
||||
|
||||
// divide block work by [M, N]
|
||||
const auto block_work_idx =
|
||||
block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
|
||||
|
||||
if(!block_2_etile_map.ValidCTileIndex(
|
||||
block_work_idx,
|
||||
make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
|
||||
e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
// HACK: this force m/n_block_data_idx_on_grid into SGPR
|
||||
const index_t m_block_data_idx_on_grid =
|
||||
__builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
|
||||
|
||||
const index_t n_block_data_idx_on_grid =
|
||||
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
|
||||
|
||||
// lds max alignment
|
||||
constexpr auto max_lds_align = math::lcm(AK1, BK1);
|
||||
|
||||
// A matrix in LDS memory, dst of blockwise copy
|
||||
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
|
||||
|
||||
// B matrix in LDS memory, dst of blockwise copy
|
||||
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
|
||||
|
||||
// A matrix blockwise copy
|
||||
auto a_blockwise_copy =
|
||||
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
|
||||
AElementwiseOperation,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
Sequence<AK0, MPerBlock, AK1>,
|
||||
ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
ABlockTransferThreadClusterArrangeOrder,
|
||||
FloatAB,
|
||||
FloatAB,
|
||||
decltype(a_grid_desc_ak0_m_ak1),
|
||||
decltype(a_block_desc_ak0_m_ak1),
|
||||
ABlockTransferSrcAccessOrder,
|
||||
Sequence<1, 0, 2>,
|
||||
ABlockTransferSrcVectorDim,
|
||||
2,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
ABlockTransferDstScalarPerVector_AK1,
|
||||
1,
|
||||
1,
|
||||
AThreadTransferSrcResetCoordinateAfterRun,
|
||||
true,
|
||||
NumGemmKPrefetchStage>(
|
||||
a_grid_desc_ak0_m_ak1,
|
||||
make_multi_index(0, m_block_data_idx_on_grid, 0),
|
||||
a_element_op,
|
||||
a_block_desc_ak0_m_ak1,
|
||||
make_multi_index(0, 0, 0),
|
||||
ck::tensor_operation::element_wise::PassThrough{});
|
||||
|
||||
// B matrix blockwise copy
|
||||
auto b_blockwise_copy =
|
||||
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
|
||||
BElementwiseOperation,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
Sequence<BK0, NPerBlock, BK1>,
|
||||
BBlockTransferThreadClusterLengths_BK0_N_BK1,
|
||||
BBlockTransferThreadClusterArrangeOrder,
|
||||
FloatAB,
|
||||
FloatAB,
|
||||
decltype(b_grid_desc_bk0_n_bk1),
|
||||
decltype(b_block_desc_bk0_n_bk1),
|
||||
BBlockTransferSrcAccessOrder,
|
||||
Sequence<1, 0, 2>,
|
||||
BBlockTransferSrcVectorDim,
|
||||
2,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
BBlockTransferDstScalarPerVector_BK1,
|
||||
1,
|
||||
1,
|
||||
BThreadTransferSrcResetCoordinateAfterRun,
|
||||
true,
|
||||
NumGemmKPrefetchStage>(
|
||||
b_grid_desc_bk0_n_bk1,
|
||||
make_multi_index(0, n_block_data_idx_on_grid, 0),
|
||||
b_element_op,
|
||||
b_block_desc_bk0_n_bk1,
|
||||
make_multi_index(0, 0, 0),
|
||||
ck::tensor_operation::element_wise::PassThrough{});
|
||||
|
||||
// GEMM definition
|
||||
// c_mtx += transpose(a_mtx) * b_mtx
|
||||
// a_mtx[K0PerBlock, MPerBlock] is in LDS
|
||||
// b_mtx[K0PerBlock, NPerBlock] is in LDS
|
||||
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
|
||||
// register
|
||||
// sanity check
|
||||
constexpr index_t KPack = math::max(
|
||||
math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
|
||||
|
||||
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
|
||||
BlockSize,
|
||||
FloatAB,
|
||||
FloatGemmAcc,
|
||||
decltype(a_block_desc_ak0_m_ak1),
|
||||
decltype(b_block_desc_bk0_n_bk1),
|
||||
MPerXdl,
|
||||
NPerXdl,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
KPack,
|
||||
LoopSched>();
|
||||
|
||||
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
|
||||
|
||||
// LDS allocation for A and B: be careful of alignment
|
||||
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
|
||||
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
|
||||
|
||||
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
|
||||
static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
|
||||
|
||||
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
|
||||
static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
|
||||
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
|
||||
|
||||
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
|
||||
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
|
||||
|
||||
// gridwise GEMM pipeline
|
||||
const auto gridwise_gemm_pipeline =
|
||||
GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
|
||||
|
||||
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
|
||||
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
|
||||
KPerBlock);
|
||||
|
||||
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
|
||||
a_block_desc_ak0_m_ak1,
|
||||
a_blockwise_copy,
|
||||
a_grid_buf,
|
||||
a_block_buf,
|
||||
a_block_slice_copy_step,
|
||||
b_grid_desc_bk0_n_bk1,
|
||||
b_block_desc_bk0_n_bk1,
|
||||
b_blockwise_copy,
|
||||
b_grid_buf,
|
||||
b_block_buf,
|
||||
b_block_slice_copy_step,
|
||||
blockwise_gemm,
|
||||
c_thread_buf,
|
||||
num_k_block_main_loop);
|
||||
|
||||
// shuffle C and write out
|
||||
{
|
||||
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
|
||||
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
|
||||
"wrong!");
|
||||
|
||||
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
|
||||
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
|
||||
|
||||
// TODO: hacky, fix it!
|
||||
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
|
||||
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
|
||||
|
||||
// TODO: hacky, fix it!
|
||||
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
|
||||
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
|
||||
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
|
||||
|
||||
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
|
||||
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
|
||||
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
|
||||
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
|
||||
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
|
||||
constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
|
||||
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
|
||||
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
|
||||
|
||||
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
|
||||
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
|
||||
|
||||
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
|
||||
static_cast<FloatCShuffle*>(p_shared),
|
||||
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
|
||||
|
||||
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
|
||||
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
|
||||
make_tuple(
|
||||
make_freeze_transform(I0),
|
||||
make_unmerge_transform(make_tuple(
|
||||
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
|
||||
M1, // M1 = MWave
|
||||
M2, // M2 * M3 * M4 = MPerXdl
|
||||
M3,
|
||||
M4)),
|
||||
make_freeze_transform(I0),
|
||||
make_unmerge_transform(make_tuple(
|
||||
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
|
||||
N1, // N1 = NWave
|
||||
N2))), // N2 = NPerXdl
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
|
||||
make_tuple(
|
||||
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
|
||||
|
||||
// calculate origin of thread output tensor on global memory
|
||||
// blockwise GEMM c matrix starting index
|
||||
const auto c_thread_mtx_on_block =
|
||||
blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
|
||||
|
||||
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
|
||||
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
|
||||
|
||||
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
|
||||
make_single_stage_tensor_adaptor(
|
||||
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
|
||||
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
|
||||
make_tuple(Sequence<0>{}));
|
||||
|
||||
const auto m_thread_data_on_block_idx =
|
||||
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
|
||||
make_multi_index(m_thread_data_on_block));
|
||||
|
||||
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
|
||||
make_single_stage_tensor_adaptor(
|
||||
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
|
||||
make_tuple(Sequence<0, 1, 2>{}),
|
||||
make_tuple(Sequence<0>{}));
|
||||
|
||||
const auto n_thread_data_on_block_idx =
|
||||
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
|
||||
make_multi_index(n_thread_data_on_block));
|
||||
|
||||
// shuffle: threadwise copy C from VGPR to LDS
|
||||
auto c_thread_copy_vgpr_to_lds =
|
||||
ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
|
||||
FloatCShuffle,
|
||||
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
|
||||
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
Sequence<CShuffleMXdlPerWavePerShuffle,
|
||||
CShuffleNXdlPerWavePerShuffle,
|
||||
I1,
|
||||
I1,
|
||||
M2,
|
||||
I1,
|
||||
M4,
|
||||
I1>,
|
||||
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
|
||||
7,
|
||||
1,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
1,
|
||||
true>{
|
||||
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
|
||||
make_multi_index(0,
|
||||
0,
|
||||
m_thread_data_on_block_idx[I1],
|
||||
n_thread_data_on_block_idx[I1],
|
||||
m_thread_data_on_block_idx[I2],
|
||||
m_thread_data_on_block_idx[I3],
|
||||
m_thread_data_on_block_idx[I4],
|
||||
n_thread_data_on_block_idx[I2]),
|
||||
ck::tensor_operation::element_wise::PassThrough{}};
|
||||
|
||||
// tuple of reference to C/Ds tensor descriptors
|
||||
const auto c_ds_desc_refs = concat_tuple_of_reference(
|
||||
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
|
||||
generate_tie(
|
||||
[&](auto i) -> const auto& // return type should be reference
|
||||
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
|
||||
Number<NumDTensor>{}));
|
||||
|
||||
// tuple of reference to C/Ds tensor descriptors
|
||||
const auto c_ds_buf_refs = concat_tuple_of_reference(
|
||||
tie(c_shuffle_block_buf),
|
||||
generate_tie(
|
||||
[&](auto i) -> const auto& // return type should be reference
|
||||
{ return ds_grid_buf[i]; },
|
||||
Number<NumDTensor>{}));
|
||||
|
||||
// tuple of starting index of C/Ds blockwise copy
|
||||
const auto idx_c_ds_block_begin = container_concat(
|
||||
make_tuple(make_multi_index(0, 0, 0, 0)),
|
||||
generate_tuple(
|
||||
[&](auto) {
|
||||
return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
|
||||
},
|
||||
Number<NumDTensor>{}));
|
||||
|
||||
// blockwise copy C/D/E between LDS and global
|
||||
auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
|
||||
ThisThreadBlock,
|
||||
decltype(container_concat(make_tuple(FloatCShuffle{}), DsDataType{})),
|
||||
Tuple<FloatE>,
|
||||
decltype(c_ds_desc_refs),
|
||||
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
|
||||
CDEElementwiseOperation,
|
||||
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
|
||||
// support arbitray type
|
||||
Sequence<1,
|
||||
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
|
||||
1,
|
||||
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
|
||||
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
|
||||
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
|
||||
3, // index_t VectorDim,
|
||||
CDEShuffleBlockTransferScalarPerVector_NPerBlock,
|
||||
sequence_merge_t<
|
||||
Sequence<true>,
|
||||
uniform_sequence_gen_t<NumDTensor,
|
||||
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
|
||||
Sequence<false>> // ThreadTransferDstResetCoordinateAfterRunFlags
|
||||
{c_ds_desc_refs,
|
||||
idx_c_ds_block_begin,
|
||||
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
|
||||
make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
|
||||
cde_element_op};
|
||||
|
||||
// space filling curve for threadwise C in VGPR before shuffle
|
||||
constexpr auto sfc_c_vgpr =
|
||||
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
|
||||
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
|
||||
Sequence<CShuffleMXdlPerWavePerShuffle,
|
||||
CShuffleNXdlPerWavePerShuffle,
|
||||
1,
|
||||
1,
|
||||
M2,
|
||||
1,
|
||||
M4,
|
||||
1>>{};
|
||||
|
||||
// space filling curve for shuffled blockwise C/D/E
|
||||
constexpr auto sfc_cde_block =
|
||||
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
|
||||
Sequence<0, 2, 1, 3>,
|
||||
Sequence<1,
|
||||
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
|
||||
1,
|
||||
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
|
||||
|
||||
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
|
||||
|
||||
static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
|
||||
|
||||
static_for<0, num_access, 1>{}([&](auto access_id) {
|
||||
// make sure it's safe to write to LDS
|
||||
block_sync_lds();
|
||||
|
||||
// each thread write its data from VGPR to LDS
|
||||
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
|
||||
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
|
||||
c_thread_buf,
|
||||
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
|
||||
c_shuffle_block_buf);
|
||||
|
||||
// make sure it's safe to read from LDS
|
||||
block_sync_lds();
|
||||
|
||||
// each block copy its data from LDS to global
|
||||
cde_block_copy_lds_and_global.Run(
|
||||
c_ds_desc_refs,
|
||||
c_ds_buf_refs,
|
||||
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
|
||||
tie(e_grid_buf));
|
||||
|
||||
if constexpr(access_id < num_access - 1)
|
||||
{
|
||||
constexpr auto cde_lds_and_global_step =
|
||||
sfc_cde_block.GetForwardStep(access_id);
|
||||
|
||||
// move on Ds
|
||||
static_for<0, NumDTensor, 1>{}([&](auto i) {
|
||||
cde_block_copy_lds_and_global.MoveSrcSliceWindow(
|
||||
c_ds_desc_refs, i + I1, cde_lds_and_global_step);
|
||||
});
|
||||
|
||||
// move on E
|
||||
cde_block_copy_lds_and_global.MoveDstSliceWindow(
|
||||
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
|
||||
I0,
|
||||
cde_lds_and_global_step);
|
||||
}
|
||||
});
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user