mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-20 06:49:15 +00:00
WMMA support for batched_gemm_reduce (#3332)
Summary: - added new device impl of Batched GEMM Reduce for WMMA - added instance library - added WMMA impl to the Batched GEMM Reduce tests
This commit is contained in:
Wojciech Laskowski
committed by
GitHub
GitHub
parent
0727e85e52
commit
b09121f860
@@ -0,0 +1,799 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#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_gemm_reduce.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_wmma_cshuffle_v3.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
#include "ck/host_utility/device_prop.hpp"
|
||||
#include "ck/host_utility/kernel_launch.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
template <typename GridwiseGemm,
|
||||
typename ReduceTrait,
|
||||
typename ComputePtrOffsetOfStridedBatch,
|
||||
bool HasMainKBlockLoop,
|
||||
InMemoryDataOperationEnum EGlobalMemoryDataOperation,
|
||||
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_batched_gemm_reduce_wmma_cshuffle_v3(
|
||||
typename GridwiseGemm::Argument karg,
|
||||
typename ReduceTrait::ReducePtrsGlobal_ p_reduces_grid,
|
||||
const typename ReduceTrait::ReduceInElementwiseOperations_ reduce_in_element_ops,
|
||||
const typename ReduceTrait::ReduceAccElementwiseOperations_ reduce_out_element_ops,
|
||||
const ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch)
|
||||
{
|
||||
#if(defined(__gfx11__) || defined(__gfx12__))
|
||||
#if defined(__gfx11__)
|
||||
// gfx11 does not support *_atomic_pk_add_f16/bf16 instructions
|
||||
using e_data_type = remove_cvref_t<remove_pointer_t<decltype(karg.p_e_grid)>>;
|
||||
if constexpr(!(EGlobalMemoryDataOperation == InMemoryDataOperationEnum::AtomicAdd &&
|
||||
(std::is_same_v<e_data_type, ck::half_t> ||
|
||||
std::is_same_v<e_data_type, ck::bhalf_t>)))
|
||||
{
|
||||
#endif
|
||||
using EpilogueType = typename GridwiseGemm::template EpilogueReduceCShuffle<ReduceTrait>;
|
||||
constexpr index_t LDS_size =
|
||||
GridwiseGemm::template GetSharedMemoryNumberOfByte<EpilogueType>();
|
||||
__shared__ char p_shared[LDS_size];
|
||||
|
||||
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
|
||||
|
||||
const index_t g_idx = amd_wave_read_first_lane(blockIdx.y);
|
||||
|
||||
const long_index_t a_batch_offset =
|
||||
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
|
||||
const long_index_t b_batch_offset =
|
||||
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
|
||||
const long_index_t c_batch_offset =
|
||||
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx));
|
||||
|
||||
auto reduces_batch = p_reduces_grid;
|
||||
compute_ptr_offset_of_batch.OffsetReducePtrs(g_idx, reduces_batch);
|
||||
|
||||
typename GridwiseGemm::AsGridPointer p_as_grid_shift;
|
||||
static_for<0, GridwiseGemm::NumATensor, 1>{}([&](auto i) {
|
||||
using ADataType_ =
|
||||
remove_cvref_t<tuple_element_t<i.value, typename GridwiseGemm::AsDataType_>>;
|
||||
p_as_grid_shift(i) = static_cast<const ADataType_*>(karg.p_as_grid[i]) +
|
||||
splitk_batch_offset.a_k_split_offset[i] + a_batch_offset;
|
||||
});
|
||||
|
||||
typename GridwiseGemm::BsGridPointer p_bs_grid_shift;
|
||||
static_for<0, GridwiseGemm::NumBTensor, 1>{}([&](auto i) {
|
||||
using BDataType_ =
|
||||
remove_cvref_t<tuple_element_t<i.value, typename GridwiseGemm::BsDataType_>>;
|
||||
p_bs_grid_shift(i) = static_cast<const BDataType_*>(karg.p_bs_grid[i]) +
|
||||
splitk_batch_offset.b_k_split_offset[i] + b_batch_offset;
|
||||
});
|
||||
|
||||
auto epilogue_args = EpilogueType(reduces_batch,
|
||||
reduce_in_element_ops,
|
||||
reduce_out_element_ops,
|
||||
karg.M,
|
||||
tensor_operation::element_wise::PassThrough{});
|
||||
|
||||
GridwiseGemm::template Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, TailNum>(
|
||||
p_as_grid_shift,
|
||||
p_bs_grid_shift,
|
||||
karg.p_ds_grid,
|
||||
karg.p_e_grid + splitk_batch_offset.c_reduce_offset + c_batch_offset,
|
||||
p_shared,
|
||||
karg,
|
||||
karg.a_element_op,
|
||||
karg.b_element_op,
|
||||
karg.cde_element_op,
|
||||
epilogue_args);
|
||||
#if defined(__gfx11__)
|
||||
}
|
||||
#endif
|
||||
#else
|
||||
ignore = karg;
|
||||
ignore = p_reduces_grid;
|
||||
ignore = reduce_in_element_ops;
|
||||
ignore = reduce_out_element_ops;
|
||||
ignore = compute_ptr_offset_of_batch;
|
||||
#endif
|
||||
}
|
||||
|
||||
} // namespace ck
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename ELayout,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename EDataType,
|
||||
typename AccDataType,
|
||||
typename CShuffleDataType,
|
||||
typename ReduceAccDataType, // Reduce
|
||||
typename ReducePtrsGlobal, // Reduce
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CElementwiseOperation,
|
||||
typename ReduceOperations, // Reduce
|
||||
typename ReduceInElementwiseOperations, // Reduce
|
||||
typename ReduceAccElementwiseOperations, // Reduce
|
||||
typename ReduceGlobalMemoryDataOperation, // Reduce
|
||||
GemmSpecialization GemmSpec,
|
||||
index_t BlockSize,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t AK1,
|
||||
index_t BK1,
|
||||
index_t MPerWmma,
|
||||
index_t NPerWmma,
|
||||
index_t MRepeat,
|
||||
index_t NRepeat,
|
||||
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
typename ABlockTransferThreadClusterArrangeOrder,
|
||||
typename ABlockTransferSrcAccessOrder,
|
||||
index_t ABlockTransferSrcVectorDim,
|
||||
index_t ABlockTransferSrcScalarPerVector,
|
||||
index_t ABlockTransferDstScalarPerVector_AK1,
|
||||
index_t ABlockLdsExtraM,
|
||||
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
|
||||
typename BBlockTransferThreadClusterArrangeOrder,
|
||||
typename BBlockTransferSrcAccessOrder,
|
||||
index_t BBlockTransferSrcVectorDim,
|
||||
index_t BBlockTransferSrcScalarPerVector,
|
||||
index_t BBlockTransferDstScalarPerVector_BK1,
|
||||
index_t BBlockLdsExtraN,
|
||||
index_t CShuffleMRepeatPerShuffle,
|
||||
index_t CShuffleNRepeatPerShuffle,
|
||||
typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
index_t CDEShuffleBlockTransferScalarPerVector,
|
||||
typename CReduceThreadClusterLengths_MPerBlock_NPerBlock, // Reduce
|
||||
index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock, // Reduce
|
||||
index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock, // Reduce
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
|
||||
typename ComputeTypeA = EDataType,
|
||||
typename ComputeTypeB = ComputeTypeA,
|
||||
bool PermuteA = false,
|
||||
bool PermuteB = false>
|
||||
struct DeviceBatchedGemmReduce_Wmma_CShuffleV3
|
||||
: public DeviceGemmReduce<0, ReduceOperations::Size()>
|
||||
{
|
||||
using DeviceOp = DeviceBatchedGemmReduce_Wmma_CShuffleV3;
|
||||
|
||||
static_assert(PermuteA == false,
|
||||
"Permute A functionality not supported by DeviceBatchedGemm operations.\n");
|
||||
static_assert(PermuteB == false,
|
||||
"Permute B functionality not supported by DeviceBatchedGemm operations.\n");
|
||||
|
||||
using CDEShuffleBlockTransferScalarPerVectors =
|
||||
Sequence<CDEShuffleBlockTransferScalarPerVector,
|
||||
CDEShuffleBlockTransferScalarPerVector,
|
||||
CDEShuffleBlockTransferScalarPerVector>;
|
||||
|
||||
using GridwiseGemm = GridwiseGemm_wmma_cshuffle_v3<
|
||||
ALayout,
|
||||
BLayout,
|
||||
Tuple<>,
|
||||
ELayout,
|
||||
Tuple<ADataType>,
|
||||
Tuple<BDataType>,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
Tuple<>,
|
||||
EDataType,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CElementwiseOperation,
|
||||
GemmSpec,
|
||||
BlockSize,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
AK1,
|
||||
BK1,
|
||||
MPerWmma,
|
||||
NPerWmma,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
ABlockTransferThreadClusterArrangeOrder,
|
||||
ABlockTransferSrcAccessOrder,
|
||||
ABlockTransferSrcVectorDim,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
ABlockTransferDstScalarPerVector_AK1,
|
||||
false,
|
||||
ABlockLdsExtraM,
|
||||
BBlockTransferThreadClusterLengths_BK0_N_BK1,
|
||||
BBlockTransferThreadClusterArrangeOrder,
|
||||
BBlockTransferSrcAccessOrder,
|
||||
BBlockTransferSrcVectorDim,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
BBlockTransferDstScalarPerVector_BK1,
|
||||
false,
|
||||
BBlockLdsExtraN,
|
||||
CShuffleMRepeatPerShuffle,
|
||||
CShuffleNRepeatPerShuffle,
|
||||
CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
CDEShuffleBlockTransferScalarPerVectors,
|
||||
BlkGemmPipeSched,
|
||||
BlkGemmPipelineVer,
|
||||
ComputeTypeA,
|
||||
ComputeTypeB,
|
||||
PermuteA,
|
||||
PermuteB,
|
||||
false,
|
||||
false,
|
||||
true>;
|
||||
|
||||
using ReduceTrait = ReduceTrait_<ReduceAccDataType,
|
||||
ReducePtrsGlobal,
|
||||
tensor_operation::element_wise::PassThrough,
|
||||
ReduceOperations,
|
||||
ReduceInElementwiseOperations,
|
||||
ReduceAccElementwiseOperations,
|
||||
ReduceGlobalMemoryDataOperation,
|
||||
CReduceThreadClusterLengths_MPerBlock_NPerBlock,
|
||||
CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
|
||||
CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock>;
|
||||
|
||||
static constexpr index_t NumReduce = ReduceOperations::Size();
|
||||
|
||||
struct ComputePtrOffsetOfStridedBatch
|
||||
{
|
||||
ComputePtrOffsetOfStridedBatch(long_index_t BatchStrideA,
|
||||
long_index_t BatchStrideB,
|
||||
long_index_t BatchStrideC,
|
||||
std::array<long_index_t, NumReduce> BatchStrideReduce)
|
||||
: BatchStrideA_{BatchStrideA},
|
||||
BatchStrideB_{BatchStrideB},
|
||||
BatchStrideC_{BatchStrideC},
|
||||
BatchStrideReduce_{BatchStrideReduce}
|
||||
{
|
||||
}
|
||||
|
||||
__host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
|
||||
{
|
||||
return g_idx * BatchStrideA_;
|
||||
}
|
||||
|
||||
__host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
|
||||
{
|
||||
return g_idx * BatchStrideB_;
|
||||
}
|
||||
|
||||
__host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
|
||||
{
|
||||
return g_idx * BatchStrideC_;
|
||||
}
|
||||
|
||||
template <typename ReducePtrs>
|
||||
__host__ __device__ void OffsetReducePtrs(index_t g_idx, ReducePtrs& ptrs) const
|
||||
{
|
||||
static_for<0, NumReduce, 1>{}(
|
||||
[&](auto I) { ptrs(I) = ptrs(I) + g_idx * BatchStrideReduce_[I.value]; });
|
||||
}
|
||||
|
||||
private:
|
||||
long_index_t BatchStrideA_;
|
||||
long_index_t BatchStrideB_;
|
||||
long_index_t BatchStrideC_;
|
||||
std::array<long_index_t, NumReduce> BatchStrideReduce_{};
|
||||
};
|
||||
|
||||
private:
|
||||
static long_index_t ComputeABatchStride(index_t MRaw, index_t KRaw, index_t StrideA)
|
||||
{
|
||||
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
|
||||
{
|
||||
return static_cast<long_index_t>(MRaw) * StrideA;
|
||||
}
|
||||
else
|
||||
{
|
||||
return static_cast<long_index_t>(KRaw) * StrideA;
|
||||
}
|
||||
}
|
||||
|
||||
static long_index_t ComputeBBatchStride(index_t KRaw, index_t NRaw, index_t StrideB)
|
||||
{
|
||||
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
|
||||
{
|
||||
return static_cast<long_index_t>(KRaw) * StrideB;
|
||||
}
|
||||
else
|
||||
{
|
||||
return static_cast<long_index_t>(NRaw) * StrideB;
|
||||
}
|
||||
}
|
||||
|
||||
static long_index_t ComputeCBatchStride(index_t MRaw, index_t NRaw, index_t StrideC)
|
||||
{
|
||||
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ELayout>)
|
||||
{
|
||||
return static_cast<long_index_t>(MRaw) * StrideC;
|
||||
}
|
||||
else
|
||||
{
|
||||
return static_cast<long_index_t>(NRaw) * StrideC;
|
||||
}
|
||||
}
|
||||
|
||||
public:
|
||||
struct Argument : public BaseArgument
|
||||
{
|
||||
Argument(const ADataType* p_a_grid,
|
||||
const BDataType* p_b_grid,
|
||||
EDataType* p_e_grid,
|
||||
ReducePtrsGlobal p_reduces_grid,
|
||||
index_t MRaw,
|
||||
index_t NRaw,
|
||||
index_t KRaw,
|
||||
index_t StrideA,
|
||||
index_t StrideB,
|
||||
index_t StrideC,
|
||||
index_t Batch,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CElementwiseOperation c_element_op,
|
||||
ReduceInElementwiseOperations reduce_in_element_ops,
|
||||
ReduceAccElementwiseOperations reduce_out_element_ops,
|
||||
std::array<long_index_t, NumReduce> batch_stride_reduce)
|
||||
: p_a_grid_{p_a_grid},
|
||||
p_b_grid_{p_b_grid},
|
||||
p_e_grid_{p_e_grid},
|
||||
p_reduces_grid_{p_reduces_grid},
|
||||
MRaw_{MRaw},
|
||||
NRaw_{NRaw},
|
||||
KRaw_{KRaw},
|
||||
StrideA_{StrideA},
|
||||
StrideB_{StrideB},
|
||||
StrideC_{StrideC},
|
||||
Batch_{Batch},
|
||||
a_element_op_{a_element_op},
|
||||
b_element_op_{b_element_op},
|
||||
c_element_op_{c_element_op},
|
||||
reduce_in_element_ops_{reduce_in_element_ops},
|
||||
reduce_out_element_ops_{reduce_out_element_ops},
|
||||
batch_stride_reduce_{batch_stride_reduce},
|
||||
compute_ptr_offset_of_batch_(
|
||||
ComputePtrOffsetOfStridedBatch{ComputeABatchStride(MRaw, KRaw, StrideA),
|
||||
ComputeBBatchStride(KRaw, NRaw, StrideB),
|
||||
ComputeCBatchStride(MRaw, NRaw, StrideC),
|
||||
batch_stride_reduce})
|
||||
{
|
||||
}
|
||||
|
||||
const ADataType* p_a_grid_;
|
||||
const BDataType* p_b_grid_;
|
||||
EDataType* p_e_grid_;
|
||||
ReducePtrsGlobal p_reduces_grid_;
|
||||
index_t MRaw_;
|
||||
index_t NRaw_;
|
||||
index_t KRaw_;
|
||||
index_t StrideA_;
|
||||
index_t StrideB_;
|
||||
index_t StrideC_;
|
||||
index_t Batch_;
|
||||
AElementwiseOperation a_element_op_;
|
||||
BElementwiseOperation b_element_op_;
|
||||
CElementwiseOperation c_element_op_;
|
||||
ReduceInElementwiseOperations reduce_in_element_ops_;
|
||||
ReduceAccElementwiseOperations reduce_out_element_ops_;
|
||||
std::array<long_index_t, NumReduce> batch_stride_reduce_{};
|
||||
ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
|
||||
};
|
||||
|
||||
struct Invoker : public BaseInvoker
|
||||
{
|
||||
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
|
||||
{
|
||||
typename GridwiseGemm::Argument gemm_arg{
|
||||
std::array<const void*, 1>{arg.p_a_grid_},
|
||||
std::array<const void*, 1>{arg.p_b_grid_},
|
||||
std::array<const void*, 0>{},
|
||||
static_cast<EDataType*>(arg.p_e_grid_),
|
||||
arg.MRaw_,
|
||||
arg.NRaw_,
|
||||
arg.KRaw_,
|
||||
std::array<index_t, 1>{arg.StrideA_}, // StrideAs
|
||||
std::array<index_t, 1>{arg.StrideB_}, // StrideBs
|
||||
std::array<index_t, 0>{}, // StrideDs
|
||||
arg.StrideC_, // StrideC
|
||||
1, // kbatch
|
||||
arg.a_element_op_,
|
||||
arg.b_element_op_,
|
||||
arg.c_element_op_};
|
||||
|
||||
if(stream_config.log_level_ > 0)
|
||||
{
|
||||
gemm_arg.Print();
|
||||
GridwiseGemm::BlockwiseGemmPipe::HotLoopInstList::Print();
|
||||
}
|
||||
|
||||
if(!GridwiseGemm::CheckValidity(gemm_arg, true))
|
||||
{
|
||||
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
|
||||
}
|
||||
|
||||
index_t gdx, gdy, gdz;
|
||||
std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.MRaw_, arg.NRaw_, 1);
|
||||
|
||||
gdy *= arg.Batch_;
|
||||
|
||||
float ave_time = 0;
|
||||
|
||||
const index_t K_split = (arg.KRaw_ + KPerBlock - 1) / KPerBlock * KPerBlock;
|
||||
|
||||
const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
|
||||
const TailNumber tail_num = GridwiseGemm::CalculateKBlockLoopTailNum(K_split);
|
||||
|
||||
const auto Run = [&](const auto& kernel) {
|
||||
// Note: cache flushing not supported
|
||||
|
||||
ave_time += launch_and_time_kernel(stream_config,
|
||||
kernel,
|
||||
dim3(gdx, gdy, gdz),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
gemm_arg,
|
||||
arg.p_reduces_grid_,
|
||||
arg.reduce_in_element_ops_,
|
||||
arg.reduce_out_element_ops_,
|
||||
arg.compute_ptr_offset_of_batch_);
|
||||
};
|
||||
|
||||
constexpr index_t minimum_occupancy = []() {
|
||||
if constexpr(BlkGemmPipeSched == BlockGemmPipelineScheduler::Interwave)
|
||||
{
|
||||
return 2;
|
||||
}
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
|
||||
{
|
||||
return (MPerBlock * NPerBlock / BlockSize <= 128) ? 2 : 1;
|
||||
}
|
||||
else
|
||||
{
|
||||
return 1;
|
||||
}
|
||||
}();
|
||||
|
||||
auto CreateAndRunKernel = [&](auto has_main_k_block_loop_, auto tail_number_) {
|
||||
constexpr bool has_loop = decltype(has_main_k_block_loop_)::value;
|
||||
constexpr TailNumber tn = tail_number_;
|
||||
|
||||
const auto kernel =
|
||||
kernel_batched_gemm_reduce_wmma_cshuffle_v3<GridwiseGemm,
|
||||
typename DeviceOp::ReduceTrait,
|
||||
ComputePtrOffsetOfStridedBatch,
|
||||
has_loop,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
tn>;
|
||||
|
||||
Run(kernel);
|
||||
};
|
||||
|
||||
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
|
||||
{
|
||||
if(has_main_k_block_loop && tail_num == TailNumber::Full)
|
||||
{
|
||||
CreateAndRunKernel(std::integral_constant<bool, true>{},
|
||||
std::integral_constant<TailNumber, TailNumber::Full>{});
|
||||
}
|
||||
else if(!has_main_k_block_loop && tail_num == TailNumber::Full)
|
||||
{
|
||||
CreateAndRunKernel(std::integral_constant<bool, false>{},
|
||||
std::integral_constant<TailNumber, TailNumber::Full>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
printf("Invalid has_main_k_block_loop and tail_num combination for V1!\n");
|
||||
return 0.0f;
|
||||
}
|
||||
}
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
|
||||
{
|
||||
if(has_main_k_block_loop && tail_num == TailNumber::Full)
|
||||
{
|
||||
CreateAndRunKernel(std::integral_constant<bool, true>{},
|
||||
std::integral_constant<TailNumber, TailNumber::Full>{});
|
||||
}
|
||||
else if(!has_main_k_block_loop && tail_num == TailNumber::Even)
|
||||
{
|
||||
CreateAndRunKernel(std::integral_constant<bool, false>{},
|
||||
std::integral_constant<TailNumber, TailNumber::Even>{});
|
||||
}
|
||||
else if(!has_main_k_block_loop && tail_num == TailNumber::Odd)
|
||||
{
|
||||
CreateAndRunKernel(std::integral_constant<bool, false>{},
|
||||
std::integral_constant<TailNumber, TailNumber::Odd>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
printf("Invalid has_main_k_block_loop and tail_num combination for V3!\n");
|
||||
return 0.0f;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
printf("Invalid pipeline version!\n");
|
||||
return 0.0f;
|
||||
}
|
||||
|
||||
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 constexpr bool IsValidCompilationParameter() { return true; }
|
||||
|
||||
static bool IsSupportedArgument(const Argument& arg)
|
||||
{
|
||||
if(!ck::is_gfx11_supported() && !ck::is_gfx12_supported())
|
||||
{
|
||||
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
|
||||
{
|
||||
std::cout << "Device implementation supports only gfx11 and gfx12! " << __FILE__
|
||||
<< ":" << __LINE__ << ", in function: " << __func__ << std::endl;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
if constexpr(std::is_same_v<ComputeTypeA, f8_t> || std::is_same_v<ComputeTypeA, bf8_t> ||
|
||||
std::is_same_v<ComputeTypeB, f8_t> || std::is_same_v<ComputeTypeB, bf8_t>)
|
||||
{
|
||||
if(ck::is_gfx11_supported())
|
||||
{
|
||||
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
|
||||
{
|
||||
std::cout << "FP8 and BF8 not supported on gfx11! " << __FILE__ << ":"
|
||||
<< __LINE__ << ", in function: " << __func__ << std::endl;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
if((arg.KRaw_ % AK1 != 0 || arg.KRaw_ % BK1 != 0) &&
|
||||
!(GemmSpec == GemmSpecialization::MKPadding ||
|
||||
GemmSpec == GemmSpecialization::NKPadding ||
|
||||
GemmSpec == GemmSpecialization::MNKPadding ||
|
||||
GemmSpec == GemmSpecialization::KPadding))
|
||||
{
|
||||
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
|
||||
{
|
||||
std::cout << "Without padding, K must be divisible by AK1 and BK1! " << __FILE__
|
||||
<< ":" << __LINE__ << ", in function: " << __func__ << std::endl;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
typename GridwiseGemm::Argument gemm_arg{std::array<const void*, 1>{arg.p_a_grid_},
|
||||
std::array<const void*, 1>{arg.p_b_grid_},
|
||||
std::array<const void*, 0>{},
|
||||
static_cast<EDataType*>(arg.p_e_grid_),
|
||||
arg.MRaw_,
|
||||
arg.NRaw_,
|
||||
arg.KRaw_,
|
||||
std::array<index_t, 1>{arg.StrideA_}, // StrideAs
|
||||
std::array<index_t, 1>{arg.StrideB_}, // StrideBs
|
||||
std::array<index_t, 0>{}, // StrideDs
|
||||
arg.StrideC_, // StrideC
|
||||
1, // kbatch
|
||||
arg.a_element_op_,
|
||||
arg.b_element_op_,
|
||||
arg.c_element_op_};
|
||||
|
||||
return GridwiseGemm::CheckValidity(gemm_arg, true);
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
bool IsSupportedArgument(const BaseArgument* p_arg) override
|
||||
{
|
||||
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
|
||||
}
|
||||
|
||||
static auto MakeArgument(const void* p_a,
|
||||
const void* p_b,
|
||||
const void* p_bias,
|
||||
std::array<const void*, 0> p_ds,
|
||||
void* p_e,
|
||||
std::array<void*, NumReduce> p_reduces,
|
||||
ck::index_t M,
|
||||
ck::index_t N,
|
||||
ck::index_t K,
|
||||
ck::index_t StrideA,
|
||||
ck::index_t StrideB,
|
||||
ck::index_t StrideC,
|
||||
std::array<ck::index_t, 0> StrideDs,
|
||||
std::array<void*, 3> gemm_element_ops,
|
||||
std::array<void*, 0> d_element_ops,
|
||||
std::array<void*, NumReduce> reduce_in_element_op,
|
||||
std::array<void*, NumReduce> reduce_out_element_op,
|
||||
ck::index_t Batch)
|
||||
{
|
||||
(void)p_bias;
|
||||
(void)p_ds;
|
||||
(void)StrideDs;
|
||||
(void)d_element_ops;
|
||||
|
||||
ReducePtrsGlobal reduce_tuple = generate_tuple(
|
||||
[&](auto I) {
|
||||
auto tmp = ReducePtrsGlobal{}[I];
|
||||
using T = remove_pointer_t<decltype(tmp)>;
|
||||
return static_cast<T*>(p_reduces[I.value]);
|
||||
},
|
||||
Number<NumReduce>{});
|
||||
|
||||
ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
|
||||
[&](auto I) {
|
||||
auto tmp = ReduceInElementwiseOperations{}[I];
|
||||
using T = remove_pointer_t<decltype(tmp)>;
|
||||
return *(static_cast<T*>(reduce_in_element_op[I.value]));
|
||||
},
|
||||
Number<NumReduce>{});
|
||||
|
||||
ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
|
||||
[&](auto I) {
|
||||
auto tmp = ReduceAccElementwiseOperations{}[I];
|
||||
using T = remove_pointer_t<decltype(tmp)>;
|
||||
return *(static_cast<T*>(reduce_out_element_op[I.value]));
|
||||
},
|
||||
Number<NumReduce>{});
|
||||
|
||||
AElementwiseOperation a_element_op =
|
||||
*(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
|
||||
BElementwiseOperation b_element_op =
|
||||
*(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
|
||||
CElementwiseOperation c_element_op =
|
||||
*(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
|
||||
|
||||
std::array<long_index_t, NumReduce> batch_stride_reduce{};
|
||||
static_for<0, NumReduce, 1>{}(
|
||||
[&](auto I) { batch_stride_reduce[I.value] = static_cast<long_index_t>(M); });
|
||||
|
||||
return Argument{static_cast<const ADataType*>(p_a),
|
||||
static_cast<const BDataType*>(p_b),
|
||||
static_cast<EDataType*>(p_e),
|
||||
reduce_tuple,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
Batch,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op,
|
||||
reduce_in_element_ops,
|
||||
reduce_out_element_ops,
|
||||
batch_stride_reduce};
|
||||
}
|
||||
|
||||
static auto MakeInvoker() { return Invoker{}; }
|
||||
|
||||
// polymorphic
|
||||
std::unique_ptr<BaseArgument>
|
||||
MakeArgumentPointer(const void* p_a,
|
||||
const void* p_b,
|
||||
const void* p_bias,
|
||||
std::array<const void*, 0> p_ds,
|
||||
void* p_e,
|
||||
std::array<void*, NumReduce> p_reduces,
|
||||
ck::index_t M,
|
||||
ck::index_t N,
|
||||
ck::index_t K,
|
||||
ck::index_t StrideA,
|
||||
ck::index_t StrideB,
|
||||
ck::index_t StrideC,
|
||||
std::array<ck::index_t, 0> StrideDs,
|
||||
std::array<void*, 3> gemm_element_ops,
|
||||
std::array<void*, 0> d_element_ops,
|
||||
std::array<void*, NumReduce> reduce_in_element_op,
|
||||
std::array<void*, NumReduce> reduce_out_element_op,
|
||||
ck::index_t Batch = 1) override
|
||||
{
|
||||
(void)p_bias;
|
||||
(void)p_ds;
|
||||
(void)StrideDs;
|
||||
(void)d_element_ops;
|
||||
|
||||
ReducePtrsGlobal reduce_tuple = generate_tuple(
|
||||
[&](auto I) {
|
||||
auto tmp = ReducePtrsGlobal{}[I];
|
||||
using T = remove_pointer_t<decltype(tmp)>;
|
||||
return static_cast<T*>(p_reduces[I.value]);
|
||||
},
|
||||
Number<NumReduce>{});
|
||||
|
||||
ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
|
||||
[&](auto I) {
|
||||
auto tmp = ReduceInElementwiseOperations{}[I];
|
||||
using T = remove_pointer_t<decltype(tmp)>;
|
||||
return *(static_cast<T*>(reduce_in_element_op[I.value]));
|
||||
},
|
||||
Number<NumReduce>{});
|
||||
|
||||
ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
|
||||
[&](auto I) {
|
||||
auto tmp = ReduceAccElementwiseOperations{}[I];
|
||||
using T = remove_pointer_t<decltype(tmp)>;
|
||||
return *(static_cast<T*>(reduce_out_element_op[I.value]));
|
||||
},
|
||||
Number<NumReduce>{});
|
||||
|
||||
AElementwiseOperation a_element_op =
|
||||
*(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
|
||||
BElementwiseOperation b_element_op =
|
||||
*(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
|
||||
CElementwiseOperation c_element_op =
|
||||
*(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
|
||||
|
||||
std::array<long_index_t, NumReduce> batch_stride_reduce{};
|
||||
static_for<0, NumReduce, 1>{}(
|
||||
[&](auto I) { batch_stride_reduce[I.value] = static_cast<long_index_t>(M); });
|
||||
|
||||
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
|
||||
static_cast<const BDataType*>(p_b),
|
||||
static_cast<EDataType*>(p_e),
|
||||
reduce_tuple,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
Batch,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op,
|
||||
reduce_in_element_ops,
|
||||
reduce_out_element_ops,
|
||||
batch_stride_reduce);
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
|
||||
{
|
||||
return std::make_unique<Invoker>(Invoker{});
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
std::string GetTypeString() const override
|
||||
{
|
||||
auto str = std::stringstream();
|
||||
|
||||
str << "DeviceBatchedGemmReduce_Wmma_CShuffleV3" << "<" << BlockSize << ", " << MPerBlock
|
||||
<< ", " << NPerBlock << ", " << KPerBlock << ", " << AK1 << ", " << BK1 << ", "
|
||||
<< MPerWmma << ", " << NPerWmma << ", " << MRepeat << ", " << NRepeat << ", "
|
||||
<< ABlockTransferSrcScalarPerVector << ", " << BBlockTransferSrcScalarPerVector << ", "
|
||||
<< CShuffleMRepeatPerShuffle << ", " << CShuffleNRepeatPerShuffle << ">";
|
||||
|
||||
return str.str();
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user