Skip to content

GitLab

Commit ccb94cea authored by aska-0096's avatar aska-0096
Browse files

navi3x_multipleD+example

parent 8efd363f
Hide whitespace changes
Inline Side-by-side
Showing with 1701 additions and 0 deletions
example/02_gemm_bilinear/CMakeLists.txt
View file @ ccb94cea
add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp)
example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp 0 → 100644
View file @ ccb94cea
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
struct AlphaBetaAdd
{
AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
template <typename E, typename C, typename D>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
template <>
__host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
ck::half_t& e, const float& c, const ck::half_t& d) const
{
e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
};
float alpha_;
float beta_;
};
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DLayout = Row;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AlphaBetaAdd;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceOpInstance =
ck::tensor_operation::device::DeviceGemmMultipleD_Wmma_CShuffle<ALayout,
BLayout,
ck::Tuple<DLayout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<DDataType>,
EDataType,
AccDataType,
CShuffleDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmSpec,
256,
128,
256,
8,
8,
16,
16,
4,
4,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
1,
1,
S<1, 32, 1, 8>,
8>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideD = 4096;
ck::index_t StrideE = 4096;
float alpha = 1.0f;
float beta = 1.0f;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
alpha = std::stof(argv[4]);
beta = std::stof(argv[5]);
}
else if(argc == 13)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideD = std::stoi(argv[9]);
StrideE = std::stoi(argv[10]);
alpha = std::stof(argv[11]);
beta = std::stof(argv[12]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, alpha, "
"beta\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d_device_buf.ToDevice(d_m_n.mData.data());
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{alpha, beta};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_m_n({M, N});
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CShuffleDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
}
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp 0 → 100644
View file @ ccb94cea
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, 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_d.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_d_wmma_cshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename AccDataType,
typename CShuffleDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
GemmSpecialization GemmSpec,
ck::index_t BlockSize,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t K0PerBlock,
ck::index_t K1,
ck::index_t MPerWMMA,
ck::index_t NPerWMMA,
ck::index_t MRepeat,
ck::index_t NRepeat,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
ck::index_t ABlockTransferSrcVectorDim,
ck::index_t ABlockTransferSrcScalarPerVector,
ck::index_t ABlockTransferDstScalarPerVector_K1,
bool ABlockLdsAddExtraM,
typename BBlockTransferThreadClusterLengths_K0_N_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
ck::index_t BBlockTransferSrcVectorDim,
ck::index_t BBlockTransferSrcScalarPerVector,
ck::index_t BBlockTransferDstScalarPerVector_K1,
bool BBlockLdsAddExtraN,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
ck::index_t NumPrefetch = 1,
ck::LoopScheduler LoopSched = make_default_loop_scheduler(),
ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation>
{
using DeviceOp = DeviceGemmMultipleD_Wmma_CShuffle;
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>{};
// K1 = Max Vector Access Pixels
static constexpr auto K1Number = Number<K1>{};
static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
{
assert(K % K1 == 0);
const index_t K0 = K / K1;
const auto a_grid_desc_m_k = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
}
#ifdef ENABLE_COLMAJOR
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
#endif
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_right_pad_transform(M, PadM)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
}
static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
{
assert(K % K1 == 0);
const index_t K0 = K / K1;
const auto b_grid_desc_k_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
}
template <typename ELayout_>
static auto MakeEGridDescriptor_M_N(index_t M, index_t N, index_t StrideE)
{
const auto e_grid_desc_m_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout_>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideE, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout_>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideE));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
e_grid_desc_m_n,
make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
e_grid_desc_m_n,
make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
}
static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& Ms,
const std::array<index_t, NumDTensor>& Ns,
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>(Ms[i], Ns[i], DsStride[i]);
},
Number<NumDTensor>{});
}
// Gridwise descriptor, mapping to whole given provblem.
using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
// GridwiseOp
using GridwiseOp = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle<
// DataType Family
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
// InMemory Data Descriptor
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
DsGridDesc_M_N,
EGridDesc_M_N,
// ElementwiseOp Family
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
InMemoryDataOperationEnum::Set,
// Tiling Family
MPerBlock,
NPerBlock,
K0PerBlock,
MPerWMMA,
NPerWMMA,
K1,
MRepeat,
NRepeat,
// ThreadCluster Family
BlockSize,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
false, // AThreadTransferSrcResetCoordinateAfterRun,
ABlockLdsAddExtraM,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
false, // BThreadTransferSrcResetCoordinateAfterRun,
BBlockLdsAddExtraN,
CShuffleMRepeatPerShuffle,
CShuffleNRepeatPerShuffle,
CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CDEShuffleBlockTransferScalarPerVector_NPerBlock,
NumPrefetch,
LoopSched,
PipelineVer>;
// Argument
struct Argument : public BaseArgument
{
Argument(const void* p_a_grid,
const void* p_b_grid,
std::array<const void*, NumDTensor> p_ds_grid,
void* p_e_grid,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
std::array<index_t, NumDTensor> StrideDs,
index_t StrideE,
index_t M01,
index_t N01,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
p_ds_grid_{},
p_e_grid_{static_cast<EDataType*>(p_e_grid)},
a_grid_desc_k0_m_k1_{},
b_grid_desc_k0_n_k1_{},
ds_grid_desc_m_n_{},
e_grid_desc_m_n_{},
ds_grid_desc_mblock_mperblock_nblock_nperblock{},
e_grid_desc_mblock_mperblock_nblock_nperblock{},
block_2_ctile_map_{},
M01_{M01},
N01_{N01},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
{
a_grid_desc_k0_m_k1_ = DeviceOp::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
b_grid_desc_k0_n_k1_ = DeviceOp::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
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) =
DeviceOp::MakeEGridDescriptor_M_N<DLayout>(M, N, StrideDs[i]);
});
e_grid_desc_m_n_ = DeviceOp::MakeEGridDescriptor_M_N<ELayout>(M, N, StrideE);
block_2_ctile_map_ = GridwiseOp::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_, M01, N01);
if(GridwiseOp::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
ds_grid_desc_m_n_,
e_grid_desc_m_n_,
block_2_ctile_map_))
{
ds_grid_desc_mblock_mperblock_nblock_nperblock =
GridwiseOp::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
ds_grid_desc_m_n_);
e_grid_desc_mblock_mperblock_nblock_nperblock =
GridwiseOp::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
e_grid_desc_m_n_);
}
}
// Pointers
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
typename GridwiseOp::DsGridPointer p_ds_grid_;
EDataType* p_e_grid_;
// Tensor Descriptors
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
DsGridDesc_M_N ds_grid_desc_m_n_;
EGridDesc_M_N e_grid_desc_m_n_;
typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock;
typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock;
// Block to Tile mapping
typename GridwiseOp::DefaultBlock2CTileMap block_2_ctile_map_;
// Idle
index_t M01_;
index_t N01_;
// ElementwiseOp
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_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_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
<< ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
<< arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
<< ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
<< arg.b_grid_desc_k0_n_k1_.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) << ", "
<< arg.c_grid_desc_m_n_.GetLength(I2) << "}" << std::endl;
}
#endif
if(!GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.ds_grid_desc_m_n_,
arg.e_grid_desc_m_n_,
arg.block_2_ctile_map_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_k0mk1_k0nk1_m0nm1_wmma_v1r1 has invalid setting");
}
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(GridwiseOp::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_mupltipe_d_wmma_cshuffle<
GridwiseOp,
ADataType,
BDataType,
typename GridwiseOp::DsGridPointer,
EDataType,
remove_reference_t<typename DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<typename DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
remove_reference_t<
typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
remove_reference_t<typename GridwiseOp::DefaultBlock2CTileMap>,
true>; // Last Option is W/O
ave_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_ds_grid_,
arg.p_e_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock,
arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
arg.a_element_op_,
arg.b_element_op_,
arg.cde_element_op_,
arg.block_2_ctile_map_);
}
else
{
const auto kernel = kernel_gemm_mupltipe_d_wmma_cshuffle<
GridwiseOp,
ADataType,
BDataType,
typename GridwiseOp::DsGridPointer,
EDataType,
remove_reference_t<typename DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<typename DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
remove_reference_t<
typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
remove_reference_t<typename GridwiseOp::DefaultBlock2CTileMap>,
false>;
ave_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_ds_grid_,
arg.p_e_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock,
arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
arg.a_element_op_,
arg.b_element_op_,
arg.cde_element_op_,
arg.block_2_ctile_map_);
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::get_device_name() == "gfx1100")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{
return false;
}
}
else
{
return false;
}
return GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.ds_grid_desc_m_n_,
arg.e_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 void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
std::array<ck::index_t, NumDTensor> StrideDs,
index_t StrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{p_a,
p_b,
p_ds,
p_e,
M,
N,
K,
StrideA,
StrideB,
StrideDs,
StrideE,
1,
1,
a_element_op,
b_element_op,
cde_element_op};
}
// polymorphic
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
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_a,
p_b,
p_ds,
p_e,
M,
N,
K,
StrideA,
StrideB,
StrideDs,
StrideE,
1,
1,
a_element_op,
b_element_op,
cde_element_op);
}
static auto MakeInvoker() { return Invoker{}; }
// 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 << "DeviceGemmMultipleD_Wmma_CShuffle"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< K1 << ", "
<< MPerWMMA << ", "
<< NPerWMMA << ", "
<< MRepeat << ", "
<< NRepeat
<< ">"
<< " NumPrefetch: "
<< NumPrefetch << ", "
<< "LoopScheduler: "
<< LoopSchedToString[LoopSched] << ", "
<< "PipelineVersion: "
<< PipelineVersionToString[PipelineVer];
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp 0 → 100644
View file @ ccb94cea
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/multi_index_transform_helper.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
template <typename GridwiseOp,
typename ADataType,
typename BDataType,
typename DsPointer,
typename EDataType,
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
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_mupltipe_d_wmma_cshuffle(
const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
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 AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
__shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid,
p_b_grid,
p_ds_grid,
p_e_grid,
p_shared,
a_grid_desc_k0_m_k1,
b_grid_desc_k0_n_k1,
ds_grid_desc_mblock_mperblock_nblock_nperblock,
e_grid_desc_mblock_mperblock_nblock_nperblock,
a_element_op,
b_element_op,
cde_element_op,
block_2_ctile_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_ds_grid;
ignore = p_e_grid;
ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_n_k1;
ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = a_element_op;
ignore = b_element_op;
ignore = cde_element_op;
ignore = block_2_ctile_map;
#endif // end of if (defined(__gfx1100__))
}
template < // DataType Family
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
// InMemory Data Descriptor
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename DsGridDesc_M_N,
typename EGridDesc_M_N,
// ElementwiseOp Family
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
InMemoryDataOperationEnum EGlobalMemoryDataOperation,
// Tiling Family
index_t MPerBlock,
index_t NPerBlock,
index_t K0PerBlock,
index_t MPerWmma,
index_t NPerWmma,
index_t K1Value,
index_t MRepeat,
index_t NRepeat,
// ThreadCluster Family
index_t BlockSize,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_K1,
bool AThreadTransferSrcResetCoordinateAfterRun,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_K0_N_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_K1,
bool BThreadTransferSrcResetCoordinateAfterRun,
bool BBlockLdsExtraN,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
index_t NumGemmKPrefetchStage = 1,
LoopScheduler LoopSched = make_default_loop_scheduler(),
PipelineVersion PipelineVer = PipelineVersion::v1>
struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
{
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>{};
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 K1 = Number<K1Value>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
using GridwiseGemmPipe = remove_cvref_t<decltype(
GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
__host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
{
constexpr auto max_lds_align = K1;
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_k0perblock_mperblock_k1 = [&]() {
if constexpr(ABlockLdsExtraM)
{
return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
}
}();
return a_block_desc_k0perblock_mperblock_k1;
}
__host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
{
constexpr auto max_lds_align = K1;
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_k0perblock_nperblock_k1 = [&]() {
if constexpr(BBlockLdsExtraN)
{
return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
}
}();
return b_block_desc_k0perblock_nperblock_k1;
}
__host__ __device__ static constexpr auto
// *Caution Here repeat is shuffle repeat
GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat()
{
constexpr index_t MWave = MPerBlock / (MRepeat * MPerWmma);
constexpr index_t NWave = NPerBlock / (NRepeat * NPerWmma);
constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<CShuffleMRepeatPerShuffle * MWave * MPerWmma>{},
I1,
Number<CShuffleNRepeatPerShuffle * NWave * NPerWmma>{}));
return c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat;
}
// ck::Tuple<const D0DataType*, const D1DataType*, ...>
static constexpr auto MakeDsGridPointer()
{
return generate_tuple(
[&](auto i) {
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
return static_cast<const DDataType*>(nullptr);
},
Number<NumDTensor>{});
}
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_desc_k0perblock_mperblock_k1 =
GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
constexpr auto b_block_desc_k0perblock_nperblock_k1 =
GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
constexpr auto max_lds_align = K1;
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize(), max_lds_align);
return (a_block_space_size_aligned * sizeof(ADataType) +
b_block_space_size_aligned * sizeof(BDataType));
}
// 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_K0_M_K1& a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
const DsGridDesc_M_N& ds_grid_desc_m_n,
const EGridDesc_M_N& e_grid_desc_m_n,
const Block2CTileMap& block_2_ctile_map)
{
static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
"wrong! K1 need to be known at compile-time");
static_assert((MPerBlock % (MPerWmma * MRepeat) == 0) &&
(NPerBlock % (NRepeat * NPerWmma)) == 0,
"Invalid tuning param!");
const auto M = a_grid_desc_k0_m_k1.GetLength(I1);
const auto N = b_grid_desc_k0_n_k1.GetLength(I1);
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
bool valid = true;
static_for<0, NumDTensor, 1>{}([&](auto i) {
valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
N == ds_grid_desc_m_n[i].GetLength(I1));
});
if(!valid)
{
return false;
}
if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1) &&
K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
return false;
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
return false;
// check gridwise gemm pipeline
const auto num_k_loop = K0 / K0PerBlock;
if(!GridwiseGemmPipe::IsSupported(num_k_loop))
{
return false;
}
if(!block_2_ctile_map.CheckValidity(e_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 / (K0PerBlock * K1);
return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
}
// E desc for destination in blockwise copy
__host__ __device__ static constexpr auto
MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
{
const auto M = e_grid_desc_m_n.GetLength(I0);
const auto N = e_grid_desc_m_n.GetLength(I1);
const auto MBlock = M / MPerBlock;
const auto NBlock = N / NPerBlock;
const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
e_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 e_grid_desc_mblock_mperblock_nblock_nperblock;
}
// Ds desc for source in blockwise copy
__host__ __device__ static constexpr auto
MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N& ds_grid_desc_m_n)
{
return generate_tuple(
[&](auto i) {
return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
},
Number<NumDTensor>{});
}
// return block_id to C matrix tile idx (m0, n0) mapping
__host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
const EGridDesc_M_N& e_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
{
return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
e_grid_desc_m_n);
}
using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
using DefaultBlock2CTileMap =
remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(EGridDesc_M_N{}, 1, 1))>;
using DsGridPointer = decltype(MakeDsGridPointer());
template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
__device__ static void Run(const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsGridPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
void* __restrict__ p_shared,
const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
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 AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op,
const Block2CTileMap& block_2_ctile_map)
{
// clang-format off
/*******************************************************************************/
// Memory buffer zone.
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
const auto ds_grid_buf = generate_tuple(
[&](auto i) {
return make_dynamic_buffer<AddressSpaceEnum::Global>(
p_ds_grid[i],
ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
},
Number<NumDTensor>{});
auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
/*******************************************************************************/
// BlockIdx.x -> [BlockId.m, BlockId.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(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{ return; }
// Store BlockId 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);
/*******************************************************************************/
// BlockLevel, A/B Matrix ThreadMapping in LDS, As Destinaion of BlockWise_Copy
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
constexpr auto max_lds_align = K1;
constexpr auto a_block_desc_k0perblock_mperblock_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
constexpr auto b_block_desc_k0perblock_nperblock_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1< ThisThreadBlock,
/* typename SrcElementwiseOperation, */ AElementwiseOperation,
/* typename DstElementwiseOperation, */ ck::tensor_operation::element_wise::PassThrough,
/* InMemoryDataOperationEnum DstInMemOp, */ InMemoryDataOperationEnum::Set,
/* typename BlockSliceLengths, */ Sequence<K0PerBlock, MPerBlock, K1>,
/* typename ThreadClusterLengths, */ ABlockTransferThreadClusterLengths_K0_M_K1,
/* typename ThreadClusterArrangeOrder, */ ABlockTransferThreadClusterArrangeOrder,
/* typename SrcData, */ ADataType,
/* typename DstData, */ ADataType,
/* typename SrcDesc, */ decltype(a_grid_desc_k0_m_k1),
/* typename DstDesc, */ decltype(a_block_desc_k0perblock_mperblock_k1),
/* typename SrcDimAccessOrder, */ ABlockTransferSrcAccessOrder,
/* typename DstDimAccessOrder, */ Sequence<0, 1, 2>,
/* index_t SrcVectorDim, */ ABlockTransferSrcVectorDim,
/* index_t DstVectorDim, */ 2,
/* index_t SrcScalarPerVector, */ ABlockTransferSrcScalarPerVector,
/* index_t DstScalarPerVector, */ ABlockTransferDstScalarPerVector_K1,
/* index_t SrcScalarStrideInVector, */ 1,
/* index_t DstScalarStrideInVector, */ 1,
/* bool ThreadTransferSrcResetCoordinateAfterRun, */ AThreadTransferSrcResetCoordinateAfterRun,
/* bool ThreadTransferDstResetCoordinateAfterRun, */ true>(
a_grid_desc_k0_m_k1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_k0perblock_mperblock_k1,
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<K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BDataType,
BDataType,
decltype(b_grid_desc_k0_n_k1),
decltype(b_block_desc_k0perblock_nperblock_k1),
BBlockTransferSrcAccessOrder,
Sequence<0, 1, 2>,
BBlockTransferSrcVectorDim,
2,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true>(
b_grid_desc_k0_n_k1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_k0perblock_nperblock_k1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
/*******************************************************************************/
// GEMM
constexpr auto WmmaK = 16;
constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
auto blockwise_gemm =
BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO<BlockSize,
ADataType,
BDataType,
AccDataType,
decltype(a_block_desc_k0perblock_mperblock_k1),
decltype(b_block_desc_k0perblock_nperblock_k1),
MPerWmma,
NPerWmma,
MRepeat,
NRepeat,
KPack>{};
// Prepare Register for C matrix
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
/*******************************************************************************/
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize(), max_lds_align);
// LDS allocation for A and B: be careful of alignment
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<ADataType*>(p_shared), a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<BDataType*>(p_shared) + a_block_space_size_aligned, b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize());
// Shift Per SUB_K
constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
// gridwise GEMM pipeline
const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
a_block_desc_k0perblock_mperblock_k1,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b_grid_desc_k0_n_k1,
b_block_desc_k0perblock_nperblock_k1,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
blockwise_gemm,
c_thread_buf,
K0BlockMainLoop);
/*******************************************************************************/
// write out to C, implement shuffle
{
constexpr auto c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs =
blockwise_gemm.GetCThreadDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
// This API Provide All dimension (size) you need
constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp =
blockwise_gemm.GetCBlockDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
constexpr auto MWave = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I1);
constexpr auto MSubGroup = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I2);
constexpr auto NWave = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I4);
constexpr auto NThreadPerSubGroup = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I5);
constexpr auto MAccVgprs = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I6);
// LDS descriptor, shuffle and write out in MRepeat x NRepeat times
constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<CShuffleDataType*>(p_shared),
c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat.GetElementSpaceSize());
constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs = transform_tensor_descriptor(
c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMRepeatPerShuffle>{}, // MRepeat per shuffle repeat
MWave, // MWave
MSubGroup, // MSubGroup * MAccVgprs = MPerWmma
MAccVgprs)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNRepeatPerShuffle>{}, // NRepeat per shuffle repeat
NWave, // NWave
NThreadPerSubGroup))), // NThreadPerSubGroup = NPerWmma
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<>{}, Sequence<0, 1, 2, 6>{}, Sequence<>{}, Sequence<3, 4, 5>{}));
// 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);
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_mrepeat_mwave_msubgroup_maccvgprs_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(MRepeat, MWave, MSubGroup, MAccVgprs))),
make_tuple(Sequence<0, 1, 2, 3>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(NRepeat, NWave, NThreadPerSubGroup))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx = m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_idx = n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_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<AccDataType,
CShuffleDataType,
decltype(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
decltype(c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMRepeatPerShuffle,
I1,
I1,
CShuffleNRepeatPerShuffle,
I1,
I1,
MAccVgprs>,
Sequence<0, 1, 2, 3, 4, 5, 6>,
6,
1, // vector write pixel
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
make_multi_index(0,
m_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
0,
n_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3]),
ck::tensor_operation::element_wise::PassThrough{}};
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat),
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor buffers
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin = container_concat(
make_tuple(make_multi_index(0, 0, 0, 0)),
generate_tuple(
[&](auto) {
return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
},
Number<NumDTensor>{}));
// shuffle: blockwise copy C from LDS to global
auto cde_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v7<
ThisThreadBlock, // ThreadGroup
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CDEElementwiseOperation, // ElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // DstInMemOp,
Sequence<1,
CShuffleMRepeatPerShuffle * MWave * MPerWmma,
1,
CShuffleNRepeatPerShuffle * NWave * NPerWmma>, // BlockSliceLengths,
CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CDEShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // bool ThreadTransferSrcResetCoordinateAfterRun,
Sequence<false>> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
cde_element_op};
// space filling curve for local reg & global memory
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MRepeat, 1, 1, NRepeat, 1, 1, MAccVgprs>,
Sequence<0, 1, 2, 3, 4, 5, 6>,
Sequence<CShuffleMRepeatPerShuffle,
1,
1,
CShuffleNRepeatPerShuffle,
1,
1,
MAccVgprs>>{};
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_cde_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMRepeatPerShuffle * MWave * MPerWmma,
1,
CShuffleNRepeatPerShuffle * NWave * NPerWmma>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_cde_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_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
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
cde_shuffle_block_copy_lds_to_global.Run(
c_ds_desc_refs,
c_ds_buf_refs,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
tie(e_grid_buf));
if constexpr(access_id < num_access - 1)
{
constexpr auto cde_global_step = sfc_cde_global.GetForwardStep(access_id);
// move on Ds
static_for<0, NumDTensor, 1>{}([&](auto i) {
cde_shuffle_block_copy_lds_to_global.MoveSrcSliceWindow(
c_ds_desc_refs, i + I1, cde_global_step);
});
// move on E
cde_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
I0,
cde_global_step);
}
});
}
// clang-format on
}
};
} // 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