mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-20 21:09:08 +00:00
Merge commit '3e777217551c82a47eb9540791fb5542f2704e63' into develop
This commit is contained in:
assistant-librarian[bot]
@@ -0,0 +1,899 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <iostream>
|
||||
#include <sstream>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
|
||||
#include "ck/utility/env.hpp"
|
||||
#include "ck/host_utility/hip_check_error.hpp"
|
||||
#include "ck/utility/common_header.hpp"
|
||||
#include "ck/utility/tuple.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#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_multi_abd_fixed_nk.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_wmma_cshuffle_v3.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,
|
||||
bool HasMainKBlockLoop,
|
||||
InMemoryDataOperationEnum EGlobalMemoryDataOperation,
|
||||
typename AsLayout,
|
||||
typename BsLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename Block2ETileMap,
|
||||
typename GroupedGemmBlock2ETileMap,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CDEElementwiseOperation,
|
||||
GemmSpecialization GemmSpec,
|
||||
index_t MinimumOccupancy = 1,
|
||||
TailNumber TailNum = TailNumber::Full>
|
||||
__global__ void
|
||||
#if CK_USE_LAUNCH_BOUNDS
|
||||
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
|
||||
#endif
|
||||
kernel_grouped_gemm_wmma_fixed_nk(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
|
||||
const index_t group_count,
|
||||
const index_t grid_size_grp,
|
||||
const AElementwiseOperation a_element_op,
|
||||
const BElementwiseOperation b_element_op,
|
||||
const CDEElementwiseOperation cde_element_op)
|
||||
{
|
||||
#if defined(__gfx11__) || defined(__gfx12__)
|
||||
__shared__ char p_shared[GridwiseGemm::template GetSharedMemoryNumberOfByte<
|
||||
typename GridwiseGemm::EpilogueCShuffle>()];
|
||||
|
||||
const index_t KBatch = 1;
|
||||
|
||||
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;
|
||||
|
||||
auto karg = gemm_desc_ptr[group_id];
|
||||
|
||||
if(karg.M == 0 || karg.N == 0 || karg.K == 0)
|
||||
return;
|
||||
|
||||
#if defined(__gfx11__)
|
||||
// gfx11 does not support *_atomic_pk_add_f16/bf16 instructions
|
||||
if constexpr(!(EGlobalMemoryDataOperation == InMemoryDataOperationEnum::AtomicAdd &&
|
||||
(std::is_same_v<typename GridwiseGemm::EDataType_, ck::half_t> ||
|
||||
std::is_same_v<typename GridwiseGemm::EDataType_, ck::bhalf_t>)))
|
||||
#endif
|
||||
{
|
||||
|
||||
typename GridwiseGemm::Problem problem(karg.M,
|
||||
karg.N,
|
||||
karg.K,
|
||||
karg.StrideAs,
|
||||
karg.StrideBs,
|
||||
karg.StrideDs,
|
||||
karg.StrideE,
|
||||
KBatch);
|
||||
|
||||
const auto e_grid_desc_m_n = GridwiseGemm::template MakeDEGridDescriptor_M_N<ELayout>(
|
||||
problem.M, problem.MPadded, problem.N, problem.NPadded, problem.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 NumATensor = GridwiseGemm::AsGridPointer::Size();
|
||||
constexpr auto NumBTensor = GridwiseGemm::BsGridPointer::Size();
|
||||
constexpr auto NumDTensor = GridwiseGemm::DsGridPointer::Size();
|
||||
|
||||
typename GridwiseGemm::AsGridPointer p_as_grid_;
|
||||
typename GridwiseGemm::BsGridPointer p_bs_grid_;
|
||||
typename GridwiseGemm::DsGridPointer p_ds_grid_;
|
||||
|
||||
static_for<0, NumATensor, 1>{}([&](auto i) {
|
||||
using ADataType = remove_cvref_t<decltype(p_as_grid_(i))>;
|
||||
p_as_grid_(i) = static_cast<ADataType>(karg.p_as_grid[i]);
|
||||
});
|
||||
|
||||
static_for<0, NumBTensor, 1>{}([&](auto i) {
|
||||
using BDataType = remove_cvref_t<decltype(p_bs_grid_(i))>;
|
||||
p_bs_grid_(i) = static_cast<BDataType>(karg.p_bs_grid[i]);
|
||||
});
|
||||
|
||||
static_for<0, NumDTensor, 1>{}([&](auto i) {
|
||||
using DDataType = remove_cvref_t<decltype(p_ds_grid_(i))>;
|
||||
p_ds_grid_(i) = static_cast<DDataType>(karg.p_ds_grid[i]);
|
||||
});
|
||||
|
||||
index_t id_off = 0;
|
||||
index_t id_local = get_block_1d_id() - BlockStart;
|
||||
|
||||
while(id_local < local_grid_size)
|
||||
{
|
||||
const auto block_2_etile_map =
|
||||
GroupedGemmBlock2ETileMap(local_b2e_tile_map, BlockStart, id_off);
|
||||
|
||||
auto epilogue_args = typename GridwiseGemm::EpilogueCShuffle{};
|
||||
|
||||
GridwiseGemm::template Run<HasMainKBlockLoop,
|
||||
EGlobalMemoryDataOperation,
|
||||
TailNum,
|
||||
decltype(block_2_etile_map),
|
||||
decltype(epilogue_args),
|
||||
1,
|
||||
2>(
|
||||
p_as_grid_,
|
||||
p_bs_grid_,
|
||||
p_ds_grid_,
|
||||
static_cast<typename GridwiseGemm::EDataType_*>(karg.p_e_grid),
|
||||
p_shared,
|
||||
problem,
|
||||
block_2_etile_map,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op,
|
||||
epilogue_args);
|
||||
|
||||
id_off += grid_size_grp;
|
||||
id_local += grid_size_grp;
|
||||
}
|
||||
}
|
||||
#else
|
||||
ignore = gemm_descs_const;
|
||||
ignore = group_count;
|
||||
ignore = grid_size_grp;
|
||||
ignore = a_element_op;
|
||||
ignore = b_element_op;
|
||||
ignore = cde_element_op;
|
||||
#endif
|
||||
}
|
||||
|
||||
template <typename AsLayout,
|
||||
typename BsLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename AsDataType,
|
||||
typename BsDataType,
|
||||
typename AccDataType,
|
||||
typename CShuffleDataType,
|
||||
typename DsDataType,
|
||||
typename EDataType,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CDEElementwiseOperation,
|
||||
GemmSpecialization GemmSpec,
|
||||
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 MPerWmma,
|
||||
ck::index_t NPerWmma,
|
||||
ck::index_t MRepeat,
|
||||
ck::index_t NRepeat,
|
||||
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 CShuffleMRepeatPerShuffle,
|
||||
index_t CShuffleNRepeatPerShuffle,
|
||||
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
index_t CDEBlockTransferScalarPerVector_NPerBlock,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
|
||||
typename ComputeTypeA = EDataType,
|
||||
typename ComputeTypeB = ComputeTypeA,
|
||||
bool PermuteA = false,
|
||||
bool PermuteB = false>
|
||||
struct DeviceGroupedGemm_Wmma_Multi_ABD_Fixed_NK
|
||||
: public DeviceGroupedGemmMultiABDFixedNK<AsLayout,
|
||||
BsLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
AsDataType,
|
||||
BsDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CDEElementwiseOperation>
|
||||
{
|
||||
using DeviceOp = DeviceGroupedGemm_Wmma_Multi_ABD_Fixed_NK;
|
||||
|
||||
static constexpr index_t NumATensor = AsDataType::Size();
|
||||
static constexpr index_t NumBTensor = BsDataType::Size();
|
||||
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>{};
|
||||
|
||||
// Note: Pass multiple layout but then using only the first one
|
||||
// This is to replicate xdl functionality but it should be extended
|
||||
using ALayout = remove_cvref_t<tuple_element_t<0, AsLayout>>;
|
||||
using BLayout = remove_cvref_t<tuple_element_t<0, BsLayout>>;
|
||||
|
||||
using GridwiseGemm = GridwiseGemm_wmma_cshuffle_v3<
|
||||
ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
AsDataType,
|
||||
BsDataType,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CDEElementwiseOperation,
|
||||
GemmSpec,
|
||||
BlockSize,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
AK1,
|
||||
BK1,
|
||||
MPerWmma,
|
||||
NPerWmma,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
ABlockTransferThreadClusterLengths_K0_M_K1,
|
||||
ABlockTransferThreadClusterArrangeOrder,
|
||||
ABlockTransferSrcAccessOrder,
|
||||
ABlockTransferSrcVectorDim,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
ABlockTransferDstScalarPerVector_K1,
|
||||
false,
|
||||
ABlockLdsExtraM,
|
||||
BBlockTransferThreadClusterLengths_K0_N_K1,
|
||||
BBlockTransferThreadClusterArrangeOrder,
|
||||
BBlockTransferSrcAccessOrder,
|
||||
BBlockTransferSrcVectorDim,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
BBlockTransferDstScalarPerVector_K1,
|
||||
false,
|
||||
BBlockLdsExtraN,
|
||||
CShuffleMRepeatPerShuffle,
|
||||
CShuffleNRepeatPerShuffle,
|
||||
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
typename uniform_sequence_gen<NumDTensor + 1,
|
||||
CDEBlockTransferScalarPerVector_NPerBlock>::type,
|
||||
BlkGemmPipeSched,
|
||||
BlkGemmPipelineVer,
|
||||
ComputeTypeA,
|
||||
ComputeTypeB,
|
||||
false,
|
||||
false>;
|
||||
|
||||
// TODO: Block to tile mappings could potentially moved out to avoid code duplications between
|
||||
// different device implementations.
|
||||
|
||||
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
|
||||
{
|
||||
__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>;
|
||||
|
||||
static constexpr index_t DefaultKBatch = 1; // implementation only supports KBatch == 1
|
||||
using KernelArgument = typename GridwiseGemm::Argument;
|
||||
|
||||
using GemmTransKernelArg =
|
||||
GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
|
||||
|
||||
static constexpr bool CalculateHasMainKBlockLoop(const GemmTransKernelArg& karg,
|
||||
index_t k_batch)
|
||||
{
|
||||
index_t k_grain = k_batch * KPerBlock;
|
||||
index_t K_split = (karg.K + k_grain - 1) / k_batch;
|
||||
return GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
|
||||
}
|
||||
|
||||
// Argument
|
||||
struct Argument : public BaseArgument
|
||||
{
|
||||
|
||||
Argument(std::vector<std::array<const void*, NumATensor>>& p_As,
|
||||
std::vector<std::array<const void*, NumBTensor>>& p_Bs,
|
||||
std::vector<std::array<const void*, NumDTensor>>& p_Ds,
|
||||
std::vector<void*>& p_Es,
|
||||
std::vector<GemmMultiABDDesc>& gemm_descs,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation c_element_op)
|
||||
: Argument(p_As,
|
||||
p_Bs,
|
||||
p_Ds,
|
||||
p_Es,
|
||||
gemm_descs,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op,
|
||||
DefaultKBatch)
|
||||
{
|
||||
// TODO: use occupancy api to calculate appropriate batch size.
|
||||
}
|
||||
|
||||
// Client is expected to manually copy the kernel arguments to the device therefore there is
|
||||
// no point in setting tensor device pointers for the argument structure.
|
||||
Argument(std::vector<std::array<const void*, NumATensor>>&,
|
||||
std::vector<std::array<const void*, NumBTensor>>&,
|
||||
std::vector<std::array<const void*, NumDTensor>>&,
|
||||
std::vector<void*>&,
|
||||
std::vector<GemmMultiABDDesc>& gemm_descs,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation c_element_op,
|
||||
index_t kbatch)
|
||||
: group_count_{ck::type_convert<ck::index_t>(gemm_descs.size())},
|
||||
a_element_op_{a_element_op},
|
||||
b_element_op_{b_element_op},
|
||||
c_element_op_{c_element_op},
|
||||
grouped_gemm_kernel_args_dev{nullptr},
|
||||
gemm_kernel_host_args_{nullptr},
|
||||
grid_size_{0},
|
||||
k_batch_{kbatch}
|
||||
{
|
||||
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 fixed_N = gemm_descs[0].N_;
|
||||
const index_t fixed_K = gemm_descs[0].K_;
|
||||
|
||||
for(std::size_t g = 0; g < gemm_descs.size(); g++)
|
||||
{
|
||||
const index_t M = gemm_descs[g].M_;
|
||||
const index_t N = gemm_descs[g].N_;
|
||||
const index_t K = gemm_descs[g].K_;
|
||||
|
||||
if(M != sum_of_m || N != fixed_N || K != fixed_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);
|
||||
|
||||
// pointer
|
||||
std::array<const void*, NumATensor> p_as_grid;
|
||||
std::array<const void*, NumBTensor> p_bs_grid;
|
||||
std::array<const void*, NumDTensor> p_ds_grid;
|
||||
|
||||
static_for<0, NumATensor, 1>{}([&](auto i) { p_as_grid[i] = nullptr; });
|
||||
static_for<0, NumBTensor, 1>{}([&](auto i) { p_bs_grid[i] = nullptr; });
|
||||
static_for<0, NumDTensor, 1>{}([&](auto i) { p_ds_grid[i] = nullptr; });
|
||||
|
||||
std::array<index_t, NumATensor> StrideAs;
|
||||
std::array<index_t, NumBTensor> StrideBs;
|
||||
std::array<index_t, NumDTensor> StrideDs;
|
||||
|
||||
const index_t StrideE = gemm_descs[g].stride_C_;
|
||||
|
||||
if(gemm_descs[g].stride_As_.size() != NumATensor)
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! gemm_descs[i].stride_As_.size() does not match NumATensor");
|
||||
}
|
||||
|
||||
static_for<0, NumATensor, 1>{}(
|
||||
[&](auto j) { StrideAs[j] = gemm_descs[g].stride_As_[j]; });
|
||||
|
||||
if(gemm_descs[g].stride_Bs_.size() != NumBTensor)
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! gemm_descs[i].stride_Bs_.size() does not match NumBTensor");
|
||||
}
|
||||
|
||||
static_for<0, NumBTensor, 1>{}(
|
||||
[&](auto j) { StrideBs[j] = gemm_descs[g].stride_Bs_[j]; });
|
||||
|
||||
if(gemm_descs[g].stride_Ds_.size() != NumDTensor)
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! gemm_descs[i].stride_Ds_.size() does not match NumDTensor");
|
||||
}
|
||||
|
||||
static_for<0, NumDTensor, 1>{}(
|
||||
[&](auto j) { StrideDs[j] = gemm_descs[g].stride_Ds_[j]; });
|
||||
|
||||
const auto e_grid_desc_m_n =
|
||||
GridwiseGemm::template MakeDEGridDescriptor_M_N<ELayout>(
|
||||
AverM, AverM, N, 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!");
|
||||
}
|
||||
|
||||
const index_t block_start = grid_size_;
|
||||
|
||||
grid_size_ += grid_size_grp_;
|
||||
|
||||
if(!local_b2c_tile_map.CheckValidity(e_grid_desc_m_n))
|
||||
{
|
||||
throw std::runtime_error("wrong! block_2_etile_map validation failed");
|
||||
}
|
||||
|
||||
auto grouped_block_2_ctile_map =
|
||||
GroupedGemmBlock2ETileMap(local_b2c_tile_map, block_start);
|
||||
|
||||
auto karg = GemmTransKernelArg({p_as_grid,
|
||||
p_bs_grid,
|
||||
p_ds_grid,
|
||||
nullptr,
|
||||
AverM,
|
||||
N,
|
||||
K,
|
||||
StrideAs,
|
||||
StrideBs,
|
||||
StrideDs,
|
||||
StrideE});
|
||||
|
||||
gemm_desc_kernel_arg_.emplace_back(std::move(karg));
|
||||
|
||||
group_id++;
|
||||
}
|
||||
}
|
||||
|
||||
void UpdateKBatch(index_t) {}
|
||||
|
||||
index_t group_count_;
|
||||
|
||||
AElementwiseOperation a_element_op_;
|
||||
BElementwiseOperation b_element_op_;
|
||||
CDEElementwiseOperation c_element_op_;
|
||||
|
||||
std::vector<GemmTransKernelArg> 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;
|
||||
void* gemm_kernel_host_args_;
|
||||
index_t grid_size_;
|
||||
index_t grid_size_grp_;
|
||||
index_t sum_of_m;
|
||||
|
||||
index_t k_batch_;
|
||||
};
|
||||
|
||||
// Invoker
|
||||
struct Invoker : public BaseInvoker
|
||||
{
|
||||
using Argument = DeviceOp::Argument;
|
||||
|
||||
float RunImp(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
|
||||
{
|
||||
if(arg.grouped_gemm_kernel_args_dev == nullptr)
|
||||
{
|
||||
throw std::runtime_error("wrong! grouped_gemm_kernel_args_dev is nullptr");
|
||||
}
|
||||
|
||||
if(arg.k_batch_ != 1)
|
||||
{
|
||||
throw std::runtime_error("Split K functionality is not supported for wmma multi "
|
||||
"abd fixed nk implementation.");
|
||||
}
|
||||
|
||||
float ave_time = 0;
|
||||
|
||||
auto launch_kernel = [&](auto e_global_memory_operation_) {
|
||||
const auto kernel = kernel_grouped_gemm_wmma_fixed_nk<GridwiseGemm,
|
||||
GemmTransKernelArg,
|
||||
true, // has_main_k_block_loop
|
||||
e_global_memory_operation_,
|
||||
AsLayout,
|
||||
BsLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
Block2ETileMap,
|
||||
GroupedGemmBlock2ETileMap,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CDEElementwiseOperation,
|
||||
GemmSpec>;
|
||||
|
||||
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),
|
||||
arg.gemm_desc_kernel_arg_.size(),
|
||||
arg.grid_size_grp_,
|
||||
arg.a_element_op_,
|
||||
arg.b_element_op_,
|
||||
arg.c_element_op_);
|
||||
};
|
||||
|
||||
constexpr auto Set = InMemoryDataOperationEnum::Set;
|
||||
ave_time = launch_kernel(integral_constant<InMemoryDataOperationEnum, Set>{});
|
||||
|
||||
return ave_time;
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
float Run(const BaseArgument* p_arg,
|
||||
const StreamConfig& stream_config = StreamConfig{}) override
|
||||
{
|
||||
return RunImp(*dynamic_cast<const Argument*>(p_arg), stream_config);
|
||||
}
|
||||
};
|
||||
|
||||
static bool IsSupportedArgument(const Argument& arg)
|
||||
{
|
||||
if(!ck::is_gfx11_supported() && !ck::is_gfx12_supported())
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
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);
|
||||
}
|
||||
}
|
||||
|
||||
for(index_t i = 0; i < arg.group_count_; i++)
|
||||
{
|
||||
if(CalculateHasMainKBlockLoop(arg.gemm_desc_kernel_arg_[i], arg.k_batch_) != true)
|
||||
{
|
||||
supported = false;
|
||||
}
|
||||
}
|
||||
|
||||
return supported;
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
bool IsSupportedArgument(const BaseArgument* p_arg) override
|
||||
{
|
||||
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
|
||||
}
|
||||
|
||||
static auto MakeArgument(std::vector<std::array<const void*, NumATensor>>& p_As,
|
||||
std::vector<std::array<const void*, NumBTensor>>& p_Bs,
|
||||
std::vector<std::array<const void*, NumDTensor>>& p_Ds,
|
||||
std::vector<void*>& p_Es,
|
||||
std::vector<GemmMultiABDDesc> gemm_descs,
|
||||
AElementwiseOperation a_element_op = AElementwiseOperation{},
|
||||
BElementwiseOperation b_element_op = BElementwiseOperation{},
|
||||
CDEElementwiseOperation c_element_op = CDEElementwiseOperation{})
|
||||
{
|
||||
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<std::array<const void*, NumATensor>>& p_As,
|
||||
std::vector<std::array<const void*, NumBTensor>>& p_Bs,
|
||||
std::vector<std::array<const void*, NumDTensor>>& p_Ds,
|
||||
std::vector<void*>& p_Es,
|
||||
std::vector<GemmMultiABDDesc>& gemm_descs,
|
||||
AElementwiseOperation a_element_op = AElementwiseOperation{},
|
||||
BElementwiseOperation b_element_op = BElementwiseOperation{},
|
||||
CDEElementwiseOperation c_element_op = CDEElementwiseOperation{}) 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_Wmma_Fixed_Nk"
|
||||
<< "<"
|
||||
<< BlockSize << ", "
|
||||
<< MPerBlock << ", "
|
||||
<< NPerBlock << ", "
|
||||
<< KPerBlock << ", "
|
||||
<< AK1 << ", "
|
||||
<< BK1 << ", "
|
||||
<< MPerWmma << ", "
|
||||
<< NPerWmma << ", "
|
||||
<< ABlockTransferSrcScalarPerVector << ", "
|
||||
<< BBlockTransferSrcScalarPerVector << ", "
|
||||
<< CShuffleMRepeatPerShuffle << ", "
|
||||
<< CShuffleNRepeatPerShuffle << ", "
|
||||
<< getGemmSpecializationString(GemmSpec)
|
||||
<< ">";
|
||||
// clang-format on
|
||||
|
||||
return str.str();
|
||||
}
|
||||
|
||||
static void SetElementwiseOps(Argument& arg,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation c_element_op)
|
||||
{
|
||||
arg.a_element_op_ = a_element_op;
|
||||
arg.b_element_op_ = b_element_op;
|
||||
arg.c_element_op_ = c_element_op;
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
void SetElementwiseOps(BaseArgument* p_arg,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation c_element_op) const override
|
||||
{
|
||||
|
||||
SetElementwiseOps(
|
||||
*dynamic_cast<Argument*>(p_arg), a_element_op, b_element_op, c_element_op);
|
||||
}
|
||||
|
||||
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 GetDeviceKernelArgSize(const BaseArgument* p_arg) const override
|
||||
{
|
||||
auto arg = *dynamic_cast<const Argument*>(p_arg);
|
||||
|
||||
return arg.group_count_ *
|
||||
sizeof(GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>);
|
||||
}
|
||||
|
||||
size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
|
||||
{
|
||||
auto p_arg_ = dynamic_cast<const Argument*>(p_arg);
|
||||
if(p_arg_)
|
||||
{
|
||||
return p_arg_->gemm_desc_kernel_arg_.size() * sizeof(GemmTransKernelArg);
|
||||
}
|
||||
else
|
||||
throw std::runtime_error(
|
||||
"The argument pointer is not an object of "
|
||||
"DeviceGroupedGemm_Wmma_Multi_ABD_Fixed_NK::Argument structure!");
|
||||
}
|
||||
|
||||
void SetWorkSpacePointer(BaseArgument* p_arg,
|
||||
void* p_workspace,
|
||||
const StreamConfig& stream_config = StreamConfig{}) const override
|
||||
{
|
||||
auto p_arg_ = dynamic_cast<Argument*>(p_arg);
|
||||
p_arg_->p_workspace_ = p_workspace;
|
||||
|
||||
hip_check_error(
|
||||
hipMemsetAsync(p_workspace, 0, GetWorkSpaceSize(p_arg), stream_config.stream_id_));
|
||||
}
|
||||
|
||||
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);
|
||||
}
|
||||
|
||||
void SetHostKernelArgsPointer(BaseArgument* p_arg, void* p_host_kernel_args) const
|
||||
{
|
||||
Argument* pArg_ = dynamic_cast<Argument*>(p_arg);
|
||||
if(!pArg_)
|
||||
{
|
||||
throw std::runtime_error("Failed to cast argument pointer!");
|
||||
}
|
||||
|
||||
pArg_->gemm_kernel_host_args_ = p_host_kernel_args;
|
||||
std::copy(pArg_->gemm_desc_kernel_arg_.begin(),
|
||||
pArg_->gemm_desc_kernel_arg_.end(),
|
||||
static_cast<GemmTransKernelArg*>(pArg_->gemm_kernel_host_args_));
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -605,7 +605,7 @@ struct DeviceGroupedGemm_Xdl_Multi_ABD_Fixed_NK
|
||||
|
||||
if(arg.grouped_gemm_kernel_args_dev == nullptr)
|
||||
{
|
||||
throw std::runtime_error("wrong! grouped_gemm_kernel_args_dev is nullpr");
|
||||
throw std::runtime_error("wrong! grouped_gemm_kernel_args_dev is nullptr");
|
||||
}
|
||||
|
||||
float ave_time = 0;
|
||||
@@ -688,6 +688,11 @@ struct DeviceGroupedGemm_Xdl_Multi_ABD_Fixed_NK
|
||||
|
||||
static bool IsSupportedArgument(const Argument& arg)
|
||||
{
|
||||
if(!ck::is_xdl_wmma_supported<ComputeType, ComputeType, MPerXDL, NPerXDL>())
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
// Split-K autodeduction is not supported
|
||||
if(arg.k_batch_ < 1)
|
||||
{
|
||||
@@ -720,6 +725,26 @@ struct DeviceGroupedGemm_Xdl_Multi_ABD_Fixed_NK
|
||||
}
|
||||
}
|
||||
|
||||
for(index_t i = 0; i < arg.group_count_; i++)
|
||||
{
|
||||
if(get_warp_size() == 64)
|
||||
{
|
||||
if(GridwiseGemm64::CalculateHasMainKBlockLoop(arg.gemm_desc_kernel_arg_[i].K_) !=
|
||||
true)
|
||||
{
|
||||
supported = false;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
if(GridwiseGemm32::CalculateHasMainKBlockLoop(arg.gemm_desc_kernel_arg_[i].K_) !=
|
||||
true)
|
||||
{
|
||||
supported = false;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return supported;
|
||||
}
|
||||
|
||||
|
||||
@@ -696,7 +696,7 @@ struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
|
||||
|
||||
if(arg.grouped_gemm_kernel_args_dev == nullptr)
|
||||
{
|
||||
throw std::runtime_error("wrong! grouped_gemm_kernel_args_dev is nullpr");
|
||||
throw std::runtime_error("wrong! grouped_gemm_kernel_args_dev is nullptr");
|
||||
}
|
||||
|
||||
float ave_time = 0;
|
||||
|
||||
@@ -333,6 +333,7 @@ struct GridwiseGemm_wmma_cshuffle_v3
|
||||
using typename Base::DsGridPointer;
|
||||
using AsDataType_ = AsDataType;
|
||||
using BsDataType_ = BsDataType;
|
||||
using EDataType_ = EDataType;
|
||||
|
||||
struct Problem
|
||||
{
|
||||
|
||||
@@ -48,6 +48,15 @@ __host__ __device__ constexpr auto concat_tuple_of_reference(const Tuple<X&...>&
|
||||
ty);
|
||||
}
|
||||
|
||||
template <typename... X, typename... Y>
|
||||
auto concat_tuple_of_reference(ck::Tuple<X&...>& tx, ck::Tuple<Y&...>& ty)
|
||||
{
|
||||
return ck::unpack2(
|
||||
[&](auto&&... zs) { return ck::Tuple<decltype(zs)...>{ck::forward<decltype(zs)>(zs)...}; },
|
||||
tx,
|
||||
ty);
|
||||
}
|
||||
|
||||
template <typename... X, typename... Y>
|
||||
__host__ __device__ constexpr auto concat_tuple(const Tuple<X...>& tx, const Tuple<Y...>& ty)
|
||||
{
|
||||
|
||||
@@ -380,9 +380,18 @@ struct QuantGemmKernel
|
||||
__device__ SplitKBatchOffset(const QuantGemmKernelArgs& kargs,
|
||||
const std::size_t k_id = blockIdx.z)
|
||||
{
|
||||
constexpr auto K1 = GemmPipeline::BlockGemmShape::WarpTile::at(I2);
|
||||
const index_t K_t = amd_wave_read_first_lane(kargs.k_batch * K1);
|
||||
const index_t KRead = amd_wave_read_first_lane((kargs.K + K_t - 1) / K_t * K1);
|
||||
constexpr auto K1 =
|
||||
GemmPipeline::BlockGemmShape::WarpTile::at(I2); // smallest unit of K work per block
|
||||
const index_t K_t = amd_wave_read_first_lane(
|
||||
kargs.k_batch * K1); // amount of K elements consumed if every split-K batch
|
||||
// performs exactly one "unit" (K1)
|
||||
const index_t KRead = amd_wave_read_first_lane(
|
||||
(kargs.K + K_t - 1) / K_t * K1); // total k elements to be read in this batch
|
||||
// offset not necessarily = KRead, because B can have packed elements (e.g. fp8i4)
|
||||
constexpr index_t BPackedSize =
|
||||
ck_tile::numeric_traits<remove_cvref_t<BDataType>>::PackedSize;
|
||||
const index_t b_k_offset_elements =
|
||||
amd_wave_read_first_lane(k_id * KRead / BPackedSize);
|
||||
|
||||
if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
|
||||
{
|
||||
@@ -395,11 +404,11 @@ struct QuantGemmKernel
|
||||
|
||||
if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
|
||||
{
|
||||
b_k_split_offset = amd_wave_read_first_lane(k_id * KRead * kargs.stride_B);
|
||||
b_k_split_offset = amd_wave_read_first_lane(b_k_offset_elements * kargs.stride_B);
|
||||
}
|
||||
else if constexpr(std::is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
|
||||
{
|
||||
b_k_split_offset = amd_wave_read_first_lane(k_id * KRead);
|
||||
b_k_split_offset = amd_wave_read_first_lane(b_k_offset_elements);
|
||||
}
|
||||
|
||||
if(k_id < static_cast<uint32_t>(kargs.k_batch - 1))
|
||||
@@ -410,10 +419,47 @@ struct QuantGemmKernel
|
||||
{
|
||||
splitted_k = amd_wave_read_first_lane(kargs.K - KRead * (kargs.k_batch - 1));
|
||||
}
|
||||
|
||||
// Compute BQ offset for BQuantGrouped mode (non-preshuffle only)
|
||||
// Note: With the alignment validation in IsSupportedArgument, KRead is always
|
||||
// a multiple of BQuantGroupSize::kK, so bq_k_split_offset will be correctly aligned.
|
||||
if constexpr(kQuantType == QuantType::BQuantGrouped && !BPreshuffleQuant)
|
||||
{
|
||||
using BQuantGroupSize = remove_cvref_t<typename GemmPipeline::BQuantGroupSize>;
|
||||
// Compute the K offset for this batch (in terms of K elements)
|
||||
const index_t k_offset = amd_wave_read_first_lane(k_id * KRead);
|
||||
// Convert K offset to BQ group offset (logical offset in K/kK dimension)
|
||||
bq_group_offset = amd_wave_read_first_lane(k_offset / BQuantGroupSize::kK);
|
||||
|
||||
// BQ tensor layout:
|
||||
// RowMajor: [K/kK, N/kN] with stride [N/kN, 1]
|
||||
// ColumnMajor: [N/kN, K/kK] with stride [K/kK, 1]
|
||||
if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, BQLayout>)
|
||||
{
|
||||
// For RowMajor BQ, K is the row dimension
|
||||
// offset = bq_group_offset * stride_BQ
|
||||
const index_t stride_bq =
|
||||
amd_wave_read_first_lane(integer_divide_ceil(kargs.N, BQuantGroupSize::kN));
|
||||
bq_k_split_offset = amd_wave_read_first_lane(bq_group_offset * stride_bq);
|
||||
}
|
||||
else if constexpr(std::is_same_v<tensor_layout::gemm::ColumnMajor, BQLayout>)
|
||||
{
|
||||
// For ColumnMajor BQ, K is the column dimension
|
||||
// offset = bq_group_offset
|
||||
bq_k_split_offset = amd_wave_read_first_lane(bq_group_offset);
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
bq_group_offset = 0;
|
||||
bq_k_split_offset = 0;
|
||||
}
|
||||
}
|
||||
|
||||
index_t a_k_split_offset;
|
||||
index_t b_k_split_offset;
|
||||
index_t bq_group_offset; // Logical offset in K-groups (K/kK dimension)
|
||||
index_t bq_k_split_offset; // Memory pointer offset (accounting for layout/stride)
|
||||
index_t splitted_k;
|
||||
};
|
||||
|
||||
@@ -805,10 +851,13 @@ struct QuantGemmKernel
|
||||
|
||||
CK_TILE_DEVICE static auto MakeBQBlockWindow(const BQDataType* bq_ptr,
|
||||
const QuantGemmKernelArgs& kargs,
|
||||
const index_t bq_group_offset,
|
||||
const index_t i_m,
|
||||
const index_t i_n)
|
||||
{
|
||||
// Step 1: Create tensor view for BQ
|
||||
// Note: For split-K, the bq_ptr is already offset by bq_k_split_offset (pointer offset).
|
||||
// The dimension should use the remaining K-groups from this offset position.
|
||||
const auto& bq_tensor_view = [&]() {
|
||||
if constexpr(kQuantType == QuantType::RowColQuant)
|
||||
{
|
||||
@@ -850,11 +899,12 @@ struct QuantGemmKernel
|
||||
"ABQuantGrouped requires ColumnMajor BQ layout");
|
||||
}
|
||||
|
||||
using BQuantGroupSize = remove_cvref_t<typename GemmPipeline::BQuantGroupSize>;
|
||||
if constexpr(std::is_same_v<BQLayout, tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
return make_naive_tensor_view<address_space_enum::global>(
|
||||
bq_ptr,
|
||||
make_tuple(integer_divide_ceil(kargs.K, BQuantGroupSize::kK),
|
||||
make_tuple(kargs.QK_B - bq_group_offset,
|
||||
integer_divide_ceil(kargs.N, BQuantGroupSize::kN)),
|
||||
make_tuple(integer_divide_ceil(kargs.N, BQuantGroupSize::kN), 1),
|
||||
number<GemmPipeline::GetVectorSizeBQ()>{},
|
||||
@@ -865,8 +915,8 @@ struct QuantGemmKernel
|
||||
return make_naive_tensor_view<address_space_enum::global>(
|
||||
bq_ptr,
|
||||
make_tuple(integer_divide_ceil(kargs.N, BQuantGroupSize::kN),
|
||||
integer_divide_ceil(kargs.K, BQuantGroupSize::kK)),
|
||||
make_tuple(integer_divide_ceil(kargs.K, BQuantGroupSize::kK), 1),
|
||||
kargs.QK_B - bq_group_offset),
|
||||
make_tuple(kargs.QK_B, 1),
|
||||
number<GemmPipeline::GetVectorSizeBQ()>{},
|
||||
number<1>{});
|
||||
}
|
||||
@@ -1047,13 +1097,61 @@ struct QuantGemmKernel
|
||||
|
||||
CK_TILE_HOST static bool IsSupportedArgument(const QuantGemmKernelArgs& kargs)
|
||||
{
|
||||
// Split-K is supported for BQuantGrouped mode without preshuffle
|
||||
if(kargs.k_batch != 1)
|
||||
{
|
||||
if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
|
||||
constexpr bool is_bquant_non_preshuffle =
|
||||
(kQuantType == QuantType::BQuantGrouped) && !BPreshuffleQuant;
|
||||
if constexpr(!is_bquant_non_preshuffle)
|
||||
{
|
||||
CK_TILE_ERROR("Conditions not met for Kbatch >1 !");
|
||||
if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
|
||||
{
|
||||
CK_TILE_ERROR("Conditions not met for Kbatch >1 ! "
|
||||
"Split-K only supported for BQuantGrouped without preshuffle.");
|
||||
}
|
||||
return false;
|
||||
}
|
||||
else
|
||||
{
|
||||
using BQuantGroupSize = remove_cvref_t<typename GemmPipeline::BQuantGroupSize>;
|
||||
constexpr auto K1 = GemmPipeline::BlockGemmShape::WarpTile::at(I2);
|
||||
const index_t K_t = kargs.k_batch * K1;
|
||||
const index_t KRead = (kargs.K + K_t - 1) / K_t * K1;
|
||||
constexpr index_t BPackedSize =
|
||||
ck_tile::numeric_traits<remove_cvref_t<BDataType>>::PackedSize;
|
||||
|
||||
// Constraint 1: KRead must align with B packing requirements.
|
||||
// For packed data types, multiple K elements are stored in each storage unit.
|
||||
// Split-K advances the B pointer by (KRead / BPackedSize) storage units per batch.
|
||||
// If KRead is not divisible by BPackedSize, this division produces a fractional
|
||||
// offset, making it impossible to start reading from a valid storage unit boundary.
|
||||
if(KRead % BPackedSize != 0)
|
||||
{
|
||||
if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
|
||||
{
|
||||
CK_TILE_ERROR("KRead must be a multiple of B packed size for split-K!");
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
// Constraint 2: KRead must align with quantization group boundaries.
|
||||
// Each split-K batch reads KRead consecutive K elements. If KRead is not
|
||||
// a multiple of BQuantGroupSize::kK, the batch will span partial quantization
|
||||
// groups, requiring split access to a quantization scale. This violates the
|
||||
// atomic processing requirement where each batch must work with complete groups.
|
||||
if(KRead % BQuantGroupSize::kK != 0)
|
||||
{
|
||||
if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
|
||||
{
|
||||
CK_TILE_ERROR("Split-K batch size must be aligned with quantization group "
|
||||
"size! KRead=" +
|
||||
std::to_string(KRead) +
|
||||
" is not divisible by BQuantGroupSize::kK=" +
|
||||
std::to_string(BQuantGroupSize::kK));
|
||||
}
|
||||
return false;
|
||||
}
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
|
||||
@@ -1215,7 +1313,10 @@ struct QuantGemmKernel
|
||||
const auto& b_block_window =
|
||||
MakeBBlockWindow(b_ptr, kargs, splitk_batch_offset.splitted_k, block_idx_n);
|
||||
const auto& aq_block_window = MakeAQBlockWindow(aq_ptr, kargs, block_idx_m, block_idx_n);
|
||||
const auto& bq_block_window = MakeBQBlockWindow(bq_ptr, kargs, block_idx_m, block_idx_n);
|
||||
// Note: Pass bq_group_offset so the tensor view dimension reflects
|
||||
// the remaining K-groups from the split-K offset position.
|
||||
const auto& bq_block_window = MakeBQBlockWindow(
|
||||
bq_ptr, kargs, splitk_batch_offset.bq_group_offset, block_idx_m, block_idx_n);
|
||||
|
||||
const index_t num_loop =
|
||||
amd_wave_read_first_lane(TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
|
||||
@@ -1343,8 +1444,9 @@ struct QuantGemmKernel
|
||||
const BDataType* b_ptr =
|
||||
static_cast<const BDataType*>(kargs.b_ptr) + splitk_batch_offset.b_k_split_offset;
|
||||
const AQDataType* aq_ptr = static_cast<const AQDataType*>(kargs.aq_ptr);
|
||||
const BQDataType* bq_ptr = static_cast<const BQDataType*>(kargs.bq_ptr);
|
||||
CDataType* c_ptr = static_cast<CDataType*>(kargs.c_ptr);
|
||||
const BQDataType* bq_ptr =
|
||||
static_cast<const BQDataType*>(kargs.bq_ptr) + splitk_batch_offset.bq_k_split_offset;
|
||||
CDataType* c_ptr = static_cast<CDataType*>(kargs.c_ptr);
|
||||
|
||||
// allocate LDS
|
||||
__shared__ char smem_ptr[GetSmemSize()];
|
||||
|
||||
@@ -387,8 +387,8 @@ struct QuantGroupedGemmKernel
|
||||
Base::MakeABlockWindow(a_ptr, kargs, splitk_batch_offset.splitted_k, block_idx_m);
|
||||
const auto& b_block_window =
|
||||
Base::MakeBBlockWindow(b_ptr, kargs, splitk_batch_offset.splitted_k, block_idx_n);
|
||||
const auto& bq_block_window =
|
||||
Base::MakeBQBlockWindow(bq_ptr, kargs, block_idx_m, block_idx_n);
|
||||
const auto& bq_block_window = Base::MakeBQBlockWindow(
|
||||
bq_ptr, kargs, splitk_batch_offset.bq_group_offset, block_idx_m, block_idx_n);
|
||||
|
||||
const index_t num_loop = __builtin_amdgcn_readfirstlane(
|
||||
TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
|
||||
@@ -453,8 +453,8 @@ struct QuantGroupedGemmKernel
|
||||
Base::MakeBBlockWindow(b_ptr, kargs, splitk_batch_offset.splitted_k, block_idx_n);
|
||||
const auto& aq_block_window =
|
||||
Base::MakeAQBlockWindow(aq_ptr, kargs, block_idx_m, block_idx_n);
|
||||
const auto& bq_block_window =
|
||||
Base::MakeBQBlockWindow(bq_ptr, kargs, block_idx_m, block_idx_n);
|
||||
const auto& bq_block_window = Base::MakeBQBlockWindow(
|
||||
bq_ptr, kargs, splitk_batch_offset.bq_group_offset, block_idx_m, block_idx_n);
|
||||
|
||||
// Get hot-loop and tail configuration
|
||||
const index_t num_loop = __builtin_amdgcn_readfirstlane(
|
||||
|
||||
Reference in New Issue
Block a user