ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 02:02:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

Grouped GEMM for fp16 (#126)

Browse Source 
* init of grouped_gemm

* 2 gemm test

* perf test

* clean

* wrap desc into a struct

* test cast static_arr to pointer

* add ptr to GemmDesc

* add grouped gemm profiler

* fixed mem issue with unique_ptr

* clean

* clean

* finished ckprofiler

* Update README.md

* readme

* fixed readme

* add example

* improve code

* fixed comments: reserve, seperate ptr and gemm_shapes

* merge group and non-group

* fixed comments: replace push_back with emplace_back to avoid copy constructor

* fixed comments: unified blk2ctile; add test

* ci fix

* fixed ci

* fixed ci

* fixed ci

[ROCm/composable_kernel commit: 716f1c7fb1]
This commit is contained in:
zjing14
2022-03-22 18:18:18 -05:00
committed by GitHub
parent d8ecdd6bd3
commit 94dadbf4ed
20 changed files with 1917 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
example/15_grouped_gemm/CMakeLists.txt Normal file
View File

@@ -0,0 +1 @@
add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)

58
example/15_grouped_gemm/README.md Normal file
View File

@@ -0,0 +1,58 @@
# Instructions for ```grouped_gemm_xdl``` Example
## Docker script
```bash
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
```
## Build ```grouped_gemm_xdl```
```bash
mkdir build && cd build
```
```bash
# Need to specify target ID, example below is gfx908
cmake \
-D BUILD_DEV=OFF \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
..
```
```bash
make -j example_grouped_gemm_xdl_fp16
```
## Run ```grouped_gemm_xdl```
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1)
./bin/example_grouped_gemm_xdl_fp16 0 1 5
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
gemm[0] a_m_k: dim 2, lengths {256, 64}, strides {64, 1} b_k_n: dim 2, lengths {64, 128}, strides {1, 64} c_m_n: dim 2, lengths {256, 128}, strides {128, 1}
gemm[1] a_m_k: dim 2, lengths {512, 128}, strides {128, 1} b_k_n: dim 2, lengths {128, 256}, strides {1, 128} c_m_n: dim 2, lengths {512, 256}, strides {256, 1}
gemm[2] a_m_k: dim 2, lengths {768, 192}, strides {192, 1} b_k_n: dim 2, lengths {192, 384}, strides {1, 192} c_m_n: dim 2, lengths {768, 384}, strides {384, 1}
gemm[3] a_m_k: dim 2, lengths {1024, 256}, strides {256, 1} b_k_n: dim 2, lengths {256, 512}, strides {1, 256} c_m_n: dim 2, lengths {1024, 512}, strides {512, 1}
group: 0 arg.a_grid_desc_k0_m_k1_{8, 256, 8}, arg.b_grid_desc_k0_n_k1_{8, 128, 8}, arg.c_grid_desc_m_n_{ 256, 128}
group: 1 arg.a_grid_desc_k0_m_k1_{16, 512, 8}, arg.b_grid_desc_k0_n_k1_{16, 256, 8}, arg.c_grid_desc_m_n_{ 512, 256}
group: 2 arg.a_grid_desc_k0_m_k1_{24, 768, 8}, arg.b_grid_desc_k0_n_k1_{24, 384, 8}, arg.c_grid_desc_m_n_{ 768, 384}
group: 3 arg.a_grid_desc_k0_m_k1_{32, 1024, 8}, arg.b_grid_desc_k0_n_k1_{32, 512, 8}, arg.c_grid_desc_m_n_{ 1024, 512}
launch_and_time_kernel: grid_dim {30, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 0.037887 ms, 11.0706 TFlops, 90.8132 GB/s, DeviceGroupedGemmXdl<256, 256, 128, 4, 8, 32, 32, 4, 2>
```

234
example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp Normal file
View File

@@ -0,0 +1,234 @@
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_grouped_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
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 = ck::half_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using AccDataType = float;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
// static constexpr auto GemmMNPadding =
// ck::tensor_operation::device::GemmSpecialization_t::MNPadding;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmXdl
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| Num|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| |
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, 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, 7, 1, 1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
exit(0);
}
int group_count = 4;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmShape> gemm_shapes;
std::vector<const void*> p_a, p_b;
std::vector<void*> p_c;
gemm_shapes.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
int M = 256 + 256 * i;
int N = 128 + 128 * i;
int K = 64 + 64 * i;
gemm_shapes.push_back({M, N, K, K, K, N});
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
std::vector<Tensor<ADataType>> a_tensors;
;
std::vector<Tensor<BDataType>> b_tensors;
std::vector<Tensor<CDataType>> c_host_tensors;
std::vector<Tensor<CDataType>> c_device_tensors;
a_tensors.reserve(group_count);
b_tensors.reserve(group_count);
c_host_tensors.reserve(group_count);
c_device_tensors.reserve(group_count);
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
a_tensors_device.reserve(group_count);
b_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < gemm_shapes.size(); i++)
{
a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{})));
b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
gemm_shapes[i].K, gemm_shapes[i].N, gemm_shapes[i].StrideB, BLayout{})));
c_host_tensors.push_back(Tensor<CDataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
c_device_tensors.push_back(Tensor<CDataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
<< " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
<< std::endl;
flop += std::size_t(2) * gemm_shapes[i].M * gemm_shapes[i].K * gemm_shapes[i].N;
num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
sizeof(CDataType) * c_device_tensors[i].mDesc.GetElementSize();
switch(init_method)
{
case 0: break;
case 1:
a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
case 2:
a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
}
}
for(int i = 0; i < gemm_shapes.size(); i++)
{
a_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize()));
b_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize()));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_device_tensors[i].mDesc.GetElementSize()));
a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
p_b.push_back(b_tensors_device[i]->GetDeviceBuffer());
p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument =
gemm.MakeArgument(p_a, p_b, p_c, gemm_shapes, a_element_op, b_element_op, c_element_op);
if(!gemm.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, nrepeat);
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, "
<< gemm.GetTypeString() << std::endl;
if(do_verification)
{
for(int i = 0; i < gemm_shapes.size(); i++)
{
c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
b_tensors[i],
c_host_tensors[i],
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
check_error(c_host_tensors[i], c_device_tensors[i]);
}
}
return 0;
}

1
example/CMakeLists.txt
View File

@@ -39,3 +39,4 @@ add_subdirectory(11_conv2d_bwd_wgt)
add_subdirectory(12_reduce)
add_subdirectory(13_pool2d_fwd)
add_subdirectory(14_gemm_xdl_requant_relu_requant)
add_subdirectory(15_grouped_gemm)

29
include/ck/tensor_operation/gpu/device/device_gemm.hpp
View File

@@ -8,6 +8,12 @@ namespace ck {
namespace tensor_operation {
namespace device {
struct GemmShape
{
ck::index_t M, N, K;
ck::index_t StrideA, StrideB, StrideC;
};
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
@@ -65,6 +71,29 @@ template <typename AElementwiseOperation,
using DeviceGemmPtr = std::unique_ptr<
DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct DeviceGroupedGemm : public BaseOperator
{
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(std::vector<const void*>& p_a,
std::vector<const void*>& p_b,
std::vector<void*>& p_c,
std::vector<GemmShape>& gemm_shapes,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
ck::index_t KBatch = 1) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
using DeviceGroupedGemmPtr = std::unique_ptr<
DeviceGroupedGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck

562
include/ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp Normal file
View File

@@ -0,0 +1,562 @@
#ifndef DEVICE_GROUPED_GEMM_XDL_HPP
#define DEVICE_GROUPED_GEMM_XDL_HPP
#include <iostream>
#include <sstream>
#include "device.hpp"
#include "device_base.hpp"
#include "device_gemm.hpp"
#include "common_header.hpp"
#include "tensor_layout.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_xdlops_v2r3.hpp"
#include "gemm_specialization.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
GemmSpecialization_t GemmSpecialization,
ck::index_t BlockSize,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t K0PerBlock,
ck::index_t K1,
ck::index_t MPerXDL,
ck::index_t NPerXDL,
ck::index_t MXdlPerWave,
ck::index_t NXdlPerWave,
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,
ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector,
ck::index_t NumPrefetch = 1,
ck::index_t MaxGroupCount = 10>
struct DeviceGroupedGemmXdl
: public DeviceGroupedGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
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));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
}();
if constexpr(GemmSpecialization == GemmSpecialization_t::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(GemmSpecialization == GemmSpecialization_t::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>{}));
}
}
static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
{
const auto c_grid_desc_m_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
}
}();
if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
c_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(
c_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>{}));
}
}
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 CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
BlockSize,
ADataType, // TODO: distinguish A/B datatype
AccDataType,
CDataType,
InMemoryDataOperationEnum_t::Set,
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
CGridDesc_M_N,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
MPerBlock,
NPerBlock,
K0PerBlock,
MPerXDL,
NPerXDL,
K1,
MXdlPerWave,
NXdlPerWave,
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,
Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
NumPrefetch>;
struct GroupedGemmBlock2CTileMap
{
GroupedGemmBlock2CTileMap()
{
block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1);
BlockStart_ = -1;
}
GroupedGemmBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n,
index_t M01,
index_t N01,
ck::index_t BlockStart)
{
block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n, M01, N01);
BlockStart_ = BlockStart;
}
template <typename TopIdx>
__host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
{
return block_2_ctile_map_.CalculateBottomIndex(
make_multi_index(idx_top[I0] - BlockStart_));
}
private:
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
ck::index_t BlockStart_;
};
struct GemmDescKernelArg
{
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
GroupedGemmBlock2CTileMap grouped_gemm_block_2_ctile_map_;
const ADataType* a_ptr;
const BDataType* b_ptr;
CDataType* c_ptr;
ck::index_t BlockStart_, BlockEnd_;
};
// Argument
struct Argument : public BaseArgument
{
Argument(std::vector<const void*>& p_a,
std::vector<const void*>& p_b,
std::vector<void*>& p_c,
std::vector<GemmShape>& gemm_shapes,
index_t M01,
index_t N01,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
: M01_{M01},
N01_{N01},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
c_element_op_{c_element_op}
{
grid_size_ = 0;
group_count_ = static_cast<int>(gemm_shapes.size());
if(!(group_count_ == p_a.size() && group_count_ == p_b.size() &&
group_count_ == p_c.size()))
{
throw std::runtime_error("wrong! group_count_ != P_a/b/c.size");
}
gemm_desc_kernel_arg_.reserve(group_count_);
for(index_t i = 0; i < gemm_shapes.size(); i++)
{
const index_t M = gemm_shapes[i].M;
const index_t N = gemm_shapes[i].N;
const index_t K = gemm_shapes[i].K;
const index_t StrideA = gemm_shapes[i].StrideA;
const index_t StrideB = gemm_shapes[i].StrideB;
const index_t StrideC = gemm_shapes[i].StrideC;
const auto a_grid_desc_k0_m_k1_ =
DeviceGroupedGemmXdl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
const auto b_grid_desc_k0_n_k1_ =
DeviceGroupedGemmXdl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
const auto c_grid_desc_m_n_ =
DeviceGroupedGemmXdl::MakeCGridDescriptor_M_N(M, N, StrideC);
const index_t grid_size_grp = GridwiseGemm::CalculateGridSize(c_grid_desc_m_n_);
const index_t BlockStart = grid_size_;
const index_t BlockEnd = grid_size_ + grid_size_grp;
grid_size_ += grid_size_grp;
if(GridwiseGemm::CheckValidity(
a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
{
const auto c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
const auto grouped_gemm_block_2_ctile_map_ =
GroupedGemmBlock2CTileMap(c_grid_desc_m_n_, M01, N01, BlockStart);
gemm_desc_kernel_arg_.push_back(
GemmDescKernelArg{a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
c_grid_desc_m_n_,
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
grouped_gemm_block_2_ctile_map_,
static_cast<const ADataType*>(p_a[i]),
static_cast<const BDataType*>(p_b[i]),
static_cast<CDataType*>(p_c[i]),
BlockStart,
BlockEnd});
}
}
}
// private:
index_t M01_;
index_t N01_;
index_t group_count_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
std::vector<GemmDescKernelArg> gemm_desc_kernel_arg_;
index_t grid_size_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceGroupedGemmXdl::Argument;
float Run(const Argument& arg, int nrepeat = 1)
{
StaticallyIndexedArray<GemmDescKernelArg, MaxGroupCount> gemm_desc_kernel_arg_arg;
bool has_main_k0_block_loop = true;
static_for<0, MaxGroupCount, 1>{}([&](auto i) {
if(i < arg.gemm_desc_kernel_arg_.size())
{
gemm_desc_kernel_arg_arg(i) = arg.gemm_desc_kernel_arg_[i];
std::cout << "group: " << i << " arg.a_grid_desc_k0_m_k1_{"
<< gemm_desc_kernel_arg_arg[i].a_grid_desc_k0_m_k1_.GetLength(I0)
<< ", "
<< gemm_desc_kernel_arg_arg[i].a_grid_desc_k0_m_k1_.GetLength(I1)
<< ", "
<< gemm_desc_kernel_arg_arg[i].a_grid_desc_k0_m_k1_.GetLength(I2)
<< "}";
std::cout << ", arg.b_grid_desc_k0_n_k1_{"
<< gemm_desc_kernel_arg_arg[i].b_grid_desc_k0_n_k1_.GetLength(I0)
<< ", "
<< gemm_desc_kernel_arg_arg[i].b_grid_desc_k0_n_k1_.GetLength(I1)
<< ", "
<< gemm_desc_kernel_arg_arg[i].b_grid_desc_k0_n_k1_.GetLength(I2)
<< "}";
std::cout << ", arg.c_grid_desc_m_n_{ "
<< gemm_desc_kernel_arg_arg[i].c_grid_desc_m_n_.GetLength(I0) << ", "
<< gemm_desc_kernel_arg_arg[i].c_grid_desc_m_n_.GetLength(I1) << "}"
<< std::endl;
if(!GridwiseGemm::CheckValidity(
gemm_desc_kernel_arg_arg[i].a_grid_desc_k0_m_k1_,
gemm_desc_kernel_arg_arg[i].b_grid_desc_k0_n_k1_,
gemm_desc_kernel_arg_arg[i].c_grid_desc_m_n_,
arg.M01_,
arg.N01_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
}
const auto K0 = gemm_desc_kernel_arg_arg[i].a_grid_desc_k0_m_k1_.GetLength(I0);
if(GridwiseGemm::CalculateHasMainK0BlockLoop(K0) != has_main_k0_block_loop)
{
throw std::runtime_error("wrong! not all gemm has_main_k0_block_loop");
}
}
});
float ave_time = 0;
if(has_main_k0_block_loop)
{
const auto kernel =
kernel_grouped_gemm_xdlops_v2r3<GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<GemmDescKernelArg>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
true,
MaxGroupCount>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(arg.grid_size_),
dim3(BlockSize),
0,
gemm_desc_kernel_arg_arg,
arg.gemm_desc_kernel_arg_.size(),
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_);
}
else
{
const auto kernel =
kernel_grouped_gemm_xdlops_v2r3<GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<GemmDescKernelArg>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
false,
MaxGroupCount>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(arg.grid_size_),
dim3(BlockSize),
0,
gemm_desc_kernel_arg_arg,
arg.gemm_desc_kernel_arg_.size(),
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_);
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if(arg.gemm_desc_kernel_arg_.size() != arg.group_count_)
return false;
else
return true;
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(std::vector<const void*>& p_a,
std::vector<const void*>& p_b,
std::vector<void*>& p_c,
std::vector<GemmShape> gemm_shapes,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
return Argument{p_a, p_b, p_c, gemm_shapes, 1, 1, a_element_op, b_element_op, c_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(std::vector<const void*>& p_a,
std::vector<const void*>& p_b,
std::vector<void*>& p_c,
std::vector<GemmShape>& gemm_shapes,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t /* KBatch */ = 1) override
{
return std::make_unique<Argument>(
p_a, p_b, p_c, gemm_shapes, 1, 1, a_element_op, b_element_op, c_element_op);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceGroupedGemmXdl"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< K1 << ", "
<< MPerXDL << ", "
<< NPerXDL << ", "
<< MXdlPerWave << ", "
<< NXdlPerWave
<< ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif

74
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
View File

@@ -54,6 +54,80 @@ __global__ void
block_2_ctile_map);
}
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename GemmDesc,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
bool HasMainK0BlockLoop,
index_t MaxGroupCount>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_grouped_gemm_xdlops_v2r3(
const StaticallyIndexedArray<GemmDesc, MaxGroupCount> gemm_desc_,
const index_t group_count,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op)
{
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t block_id = get_block_1d_id();
#if 1
static_for<0, MaxGroupCount, 1>{}([&](auto i) {
if(block_id >= gemm_desc_[i].BlockStart_ && block_id < gemm_desc_[i].BlockEnd_ &&
i < group_count)
{
auto group_id = i;
GridwiseGemm::template Run<HasMainK0BlockLoop>(
gemm_desc_[group_id].a_ptr,
gemm_desc_[group_id].b_ptr,
gemm_desc_[group_id].c_ptr,
p_shared,
gemm_desc_[group_id].a_grid_desc_k0_m_k1_,
gemm_desc_[group_id].b_grid_desc_k0_n_k1_,
gemm_desc_[group_id].c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
a_element_op,
b_element_op,
c_element_op,
gemm_desc_[group_id].grouped_gemm_block_2_ctile_map_);
}
});
#else
const auto gemm_desc_ptr = reinterpret_cast<const GemmDesc*>(&gemm_desc_);
index_t group_id = 0;
static_for<0, MaxGroupCount, 1>{}([&](auto i) {
group_id = (block_id >= gemm_desc_[i].BlockStart && block_id < gemm_desc_[i].BlockEnd &&
i < group_count)
? i
: group_id;
});
const index_t block_id_grp = block_id - gemm_desc_ptr[group_id].BlockStart;
GridwiseGemm::template Run<HasMainK0BlockLoop>(
gemm_desc_ptr[group_id].a_ptr,
gemm_desc_ptr[group_id].b_ptr,
gemm_desc_ptr[group_id].c_ptr,
p_shared,
gemm_desc_ptr[group_id].a_grid_desc_k0_m_k1_,
gemm_desc_ptr[group_id].b_grid_desc_k0_n_k1_,
gemm_desc_ptr[group_id].c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
a_element_op,
b_element_op,
c_element_op,
gemm_desc_ptr[group_id].block_2_ctile_map_,
block_id_grp);
#endif
}
template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,

1
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View File

@@ -28,3 +28,4 @@ add_subdirectory(conv2d_fwd_bias_relu_add)
add_subdirectory(conv2d_fwd_bias_relu_atomic_add)
add_subdirectory(conv2d_bwd_data)
add_subdirectory(reduce)
add_subdirectory(grouped_gemm)

15
library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt Normal file
View File

@@ -0,0 +1,15 @@
# device_grouped_gemm_instance
set(DEVICE_GROUPED_GEMM_INSTANCE_SOURCE
device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp;
device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp;
device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp;
device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp;
)
add_library(device_grouped_gemm_instance SHARED ${DEVICE_GROUPED_GEMM_INSTANCE_SOURCE})
target_compile_features(device_grouped_gemm_instance PUBLIC)
set_target_properties(device_grouped_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
install(TARGETS device_grouped_gemm_instance LIBRARY DESTINATION lib)
clang_tidy_check(device_grouped_gemm_instance)

53
library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,53 @@
#include <stdlib.h>
#include "config.hpp"
#include "device_grouped_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_grouped_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances = std::tuple<
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>
// clang-format on
>;
void add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGroupedGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances{});
}
} // namespace device_grouped_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

53
library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp Normal file
View File

@@ -0,0 +1,53 @@
#include <stdlib.h>
#include "config.hpp"
#include "device_grouped_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_grouped_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
using device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances = std::tuple<
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>
// clang-format on
>;
void add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGroupedGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances{});
}
} // namespace device_grouped_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

62
library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,62 @@
#include <stdlib.h>
#include "config.hpp"
#include "device_grouped_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_grouped_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances = std::tuple<
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 32, 256, 4, 8, 32, 32, 1, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 32, 64, 4, 8, 32, 32, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 32, 32, 4, 8, 32, 32, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 16, 256, 4, 8, 16, 16, 1, 8, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 16, 128, 4, 8, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 16, 64, 4, 8, 16, 16, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 16, 32, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>
// clang-format on
>;
void add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGroupedGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances{});
}
} // namespace device_grouped_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

73
library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp Normal file
View File

@@ -0,0 +1,73 @@
#include <stdlib.h>
#include "config.hpp"
#include "device_grouped_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_grouped_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization_t::MNPadding;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances = std::tuple<
// clang-format off
//##################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//##################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//##################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//##################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, 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, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 8, 32, 32, 2, 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, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, 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, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, 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, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, 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, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>
// clang-format on
>;
// irregular tile size
using device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
// clang-format off
//##################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//##################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//##################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//##################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 144, 8, 8, 16, 16, 2, 9, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 8, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1>,
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 144, 4, 8, 16, 16, 2, 9, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1>
// clang-format on
>;
void add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGroupedGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances{});
add_device_operation_instances(
instances, device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
}
} // namespace device_grouped_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

3
profiler/CMakeLists.txt
View File

@@ -33,6 +33,7 @@ set(PROFILER_SOURCE
src/profile_conv_fwd_bias_relu_atomic_add.cpp
src/profile_conv_bwd_data.cpp
src/profile_reduce.cpp
src/profile_grouped_gemm.cpp
)
add_executable(ckProfiler ${PROFILER_SOURCE})
@@ -49,3 +50,5 @@ target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_add_instanc
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_atomic_add_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_data_instance)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)

314
profiler/include/profile_grouped_gemm_impl.hpp Normal file
View File

@@ -0,0 +1,314 @@
#pragma once
#include <iomanip>
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "device_gemm.hpp"
#include "reference_gemm.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_grouped_gemm_instance {
using DeviceGroupedGemmNoOpPtr = ck::tensor_operation::device::DeviceGroupedGemmPtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGroupedGemmNoOpPtr>&);
void add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGroupedGemmNoOpPtr>&);
void add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGroupedGemmNoOpPtr>&);
void add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGroupedGemmNoOpPtr>&);
} // namespace device_grouped_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_grouped_gemm_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
std::vector<int> Ms,
std::vector<int> Ns,
std::vector<int> Ks,
std::vector<int> StrideAs,
std::vector<int> StrideBs,
std::vector<int> StrideCs)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
int group_count = Ms.size();
if(!(group_count == Ns.size() && group_count == Ks.size() && group_count == StrideAs.size() &&
group_count == StrideBs.size() && group_count == StrideCs.size()))
{
throw std::runtime_error("wrong! inconsistent M/N/Ks, StrideA/B/Cs size\n");
}
std::vector<Tensor<ADataType>> a_m_k;
std::vector<Tensor<BDataType>> b_k_n;
std::vector<Tensor<CDataType>> c_m_n_device_results;
for(int i = 0; i < Ms.size(); i++)
{
a_m_k.push_back(
Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
b_k_n.push_back(
Tensor<BDataType>(f_host_tensor_descriptor(Ks[i], Ns[i], StrideBs[i], BLayout{})));
c_m_n_device_results.push_back(
Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n[" << i
<< "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
<< "]:" << c_m_n_device_results[i].mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency();
switch(init_method)
{
case 0: break;
case 1:
a_m_k[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
break;
default:
a_m_k[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
c_m_n_device_results[i].GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
// if(do_verification)
// {
// }
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_device_buf, b_device_buf, c_device_buf;
a_device_buf.reserve(group_count);
b_device_buf.reserve(group_count);
c_device_buf.reserve(group_count);
std::vector<const void*> p_a, p_b;
std::vector<void*> p_c;
p_a.reserve(group_count);
p_b.reserve(group_count);
p_c.reserve(group_count);
std::vector<ck::tensor_operation::device::GemmShape> gemm_shapes;
gemm_shapes.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
a_device_buf.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSize()));
b_device_buf.emplace_back(
std::make_unique<DeviceMem>(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSize()));
c_device_buf.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_m_n_device_results[i].mDesc.GetElementSize()));
a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
c_device_buf[i]->ToDevice(c_m_n_device_results[i].mData.data());
gemm_shapes.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i]});
p_a.push_back(a_device_buf[i]->GetDeviceBuffer());
p_b.push_back(b_device_buf[i]->GetDeviceBuffer());
p_c.push_back(c_device_buf[i]->GetDeviceBuffer());
}
// add device GEMM instances
std::vector<
ck::tensor_operation::device::device_grouped_gemm_instance::DeviceGroupedGemmNoOpPtr>
gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
{
auto argument_ptr =
gemm_ptr->MakeArgumentPointer(p_a,
p_b,
p_c,
gemm_shapes,
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{});
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < gemm_shapes.size(); i++)
{
flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
num_btype += sizeof(ADataType) * Ms[i] * Ks[i] + sizeof(BDataType) * Ks[i] * Ns[i] +
sizeof(CDataType) * Ms[i] * Ns[i];
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
for(int i = 0; i < gemm_shapes.size(); i++)
{
c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
Tensor<CDataType> c_m_n_host_result(
f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{}));
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_m_k[i],
b_k_n[i],
c_m_n_host_result,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
check_error(c_m_n_host_result, c_m_n_device_results[i]);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k[i].mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n[i].mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_m_n_device_results[i].mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
}
}
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
} // namespace profiler
} // namespace profiler
} // namespace ck

157
profiler/src/profile_grouped_gemm.cpp Normal file
View File

@@ -0,0 +1,157 @@
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "profile_grouped_gemm_impl.hpp"
enum GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
};
std::vector<int> argToIntArray(char* input)
{
std::vector<int> out;
std::istringstream in(input);
std::string item;
while(std::getline(in, item, ','))
{
out.push_back(std::stoi(item));
}
return out;
}
int profile_grouped_gemm(int argc, char* argv[])
{
if(!(argc == 14))
{
printf("arg1: tensor operation (grouped_gemm: Grouped GEMM)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg8 to 13: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
"64,64 64,64 128,128)\n");
exit(1);
}
const int data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const int layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const int nrepeat = std::stoi(argv[7]);
const auto Ms = argToIntArray(argv[8]);
const auto Ns = argToIntArray(argv[9]);
const auto Ks = argToIntArray(argv[10]);
const auto StrideAs = argToIntArray(argv[11]);
const auto StrideBs = argToIntArray(argv[12]);
const auto StrideCs = argToIntArray(argv[13]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
nrepeat,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
nrepeat,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
nrepeat,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
nrepeat,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs);
}
else
{
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
return 1;
}

10
profiler/src/profiler.cpp
View File

@@ -15,9 +15,11 @@ int profile_conv_fwd_bias_relu_add(int, char*[]);
int profile_conv_fwd_bias_relu_atomic_add(int, char*[]);
int profile_conv_bwd_data(int, char*[]);
int profile_reduce(int, char*[]);
int profile_grouped_gemm(int, char*[]);
int main(int argc, char* argv[])
{
#if 0
if(strcmp(argv[1], "gemm") == 0)
{
return profile_gemm(argc, argv);
@@ -62,6 +64,10 @@ int main(int argc, char* argv[])
{
return profile_reduce(argc, argv);
}
else if(strcmp(argv[1], "grouped_gemm") == 0)
{
return profile_grouped_gemm(argc, argv);
}
else
{
// clang-format off
@@ -74,9 +80,13 @@ int main(int argc, char* argv[])
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
" conv_fwd_bias_relu_atomic_add: ForwardConvolution+Bias+ReLU+AtomicAdd\n"
" conv_bwd: BackwardConvolution\n"
" grouped_gemm: Grouped Gemm\n"
" reduce: REDUCE\n");
// clang-format on
return 0;
}
#else
profile_grouped_gemm(argc, argv);
#endif
}

1
test/CMakeLists.txt
View File

@@ -35,6 +35,7 @@ add_subdirectory(space_filling_curve)
add_subdirectory(conv_util)
add_subdirectory(reference_conv_fwd)
add_subdirectory(gemm)
add_subdirectory(grouped_gemm)
add_subdirectory(gemm_split_k)
add_subdirectory(conv2d_fwd)
add_subdirectory(convnd_fwd)

3
test/grouped_gemm/CMakeLists.txt Normal file
View File

@@ -0,0 +1,3 @@
add_test_executable(test_grouped_gemm_fp16 grouped_gemm_fp16.cpp)
target_link_libraries(test_grouped_gemm_fp16 PRIVATE host_tensor)
target_link_libraries(test_grouped_gemm_fp16 PRIVATE device_grouped_gemm_instance)

213
test/grouped_gemm/grouped_gemm_fp16.cpp Normal file
View File

@@ -0,0 +1,213 @@
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_grouped_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
#include "test_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGroupedGemmPtr_ = ck::tensor_operation::device::DeviceGroupedGemmPtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_grouped_gemm_instance {
void add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGroupedGemmPtr_>&);
}
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace {
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using AccDataType = float;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
template <typename T>
static bool check_err(const Tensor<T>& ref, const Tensor<T>& result)
{
float max_diff = 1e-2;
for(int i = 0; i < ref.mData.size(); ++i)
{
float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
if(max_diff < diff)
{
std::cout << double(ref.mData[i]) << "," << double(result.mData[i]) << std::endl;
return false;
}
}
return true;
}
bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
{
int group_count = 4;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmShape> gemm_shapes;
std::vector<const void*> p_a, p_b;
std::vector<void*> p_c;
gemm_shapes.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
int M = 256 + 256 * i;
int N = 128 + 128 * i;
int K = 128 + 64 * i;
int AStride = std::is_same<ck::tensor_layout::gemm::RowMajor, ALayout>::value ? K : M;
int BStride = std::is_same<ck::tensor_layout::gemm::RowMajor, BLayout>::value ? N : K;
int CStride = std::is_same<ck::tensor_layout::gemm::RowMajor, CLayout>::value ? N : M;
gemm_shapes.push_back({M, N, K, AStride, BStride, CStride});
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
std::vector<Tensor<ADataType>> a_tensors;
;
std::vector<Tensor<BDataType>> b_tensors;
std::vector<Tensor<CDataType>> c_host_tensors;
std::vector<Tensor<CDataType>> c_device_tensors;
a_tensors.reserve(group_count);
b_tensors.reserve(group_count);
c_host_tensors.reserve(group_count);
c_device_tensors.reserve(group_count);
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
a_tensors_device.reserve(group_count);
b_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
for(int i = 0; i < gemm_shapes.size(); i++)
{
a_tensors.emplace_back(Tensor<ADataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{})));
b_tensors.emplace_back(Tensor<BDataType>(f_host_tensor_descriptor(
gemm_shapes[i].K, gemm_shapes[i].N, gemm_shapes[i].StrideB, BLayout{})));
c_host_tensors.emplace_back(Tensor<CDataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
c_device_tensors.emplace_back(Tensor<CDataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
}
for(int i = 0; i < gemm_shapes.size(); i++)
{
a_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize()));
b_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize()));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_device_tensors[i].mDesc.GetElementSize()));
a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
p_b.push_back(b_tensors_device[i]->GetDeviceBuffer());
p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
}
auto a_element_op = PassThrough{};
auto b_element_op = PassThrough{};
auto c_element_op = PassThrough{};
// do GEMM
auto invoker_ptr = groupedGemmPtr->MakeInvokerPointer();
auto argument_ptr = groupedGemmPtr->MakeArgumentPointer(
p_a, p_b, p_c, gemm_shapes, a_element_op, b_element_op, c_element_op);
invoker_ptr->Run(argument_ptr.get());
for(int i = 0; i < gemm_shapes.size(); i++)
{
c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
b_tensors[i],
c_host_tensors[i],
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
bool res = check_err(c_device_tensors[i], c_host_tensors[i]);
std::cout << "group_id: " << i << (res ? " SUCCESS" : " FAILURE") << std::endl;
if(!res)
return false;
}
return true;
}
} // anonymous namespace
int main()
{
std::vector<DeviceGroupedGemmPtr_> groupedGemmPtrs;
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(groupedGemmPtrs);
bool res = true;
for(auto& gemmPtr : groupedGemmPtrs)
{
res &= TestGroupedGemm(gemmPtr);
}
std::cout << "TestGroupedGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
}