mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-05 06:01:23 +00:00
Add multiple A/B support (#906)
* add gridwise_multi_abd * move element_op into RunRead * merge element_wise op with data read * add multiABD example * allow packed elementwise_op * changed example * clean * clean * add is_detected * fix * minor fix * add scaleAdd_vec4 example --------- Co-authored-by: Jing Zhang <jizha@amd.com>
This commit is contained in:
zjing14
@@ -0,0 +1,214 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/utility/common_header.hpp"
|
||||
#include "ck/tensor_description/tensor_descriptor.hpp"
|
||||
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
|
||||
#include "ck/tensor_description/cluster_descriptor.hpp"
|
||||
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp"
|
||||
#include "ck/utility/is_detected.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
// Thread-group level multi-source, multi-destination tensor slice data movement
|
||||
// Assume:
|
||||
// 1. All sources and destinations are DynamicBuffer
|
||||
// 2. Same VectorDim and ScalerPerVector for all sources and destinations
|
||||
// 3. DstInMemOps are per destination tensor
|
||||
// 4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
|
||||
// 5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
|
||||
//
|
||||
// Does following things to avoid scratch memory issue
|
||||
// 1. Pass tensor descritpors by reference (or tuple of references)
|
||||
// 2. Does not keep reference to tensor descriptor
|
||||
// 3. Does not construct new tensor coordinate when call Run()
|
||||
template <typename ThreadGroup,
|
||||
typename SrcDatas,
|
||||
typename DstDatas,
|
||||
typename SrcDescs,
|
||||
typename DstDescs,
|
||||
typename ElementwiseOperation,
|
||||
typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
|
||||
typename SliceLengths,
|
||||
typename ThreadClusterLengths,
|
||||
typename ThreadClusterArrangeOrder,
|
||||
typename SrcDimAccessOrder,
|
||||
typename DstDimAccessOrder,
|
||||
index_t SrcVectorDim,
|
||||
index_t DstVectorDim,
|
||||
index_t SrcScalarPerVector,
|
||||
index_t DstScalarPerVector,
|
||||
typename ThreadTransferSrcResetCoordinateAfterRunFlags,
|
||||
typename ThreadTransferDstResetCoordinateAfterRunFlags>
|
||||
struct ThreadGroupTensorSliceTransfer_v7r2
|
||||
{
|
||||
static constexpr index_t nDim =
|
||||
remove_cvref_t<tuple_element_t<0, SrcDescs>>::GetNumOfDimension();
|
||||
|
||||
static constexpr index_t nSrc = remove_cvref_t<SrcDescs>::Size();
|
||||
static constexpr index_t nDst = remove_cvref_t<DstDescs>::Size();
|
||||
|
||||
using Index = MultiIndex<nDim>;
|
||||
|
||||
static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
|
||||
|
||||
__device__ constexpr ThreadGroupTensorSliceTransfer_v7r2(
|
||||
const SrcDescs& src_descs,
|
||||
const StaticallyIndexedArray<Index, nSrc>& src_block_slice_origins,
|
||||
const DstDescs& dst_descs,
|
||||
const StaticallyIndexedArray<Index, nDst>& dst_block_slice_origins,
|
||||
const ElementwiseOperation& element_op)
|
||||
: threadwise_transfer_(src_descs,
|
||||
StaticallyIndexedArray<Index, nSrc>{},
|
||||
dst_descs,
|
||||
StaticallyIndexedArray<Index, nDst>{},
|
||||
element_op)
|
||||
{
|
||||
static_assert(nSrc == SrcDatas::Size() && nSrc == SrcDescs::Size() &&
|
||||
nSrc == ThreadTransferSrcResetCoordinateAfterRunFlags::Size() &&
|
||||
nDst == DstDatas::Size() && nDst == DstDescs::Size() &&
|
||||
nDst == ThreadTransferDstResetCoordinateAfterRunFlags::Size(),
|
||||
"wrong!");
|
||||
|
||||
static_for<0, nSrc, 1>{}([&](auto i) {
|
||||
static_assert(
|
||||
nDim == remove_cvref_t<tuple_element_t<i.value, SrcDescs>>::GetNumOfDimension(),
|
||||
"wrong!");
|
||||
});
|
||||
|
||||
static_for<0, nDst, 1>{}([&](auto i) {
|
||||
static_assert(
|
||||
nDim == remove_cvref_t<tuple_element_t<i.value, DstDescs>>::GetNumOfDimension(),
|
||||
"wrong!");
|
||||
});
|
||||
|
||||
static_assert(nDim == ThreadClusterLengths::Size() &&
|
||||
nDim == ThreadClusterArrangeOrder::Size() &&
|
||||
nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(),
|
||||
"wrong! nDim not consistent");
|
||||
|
||||
static_assert(
|
||||
is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
|
||||
"wrong! threads should be mapped to cover entire slicing window");
|
||||
|
||||
static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
|
||||
"wrong! ThreadGroup::GetNumOfThread() too small");
|
||||
|
||||
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
|
||||
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
|
||||
{
|
||||
const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
|
||||
make_multi_index(get_thread_local_1d_id()));
|
||||
|
||||
const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
|
||||
|
||||
const auto src_thread_slice_origins = generate_tuple(
|
||||
[&](auto i) { return src_block_slice_origins[i] + thread_data_idx_begin; },
|
||||
Number<nSrc>{});
|
||||
|
||||
const auto dst_thread_slice_origins = generate_tuple(
|
||||
[&](auto i) { return dst_block_slice_origins[i] + thread_data_idx_begin; },
|
||||
Number<nDst>{});
|
||||
|
||||
threadwise_transfer_.SetSrcSliceOrigins(src_descs, src_thread_slice_origins);
|
||||
threadwise_transfer_.SetDstSliceOrigins(dst_descs, dst_thread_slice_origins);
|
||||
}
|
||||
}
|
||||
|
||||
template <typename SrcBuffers>
|
||||
__device__ void RunRead(const SrcDescs& src_descs, const SrcBuffers& src_bufs)
|
||||
{
|
||||
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
|
||||
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
|
||||
{
|
||||
threadwise_transfer_.RunRead(src_descs, src_bufs);
|
||||
}
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
using is_tuple = decltype(std::declval<T&>().IsTuple());
|
||||
|
||||
template <typename DstBuffers>
|
||||
__device__ void RunWrite(const DstDescs& dst_descs, DstBuffers dst_bufs)
|
||||
{
|
||||
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
|
||||
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
|
||||
{
|
||||
if constexpr(is_detected<is_tuple, decltype(dst_bufs)>::value)
|
||||
threadwise_transfer_.RunWrite(dst_descs, dst_bufs);
|
||||
else
|
||||
threadwise_transfer_.RunWrite(dst_descs, tie(dst_bufs));
|
||||
}
|
||||
}
|
||||
|
||||
template <typename SrcBuffers, typename DstBuffers>
|
||||
__device__ void Run(const SrcDescs& src_descs,
|
||||
const SrcBuffers& src_bufs,
|
||||
const DstDescs& dst_descs,
|
||||
DstBuffers dst_bufs)
|
||||
{
|
||||
RunRead(src_descs, src_bufs);
|
||||
RunWrite(dst_descs, dst_bufs);
|
||||
}
|
||||
|
||||
template <index_t ISrc>
|
||||
__device__ void
|
||||
MoveSrcSliceWindow(const SrcDescs& src_descs, Number<ISrc> iSrc, const Index& step)
|
||||
{
|
||||
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
|
||||
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
|
||||
{
|
||||
threadwise_transfer_.MoveSrcSliceWindow(src_descs, iSrc, step);
|
||||
}
|
||||
}
|
||||
|
||||
__device__ void MoveSrcSliceWindow(const SrcDescs& src_descs, const Index& step)
|
||||
{
|
||||
static_for<0, SrcDescs::Size(), 1>{}(
|
||||
[&](auto i) { MoveSrcSliceWindow(src_descs, i, step); });
|
||||
}
|
||||
|
||||
template <index_t IDst>
|
||||
__device__ void
|
||||
MoveDstSliceWindow(const DstDescs& dst_descs, Number<IDst> iDst, const Index& step)
|
||||
{
|
||||
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
|
||||
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
|
||||
{
|
||||
threadwise_transfer_.MoveDstSliceWindow(dst_descs, iDst, step);
|
||||
}
|
||||
}
|
||||
|
||||
__device__ void MoveDstSliceWindow(const DstDescs& dst_descs, const Index& step)
|
||||
{
|
||||
static_for<0, DstDescs::Size(), 1>{}(
|
||||
[&](auto i) { MoveDstSliceWindow(dst_descs, i, step); });
|
||||
}
|
||||
|
||||
private:
|
||||
static constexpr auto thread_cluster_desc_ =
|
||||
make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
|
||||
|
||||
using ThreadwiseTransfer =
|
||||
ThreadwiseTensorSliceTransfer_v7r2<SrcDatas,
|
||||
DstDatas,
|
||||
SrcDescs,
|
||||
DstDescs,
|
||||
ElementwiseOperation,
|
||||
DstInMemOps,
|
||||
decltype(thread_slice_lengths),
|
||||
SrcDimAccessOrder,
|
||||
DstDimAccessOrder,
|
||||
SrcVectorDim,
|
||||
DstVectorDim,
|
||||
SrcScalarPerVector,
|
||||
DstScalarPerVector,
|
||||
ThreadTransferSrcResetCoordinateAfterRunFlags,
|
||||
ThreadTransferDstResetCoordinateAfterRunFlags>;
|
||||
|
||||
ThreadwiseTransfer threadwise_transfer_;
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,60 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <array>
|
||||
|
||||
#include "ck/tensor_operation/gpu/device/device_base.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
|
||||
// GEMM:
|
||||
// input : A0[M, K], B0[K, N],
|
||||
// input : D0[M, N], D1[M, N], ...
|
||||
// output : E[M, N]
|
||||
// C = a_op(A) * b_op(B)
|
||||
// E = cde_op(C, D0, D1, ...)
|
||||
// Assume:
|
||||
// D0, D1, ... and E have the same layout
|
||||
template <typename AsLayout,
|
||||
typename BsLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename AsDataType,
|
||||
typename BsDataType,
|
||||
typename DsDataType,
|
||||
typename EDataType,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CDEElementwiseOperation>
|
||||
struct DeviceGemmMultipleABD : public BaseOperator
|
||||
{
|
||||
static constexpr index_t NumATensor = AsDataType::Size();
|
||||
static constexpr index_t NumBTensor = BsDataType::Size();
|
||||
static constexpr index_t NumDTensor = DsDataType::Size();
|
||||
|
||||
virtual std::unique_ptr<BaseArgument>
|
||||
MakeArgumentPointer(std::array<const void*, NumATensor> p_as,
|
||||
std::array<const void*, NumBTensor> p_bs,
|
||||
std::array<const void*, NumDTensor> p_ds,
|
||||
void* p_e,
|
||||
ck::index_t M,
|
||||
ck::index_t N,
|
||||
ck::index_t K,
|
||||
std::array<ck::index_t, NumATensor> StrideAs,
|
||||
std::array<ck::index_t, NumBTensor> StrideBs,
|
||||
std::array<ck::index_t, NumDTensor> StrideDs,
|
||||
ck::index_t StrideE,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation cde_element_op) = 0;
|
||||
|
||||
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,766 @@
|
||||
// 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_gemm_multiple_abd.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
|
||||
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp"
|
||||
#include "ck/host_utility/device_prop.hpp"
|
||||
#include "ck/host_utility/kernel_launch.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
template <typename GridwiseGemm,
|
||||
typename AsPointer,
|
||||
typename BsPointer,
|
||||
typename DsPointer,
|
||||
typename EDataType,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CDEElementwiseOperation,
|
||||
typename AsGridDesc_AK0_M_AK1,
|
||||
typename BsGridDesc_BK0_N_BK1,
|
||||
typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
typename Block2ETileMap,
|
||||
bool HasMainKBlockLoop>
|
||||
__global__ void
|
||||
#if CK_USE_LAUNCH_BOUNDS
|
||||
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
|
||||
#endif
|
||||
kernel_gemm_multiple_abd_xdl_cshuffle(
|
||||
AsPointer p_as_grid,
|
||||
BsPointer p_bs_grid,
|
||||
DsPointer p_ds_grid,
|
||||
EDataType* __restrict__ p_e_grid,
|
||||
const AElementwiseOperation a_element_op,
|
||||
const BElementwiseOperation b_element_op,
|
||||
const CDEElementwiseOperation cde_element_op,
|
||||
const AsGridDesc_AK0_M_AK1 as_grid_desc_ak0_m_ak1,
|
||||
const BsGridDesc_BK0_N_BK1 bs_grid_desc_bk0_n_bk1,
|
||||
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
|
||||
ds_grid_desc_mblock_mperblock_nblock_nperblock,
|
||||
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
|
||||
e_grid_desc_mblock_mperblock_nblock_nperblock,
|
||||
const Block2ETileMap block_2_etile_map)
|
||||
{
|
||||
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
|
||||
defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
|
||||
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
|
||||
|
||||
GridwiseGemm::template Run<HasMainKBlockLoop>(p_as_grid,
|
||||
p_bs_grid,
|
||||
p_ds_grid,
|
||||
p_e_grid,
|
||||
p_shared,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op,
|
||||
as_grid_desc_ak0_m_ak1,
|
||||
bs_grid_desc_bk0_n_bk1,
|
||||
ds_grid_desc_mblock_mperblock_nblock_nperblock,
|
||||
e_grid_desc_mblock_mperblock_nblock_nperblock,
|
||||
block_2_etile_map);
|
||||
#else
|
||||
ignore = p_as_grid;
|
||||
ignore = p_bs_grid;
|
||||
ignore = p_ds_grid;
|
||||
ignore = p_e_grid;
|
||||
ignore = a_element_op;
|
||||
ignore = b_element_op;
|
||||
ignore = cde_element_op;
|
||||
ignore = as_grid_desc_ak0_m_ak1;
|
||||
ignore = bs_grid_desc_bk0_n_bk1;
|
||||
ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
|
||||
ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
|
||||
ignore = block_2_etile_map;
|
||||
#endif
|
||||
}
|
||||
|
||||
} // namespace ck
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
|
||||
// GEMM:
|
||||
// input : A[M, K]
|
||||
// input : B[N, K]
|
||||
// input : D0[M, N], D1[M, N], ...
|
||||
// output : E[M, N]
|
||||
// C = a_op(A) * b_op(B)
|
||||
// E = cde_op(C, D0, D1, ...)
|
||||
// Assume:
|
||||
// D0, D1, ... and E have the same layout
|
||||
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,
|
||||
index_t NumGemmKPrefetchStage,
|
||||
index_t BlockSize,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t AK1,
|
||||
index_t BK1,
|
||||
index_t MPerXDL,
|
||||
index_t NPerXDL,
|
||||
index_t MXdlPerWave,
|
||||
index_t NXdlPerWave,
|
||||
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
typename ABlockTransferThreadClusterArrangeOrder,
|
||||
typename ABlockTransferSrcAccessOrder,
|
||||
index_t ABlockTransferSrcVectorDim,
|
||||
index_t ABlockTransferSrcScalarPerVector,
|
||||
index_t ABlockTransferDstScalarPerVector_AK1,
|
||||
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 CShuffleMXdlPerWavePerShuffle,
|
||||
index_t CShuffleNXdlPerWavePerShuffle,
|
||||
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
index_t CDEBlockTransferScalarPerVector_NPerBlock,
|
||||
LoopScheduler LoopSched = make_default_loop_scheduler(),
|
||||
PipelineVersion PipelineVer = PipelineVersion::v1>
|
||||
struct DeviceGemmMultipleABD_Xdl_CShuffle : public DeviceGemmMultipleABD<AsLayout,
|
||||
BsLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
AsDataType,
|
||||
BsDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CDEElementwiseOperation>
|
||||
{
|
||||
using DeviceOp = DeviceGemmMultipleABD_Xdl_CShuffle;
|
||||
|
||||
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>{};
|
||||
static constexpr auto I3 = Number<3>{};
|
||||
|
||||
#if 0
|
||||
static constexpr auto matrix_padder =
|
||||
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
|
||||
|
||||
static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideAs)
|
||||
{
|
||||
const auto a_grid_desc_mraw_kraw = [&]() {
|
||||
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, AsLayout>)
|
||||
{
|
||||
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
|
||||
make_tuple(StrideAs, I1));
|
||||
}
|
||||
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, AsLayout>)
|
||||
{
|
||||
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
|
||||
make_tuple(I1, StrideAs));
|
||||
}
|
||||
}();
|
||||
|
||||
return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
|
||||
}
|
||||
|
||||
static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideBs)
|
||||
{
|
||||
const auto b_grid_desc_nraw_kraw = [&]() {
|
||||
if constexpr(is_same<tensor_layout::gemm::RowMajor, BsLayout>::value)
|
||||
{
|
||||
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
|
||||
make_tuple(I1, StrideBs));
|
||||
}
|
||||
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BsLayout>::value)
|
||||
{
|
||||
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
|
||||
make_tuple(StrideBs, I1));
|
||||
}
|
||||
}();
|
||||
|
||||
return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
|
||||
}
|
||||
|
||||
template <typename ELay>
|
||||
static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
|
||||
{
|
||||
const auto e_grid_desc_mraw_nraw = [&]() {
|
||||
if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
|
||||
{
|
||||
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
|
||||
make_tuple(StrideE, I1));
|
||||
}
|
||||
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
|
||||
{
|
||||
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
|
||||
make_tuple(I1, StrideE));
|
||||
}
|
||||
}();
|
||||
|
||||
return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
|
||||
}
|
||||
|
||||
static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
|
||||
const std::array<index_t, NumDTensor>& NRaws,
|
||||
const std::array<index_t, NumDTensor>& DsStride)
|
||||
{
|
||||
return generate_tuple(
|
||||
[&](auto i) {
|
||||
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
|
||||
|
||||
return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
|
||||
},
|
||||
Number<NumDTensor>{});
|
||||
}
|
||||
#endif
|
||||
using ComputeDataType = EDataType;
|
||||
|
||||
// GridwiseGemm
|
||||
using GridwiseGemm = GridwiseGemmMultipleABD_xdl_cshuffle<
|
||||
AsDataType,
|
||||
BsDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CDEElementwiseOperation,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
NumGemmKPrefetchStage,
|
||||
BlockSize,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
AK1,
|
||||
BK1,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
ABlockTransferThreadClusterArrangeOrder,
|
||||
ABlockTransferSrcAccessOrder,
|
||||
ABlockTransferSrcVectorDim,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
ABlockTransferDstScalarPerVector_AK1,
|
||||
false,
|
||||
ABlockLdsExtraM,
|
||||
BBlockTransferThreadClusterLengths_BK0_N_BK1,
|
||||
BBlockTransferThreadClusterArrangeOrder,
|
||||
BBlockTransferSrcAccessOrder,
|
||||
BBlockTransferSrcVectorDim,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
BBlockTransferDstScalarPerVector_BK1,
|
||||
false,
|
||||
BBlockLdsExtraN,
|
||||
CShuffleMXdlPerWavePerShuffle,
|
||||
CShuffleNXdlPerWavePerShuffle,
|
||||
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
CDEBlockTransferScalarPerVector_NPerBlock,
|
||||
LoopSched,
|
||||
PipelineVer>;
|
||||
|
||||
// desc for problem definition
|
||||
using AsGridDesc_M_K =
|
||||
remove_cvref_t<decltype(GridwiseGemm::template MakeAsGridDescriptor_M_K<AsLayout, GemmSpec>(
|
||||
{}, {}, {}))>;
|
||||
using BsGridDesc_N_K =
|
||||
remove_cvref_t<decltype(GridwiseGemm::template MakeBsGridDescriptor_N_K<BsLayout, GemmSpec>(
|
||||
{}, {}, {}))>;
|
||||
using DsGridDesc_M_N =
|
||||
remove_cvref_t<decltype(GridwiseGemm::template MakeDsGridDescriptor_M_N<DsLayout, GemmSpec>(
|
||||
{}, {}, {}))>;
|
||||
using EGridDesc_M_N =
|
||||
decltype(GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(1, 1, 1));
|
||||
|
||||
// desc for blockwise copy
|
||||
using AsGridDesc_AK0_M_AK1 =
|
||||
remove_cvref_t<decltype(GridwiseGemm::MakeAsGridDescriptor_AK0_M_AK1(AsGridDesc_M_K{}))>;
|
||||
using BsGridDesc_BK0_N_BK1 =
|
||||
remove_cvref_t<decltype(GridwiseGemm::MakeBsGridDescriptor_BK0_N_BK1(BsGridDesc_N_K{}))>;
|
||||
using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
|
||||
decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
|
||||
DsGridDesc_M_N{}))>;
|
||||
using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
|
||||
remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
|
||||
EGridDesc_M_N{}))>;
|
||||
|
||||
// block-to-e-tile map
|
||||
using Block2ETileMap =
|
||||
remove_cvref_t<decltype(GridwiseGemm::MakeBlock2ETileMap(EGridDesc_M_N{}))>;
|
||||
|
||||
// Argument
|
||||
struct Argument : public BaseArgument
|
||||
{
|
||||
Argument(std::array<const void*, NumATensor> p_as_grid,
|
||||
std::array<const void*, NumBTensor> p_bs_grid,
|
||||
std::array<const void*, NumDTensor> p_ds_grid,
|
||||
void* p_e_grid,
|
||||
index_t MRaw,
|
||||
index_t NRaw,
|
||||
index_t KRaw,
|
||||
std::array<index_t, NumATensor> StrideAs,
|
||||
std::array<index_t, NumBTensor> StrideBs,
|
||||
std::array<index_t, NumDTensor> StrideDs,
|
||||
index_t StrideE,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation cde_element_op)
|
||||
: p_as_grid_{},
|
||||
p_bs_grid_{},
|
||||
p_ds_grid_{},
|
||||
p_e_grid_{static_cast<EDataType*>(p_e_grid)},
|
||||
as_grid_desc_m_k_{},
|
||||
bs_grid_desc_n_k_{},
|
||||
ds_grid_desc_m_n_{},
|
||||
e_grid_desc_m_n_{GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(
|
||||
MRaw, NRaw, StrideE)},
|
||||
as_grid_desc_ak0_m_ak1_{},
|
||||
bs_grid_desc_bk0_n_bk1_{},
|
||||
ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
|
||||
e_grid_desc_mblock_mperblock_nblock_nperblock_{},
|
||||
block_2_etile_map_{GridwiseGemm::MakeBlock2ETileMap(e_grid_desc_m_n_)},
|
||||
a_element_op_{a_element_op},
|
||||
b_element_op_{b_element_op},
|
||||
cde_element_op_{cde_element_op},
|
||||
MRaw_{MRaw},
|
||||
NRaw_{NRaw},
|
||||
KRaw_{KRaw}
|
||||
{
|
||||
// populate pointer, desc for As
|
||||
static_for<0, NumATensor, 1>{}([&](auto i) {
|
||||
using ALayout = remove_cvref_t<tuple_element_t<i.value, AsLayout>>;
|
||||
using ADataType = remove_cvref_t<tuple_element_t<i.value, AsDataType>>;
|
||||
|
||||
// A pointer
|
||||
p_as_grid_(i) = static_cast<const ADataType*>(p_as_grid[i]);
|
||||
|
||||
// A desc
|
||||
as_grid_desc_m_k_(i) =
|
||||
GridwiseGemm::template MakeAGridDescriptor_M_K<ALayout, GemmSpec>(
|
||||
MRaw, KRaw, StrideAs[i]);
|
||||
});
|
||||
|
||||
// populate pointer, desc for Bs
|
||||
static_for<0, NumBTensor, 1>{}([&](auto i) {
|
||||
using BLayout = remove_cvref_t<tuple_element_t<i.value, BsLayout>>;
|
||||
using BDataType = remove_cvref_t<tuple_element_t<i.value, BsDataType>>;
|
||||
|
||||
// B pointer
|
||||
p_bs_grid_(i) = static_cast<const BDataType*>(p_bs_grid[i]);
|
||||
|
||||
// B desc
|
||||
bs_grid_desc_n_k_(i) =
|
||||
GridwiseGemm::template MakeBGridDescriptor_N_K<BLayout, GemmSpec>(
|
||||
KRaw, NRaw, StrideBs[i]);
|
||||
});
|
||||
|
||||
// populate pointer, desc for Ds
|
||||
static_for<0, NumDTensor, 1>{}([&](auto i) {
|
||||
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
|
||||
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
|
||||
|
||||
// D pointer
|
||||
p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
|
||||
|
||||
// D desc
|
||||
ds_grid_desc_m_n_(i) =
|
||||
GridwiseGemm::template MakeEGridDescriptor_M_N<DLayout, GemmSpec>(
|
||||
MRaw, NRaw, StrideDs[i]);
|
||||
});
|
||||
|
||||
// populate desc for Ds/E
|
||||
if(GridwiseGemm::CheckValidity(as_grid_desc_m_k_,
|
||||
bs_grid_desc_n_k_,
|
||||
ds_grid_desc_m_n_,
|
||||
e_grid_desc_m_n_,
|
||||
block_2_etile_map_))
|
||||
{
|
||||
as_grid_desc_ak0_m_ak1_ =
|
||||
GridwiseGemm::MakeAsGridDescriptor_AK0_M_AK1(as_grid_desc_m_k_);
|
||||
|
||||
bs_grid_desc_bk0_n_bk1_ =
|
||||
GridwiseGemm::MakeBsGridDescriptor_BK0_N_BK1(bs_grid_desc_n_k_);
|
||||
|
||||
ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
|
||||
GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
|
||||
ds_grid_desc_m_n_);
|
||||
|
||||
e_grid_desc_mblock_mperblock_nblock_nperblock_ =
|
||||
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
|
||||
e_grid_desc_m_n_);
|
||||
}
|
||||
}
|
||||
|
||||
void Print() const
|
||||
{
|
||||
// std::cout << "A[M, K]: " << as_grid_desc_m_k_ << std::endl;
|
||||
// std::cout << "B[N, K]: " << bs_grid_desc_n_k_ << std::endl;
|
||||
// static_for<0, NumDTensor, 1>{}(
|
||||
//[&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
|
||||
// std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
|
||||
}
|
||||
|
||||
// private:
|
||||
// pointers
|
||||
typename GridwiseGemm::AsGridPointer p_as_grid_;
|
||||
typename GridwiseGemm::BsGridPointer p_bs_grid_;
|
||||
typename GridwiseGemm::DsGridPointer p_ds_grid_;
|
||||
EDataType* p_e_grid_;
|
||||
|
||||
// tensor descriptors for problem definiton
|
||||
AsGridDesc_M_K as_grid_desc_m_k_;
|
||||
BsGridDesc_N_K bs_grid_desc_n_k_;
|
||||
DsGridDesc_M_N ds_grid_desc_m_n_;
|
||||
EGridDesc_M_N e_grid_desc_m_n_;
|
||||
|
||||
// tensor descriptors for block/thread-wise copy
|
||||
AsGridDesc_AK0_M_AK1 as_grid_desc_ak0_m_ak1_;
|
||||
BsGridDesc_BK0_N_BK1 bs_grid_desc_bk0_n_bk1_;
|
||||
DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
|
||||
ds_grid_desc_mblock_mperblock_nblock_nperblock_;
|
||||
EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
|
||||
|
||||
// block-to-e-tile map
|
||||
Block2ETileMap block_2_etile_map_;
|
||||
|
||||
// element-wise op
|
||||
AElementwiseOperation a_element_op_;
|
||||
BElementwiseOperation b_element_op_;
|
||||
CDEElementwiseOperation cde_element_op_;
|
||||
|
||||
// for checking vector load/store
|
||||
index_t MRaw_;
|
||||
index_t NRaw_;
|
||||
index_t KRaw_;
|
||||
};
|
||||
|
||||
// Invoker
|
||||
struct Invoker : public BaseInvoker
|
||||
{
|
||||
using Argument = DeviceOp::Argument;
|
||||
|
||||
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
|
||||
{
|
||||
if(!GridwiseGemm::CheckValidity(arg.as_grid_desc_m_k_,
|
||||
arg.bs_grid_desc_n_k_,
|
||||
arg.ds_grid_desc_m_n_,
|
||||
arg.e_grid_desc_m_n_,
|
||||
arg.block_2_etile_map_))
|
||||
{
|
||||
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
|
||||
}
|
||||
|
||||
const index_t grid_size =
|
||||
arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
|
||||
|
||||
auto launch_kernel = [&](auto has_main_k_block_loop) {
|
||||
constexpr bool has_main_loop = has_main_k_block_loop.value;
|
||||
|
||||
const auto kernel = kernel_gemm_multiple_abd_xdl_cshuffle<
|
||||
GridwiseGemm,
|
||||
typename GridwiseGemm::AsGridPointer,
|
||||
typename GridwiseGemm::BsGridPointer,
|
||||
typename GridwiseGemm::DsGridPointer,
|
||||
EDataType,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CDEElementwiseOperation,
|
||||
DeviceOp::AsGridDesc_AK0_M_AK1,
|
||||
DeviceOp::BsGridDesc_BK0_N_BK1,
|
||||
DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
DeviceOp::Block2ETileMap,
|
||||
has_main_loop>;
|
||||
|
||||
return launch_and_time_kernel(stream_config,
|
||||
kernel,
|
||||
dim3(grid_size),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
arg.p_as_grid_,
|
||||
arg.p_bs_grid_,
|
||||
arg.p_ds_grid_,
|
||||
arg.p_e_grid_,
|
||||
arg.a_element_op_,
|
||||
arg.b_element_op_,
|
||||
arg.cde_element_op_,
|
||||
arg.as_grid_desc_ak0_m_ak1_,
|
||||
arg.bs_grid_desc_bk0_n_bk1_,
|
||||
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
|
||||
arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
|
||||
arg.block_2_etile_map_);
|
||||
};
|
||||
|
||||
const auto K = arg.as_grid_desc_m_k_[I0].GetLength(I1);
|
||||
|
||||
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
|
||||
{
|
||||
return launch_kernel(integral_constant<bool, true>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
return launch_kernel(integral_constant<bool, false>{});
|
||||
}
|
||||
}
|
||||
|
||||
// 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::is_xdl_supported())
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
// check vector load/store
|
||||
{
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
bool all_valid = true;
|
||||
|
||||
static_for<0, NumATensor, 1>{}([&](auto i) {
|
||||
using ALayout = remove_cvref_t<tuple_element_t<i.value, AsLayout>>;
|
||||
// check vector load of A
|
||||
if constexpr(is_same_v<ALayout, Row> && ABlockTransferSrcVectorDim == 2)
|
||||
{
|
||||
if(arg.KRaw_ % ABlockTransferSrcScalarPerVector != 0)
|
||||
{
|
||||
all_valid = false;
|
||||
}
|
||||
}
|
||||
else if constexpr(is_same_v<ALayout, Col> && ABlockTransferSrcVectorDim == 1)
|
||||
{
|
||||
// FIXME: not rigorous
|
||||
if(arg.MRaw_ % ABlockTransferSrcScalarPerVector != 0)
|
||||
{
|
||||
all_valid = false;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
all_valid = false;
|
||||
}
|
||||
});
|
||||
|
||||
static_for<0, NumBTensor, 1>{}([&](auto i) {
|
||||
using BLayout = remove_cvref_t<tuple_element_t<i.value, BsLayout>>;
|
||||
// check vector laod of B
|
||||
if constexpr(is_same_v<BLayout, Col> && BBlockTransferSrcVectorDim == 2)
|
||||
{
|
||||
if(arg.KRaw_ % BBlockTransferSrcScalarPerVector != 0)
|
||||
{
|
||||
all_valid = false;
|
||||
}
|
||||
}
|
||||
else if constexpr(is_same_v<BLayout, Row> && BBlockTransferSrcVectorDim == 1)
|
||||
{
|
||||
// FIXME: not rigorous
|
||||
if(arg.NRaw_ % BBlockTransferSrcScalarPerVector != 0)
|
||||
{
|
||||
all_valid = false;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
all_valid = false;
|
||||
}
|
||||
});
|
||||
|
||||
// check vector load of Ds
|
||||
// only support RowMajor for now
|
||||
|
||||
static_for<0, NumDTensor, 1>{}([&](auto i) {
|
||||
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
|
||||
|
||||
if constexpr(!is_same_v<DLayout, Row>)
|
||||
{
|
||||
all_valid = false;
|
||||
}
|
||||
});
|
||||
|
||||
if(!all_valid)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
// check vector store of E
|
||||
// only support RowMajor for now
|
||||
if constexpr(is_same_v<ELayout, Row>)
|
||||
{
|
||||
if(arg.NRaw_ % CDEBlockTransferScalarPerVector_NPerBlock != 0)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
return GridwiseGemm::CheckValidity(arg.as_grid_desc_m_k_,
|
||||
arg.bs_grid_desc_n_k_,
|
||||
arg.ds_grid_desc_m_n_,
|
||||
arg.e_grid_desc_m_n_,
|
||||
arg.block_2_etile_map_);
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
bool IsSupportedArgument(const BaseArgument* p_arg) override
|
||||
{
|
||||
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
|
||||
}
|
||||
|
||||
static auto MakeArgument(std::array<const void*, NumATensor> p_as,
|
||||
std::array<const void*, NumBTensor> p_bs,
|
||||
std::array<const void*, NumDTensor> p_ds,
|
||||
void* p_e,
|
||||
index_t MRaw,
|
||||
index_t NRaw,
|
||||
index_t KRaw,
|
||||
std::array<index_t, NumATensor> StrideAs,
|
||||
std::array<index_t, NumBTensor> StrideBs,
|
||||
std::array<index_t, NumDTensor> StrideDs,
|
||||
index_t StrideE,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation cde_element_op)
|
||||
{
|
||||
return Argument{p_as,
|
||||
p_bs,
|
||||
p_ds,
|
||||
p_e,
|
||||
MRaw,
|
||||
NRaw,
|
||||
KRaw,
|
||||
StrideAs,
|
||||
StrideBs,
|
||||
StrideDs,
|
||||
StrideE,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op};
|
||||
}
|
||||
|
||||
static auto MakeInvoker() { return Invoker{}; }
|
||||
|
||||
// polymorphic
|
||||
std::unique_ptr<BaseArgument>
|
||||
MakeArgumentPointer(std::array<const void*, NumATensor> p_as,
|
||||
std::array<const void*, NumBTensor> p_bs,
|
||||
std::array<const void*, NumDTensor> p_ds,
|
||||
void* p_e,
|
||||
index_t MRaw,
|
||||
index_t NRaw,
|
||||
index_t KRaw,
|
||||
std::array<ck::index_t, NumATensor> StrideAs,
|
||||
std::array<ck::index_t, NumBTensor> StrideBs,
|
||||
std::array<ck::index_t, NumDTensor> StrideDs,
|
||||
index_t StrideE,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation cde_element_op) override
|
||||
{
|
||||
return std::make_unique<Argument>(p_as,
|
||||
p_bs,
|
||||
p_ds,
|
||||
p_e,
|
||||
MRaw,
|
||||
NRaw,
|
||||
KRaw,
|
||||
StrideAs,
|
||||
StrideBs,
|
||||
StrideDs,
|
||||
StrideE,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_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();
|
||||
|
||||
std::map<LoopScheduler, std::string> LoopSchedToString{
|
||||
{LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
|
||||
|
||||
std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
|
||||
{PipelineVersion::v2, "v2"}};
|
||||
|
||||
// clang-format off
|
||||
str << "DeviceGemmMultipleABD_Xdl_CShuffle"
|
||||
<< "<"
|
||||
<< BlockSize << ", "
|
||||
<< MPerBlock << ", "
|
||||
<< NPerBlock << ", "
|
||||
<< KPerBlock << ", "
|
||||
<< AK1 << ", "
|
||||
<< BK1 << ", "
|
||||
<< MPerXDL << ", "
|
||||
<< NPerXDL << ", "
|
||||
<< MXdlPerWave << ", "
|
||||
<< NXdlPerWave << ", "
|
||||
<< ABlockTransferSrcScalarPerVector << ", "
|
||||
<< BBlockTransferSrcScalarPerVector << ", "
|
||||
<< CShuffleMXdlPerWavePerShuffle << ", "
|
||||
<< CShuffleNXdlPerWavePerShuffle << ", "
|
||||
<< getGemmSpecializationString(GemmSpec)
|
||||
<< ">"
|
||||
<< " LoopScheduler: "
|
||||
<< LoopSchedToString[LoopSched] << ", "
|
||||
<< "PipelineVersion: "
|
||||
<< PipelineVersionToString[PipelineVer];
|
||||
// clang-format on
|
||||
|
||||
return str.str();
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,420 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/utility/common_header.hpp"
|
||||
#include "ck/tensor_description/tensor_descriptor.hpp"
|
||||
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
|
||||
#include "ck/tensor_description/tensor_space_filling_curve.hpp"
|
||||
#include "ck/utility/is_detected.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
// Thread-level multi-source, multi-destination tensor slice data movement
|
||||
// Assume:
|
||||
// 1. All sources and destinations are DynamicBuffer
|
||||
// 2. Same VectorDim and ScalerPerVector for all sources and destinations
|
||||
// 3. DstInMemOps are per destination tensor
|
||||
// 4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
|
||||
// 5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
|
||||
// 6. Does not need to know src_descs and dst_descs at compile-time
|
||||
// 7. Does not need to know src_slice_origins and dst_slice_origins at compile-time,
|
||||
//
|
||||
// Does following things to avoid scratch memory issue
|
||||
// 1. Use StaticallyIndexedArray or vector_type instead of C array for thread buffer
|
||||
// 2. Pass tensor descritpors by reference (or tuple of references)
|
||||
// 3. Does not keep reference to tensor descriptor
|
||||
// 4. Does not construct new tensor coordinate when call Run()
|
||||
template <typename SrcDatas,
|
||||
typename DstDatas,
|
||||
typename SrcDescs,
|
||||
typename DstDescs,
|
||||
typename ElementwiseOperation,
|
||||
typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
|
||||
typename SliceLengths,
|
||||
typename SrcDimAccessOrder,
|
||||
typename DstDimAccessOrder,
|
||||
index_t SrcVectorDim,
|
||||
index_t DstVectorDim,
|
||||
index_t SrcScalarPerVector,
|
||||
index_t DstScalarPerVector,
|
||||
typename SrcResetCoordinateAfterRunFlags, // Sequence<bool ...>
|
||||
typename DstResetCoordinateAfterRunFlags> // Sequence<bool ...>
|
||||
struct ThreadwiseTensorSliceTransfer_v7r2
|
||||
{
|
||||
static constexpr auto I0 = Number<0>{};
|
||||
|
||||
static constexpr index_t nDim = SliceLengths::Size();
|
||||
|
||||
static constexpr index_t nSrc = SrcDescs::Size();
|
||||
static constexpr index_t nDst = DstDescs::Size();
|
||||
|
||||
using Index = MultiIndex<nDim>;
|
||||
|
||||
// return a tuple of coordiantes for a tuple of tensor
|
||||
template <typename Descs,
|
||||
typename Indices,
|
||||
enable_if_t<Descs::Size() == Indices::Size(), bool> = false>
|
||||
static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices)
|
||||
{
|
||||
return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); },
|
||||
Number<Descs::Size()>{});
|
||||
}
|
||||
|
||||
using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray<Index, nSrc>{}));
|
||||
using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray<Index, nDst>{}));
|
||||
|
||||
// scalar per access on each dim
|
||||
// FIXME: don't use lambda_scalar_per_access
|
||||
static constexpr auto src_scalar_per_access = generate_sequence(
|
||||
detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
|
||||
|
||||
using SrcSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
|
||||
SrcDimAccessOrder,
|
||||
remove_cv_t<decltype(src_scalar_per_access)>>;
|
||||
|
||||
static constexpr auto dst_scalar_per_access = generate_sequence(
|
||||
detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
|
||||
|
||||
using DstSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
|
||||
DstDimAccessOrder,
|
||||
remove_cv_t<decltype(dst_scalar_per_access)>>;
|
||||
|
||||
__device__ constexpr ThreadwiseTensorSliceTransfer_v7r2(
|
||||
const SrcDescs& src_descs,
|
||||
const StaticallyIndexedArray<Index, nSrc>& src_slice_origins,
|
||||
const DstDescs& dst_descs,
|
||||
const StaticallyIndexedArray<Index, nDst>& dst_slice_origins,
|
||||
const ElementwiseOperation& element_op)
|
||||
: src_coords_(MakeCoordinates(src_descs, src_slice_origins)),
|
||||
dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)),
|
||||
element_op_(element_op)
|
||||
{
|
||||
static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
|
||||
"wrong! cannot evenly divide");
|
||||
|
||||
static_assert(SliceLengths::At(Number<DstVectorDim>{}) % DstScalarPerVector == 0,
|
||||
"wrong! cannot evenly divide");
|
||||
}
|
||||
|
||||
template <typename Indices, enable_if_t<SrcDescs::Size() == Indices::Size(), bool> = false>
|
||||
__device__ void SetSrcSliceOrigins(const SrcDescs& src_descs,
|
||||
const Indices& src_slice_origin_idxs)
|
||||
{
|
||||
static_for<0, nSrc, 1>{}([&](auto i) {
|
||||
src_coords_(i) = make_tensor_coordinate(src_descs[i], src_slice_origin_idxs[i]);
|
||||
});
|
||||
}
|
||||
|
||||
template <typename Indices, enable_if_t<DstDescs::Size() == Indices::Size(), bool> = false>
|
||||
__device__ void SetDstSliceOrigins(const DstDescs& dst_descs,
|
||||
const Indices& dst_slice_origin_idxs)
|
||||
{
|
||||
static_for<0, nDst, 1>{}([&](auto i) {
|
||||
dst_coords_(i) = make_tensor_coordinate(dst_descs[i], dst_slice_origin_idxs[i]);
|
||||
});
|
||||
}
|
||||
|
||||
template <typename DataTypes, index_t ScalarPerVector>
|
||||
__device__ static auto generate_vectors()
|
||||
{
|
||||
auto data_types = DataTypes{};
|
||||
|
||||
constexpr index_t num = data_types.Size();
|
||||
|
||||
return generate_tuple(
|
||||
[&](auto i) {
|
||||
using DataType = remove_cvref_t<decltype(data_types[i])>;
|
||||
|
||||
return vector_type_maker_t<DataType, ScalarPerVector>{};
|
||||
},
|
||||
Number<num>{});
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
using has_vec_len = decltype(std::declval<T&>().vec_len);
|
||||
|
||||
// SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
|
||||
// SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
|
||||
template <typename SrcBuffers,
|
||||
enable_if_t<SrcDescs::Size() == SrcBuffers::Size(), bool> = false>
|
||||
__device__ void RunRead(const SrcDescs& src_descs, const SrcBuffers& src_bufs)
|
||||
{
|
||||
// loop over space-filling curve
|
||||
static_for<0, num_access, 1>{}([&](auto iAccess) {
|
||||
auto src_vectors = generate_vectors<SrcDatas, SrcScalarPerVector>();
|
||||
auto dst_vectors = generate_vectors<DstDatas, DstScalarPerVector>();
|
||||
|
||||
// copy data from src_bufs into src_vectors
|
||||
static_for<0, nSrc, 1>{}([&](auto i) {
|
||||
using src_vector_t = typename remove_cvref_t<decltype(src_vectors[i])>::type;
|
||||
|
||||
const bool is_src_valid =
|
||||
coordinate_has_valid_offset_assuming_visible_index_is_valid(src_descs[i],
|
||||
src_coords_[i]);
|
||||
|
||||
src_vectors(i).template AsType<src_vector_t>()(I0) =
|
||||
src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(),
|
||||
is_src_valid);
|
||||
});
|
||||
|
||||
if constexpr(is_detected<has_vec_len, decltype(element_op_)>::value)
|
||||
{
|
||||
constexpr auto elem_op_vec_len = decltype(element_op_)::vec_len;
|
||||
|
||||
static_assert(is_same<remove_cvref_t<decltype(elem_op_vec_len)>, index_t>::value,
|
||||
"vec_len in element_op_ type is not index_t");
|
||||
|
||||
static_assert(elem_op_vec_len == 1 || elem_op_vec_len == 2 ||
|
||||
elem_op_vec_len == 4 || elem_op_vec_len == 8,
|
||||
"vec_len in element_op_ must be 1, 2, 4, 8");
|
||||
|
||||
static_assert(SrcScalarPerVector % elem_op_vec_len == 0,
|
||||
"vec_len in element_op_ cannot be divided by SrcScalarPerVector!");
|
||||
|
||||
// apply pointwise function
|
||||
static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto i) {
|
||||
// get reference to src data
|
||||
const auto src_data_refs = generate_tie(
|
||||
// return type should be lvalue
|
||||
[&](auto iSrc) -> const auto& {
|
||||
using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
|
||||
|
||||
using elem_op_vec_t =
|
||||
typename vector_type<SrcData, elem_op_vec_len>::type;
|
||||
|
||||
return src_vectors[iSrc].template AsType<elem_op_vec_t>()[i];
|
||||
},
|
||||
Number<nSrc>{});
|
||||
|
||||
// get reference to dst data
|
||||
auto dst_data_refs = generate_tie(
|
||||
// return type should be lvalue
|
||||
[&](auto iDst) -> auto& {
|
||||
using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
|
||||
|
||||
using elem_op_vec_t =
|
||||
typename vector_type<DstData, elem_op_vec_len>::type;
|
||||
|
||||
return dst_vectors(iDst).template AsType<elem_op_vec_t>()(i);
|
||||
},
|
||||
Number<nDst>{});
|
||||
|
||||
// apply pointwise function
|
||||
// pointwise function signature:
|
||||
// element_op_(dst_data_refs[I0],
|
||||
// dst_data_refs[I1],
|
||||
// ...,
|
||||
// src_data_refs[I0],
|
||||
// src_data_refs[I1],
|
||||
// ...)
|
||||
unpack2(element_op_, dst_data_refs, src_data_refs);
|
||||
});
|
||||
}
|
||||
else
|
||||
{
|
||||
// apply pointwise function
|
||||
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
|
||||
// get reference to src data
|
||||
const auto src_data_refs = generate_tie(
|
||||
// return type should be lvalue
|
||||
[&](auto iSrc) -> const auto& {
|
||||
using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
|
||||
|
||||
return src_vectors[iSrc].template AsType<SrcData>()[i];
|
||||
},
|
||||
Number<nSrc>{});
|
||||
|
||||
// get reference to dst data
|
||||
auto dst_data_refs = generate_tie(
|
||||
// return type should be lvalue
|
||||
[&](auto iDst) -> auto& {
|
||||
using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
|
||||
|
||||
return dst_vectors(iDst).template AsType<DstData>()(i);
|
||||
},
|
||||
Number<nDst>{});
|
||||
|
||||
// apply pointwise function
|
||||
// pointwise function signature:
|
||||
// element_op_(dst_data_refs[I0],
|
||||
// dst_data_refs[I1],
|
||||
// ...,
|
||||
// src_data_refs[I0],
|
||||
// src_data_refs[I1],
|
||||
// ...)
|
||||
unpack2(element_op_, dst_data_refs, src_data_refs);
|
||||
});
|
||||
}
|
||||
|
||||
dst_vectors_tuple_(iAccess) = dst_vectors;
|
||||
|
||||
// move coordinate
|
||||
if constexpr(iAccess.value != num_access - 1)
|
||||
{
|
||||
constexpr auto forward_step = SrcSpaceFillingCurve::GetForwardStep(iAccess);
|
||||
|
||||
static_for<0, nSrc, 1>{}([&](auto i) {
|
||||
move_tensor_coordinate(src_descs[i],
|
||||
src_coords_(i),
|
||||
make_tensor_coordinate_step(src_descs[i], forward_step));
|
||||
});
|
||||
}
|
||||
});
|
||||
|
||||
// move coordinate back to slice origin (or not)
|
||||
static_for<0, nSrc, 1>{}([&](auto i) {
|
||||
if constexpr(SrcResetCoordinateAfterRunFlags::At(i))
|
||||
{
|
||||
const auto src_reset_step =
|
||||
make_tensor_coordinate_step(src_descs[i], GetSrcCoordinateResetStep());
|
||||
|
||||
move_tensor_coordinate(src_descs[i], src_coords_(i), src_reset_step);
|
||||
}
|
||||
});
|
||||
}
|
||||
|
||||
// DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
|
||||
// DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
|
||||
template <typename DstBuffers,
|
||||
enable_if_t<DstDescs::Size() == DstBuffers::Size(), bool> = false>
|
||||
__device__ void RunWrite(const DstDescs& dst_descs, DstBuffers dst_bufs)
|
||||
{
|
||||
// loop over space-filling curve
|
||||
static_for<0, num_access, 1>{}([&](auto iAccess) {
|
||||
auto dst_vectors = dst_vectors_tuple_[iAccess];
|
||||
|
||||
// copy data from buf_vectors into dst_bufs
|
||||
static_for<0, nDst, 1>{}([&](auto i) {
|
||||
using dst_vector_t = typename remove_cvref_t<decltype(dst_vectors[i])>::type;
|
||||
|
||||
const bool is_dst_valid =
|
||||
coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_descs[i],
|
||||
dst_coords_[i]);
|
||||
|
||||
constexpr InMemoryDataOperationEnum DstInMemOp =
|
||||
static_cast<InMemoryDataOperationEnum>(DstInMemOps::At(i.value));
|
||||
|
||||
dst_bufs(i).template Update<DstInMemOp, dst_vector_t>(
|
||||
dst_coords_[i].GetOffset(),
|
||||
is_dst_valid,
|
||||
dst_vectors[i].template AsType<dst_vector_t>()[I0]);
|
||||
});
|
||||
|
||||
// move coordinate
|
||||
if constexpr(iAccess.value != num_access - 1)
|
||||
{
|
||||
constexpr auto forward_step = DstSpaceFillingCurve::GetForwardStep(iAccess);
|
||||
|
||||
static_for<0, nDst, 1>{}([&](auto i) {
|
||||
move_tensor_coordinate(dst_descs[i],
|
||||
dst_coords_(i),
|
||||
make_tensor_coordinate_step(dst_descs[i], forward_step));
|
||||
});
|
||||
}
|
||||
});
|
||||
|
||||
static_for<0, nDst, 1>{}([&](auto i) {
|
||||
if constexpr(DstResetCoordinateAfterRunFlags::At(i))
|
||||
{
|
||||
const auto dst_reset_step =
|
||||
make_tensor_coordinate_step(dst_descs[i], GetDstCoordinateResetStep());
|
||||
|
||||
move_tensor_coordinate(dst_descs[i], dst_coords_(i), dst_reset_step);
|
||||
}
|
||||
});
|
||||
}
|
||||
|
||||
// SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
|
||||
// SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
|
||||
// DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
|
||||
// DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
|
||||
template <typename SrcBuffers,
|
||||
typename DstBuffers,
|
||||
enable_if_t<SrcDescs::Size() == SrcBuffers::Size() &&
|
||||
DstDescs::Size() == DstBuffers::Size(),
|
||||
bool> = false>
|
||||
__device__ void Run(const SrcDescs& src_descs,
|
||||
const SrcBuffers& src_bufs,
|
||||
const DstDescs& dst_descs,
|
||||
DstBuffers dst_bufs)
|
||||
{
|
||||
RunRead(src_descs, src_bufs);
|
||||
RunWrite(dst_descs, dst_bufs);
|
||||
}
|
||||
|
||||
__device__ static constexpr auto GetSrcCoordinateResetStep()
|
||||
{
|
||||
if constexpr(num_access == 0)
|
||||
{
|
||||
return typename SrcSpaceFillingCurve::Index{};
|
||||
}
|
||||
else
|
||||
{
|
||||
return SrcSpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
|
||||
}
|
||||
}
|
||||
|
||||
__device__ static constexpr auto GetDstCoordinateResetStep()
|
||||
{
|
||||
if constexpr(num_access == 0)
|
||||
{
|
||||
return typename DstSpaceFillingCurve::Index{};
|
||||
}
|
||||
else
|
||||
{
|
||||
return DstSpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
|
||||
}
|
||||
}
|
||||
|
||||
// src_slice_origin_step_idx need to be known at compile-time, for performance reason
|
||||
template <index_t ISrc>
|
||||
__device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
|
||||
Number<ISrc> iSrc,
|
||||
const Index& src_slice_origin_step_idx)
|
||||
{
|
||||
// if src coord was not reset by RunRead(), then need to adjust the step here
|
||||
const auto adjusted_step_idx =
|
||||
SrcResetCoordinateAfterRunFlags::At(iSrc)
|
||||
? src_slice_origin_step_idx
|
||||
: src_slice_origin_step_idx + GetSrcCoordinateResetStep();
|
||||
|
||||
// is it OK to construct a new step every time?
|
||||
const auto adjusted_step = make_tensor_coordinate_step(src_descs[iSrc], adjusted_step_idx);
|
||||
|
||||
move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step);
|
||||
}
|
||||
|
||||
// dst_slice_origin_step_idx need to be known at compile-time, for performance reason
|
||||
template <index_t IDst>
|
||||
__device__ void MoveDstSliceWindow(const DstDescs& dst_descs,
|
||||
Number<IDst> iDst,
|
||||
const Index& dst_slice_origin_step_idx)
|
||||
{
|
||||
// if dst coord was not reset by Run(), then need to adjust the step here
|
||||
const auto adjusted_step_idx =
|
||||
DstResetCoordinateAfterRunFlags::At(iDst)
|
||||
? dst_slice_origin_step_idx
|
||||
: dst_slice_origin_step_idx + GetDstCoordinateResetStep();
|
||||
|
||||
// is it OK to construct a new step every time?
|
||||
const auto adjusted_step = make_tensor_coordinate_step(dst_descs[iDst], adjusted_step_idx);
|
||||
|
||||
move_tensor_coordinate(dst_descs[iDst], dst_coords_(iDst), adjusted_step);
|
||||
}
|
||||
|
||||
private:
|
||||
using SrcVectorsType = decltype(generate_vectors<SrcDatas, SrcScalarPerVector>());
|
||||
using DstVectorsType = decltype(generate_vectors<DstDatas, DstScalarPerVector>());
|
||||
|
||||
static constexpr auto num_access = SrcSpaceFillingCurve::GetNumOfAccess();
|
||||
|
||||
StaticallyIndexedArray<DstVectorsType, num_access> dst_vectors_tuple_;
|
||||
|
||||
SrcCoords src_coords_;
|
||||
DstCoords dst_coords_;
|
||||
const ElementwiseOperation element_op_;
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user