Skip to content

GitLab

Commit 09a2b547 authored by rocking's avatar rocking
Browse files

Copy "gemm reduce" to "gemm bias add reduce"

parent b6eaf3eb
Hide whitespace changes
Inline Side-by-side
Showing with 1628 additions and 0 deletions
include/ck/tensor_operation/gpu/device/device_gemm_bias_add_reduce_xdl_cshuffle.hpp 0 → 100644
View file @ 09a2b547
#pragma once
#include <iostream>
#include <sstream>
#include "device.hpp"
#include "device_gemm_reduce.hpp"
#include "common_header.hpp"
#include "tensor_layout.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_reduce_xdl_cshuffle_v1.hpp"
#include "gemm_specialization.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
// Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
// version currently has compiler issues with register spill which further causes validation
// failures.
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename BDataType,
typename CDataType,
typename GemmAccDataType,
typename CShuffleDataType,
typename ReduceAccDataType,
typename DPtrsGlobal,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename DxsReduceOperation,
typename DxsInElementwiseOperation,
typename DxsAccElementwiseOperation,
typename DGlobalMemoryDataOperation,
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,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
LoopScheduler LoopSched = make_default_loop_scheduler()>
struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGlobal,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
DxsInElementwiseOperation,
DxsAccElementwiseOperation>
{
using DeviceOp = DeviceGemmReduce_Xdl_CShuffle;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
{
const auto a_grid_desc_mraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(StrideA, I1));
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(I1, StrideA));
}
}();
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
const auto MPad = M - MRaw;
const auto KPad = K - KRaw;
if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both M and K
assert(K % AK1 == 0);
const auto AK0 = K / AK1;
const auto a_grid_desc_m_k =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_right_pad_transform(MRaw, MPad),
make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad M, but not K
assert(KRaw % AK1 == 0);
const auto AK0 = KRaw / AK1;
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_right_pad_transform(MRaw, MPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad K, but not M
assert(K % AK1 == 0);
const auto AK0 = K / AK1;
const auto a_grid_desc_m_k = transform_tensor_descriptor(
a_grid_desc_mraw_kraw,
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(MRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else
{
// not pad M or K
assert(KRaw % AK1 == 0);
const auto AK0 = KRaw / AK1;
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(MRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
}
static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
{
const auto b_grid_desc_nraw_kraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(I1, StrideB));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(StrideB, I1));
}
}();
const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
const auto NPad = N - NRaw;
const auto KPad = K - KRaw;
if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both N and K
assert(K % BK1 == 0);
const auto BK0 = K / BK1;
const auto b_grid_desc_n_k =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_right_pad_transform(NRaw, NPad),
make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad N, but not K
assert(KRaw % BK1 == 0);
const auto BK0 = KRaw / BK1;
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad K, but not N
assert(K % BK1 == 0);
const auto BK0 = K / BK1;
const auto b_grid_desc_n_k = transform_tensor_descriptor(
b_grid_desc_nraw_kraw,
make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else
{
// not pad N or K
assert(KRaw % BK1 == 0);
const auto BK0 = KRaw / BK1;
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
}
static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
{
const auto c_grid_desc_mraw_nraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(StrideC, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(I1, StrideC));
}
}();
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
const auto MPad = M - MRaw;
const auto NPad = N - NRaw;
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M and N
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(MRaw, MPad),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad M, but not N
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad N, but not M
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
// not pad M or N
return c_grid_desc_mraw_nraw;
}
}
// assume D is packed tensor
static auto MakeDGridDescriptor_M(index_t MRaw)
{
const auto d_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(MRaw));
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto MPad = M - MRaw;
if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M
return transform_tensor_descriptor(d_grid_desc_mraw,
make_tuple(make_right_pad_transform(MRaw, MPad)),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0>{}));
}
else
{
// not pad M
return d_grid_desc_mraw;
}
}
using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
using DGridDesc_M = decltype(MakeDGridDescriptor_M(1));
// GridwiseGemm
using GridwiseGemm = GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
ADataType, // TODO: distinguish A/B datatype
GemmAccDataType,
CShuffleDataType,
CDataType,
ReduceAccDataType,
DPtrsGlobal,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
DxsReduceOperation,
DxsInElementwiseOperation,
DxsAccElementwiseOperation,
InMemoryDataOperationEnum::Set,
DGlobalMemoryDataOperation,
AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1,
CGridDesc_M_N,
DGridDesc_M,
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,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CShuffleBlockTransferScalarPerVector_NPerBlock,
CReduceThreadClusterLengths_MPerBlock_NPerBlock,
CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
LoopSched>;
// Argument
struct Argument : public BaseArgument
{
Argument(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
DPtrsGlobal p_ds_grid,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsAccElementwiseOperation dxs_out_element_op)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_c_grid_{p_c_grid},
p_ds_grid_{p_ds_grid},
a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC)},
d_grid_desc_m_{DeviceOp::MakeDGridDescriptor_M(MRaw)},
c_grid_desc_mblock_mperblock_nblock_nperblock_{},
d_grid_desc_mblock_mperblock_{},
block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
c_element_op_{c_element_op},
dxs_in_element_op_{dxs_in_element_op},
dxs_out_element_op_{dxs_out_element_op}
{
if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
b_grid_desc_bk0_n_bk1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n_);
d_grid_desc_mblock_mperblock_ =
GridwiseGemm::MakeDGridDescriptor_MBlock_MPerBlock(d_grid_desc_m_);
}
}
// private:
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
DPtrsGlobal p_ds_grid_;
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
CGridDesc_M_N c_grid_desc_m_n_;
DGridDesc_M d_grid_desc_m_;
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock_;
typename GridwiseGemm::DGridDescriptor_MBlock_MPerBlock d_grid_desc_mblock_mperblock_;
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
DxsInElementwiseOperation dxs_in_element_op_;
DxsAccElementwiseOperation dxs_out_element_op_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
#if 0
{
std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
<< arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
<< arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) << ", "
<< arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.b_grid_desc_bk0_n_bk1_{"
<< arg.b_grid_desc_bk0_n_bk1_.GetLength(I0) << ", "
<< arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) << ", "
<< arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
<< arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
std::cout << "arg.d_grid_desc_m_{ " << arg.d_grid_desc_m_.GetLength(I0) << "}"
<< std::endl;
}
#endif
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_))
{
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
}
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K =
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
float elapsed_time = 0.0f;
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_reduce_xdl_cshuffle_v1<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
DPtrsGlobal,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
DxsInElementwiseOperation,
DxsAccElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename GridwiseGemm::DGridDescriptor_MBlock_MPerBlock,
typename GridwiseGemm::DefaultBlock2CTileMap,
true>;
elapsed_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.p_ds_grid_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.dxs_in_element_op_,
arg.dxs_out_element_op_,
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.d_grid_desc_mblock_mperblock_,
arg.block_2_ctile_map_);
}
else
{
const auto kernel = kernel_gemm_reduce_xdl_cshuffle_v1<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
DPtrsGlobal,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
DxsInElementwiseOperation,
DxsAccElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename GridwiseGemm::DGridDescriptor_MBlock_MPerBlock,
typename GridwiseGemm::DefaultBlock2CTileMap,
false>;
elapsed_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.p_ds_grid_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.dxs_in_element_op_,
arg.dxs_out_element_op_,
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.d_grid_desc_mblock_mperblock_,
arg.block_2_ctile_map_);
}
return elapsed_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()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_);
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
CDataType* p_c,
DPtrsGlobal p_dxs,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsAccElementwiseOperation dxs_out_element_op)
{
return Argument{p_a,
p_b,
p_c,
p_dxs,
MRaw,
NRaw,
KRaw,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
dxs_in_element_op,
dxs_out_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
DPtrsGlobal p_dxs,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsAccElementwiseOperation dxs_out_element_op,
index_t /* KBatch */ = 1) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
p_dxs,
MRaw,
NRaw,
KRaw,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
dxs_in_element_op,
dxs_out_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 << "DeviceGemmReduce_Xdl_CShuffle"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< KPerBlock << ", "
<< AK1 << ", "
<< BK1
<< ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp 0 → 100644
View file @ 09a2b547
#pragma once
#include "common_header.hpp"
#include "multi_index_transform_helper.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "blockwise_gemm_xdlops.hpp"
#include "thread_group_tensor_slice_transfer_v4r1.hpp"
#include "thread_group_tensor_slice_transfer_v6r1.hpp"
#include "threadwise_tensor_slice_transfer.hpp"
#include "gridwise_gemm_pipeline_v1.hpp"
#include "reduction_functions_threadwise.hpp"
namespace ck {
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename DPtrsGlobal,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename DxsInElementwiseOperation,
typename DxsAccElementwiseOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename DGridDescriptor_MBlock_MPerBlock,
typename Block2CTileMap,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_reduce_xdl_cshuffle_v1(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
DPtrsGlobal p_ds_grid,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op,
const DxsInElementwiseOperation dxs_in_element_op,
const DxsAccElementwiseOperation dxs_out_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const DGridDescriptor_MBlock_MPerBlock d_grid_desc_mblock_mperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
p_b_grid,
p_c_grid,
p_ds_grid,
p_shared,
a_element_op,
b_element_op,
c_element_op,
dxs_in_element_op,
dxs_out_element_op,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
c_grid_desc_mblock_mperblock_nblock_nperblock,
d_grid_desc_mblock_mperblock,
block_2_ctile_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_c_grid;
ignore = p_ds_grid;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
ignore = dxs_in_element_op;
ignore = dxs_out_element_op;
ignore = a_grid_desc_ak0_m_ak1;
ignore = b_grid_desc_bk0_n_bk1;
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = d_grid_desc_mblock_mperblock;
ignore = block_2_ctile_map;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
template <typename FloatAB,
typename FloatGemmAcc,
typename FloatCShuffle,
typename FloatC,
typename FloatReduceAcc,
typename DPtrsGlobal,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename DxsReduceOperation,
typename DxsInElementwiseOperation,
typename DxsAccElementwiseOperation,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename DGlobalMemoryDataOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename CGridDesc_M_N,
typename DGridDesc_M,
index_t NumGemmKPrefetchStage,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1Value,
index_t BK1Value,
index_t MPerXdl,
index_t NPerXdl,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool AThreadTransferSrcResetCoordinateAfterRun,
index_t ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BThreadTransferSrcResetCoordinateAfterRun,
index_t BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
LoopScheduler LoopSched>
struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static constexpr auto I5 = Number<5>{};
static constexpr auto I6 = Number<6>{};
static constexpr auto I7 = Number<7>{};
// K1 should be Number<...>
static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
static constexpr auto AK1 = Number<AK1Value>{};
static constexpr auto BK1 = Number<BK1Value>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemmKPrefetchStage>;
__host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
{
// A matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(AK0, Number<MPerBlock>{}, AK1),
make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
}
__host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
{
// B matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(BK0, Number<NPerBlock>{}, BK1),
make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
}
__host__ __device__ static constexpr auto
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
{
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
I1,
Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
}
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1, BK1);
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
// LDS allocation for C shuffle in LDS
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
constexpr auto c_block_size =
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
sizeof(FloatAB),
c_block_size * sizeof(FloatCShuffle));
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
template <typename Block2CTileMap>
__host__ __device__ static constexpr bool
CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
const CGridDesc_M_N& c_grid_desc_m_n,
const Block2CTileMap& block_2_ctile_map)
{
// static_assert(is_known_at_compile_time<remove_cv_t<decltype(AK1)>>::value &&
// is_known_at_compile_time<remove_cv_t<decltype(BK1)>>::value,
// "wrong! K1 need to be known at compile-time");
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
"Invalid tuning param!");
const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
return false;
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
return false;
// check gridwise gemm pipeline
const auto num_k_loop = K / KPerBlock;
if(!GridwiseGemmPipe::IsSupported(num_k_loop))
{
return false;
}
if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
{
return false;
}
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
return true;
}
__host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
{
const index_t num_loop = K / KPerBlock;
return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
}
__host__ __device__ static constexpr auto
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
{
const auto M = c_grid_desc_m_n.GetLength(I0);
const auto N = c_grid_desc_m_n.GetLength(I1);
const auto MBlock = M / MPerBlock;
const auto NBlock = N / NPerBlock;
const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
return c_grid_desc_mblock_mperblock_nblock_nperblock;
}
__host__ __device__ static constexpr auto
MakeDGridDescriptor_MBlock_MPerBlock(const DGridDesc_M& d_grid_desc_m)
{
const auto M = d_grid_desc_m.GetLength(I0);
const auto MBlock = M / MPerBlock;
const auto d_grid_desc_mblock_mperblock = transform_tensor_descriptor(
d_grid_desc_m,
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{}))),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0, 1>{}));
return d_grid_desc_mblock_mperblock;
}
// return block_id to C matrix tile idx (m0, n0) mapping
__host__ __device__ static constexpr auto
MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
{
return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
c_grid_desc_m_n);
}
using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
using DGridDescriptor_MBlock_MPerBlock =
remove_cvref_t<decltype(MakeDGridDescriptor_MBlock_MPerBlock(DGridDesc_M{}))>;
using DefaultBlock2CTileMap =
remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
template <bool HasMainKBlockLoop, typename Block2CTileMap>
__device__ static void Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
DPtrsGlobal p_ds_grid,
void* __restrict__ p_shared,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CElementwiseOperation& c_element_op,
const DxsInElementwiseOperation& dxs_in_element_op,
const DxsAccElementwiseOperation& dxs_out_element_op,
const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock,
const DGridDescriptor_MBlock_MPerBlock& d_grid_desc_mblock_mperblock,
const Block2CTileMap& block_2_ctile_map)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
// divide block work by [M, N]
const auto block_work_idx =
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_ctile_map.ValidCTileIndex(
block_work_idx,
make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1, BK1);
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<AK0, MPerBlock, AK1>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
decltype(a_grid_desc_ak0_m_ak1),
decltype(a_block_desc_ak0_m_ak1),
ABlockTransferSrcAccessOrder,
Sequence<1, 0, 2>,
ABlockTransferSrcVectorDim,
2,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true,
NumGemmKPrefetchStage>(
a_grid_desc_ak0_m_ak1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_ak0_m_ak1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<BK0, NPerBlock, BK1>,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
decltype(b_grid_desc_bk0_n_bk1),
decltype(b_block_desc_bk0_n_bk1),
BBlockTransferSrcAccessOrder,
Sequence<1, 0, 2>,
BBlockTransferSrcVectorDim,
2,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true,
NumGemmKPrefetchStage>(
b_grid_desc_bk0_n_bk1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_bk0_n_bk1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx[K0PerBlock, MPerBlock] is in LDS
// b_mtx[K0PerBlock, NPerBlock] is in LDS
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
// register
// sanity check
constexpr index_t KPack = math::max(
math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
BlockSize,
FloatAB,
FloatGemmAcc,
decltype(a_block_desc_ak0_m_ak1),
decltype(b_block_desc_bk0_n_bk1),
MPerXdl,
NPerXdl,
MXdlPerWave,
NXdlPerWave,
KPack,
LoopSched>();
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
// gridwise GEMM pipeline
const auto gridwise_gemm_pipeline =
GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
a_block_desc_ak0_m_ak1,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b_grid_desc_bk0_n_bk1,
b_block_desc_bk0_n_bk1,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
blockwise_gemm,
c_thread_buf,
num_k_block_main_loop);
// shuffle C + reduction + write out
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatCShuffle*>(p_shared),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
FloatCShuffle,
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
M2,
I1,
M4,
I1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3],
m_thread_data_on_block_idx[I4],
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
// shuffle: blockwise copy C from LDS to global
auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
ThisThreadBlock, // ThreadGroup
CElementwiseOperation, // ElementwiseOperation,
CGlobalMemoryDataOperation, // DstInMemOp,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
FloatCShuffle, // typename SrcData,
FloatC, // typename DstData,
decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
true, // bool ThreadTransferSrcResetCoordinateAfterRun,
false> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
c_element_op};
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
1,
1,
M2,
1,
M4,
1>>{};
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_c_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
// TODO: this should be implemented as a blockwise reduction
// LDS c_reduce_block_desc_mperblock_nperblock
constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_pass_through_transform(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
make_freeze_transform(I0),
make_pass_through_transform(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
static_assert(CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) *
CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
BlockSize,
"wrong!");
static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) ==
0 &&
(CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
0,
"wrong!");
constexpr index_t mreduce_per_thread =
(CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0);
constexpr index_t nreduce_per_thread =
(CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1);
constexpr auto c_reduce_thread_lengths_mperblock_nperblock =
Sequence<mreduce_per_thread, nreduce_per_thread>{};
// VGPR c_reduce_thread_desc_mperblock_nperblock
constexpr auto c_reduce_thread_desc_mperblock_nperblock =
make_naive_tensor_descriptor_packed(
make_tuple(Number<mreduce_per_thread>{}, Number<nreduce_per_thread>{}));
// VGPR d_reduce_thread_desc_mperblock
constexpr auto d_reduce_thread_desc_mperblock =
make_naive_tensor_descriptor_packed(make_tuple(Number<mreduce_per_thread>{}));
// VGPR d_reduce_thread_desc_mblock_mperblock
constexpr auto d_reduce_thread_desc_mblock_mperblock =
make_naive_tensor_descriptor_packed(make_tuple(I1, Number<mreduce_per_thread>{}));
auto c_reduce_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
c_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
// reduce: threadwise copy from LDS to VGPR
constexpr auto c_reduce_thread_cluster_desc = make_cluster_descriptor(
CReduceThreadClusterLengths_MPerBlock_NPerBlock{}, Sequence<1, 0>{});
const auto c_reduce_thread_cluster_idx =
c_reduce_thread_cluster_desc.CalculateBottomIndex(
make_multi_index(get_thread_local_1d_id()));
const auto c_reduce_thread_data_idx_begin =
c_reduce_thread_cluster_idx * c_reduce_thread_lengths_mperblock_nperblock;
auto c_reduce_thread_copy_lds_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
FloatCShuffle,
FloatReduceAcc,
decltype(c_reduce_block_desc_mperblock_nperblock),
decltype(c_reduce_thread_desc_mperblock_nperblock),
decltype(c_reduce_thread_lengths_mperblock_nperblock),
Sequence<0, 1>,
1,
CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
1,
true>{c_reduce_block_desc_mperblock_nperblock, c_reduce_thread_data_idx_begin};
auto dxs_reduce_thread_copy_vgpr_to_global = generate_tuple(
[&](auto I) {
auto p_d_grid = p_ds_grid[I];
auto d_out_element_op = dxs_out_element_op[I];
return ThreadwiseTensorSliceTransfer_v1r3<
FloatReduceAcc,
remove_pointer_t<decltype(p_d_grid)>,
decltype(d_reduce_thread_desc_mblock_mperblock),
decltype(d_grid_desc_mblock_mperblock),
decltype(d_out_element_op),
Sequence<1, mreduce_per_thread>,
Sequence<0, 1>,
1,
CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
DGlobalMemoryDataOperation::At(I),
1,
false>{d_grid_desc_mblock_mperblock,
make_multi_index(block_work_idx[I0], // mblock
c_reduce_thread_data_idx_begin[I0]), // mperblock
d_out_element_op};
},
Number<p_ds_grid.Size()>{});
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
c_shuffle_block_copy_lds_to_global.Run(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
c_shuffle_block_buf,
c_grid_desc_mblock_mperblock_nblock_nperblock,
c_grid_buf);
// TODO - extract following into reduction_blockwise
{
c_reduce_thread_copy_lds_to_vgpr.Run(c_reduce_block_desc_mperblock_nperblock,
c_shuffle_block_buf,
c_reduce_thread_desc_mperblock_nperblock,
make_tuple(I0, I0),
c_reduce_thread_buf);
static_for<0, p_ds_grid.Size(), 1>{}([&](auto In) {
auto& p_d_grid = p_ds_grid[In];
auto d_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_d_grid, d_grid_desc_mblock_mperblock.GetElementSpaceSize());
auto d_thread_buf =
make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
d_reduce_thread_desc_mperblock.GetElementSpaceSize());
auto& d_in_element_op = dxs_in_element_op[In];
auto& d_reduce_thread_copy_vgpr_to_global =
dxs_reduce_thread_copy_vgpr_to_global(In);
using DReduceOperation = remove_cvref_t<decltype(DxsReduceOperation{}[In])>;
using ThreadwiseReduce =
ThreadwiseReduction<FloatReduceAcc,
decltype(c_reduce_thread_desc_mperblock_nperblock),
decltype(d_reduce_thread_desc_mperblock),
DReduceOperation,
false>;
// Global write Gemm shuffle + reduction
const auto d_zeroVal = DReduceOperation::GetReductionZeroVal();
static_for<0, mreduce_per_thread, 1>{}(
[&](auto I) { d_thread_buf(I) = d_zeroVal; });
// reduce in VGPR
static_for<0, mreduce_per_thread, 1>{}([&](auto im) {
static_for<0, nreduce_per_thread, 1>{}([&](auto in) {
constexpr auto offset =
Number<c_reduce_thread_desc_mperblock_nperblock.CalculateOffset(
make_tuple(im, in))>{};
d_in_element_op(c_reduce_thread_buf(offset),
c_reduce_thread_buf(offset));
});
});
ThreadwiseReduce::Reduce(c_reduce_thread_buf, d_thread_buf);
// copy from VGPR to Global
d_reduce_thread_copy_vgpr_to_global.Run(
d_reduce_thread_desc_mblock_mperblock,
make_tuple(I0, I0),
d_thread_buf,
d_grid_desc_mblock_mperblock,
d_grid_buf);
if constexpr(access_id < num_access - 1)
{
constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
d_reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow(
d_grid_desc_mblock_mperblock,
make_tuple(c_global_step[I0], c_global_step[I1]));
}
});
}
if constexpr(access_id < num_access - 1)
{
constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
// move on C
c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
}
});
// Reduction
}
}
};
} // namespace ck
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment