mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-03 13:11:25 +00:00
Grouped Gemm with Fixed K and N with SplitK (#818)
* move all arguments into device * add b2c_tile_map * add examples * add SetDeviceKernelArgs * dedicated fixed_nk solution * init client api * add grouped_gemm_bias example * add a instance * add instances * formatting * fixed cmake * Update EnableCompilerWarnings.cmake * Update cmake-ck-dev.sh * clean; fixed comments * fixed comment * add instances for fp32 output * add instances for fp32 output * add fp32 out client example * fixed CI * init commit for kbatch * add splitk gridwise * format * fixed * clean deviceop * clean code * finish splitk * fixed instances * change m_loops to tile_loops * add setkbatch * clean code * add splitK+bias * add instances * opt mk_nk instances * clean examples * fixed CI * remove zero * finished non-zero * clean * clean code * optimized global_barrier * fixed ci * fixed CI * removed AddBias * format * fixed CI * fixed CI * move 20_grouped_gemm to 21_grouped_gemm --------- Co-authored-by: Jing Zhang <jizha@amd.com>
This commit is contained in:
zjing14
@@ -0,0 +1,836 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <iostream>
|
||||
#include <sstream>
|
||||
|
||||
#include "ck/utility/common_header.hpp"
|
||||
#include "ck/tensor_description/tensor_descriptor.hpp"
|
||||
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp"
|
||||
#include "ck/host_utility/device_prop.hpp"
|
||||
#include "ck/host_utility/kernel_launch.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
|
||||
template <typename GridwiseGemm,
|
||||
typename GemmDesc,
|
||||
GemmSpecialization GemmSpec,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename DsDataType,
|
||||
typename Block2ETileMap,
|
||||
typename GroupedGemmBlock2ETileMap,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CDEElementwiseOperation,
|
||||
InMemoryDataOperationEnum EGlobalMemoryDataOperation,
|
||||
bool HasMainKBlockLoop>
|
||||
__global__ void
|
||||
#if CK_USE_LAUNCH_BOUNDS
|
||||
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
|
||||
#endif
|
||||
kernel_grouped_gemm_xdl_fixed_nk(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
|
||||
uint32_t* barrier_count,
|
||||
const index_t barrier_size_grp,
|
||||
const index_t group_count,
|
||||
const index_t grid_size_grp,
|
||||
const index_t KBatch,
|
||||
const AElementwiseOperation a_element_op,
|
||||
const BElementwiseOperation b_element_op,
|
||||
const CDEElementwiseOperation c_element_op)
|
||||
{
|
||||
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
|
||||
defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
|
||||
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
|
||||
|
||||
const index_t block_id = get_block_1d_id();
|
||||
|
||||
const auto gemm_desc_ptr =
|
||||
reinterpret_cast<const GemmDesc*>(cast_pointer_to_generic_address_space(gemm_descs_const));
|
||||
|
||||
const index_t group_id = block_id / grid_size_grp;
|
||||
|
||||
if(group_id >= group_count)
|
||||
return;
|
||||
|
||||
const index_t M = gemm_desc_ptr[group_id].M;
|
||||
const index_t N = gemm_desc_ptr[group_id].N;
|
||||
const index_t K = gemm_desc_ptr[group_id].K;
|
||||
|
||||
if(M * N * K == 0)
|
||||
return;
|
||||
|
||||
const auto StrideA = gemm_desc_ptr[group_id].StrideA;
|
||||
const auto StrideB = gemm_desc_ptr[group_id].StrideB;
|
||||
const auto StrideDs = gemm_desc_ptr[group_id].StrideDs;
|
||||
const auto StrideE = gemm_desc_ptr[group_id].StrideE;
|
||||
|
||||
const auto e_grid_desc_m_n =
|
||||
GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
|
||||
|
||||
const index_t BlockStart = group_id * grid_size_grp;
|
||||
|
||||
const auto local_b2e_tile_map = Block2ETileMap{e_grid_desc_m_n, KBatch};
|
||||
|
||||
const auto local_grid_size = local_b2e_tile_map.CalculateGridSize(e_grid_desc_m_n);
|
||||
|
||||
constexpr auto NumDTensor = DsDataType::Size();
|
||||
|
||||
using DsGridPointer = decltype(GridwiseGemm::MakeDsGridPointer());
|
||||
|
||||
DsGridPointer p_ds_grid_;
|
||||
|
||||
static_for<0, NumDTensor, 1>{}([&](auto i) {
|
||||
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
|
||||
// D pointer
|
||||
p_ds_grid_(i) = static_cast<const DDataType*>(gemm_desc_ptr[group_id].p_ds_grid[i]);
|
||||
});
|
||||
|
||||
index_t id_off = 0;
|
||||
index_t id_local = get_block_1d_id() - BlockStart;
|
||||
|
||||
const index_t mn_blocks = local_grid_size / KBatch;
|
||||
|
||||
while(id_local < local_grid_size)
|
||||
{
|
||||
const auto block_2_etile_map =
|
||||
GroupedGemmBlock2ETileMap(local_b2e_tile_map, BlockStart, id_off);
|
||||
|
||||
auto barrier_count_finished =
|
||||
barrier_count + group_id * barrier_size_grp + id_local % mn_blocks;
|
||||
|
||||
GridwiseGemm::template Run<HasMainKBlockLoop,
|
||||
EGlobalMemoryDataOperation,
|
||||
GemmSpec,
|
||||
ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout>(gemm_desc_ptr[group_id].p_a_grid,
|
||||
gemm_desc_ptr[group_id].p_b_grid,
|
||||
p_ds_grid_,
|
||||
gemm_desc_ptr[group_id].p_e_grid,
|
||||
p_shared,
|
||||
barrier_count_finished,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideDs,
|
||||
StrideE,
|
||||
KBatch,
|
||||
block_2_etile_map);
|
||||
|
||||
id_off += grid_size_grp;
|
||||
id_local += grid_size_grp;
|
||||
}
|
||||
#else
|
||||
ignore = gemm_descs_const;
|
||||
ignore = barrier_count;
|
||||
ignore = barrier_size_grp;
|
||||
ignore = group_count;
|
||||
ignore = grid_size_grp;
|
||||
ignore = KBatch;
|
||||
ignore = a_element_op;
|
||||
ignore = b_element_op;
|
||||
ignore = c_element_op;
|
||||
#endif
|
||||
}
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename AccDataType,
|
||||
typename CShuffleDataType,
|
||||
typename DsDataType,
|
||||
typename EDataType,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CDEElementwiseOperation,
|
||||
GemmSpecialization GemmSpec,
|
||||
ck::index_t NumPrefetch,
|
||||
ck::index_t BlockSize,
|
||||
ck::index_t MPerBlock,
|
||||
ck::index_t NPerBlock,
|
||||
ck::index_t KPerBlock,
|
||||
ck::index_t AK1,
|
||||
ck::index_t BK1,
|
||||
ck::index_t MPerXDL,
|
||||
ck::index_t NPerXDL,
|
||||
ck::index_t MXdlPerWave,
|
||||
ck::index_t NXdlPerWave,
|
||||
typename ABlockTransferThreadClusterLengths_K0_M_K1,
|
||||
typename ABlockTransferThreadClusterArrangeOrder,
|
||||
typename ABlockTransferSrcAccessOrder,
|
||||
ck::index_t ABlockTransferSrcVectorDim,
|
||||
ck::index_t ABlockTransferSrcScalarPerVector,
|
||||
ck::index_t ABlockTransferDstScalarPerVector_K1,
|
||||
bool ABlockLdsExtraM,
|
||||
typename BBlockTransferThreadClusterLengths_K0_N_K1,
|
||||
typename BBlockTransferThreadClusterArrangeOrder,
|
||||
typename BBlockTransferSrcAccessOrder,
|
||||
ck::index_t BBlockTransferSrcVectorDim,
|
||||
ck::index_t BBlockTransferSrcScalarPerVector,
|
||||
ck::index_t BBlockTransferDstScalarPerVector_K1,
|
||||
bool BBlockLdsExtraN,
|
||||
index_t CShuffleMXdlPerWavePerShuffle,
|
||||
index_t CShuffleNXdlPerWavePerShuffle,
|
||||
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
index_t CDEBlockTransferScalarPerVector_NPerBlock,
|
||||
LoopScheduler LoopSched = make_default_loop_scheduler()>
|
||||
struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CDEElementwiseOperation>
|
||||
{
|
||||
using DeviceOp = DeviceGroupedGemm_Xdl_Fixed_NK;
|
||||
|
||||
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>{};
|
||||
|
||||
// GridwiseGemm
|
||||
using GridwiseGemm = GridwiseGemmMultipleD_xdl_splitk_cshuffle<
|
||||
ADataType, // TODO: distinguish A/B datatype
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CDEElementwiseOperation,
|
||||
NumPrefetch, // NumGemmKPrefetchStage
|
||||
BlockSize,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
AK1,
|
||||
BK1,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
ABlockTransferThreadClusterLengths_K0_M_K1,
|
||||
ABlockTransferThreadClusterArrangeOrder,
|
||||
ABlockTransferSrcAccessOrder,
|
||||
ABlockTransferSrcVectorDim,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
ABlockTransferDstScalarPerVector_K1,
|
||||
false, // AThreadTransferSrcResetCoordinateAfterRun,
|
||||
ABlockLdsExtraM,
|
||||
BBlockTransferThreadClusterLengths_K0_N_K1,
|
||||
BBlockTransferThreadClusterArrangeOrder,
|
||||
BBlockTransferSrcAccessOrder,
|
||||
BBlockTransferSrcVectorDim,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
BBlockTransferDstScalarPerVector_K1,
|
||||
false, // BThreadTransferSrcResetCoordinateAfterRun,
|
||||
BBlockLdsExtraN,
|
||||
CShuffleMXdlPerWavePerShuffle,
|
||||
CShuffleNXdlPerWavePerShuffle,
|
||||
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
CDEBlockTransferScalarPerVector_NPerBlock,
|
||||
LoopSched>;
|
||||
|
||||
template <typename UnderlyingBlockToCTileMap>
|
||||
struct OffsettedBlockToCTileMapMLoops
|
||||
{
|
||||
using underlying_type = UnderlyingBlockToCTileMap;
|
||||
|
||||
__host__ __device__ OffsettedBlockToCTileMapMLoops(
|
||||
UnderlyingBlockToCTileMap block_to_ctile_map, index_t block_start, index_t id_off = 0)
|
||||
{
|
||||
block_to_ctile_map_ = block_to_ctile_map;
|
||||
block_start_ = block_start;
|
||||
id_off_ = id_off;
|
||||
}
|
||||
|
||||
template <typename TopIdx>
|
||||
__host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
|
||||
{
|
||||
auto idx_bot = block_to_ctile_map_.CalculateBottomIndex(
|
||||
make_multi_index(idx_top[Number<0>{}] - block_start_ + id_off_));
|
||||
|
||||
return make_tuple(idx_bot[Number<0>{}], idx_bot[Number<1>{}], idx_bot[Number<2>{}]);
|
||||
}
|
||||
|
||||
template <typename CTileIdx, typename CTileDim>
|
||||
__host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
|
||||
const CTileDim& c_tile_dim) const
|
||||
{
|
||||
return block_to_ctile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
|
||||
}
|
||||
|
||||
template <typename CGridDesc_M_N>
|
||||
__host__ bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
|
||||
{
|
||||
return block_to_ctile_map_.CheckValidity(c_grid_desc_m_n);
|
||||
}
|
||||
|
||||
template <typename CGridDesc_M_N>
|
||||
__host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
|
||||
{
|
||||
return block_to_ctile_map_.CalculateGridSize(c_grid_desc_m_n);
|
||||
}
|
||||
|
||||
UnderlyingBlockToCTileMap block_to_ctile_map_;
|
||||
index_t block_start_;
|
||||
index_t id_off_;
|
||||
};
|
||||
|
||||
template <index_t MPerBlock_, index_t NPerBlock_>
|
||||
struct BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops
|
||||
{
|
||||
static constexpr auto I0 = Number<0>{};
|
||||
static constexpr auto I1 = Number<1>{};
|
||||
|
||||
__host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops() = default;
|
||||
|
||||
__host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(
|
||||
const BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&) = default;
|
||||
__host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(
|
||||
BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&&) = default;
|
||||
__host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&
|
||||
operator=(const BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&) = default;
|
||||
__host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&
|
||||
operator=(BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&&) = default;
|
||||
|
||||
__host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(index_t M,
|
||||
index_t N,
|
||||
index_t KBatch,
|
||||
index_t M01 = 8)
|
||||
: M_(M), N_(N), KBatch_(KBatch), M01_(M01)
|
||||
{
|
||||
}
|
||||
|
||||
template <typename CGridDesc_M_N>
|
||||
__host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(
|
||||
const CGridDesc_M_N& c_grid_desc_m_n, index_t KBatch, index_t M01 = 8)
|
||||
: BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(
|
||||
c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1), KBatch, M01)
|
||||
{
|
||||
}
|
||||
|
||||
__host__ __device__ constexpr index_t CalculateGridSize(index_t M, index_t N) const
|
||||
{
|
||||
const auto M0 = math::integer_divide_ceil(M, MPerBlock);
|
||||
const auto N0 = math::integer_divide_ceil(N, NPerBlock);
|
||||
|
||||
return M0 * N0 * KBatch_;
|
||||
}
|
||||
|
||||
template <typename CGridDesc_M_N>
|
||||
__host__ __device__ constexpr index_t
|
||||
CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
|
||||
{
|
||||
return CalculateGridSize(c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1));
|
||||
}
|
||||
|
||||
template <typename CGridDesc_M_N>
|
||||
__host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
|
||||
{
|
||||
return true;
|
||||
}
|
||||
|
||||
template <typename TopIdx>
|
||||
__host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
|
||||
{
|
||||
auto block_1d_id = idx_top[I0];
|
||||
|
||||
const auto M0 = math::integer_divide_ceil(M_, MPerBlock_);
|
||||
const auto N0 = math::integer_divide_ceil(N_, NPerBlock_);
|
||||
|
||||
block_1d_id = block_1d_id % (M0 * N0 * KBatch_); // hide groups
|
||||
|
||||
const index_t idx_ksplit = block_1d_id / (M0 * N0);
|
||||
block_1d_id = block_1d_id % (M0 * N0);
|
||||
|
||||
index_t idx_N0 = block_1d_id % N0;
|
||||
index_t idx_M0 = block_1d_id / N0;
|
||||
|
||||
const auto M01_adapt = (idx_M0 < M0 - M0 % M01_) ? M01_ : M0 % M01_;
|
||||
|
||||
index_t idx_M00 = idx_M0 / M01_;
|
||||
index_t idx_M01 = idx_M0 % M01_;
|
||||
index_t idx_N0_M01_local = idx_N0 + idx_M01 * N0;
|
||||
|
||||
return make_tuple(idx_ksplit,
|
||||
idx_N0_M01_local % M01_adapt + idx_M00 * M01_,
|
||||
idx_N0_M01_local / M01_adapt);
|
||||
}
|
||||
|
||||
template <typename CTileIdx, typename CTileDim>
|
||||
__host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
|
||||
const CTileDim& /* c_tile_dim */) const
|
||||
{
|
||||
return true; // always valid provided that user gets grid size from CalculateGridSize()
|
||||
}
|
||||
|
||||
private:
|
||||
index_t M_;
|
||||
index_t N_;
|
||||
index_t KBatch_;
|
||||
index_t M01_;
|
||||
};
|
||||
|
||||
using Block2ETileMap = BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops<MPerBlock, NPerBlock>;
|
||||
using GroupedGemmBlock2ETileMap = OffsettedBlockToCTileMapMLoops<Block2ETileMap>;
|
||||
|
||||
struct GemmBiasTransKernelArg
|
||||
{
|
||||
// pointers
|
||||
const void* a_ptr_;
|
||||
const void* b_ptr_;
|
||||
std::array<const void*, NumDTensor> ds_ptr_;
|
||||
void* e_ptr_;
|
||||
|
||||
index_t M_, N_, K_;
|
||||
index_t StrideA_, StrideB_;
|
||||
std::array<index_t, NumDTensor> StrideDs_;
|
||||
index_t StrideE_;
|
||||
};
|
||||
|
||||
// Argument
|
||||
struct Argument : public BaseArgument
|
||||
{
|
||||
|
||||
void UpdateKBatch(index_t k_batch)
|
||||
{
|
||||
k_batch_ = k_batch;
|
||||
|
||||
if(k_batch_ < 1)
|
||||
{
|
||||
|
||||
throw std::runtime_error("wrong! k_batch must be > 0");
|
||||
}
|
||||
|
||||
const index_t AverM = math::integer_divide_ceil(sum_of_m, group_count_);
|
||||
|
||||
const index_t StrideE = gemm_desc_kernel_arg_[0].StrideE_;
|
||||
const index_t N = gemm_desc_kernel_arg_[0].N_;
|
||||
|
||||
const auto e_grid_desc_m_n =
|
||||
GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(
|
||||
AverM, N, StrideE);
|
||||
|
||||
const auto local_b2c_tile_map = Block2ETileMap{e_grid_desc_m_n, k_batch_};
|
||||
|
||||
grid_size_grp_ = local_b2c_tile_map.CalculateGridSize(e_grid_desc_m_n);
|
||||
|
||||
grid_size_ = grid_size_grp_ * group_count_;
|
||||
}
|
||||
|
||||
Argument(std::vector<const void*>&,
|
||||
std::vector<const void*>&,
|
||||
std::vector<std::array<const void*, NumDTensor>>&,
|
||||
std::vector<void*>&,
|
||||
std::vector<GemmDesc>& gemm_descs,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation c_element_op)
|
||||
: a_element_op_{a_element_op}, b_element_op_{b_element_op}, c_element_op_{c_element_op}
|
||||
{
|
||||
grid_size_ = 0;
|
||||
|
||||
k_batch_ = 1;
|
||||
|
||||
grouped_gemm_kernel_args_dev = nullptr;
|
||||
|
||||
group_count_ = ck::type_convert<ck::index_t>(gemm_descs.size());
|
||||
|
||||
gemm_desc_kernel_arg_.reserve(group_count_);
|
||||
|
||||
index_t group_id = 0;
|
||||
|
||||
sum_of_m = gemm_descs[0].M_;
|
||||
const index_t AverM = math::integer_divide_ceil(sum_of_m, group_count_);
|
||||
const index_t N = gemm_descs[0].N_;
|
||||
const index_t K = gemm_descs[0].K_;
|
||||
|
||||
for(std::size_t i = 0; i < gemm_descs.size(); i++)
|
||||
{
|
||||
if(sum_of_m != gemm_descs[i].M_ || N != gemm_descs[i].N_ || K != gemm_descs[i].K_)
|
||||
{
|
||||
throw std::runtime_error("wrong! M/N/K is not identical");
|
||||
}
|
||||
|
||||
a_mtx_mraw_kraw_.emplace_back(sum_of_m, K);
|
||||
b_mtx_nraw_kraw_.emplace_back(N, K);
|
||||
|
||||
const index_t StrideA = gemm_descs[i].stride_A_;
|
||||
const index_t StrideB = gemm_descs[i].stride_B_;
|
||||
const index_t StrideE = gemm_descs[i].stride_C_;
|
||||
|
||||
// pointer
|
||||
std::array<const void*, NumDTensor> p_ds_grid;
|
||||
|
||||
static_for<0, NumDTensor, 1>{}([&](auto j) { p_ds_grid[j] = nullptr; });
|
||||
|
||||
std::array<index_t, NumDTensor> StrideDs;
|
||||
|
||||
static_for<0, NumDTensor, 1>{}([&](auto j) {
|
||||
// using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
|
||||
|
||||
if(gemm_descs[i].stride_Ds_.size() != NumDTensor)
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! gemm_descs[i].stride_Ds_.size() does not match NumDTensor");
|
||||
}
|
||||
|
||||
StrideDs[j] = gemm_descs[i].stride_Ds_[j];
|
||||
});
|
||||
|
||||
const auto e_grid_desc_m_n =
|
||||
GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(
|
||||
AverM, N, StrideE);
|
||||
|
||||
// block-to-e-tile map
|
||||
const auto local_b2c_tile_map = Block2ETileMap{e_grid_desc_m_n, k_batch_};
|
||||
|
||||
grid_size_grp_ = local_b2c_tile_map.CalculateGridSize(e_grid_desc_m_n);
|
||||
|
||||
if(group_id * grid_size_grp_ != grid_size_)
|
||||
{
|
||||
throw std::runtime_error("wrong! grid_size_grp_ is not identical!");
|
||||
}
|
||||
|
||||
grid_size_ += grid_size_grp_;
|
||||
|
||||
// check block-to-E-tile
|
||||
if(!local_b2c_tile_map.CheckValidity(e_grid_desc_m_n))
|
||||
{
|
||||
throw std::runtime_error("wrong! block_2_etile_map validation failed");
|
||||
}
|
||||
|
||||
if(!GridwiseGemm::
|
||||
template CheckValidity<ALayout, BLayout, DsLayout, ELayout, GemmSpec>(
|
||||
AverM, N, K, StrideA, StrideB, StrideDs, StrideE, 1))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
|
||||
}
|
||||
|
||||
gemm_desc_kernel_arg_.push_back(GemmBiasTransKernelArg{
|
||||
nullptr,
|
||||
nullptr,
|
||||
p_ds_grid,
|
||||
nullptr,
|
||||
AverM,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideDs,
|
||||
StrideE,
|
||||
});
|
||||
|
||||
group_id++;
|
||||
}
|
||||
|
||||
const auto e_grid_desc_sum_m_n =
|
||||
GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(
|
||||
sum_of_m, gemm_desc_kernel_arg_[0].N_, gemm_desc_kernel_arg_[0].StrideE_);
|
||||
|
||||
const auto local_b2c_tile_map = Block2ETileMap{e_grid_desc_sum_m_n, 1};
|
||||
|
||||
barrier_size_grp_ = local_b2c_tile_map.CalculateGridSize(e_grid_desc_sum_m_n);
|
||||
}
|
||||
|
||||
// private:
|
||||
index_t group_count_;
|
||||
|
||||
AElementwiseOperation a_element_op_;
|
||||
BElementwiseOperation b_element_op_;
|
||||
CDEElementwiseOperation c_element_op_;
|
||||
|
||||
std::vector<GemmBiasTransKernelArg> gemm_desc_kernel_arg_;
|
||||
std::vector<Tuple<index_t, index_t>> a_mtx_mraw_kraw_;
|
||||
std::vector<Tuple<index_t, index_t>> b_mtx_nraw_kraw_;
|
||||
|
||||
const void* grouped_gemm_kernel_args_dev;
|
||||
|
||||
index_t grid_size_;
|
||||
index_t grid_size_grp_;
|
||||
index_t barrier_size_grp_;
|
||||
index_t sum_of_m;
|
||||
|
||||
index_t k_batch_;
|
||||
};
|
||||
|
||||
// Invoker
|
||||
struct Invoker : public BaseInvoker
|
||||
{
|
||||
using Argument = DeviceOp::Argument;
|
||||
|
||||
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
|
||||
{
|
||||
bool has_main_k_block_loop = true;
|
||||
|
||||
for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
|
||||
{
|
||||
const auto KPad =
|
||||
GridwiseGemm::CalculateKPadded(arg.gemm_desc_kernel_arg_[i].K_, arg.k_batch_);
|
||||
|
||||
if(GridwiseGemm::CalculateHasMainKBlockLoop(KPad) != has_main_k_block_loop)
|
||||
{
|
||||
throw std::runtime_error("wrong! not all gemm has_main_k_block_loop");
|
||||
}
|
||||
}
|
||||
|
||||
if(arg.grouped_gemm_kernel_args_dev == nullptr)
|
||||
{
|
||||
throw std::runtime_error("wrong! grouped_gemm_kernel_args_dev is nullpr");
|
||||
}
|
||||
|
||||
float ave_time = 0;
|
||||
|
||||
auto launch_kernel = [&](auto has_main_k_block_loop_, auto e_global_memory_operation_) {
|
||||
const auto kernel =
|
||||
kernel_grouped_gemm_xdl_fixed_nk<GridwiseGemm,
|
||||
GroupedGemmKernelArgument<NumDTensor>,
|
||||
GemmSpec,
|
||||
ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
DsDataType,
|
||||
Block2ETileMap,
|
||||
GroupedGemmBlock2ETileMap,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CDEElementwiseOperation,
|
||||
e_global_memory_operation_,
|
||||
has_main_k_block_loop_>;
|
||||
|
||||
return launch_and_time_kernel(
|
||||
stream_config,
|
||||
kernel,
|
||||
dim3(arg.grid_size_),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
cast_pointer_to_constant_address_space(arg.grouped_gemm_kernel_args_dev),
|
||||
reinterpret_cast<uint32_t*>(arg.p_workspace_),
|
||||
arg.barrier_size_grp_,
|
||||
arg.gemm_desc_kernel_arg_.size(),
|
||||
arg.grid_size_grp_,
|
||||
arg.k_batch_,
|
||||
arg.a_element_op_,
|
||||
arg.b_element_op_,
|
||||
arg.c_element_op_);
|
||||
};
|
||||
|
||||
constexpr auto AtomicAdd = InMemoryDataOperationEnum::AtomicAdd;
|
||||
constexpr auto Set = InMemoryDataOperationEnum::Set;
|
||||
|
||||
if(arg.k_batch_ > 1)
|
||||
{
|
||||
if(has_main_k_block_loop)
|
||||
{
|
||||
ave_time =
|
||||
launch_kernel(integral_constant<bool, true>{},
|
||||
integral_constant<InMemoryDataOperationEnum, AtomicAdd>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
ave_time =
|
||||
launch_kernel(integral_constant<bool, false>{},
|
||||
integral_constant<InMemoryDataOperationEnum, AtomicAdd>{});
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
if(has_main_k_block_loop)
|
||||
{
|
||||
ave_time = launch_kernel(integral_constant<bool, true>{},
|
||||
integral_constant<InMemoryDataOperationEnum, Set>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
ave_time = launch_kernel(integral_constant<bool, false>{},
|
||||
integral_constant<InMemoryDataOperationEnum, Set>{});
|
||||
}
|
||||
}
|
||||
|
||||
return ave_time;
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
float Run(const BaseArgument* p_arg,
|
||||
const StreamConfig& stream_config = StreamConfig{}) override
|
||||
{
|
||||
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
|
||||
}
|
||||
};
|
||||
|
||||
static bool IsSupportedArgument(const Argument& arg)
|
||||
{
|
||||
if(ck::type_convert<ck::index_t>(arg.gemm_desc_kernel_arg_.size()) != arg.group_count_)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
bool supported = true;
|
||||
|
||||
// If we use padding we do not support vector loads for dimensions not divisible by vector
|
||||
// load size.
|
||||
if constexpr(GemmSpec != GemmSpecialization::Default)
|
||||
{
|
||||
// [A|B]BlockTransferSrcVectorDim value define dimension in the block {K0,M,K1} layout,
|
||||
// thus we have to adapt it to the {M,K} or {N,K} layout.
|
||||
const auto a_raw_vector_dim = ABlockTransferSrcVectorDim != 1 ? 1 : 0;
|
||||
const auto b_raw_vector_dim = BBlockTransferSrcVectorDim != 1 ? 1 : 0;
|
||||
|
||||
for(index_t i = 0; i < arg.group_count_; ++i)
|
||||
{
|
||||
const auto a_vector_dim = arg.a_mtx_mraw_kraw_[i].At(Number<a_raw_vector_dim>{});
|
||||
const auto b_vector_dim = arg.b_mtx_nraw_kraw_[i].At(Number<b_raw_vector_dim>{});
|
||||
|
||||
supported = supported & (a_vector_dim % ABlockTransferSrcScalarPerVector == 0);
|
||||
supported = supported & (b_vector_dim % BBlockTransferSrcScalarPerVector == 0);
|
||||
}
|
||||
}
|
||||
|
||||
return supported;
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
bool IsSupportedArgument(const BaseArgument* p_arg) override
|
||||
{
|
||||
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
|
||||
}
|
||||
|
||||
static auto MakeArgument(std::vector<const void*>& p_As,
|
||||
std::vector<const void*>& p_Bs,
|
||||
std::vector<std::array<const void*, NumDTensor>>& p_Ds,
|
||||
std::vector<void*>& p_Es,
|
||||
std::vector<GemmDesc> gemm_descs,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation c_element_op)
|
||||
{
|
||||
return Argument{
|
||||
p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op};
|
||||
}
|
||||
|
||||
static auto MakeInvoker() { return Invoker{}; }
|
||||
|
||||
// polymorphic
|
||||
std::unique_ptr<BaseArgument>
|
||||
MakeArgumentPointer(std::vector<const void*>& p_As,
|
||||
std::vector<const void*>& p_Bs,
|
||||
std::vector<std::array<const void*, NumDTensor>>& p_Ds,
|
||||
std::vector<void*>& p_Es,
|
||||
std::vector<GemmDesc>& gemm_descs,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation c_element_op) override
|
||||
{
|
||||
return std::make_unique<Argument>(
|
||||
p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op);
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
|
||||
{
|
||||
return std::make_unique<Invoker>(Invoker{});
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
std::string GetTypeString() const override
|
||||
{
|
||||
auto str = std::stringstream();
|
||||
|
||||
// clang-format off
|
||||
str << "DeviceGroupedGemm_Xdl_Fixed_NK"
|
||||
<< "<"
|
||||
<< BlockSize << ", "
|
||||
<< MPerBlock << ", "
|
||||
<< NPerBlock << ", "
|
||||
<< KPerBlock << ", "
|
||||
<< AK1 << ", "
|
||||
<< BK1 << ", "
|
||||
<< MPerXDL << ", "
|
||||
<< NPerXDL << ", "
|
||||
<< MXdlPerWave << ", "
|
||||
<< NXdlPerWave << ", "
|
||||
<< ABlockTransferSrcScalarPerVector << ", "
|
||||
<< BBlockTransferSrcScalarPerVector << ", "
|
||||
<< CShuffleMXdlPerWavePerShuffle << ", "
|
||||
<< CShuffleNXdlPerWavePerShuffle << ", "
|
||||
<< getGemmSpecializationString(GemmSpec)
|
||||
<< ">";
|
||||
// clang-format on
|
||||
|
||||
return str.str();
|
||||
}
|
||||
|
||||
static void SetDeviceKernelArgs(Argument& arg, const void* kernel_args)
|
||||
{
|
||||
arg.grouped_gemm_kernel_args_dev = kernel_args;
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
void SetDeviceKernelArgs(BaseArgument* p_arg, const void* kernel_args) const override
|
||||
{
|
||||
return SetDeviceKernelArgs(*dynamic_cast<Argument*>(p_arg), kernel_args);
|
||||
}
|
||||
|
||||
size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
|
||||
{
|
||||
auto arg = *dynamic_cast<const Argument*>(p_arg);
|
||||
|
||||
return arg.group_count_ * arg.barrier_size_grp_ * sizeof(uint32_t);
|
||||
}
|
||||
|
||||
size_t GetDeviceKernelArgSize(const BaseArgument* p_arg) const override
|
||||
{
|
||||
auto arg = *dynamic_cast<const Argument*>(p_arg);
|
||||
|
||||
return arg.group_count_ * sizeof(GroupedGemmKernelArgument<NumDTensor>);
|
||||
}
|
||||
|
||||
void SetWorkSpacePointer(BaseArgument* p_arg, void* p_workspace) const override
|
||||
{
|
||||
auto p_arg_ = dynamic_cast<Argument*>(p_arg);
|
||||
p_arg_->p_workspace_ = p_workspace;
|
||||
|
||||
hip_check_error(hipMemset(p_workspace, 0, GetWorkSpaceSize(p_arg)));
|
||||
}
|
||||
|
||||
static void SetKBatch(Argument& arg, index_t k_batch) { arg.UpdateKBatch(k_batch); }
|
||||
|
||||
// polymorphic
|
||||
void SetKBatch(BaseArgument* p_arg, index_t k_batch) const override
|
||||
{
|
||||
return SetKBatch(*dynamic_cast<Argument*>(p_arg), k_batch);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user