Skip to content

GitLab

Commit 2b27d5fc authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into rosenrodt/gemm-layernorm

parents f689a155 fa9a0a5c
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1019 additions and 423 deletions
CMakeLists.txt
View file @ 2b27d5fc
......@@ -71,13 +71,6 @@ if( DEFINED CK_OVERRIDE_HIP_VERSION_PATCH )
endif()
message(STATUS "Build with HIP ${HIP_VERSION}")
rocm_create_package(
NAME composablekernel
DESCRIPTION "High Performance Composable Kernel for AMD GPUs"
MAINTAINER "MIOpen Kernels Dev Team <dl.MIOpen@amd.com>"
LDCONFIG
)
## tidy
include(EnableCompilerWarnings)
set(CK_TIDY_ERRORS ERRORS * -readability-inconsistent-declaration-parameter-name)
......
Jenkinsfile
View file @ 2b27d5fc
......@@ -379,23 +379,23 @@ pipeline {
}
}
}
//stage("Client App")
//{
// parallel
// {
// stage("Run Client App")
// {
// agent{ label rocmnode("gfx908")}
// environment{
// setup_args = """ -D -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " """
// execute_args = """ cd ../test/client_app && rm -rf build && mkdir build && cd build && cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" .. && make """
// }
// steps{
// buildHipClangJobAndReboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
// }
// }
// }
//}
stage("Client App")
{
parallel
{
stage("Run Client App")
{
agent{ label rocmnode("gfx908")}
environment{
setup_args = """ -D -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DCMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc .. && make -j """
}
steps{
buildHipClangJobAndReboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
}
}
}
}
stage("Performance Tests")
{
parallel
......
client_example/02_gemm_add_add_fastgelu/CMakeLists.txt 0 → 100644
View file @ 2b27d5fc
add_executable(client_gemm_add_add_fastgelu gemm_add_add_fastgelu.cpp)
target_link_libraries(client_gemm_add_add_fastgelu PRIVATE composable_kernel::device_operations)
client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp 0 → 100644
View file @ 2b27d5fc
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.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/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/device_gemm_add_add_fastgelu_instance.hpp"
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 AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddAddFastGelu;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using D0DataType = F16;
using D1DataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using D0Layout = Row;
using D1Layout = Row;
using ELayout = Row;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main(int argc, char* argv[])
{
// 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 StrideD0 = 0;
ck::index_t StrideD1 = 4096;
ck::index_t StrideE = 4096;
if(argc == 1)
{
// use default case
}
else if(argc == 9)
{
M = std::stoi(argv[1]);
N = std::stoi(argv[2]);
K = std::stoi(argv[3]);
StrideA = std::stoi(argv[4]);
StrideB = std::stoi(argv[5]);
StrideD0 = std::stoi(argv[6]);
StrideD1 = std::stoi(argv[7]);
StrideE = std::stoi(argv[8]);
}
else
{
printf("arg1 to 8: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
exit(0);
}
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
f_matrix_space_size(M, N, StrideD0, D0Layout{}));
SimpleDeviceMem d1_m_n_device_buf(sizeof(D1DataType) *
f_matrix_space_size(M, N, StrideD1, D1Layout{}));
SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
// add device op instances
const auto op_ptrs = ck::tensor_operation::device::device_gemm_instance::
get_device_gemm_add_add_fastgelu_instances<ADataType,
BDataType,
AccDataType,
D0DataType,
D1DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
D1Layout,
ELayout>();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
d1_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 2>{StrideD0, StrideD1},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
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: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
d1_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 2>{StrideD0, StrideD1},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
}
return 0;
}
client_example/03_gemm_layernorm/CMakeLists.txt 0 → 100644
View file @ 2b27d5fc
add_executable(gemm_add_add_reduce_normalize gemm_add_add_layernorm.cpp)
target_link_libraries(gemm_add_add_reduce_normalize PRIVATE composable_kernel::device_operations)
client_example/03_gemm_layernorm/gemm_add_add_layernorm.cpp 0 → 100644
View file @ 2b27d5fc
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F16;
using BDataType = F16;
using BiasDataType = F32;
using CDataType = F16;
using D0DataType = F16;
using ReduceDataType = F32;
using GammaDataType = F16;
using BetaDataType = F16;
using LayerNormOutDataType = F16;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
template <typename gemm_reduce_op_ptr>
bool RunDeviceGemmMeanSquareMean(gemm_reduce_op_ptr& p_op,
const void* p_a,
const void* p_b,
const void* p_bias,
const void* p_d0,
void* p_c,
void* p_mean,
void* p_square_mean,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int StrideD0,
bool time_kernel)
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
auto passOp = PassThrough{};
auto squareOp = UnarySquareElementOp{};
auto divOp = UnaryDivElementOp{N};
auto argument_ptr =
p_op->MakeArgumentPointer(p_a,
p_b,
p_bias,
{p_d0},
p_c,
{p_mean, p_square_mean},
M,
N,
K,
StrideA,
StrideB,
StrideC,
{StrideD0},
{&passOp, &passOp, &passOp}, // functor for a, b, c
{&passOp}, // functor for d0
{&passOp, &squareOp}, // functor for inputs of reduction
{&divOp, &divOp}); // functor for outputs of reduction
if(p_op->IsSupportedArgument(argument_ptr.get()))
{
auto invoker_ptr = p_op->MakeInvokerPointer();
// If we evaluate running time of gemm_reduce. The output may wrong.
// Because we need to initialize the reduction tensor before runing the kernel.
// However we run kernel many times for time_kernel = trie without reinitialize the out
// of reduction tensor.
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
if(time_kernel)
std::cout << "Gemm + reduce Perf: " << std::setw(10) << ave_time << " ms" << std::endl;
return true;
}
return false;
}
template <typename normalize_op_ptr>
bool RunDeviceNormalize2D(normalize_op_ptr& p_op,
const void* p_x,
const void* p_mean,
const void* p_square_mean,
const void* p_gamma,
const void* p_beta,
void* p_y,
int M,
int N,
int StrideX,
bool time_kernel)
{
std::array<const void*, 5> input = {p_x, p_mean, p_square_mean, p_gamma, p_beta};
std::array<void*, 1> output = {p_y};
auto normalize_functor = ck::tensor_operation::element_wise::Normalize{};
auto argument_ptr = p_op->MakeArgumentPointer(input,
output,
{M, N},
{{StrideX, 1}, {1, 0}, {1, 0}, {0, 1}, {0, 1}},
{{StrideX, 1}},
ck::tensor_operation::element_wise::Normalize{});
if(p_op->IsSupportedArgument(argument_ptr.get()))
{
auto invoker_ptr = p_op->MakeInvokerPointer();
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
if(time_kernel)
std::cout << "Normalize Perf: " << std::setw(10) << ave_time << " ms" << std::endl;
return true;
}
return false;
}
int main()
{
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = 1024;
ck::index_t StrideB = 1024;
ck::index_t StrideC = 1024;
ck::index_t StrideD0 = 1024;
const auto gemm_reduce_ptrs = ck::tensor_operation::device::device_gemm_instance::
get_device_gemm_add_add_mean_squaremean_instances<ADataType,
BDataType,
CDataType,
ALayout,
BLayout,
CLayout>();
const auto normalize_ptrs =
ck::tensor_operation::device::get_device_normalize_from_mean_meansquare_instances<
CDataType,
ReduceDataType,
ReduceDataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType>();
std::cout << "found " << gemm_reduce_ptrs.size()
<< " gemm_reduceMean_reduceSquareMean instances" << std::endl;
std::cout << "found " << normalize_ptrs.size() << " normalize instances" << std::endl;
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
SimpleDeviceMem bias_device_buf(sizeof(BiasDataType) * N);
SimpleDeviceMem c_device_buf(sizeof(CDataType) * f_matrix_space_size(M, N, StrideC, CLayout{}));
SimpleDeviceMem d0_device_buf(sizeof(D0DataType) *
f_matrix_space_size(M, N, StrideD0, CLayout{}));
SimpleDeviceMem reduceMean_device_buf(sizeof(ReduceDataType) * M);
SimpleDeviceMem reduceMeanSquare_device_buf(sizeof(ReduceDataType) * M);
SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
SimpleDeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) * M * N);
bool b_time_kernel = true;
bool b_only_run_first_kernel = true;
// layernorm => (1) + (2)
// (1). c = gemm(a, b), reduce_mean(c), reduce_square_mean(c)
// (2). normalize(c, mean, square_mean, gamma, beta)
for(auto& gemm_reduce_ptr : gemm_reduce_ptrs)
{
// run first available kernel
if(RunDeviceGemmMeanSquareMean(gemm_reduce_ptr,
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
bias_device_buf.GetDeviceBuffer(),
d0_device_buf.GetDeviceBuffer(),
c_device_buf.GetDeviceBuffer(),
reduceMean_device_buf.GetDeviceBuffer(),
reduceMeanSquare_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideD0,
b_time_kernel))
{
if(b_only_run_first_kernel)
break;
}
else
{
std::cout << gemm_reduce_ptr->GetTypeString() << " does not support this problem"
<< std::endl;
}
}
for(auto& normalize_ptr : normalize_ptrs)
{
if(RunDeviceNormalize2D(normalize_ptr,
c_device_buf.GetDeviceBuffer(),
reduceMean_device_buf.GetDeviceBuffer(),
reduceMeanSquare_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
layerNorm_device_buf.GetDeviceBuffer(),
M,
N,
StrideC,
b_time_kernel))
{
if(b_only_run_first_kernel)
break;
}
else
{
std::cout << normalize_ptr->GetTypeString() << " does not support this problem"
<< std::endl;
}
}
}
\ No newline at end of file
client_example/CMakeLists.txt 0 → 100644
View file @ 2b27d5fc
cmake_minimum_required(VERSION 3.15)
project(ck_app)
add_compile_options(-std=c++17)
find_package(composable_kernel 1.0.0 COMPONENTS device_operations)
find_package(hip REQUIRED PATHS /opt/rocm)
message(STATUS "Build with HIP ${hip_VERSION}")
add_subdirectory(02_gemm_add_add_fastgelu)
add_subdirectory(03_gemm_layernorm)
client_example/README.md 0 → 100644
View file @ 2b27d5fc
##
Client application links to CK library, and therefore CK library needs to be installed before building client applications.
## Docker script
```bash
docker run \
-it \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm5.1-tf2.6-dev \
/bin/bash
```
## Build
```bash
mkdir -p client_example/build
cd client_example/build
```
```bash
cmake \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
..
```
### Build client example
```bash
make -j
```
example/01_gemm/gemm_xdl_bf16.cpp
View file @ 2b27d5fc
......@@ -84,8 +84,13 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
8>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<float, float, float, float, PassThrough, PassThrough, PassThrough>;
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
int main(int argc, char* argv[])
{
......@@ -216,24 +221,17 @@ int main(int argc, char* argv[])
if(do_verification)
{
Tensor<float> a_f32_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<float> b_f32_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<float> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<float> c_m_n_device_f32_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
bf16_to_f32_(a_m_k, a_f32_m_k);
bf16_to_f32_(b_k_n, b_f32_k_n);
bf16_to_f32_(c_m_n_device_result, c_m_n_device_f32_result);
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_f32_m_k, b_f32_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
return ck::utils::check_err(c_m_n_device_f32_result.mData, c_m_n_host_result.mData) ? 0 : 1;
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
}
return 0;
......
example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp
View file @ 2b27d5fc
......@@ -33,19 +33,19 @@ using BDataType = F16;
using CDataType = F16;
using GemmAccDataType = F32;
using ReduceAccDataType = F32;
using DDataType = F64;
using DPtrsGlobal = ck::Tuple<DDataType*>;
using ReduceDataType = F64;
using ReducePtrsGlobal = ck::Tuple<ReduceDataType*>;
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;
using DsReduceOp = ck::Tuple<ck::reduce::Max>;
using DsElementOp = ck::Tuple<ck::tensor_operation::element_wise::PassThrough>;
using DGlobalMemOp =
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using ReduceOps = ck::Tuple<ck::reduce::Max>;
using ReduceElementOps = ck::Tuple<ck::tensor_operation::element_wise::PassThrough>;
using ReduceGlobalMemOps =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
static constexpr auto GemmSpecialization =
......@@ -53,11 +53,11 @@ static constexpr auto GemmSpecialization =
// clang-format off
using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, ReduceAccDataType, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DsReduceOp, DsElementOp, DsElementOp, DGlobalMemOp, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| ReduceData| A| B| C| Reduce| ReduceInEleOp| ReduceAccEleOp| Reduce| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Operation| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, ReduceAccDataType, ReducePtrsGlobal, AElementOp, BElementOp, CElementOp, ReduceOps, ReduceElementOps, ReduceElementOps, ReduceGlobalMemOps, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
......@@ -68,12 +68,12 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataTyp
BElementOp,
CElementOp>;
template <typename ADataType, typename BDataType, typename CDataType, typename DDataType>
template <typename ADataType, typename BDataType, typename CDataType, typename ReduceDataType>
void DumpGemmLayerNormPerf(float gemm_reduce_time, int M, int N, int K)
{
std::size_t gemm_flop = std::size_t(2) * M * N * K;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(DDataType) * M;
sizeof(CDataType) * M * N + sizeof(ReduceDataType) * M;
float tflops = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
float gemm_gb_per_sec = gemm_num_byte / 1.E6 / gemm_reduce_time;
......@@ -148,17 +148,17 @@ int main(int argc, char* argv[])
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<DDataType> d_m_host_result(
Tensor<ReduceDataType> reduce_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<DDataType> d_m_device_result(
Tensor<ReduceDataType> reduce_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "d_m: " << d_m_host_result.mDesc << std::endl;
std::cout << "reduce_m: " << reduce_m_host_result.mDesc << std::endl;
switch(init_method)
{
......@@ -176,35 +176,40 @@ int main(int argc, char* argv[])
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem d_device_buf(sizeof(DDataType) * d_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce_device_buf(sizeof(ReduceDataType) *
reduce_m_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto ds_element_op = DsElementOp{};
auto p_ds_global = ck::make_tuple(static_cast<DDataType*>(d_device_buf.GetDeviceBuffer()));
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto reduce_element_op = ReduceElementOps{}[ck::Number<0>{}];
std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
std::array<void*, 1> reduce_element_ops = {&reduce_element_op};
std::array<void*, 1> p_reduces = {reduce_device_buf.GetDeviceBuffer()};
// do GEMM
auto gemm = DeviceGemmReduceInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
p_ds_global,
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
nullptr,
{},
c_device_buf.GetDeviceBuffer(),
p_reduces,
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
ds_element_op,
ds_element_op);
{},
gemm_element_ops,
{},
reduce_element_ops,
reduce_element_ops);
if(!gemm.IsSupportedArgument(argument))
{
......@@ -215,7 +220,7 @@ int main(int argc, char* argv[])
// [CAUSION]: launch_and_time_kernel will not initialize D.
// If we evaluate kernel multiple time but without initialize D. Verification will fail
d_device_buf.SetValue(ck::NumericLimits<DDataType>::Lowest());
reduce_device_buf.SetValue(ck::NumericLimits<ReduceDataType>::Lowest());
invoker.Run(argument, StreamConfig{nullptr, false});
bool pass = true;
......@@ -223,7 +228,7 @@ int main(int argc, char* argv[])
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
d_device_buf.FromDevice(d_m_device_result.mData.data());
reduce_device_buf.FromDevice(reduce_m_device_result.mData.data());
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -233,27 +238,27 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
auto d_reduce_op = DsReduceOp{}[ck::Number<0>{}];
auto reduce_op = ReduceOps{}[ck::Number<0>{}];
for(int m = 0; m < M; ++m)
{
ReduceAccDataType d_acc = d_reduce_op.GetIdentityValue<ReduceAccDataType>();
ReduceAccDataType reduce_acc = reduce_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType curr_val =
ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
d_reduce_op(d_acc, curr_val);
reduce_op(reduce_acc, curr_val);
};
d_m_host_result(m) = d_acc;
reduce_m_host_result(m) = reduce_acc;
}
pass = ck::utils::check_err(c_m_n_device_result.mData,
c_m_n_host_result.mData,
"Error: Incorrect results c") &&
ck::utils::check_err(d_m_device_result.mData,
d_m_host_result.mData,
ck::utils::check_err(reduce_m_device_result.mData,
reduce_m_host_result.mData,
"Error: Incorrect results d",
1e-3,
1e-3);
......@@ -263,7 +268,7 @@ int main(int argc, char* argv[])
{
float gemm_reduceMax_ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, DDataType>(
DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, ReduceDataType>(
gemm_reduceMax_ave_time, M, N, K);
}
......
example/16_gemm_reduce/gemm_reduce_xdl_mean_squaremean_fp16.cpp
View file @ 2b27d5fc
......@@ -33,27 +33,27 @@ using BDataType = F16;
using CDataType = F16;
using GemmAccDataType = F32;
using ReduceAccDataType = F32;
using DDataType = F32;
using DPtrsGlobal = ck::Tuple<DDataType*, DDataType*>;
using ReduceDataType = F32;
using ReducePtrsGlobal = ck::Tuple<ReduceDataType*, ReduceDataType*>;
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;
using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add;
using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using ReduceOp0 = ck::reduce::Add;
using ReduceOp1 = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceOp0, ReduceOp1>;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using ReduceInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using ReduceOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using DGlobalMemOp =
using ReduceGlobalMemOps =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
......@@ -62,11 +62,11 @@ static constexpr auto GemmSpecialization =
// clang-format off
using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOps, DxsOutElementOps, DGlobalMemOp, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| ReduceDData| A| B| C| Reduce| ReduceInEleOp| ReduceOutEleOp| Reduce| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Operation| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, AElementOp, BElementOp, CElementOp, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceGlobalMemOps, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
......@@ -77,13 +77,13 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataTyp
BElementOp,
CElementOp>;
template <typename ADataType, typename BDataType, typename CDataType, typename DDataType>
template <typename ADataType, typename BDataType, typename CDataType, typename ReduceDataType>
void DumpGemmLayerNormPerf(float gemm_reduce_time, int M, int N, int K)
{
std::size_t gemm_flop = std::size_t(2) * M * N * K;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(DDataType) * M +
sizeof(DDataType) * M;
sizeof(CDataType) * M * N + sizeof(ReduceDataType) * M +
sizeof(ReduceDataType) * M;
float tflops = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
float gemm_gb_per_sec = gemm_num_byte / 1.E6 / gemm_reduce_time;
......@@ -158,22 +158,22 @@ int main(int argc, char* argv[])
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_host_result(
Tensor<ReduceDataType> reduce0_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_host_result(
Tensor<ReduceDataType> reduce1_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_device_result(
Tensor<ReduceDataType> reduce0_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_device_result(
Tensor<ReduceDataType> reduce1_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
std::cout << "reduce0_m: " << reduce0_m_host_result.mDesc << std::endl;
std::cout << "reduce1_m: " << reduce1_m_host_result.mDesc << std::endl;
switch(init_method)
{
......@@ -191,39 +191,48 @@ int main(int argc, char* argv[])
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
reduce0_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
reduce1_m_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{N, N};
auto passthrough = UnaryIdenticElementOp{};
auto square = UnarySquareElementOp{};
auto div = UnaryDivElementOp{N};
std::array<void*, 2> reduce_in_element_ops = {&passthrough, &square};
std::array<void*, 2> reduce_out_element_ops = {&div, &div};
std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
reduce1_device_buf.GetDeviceBuffer()};
// do GEMM
auto gemm = DeviceGemmReduceInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
dxs_global,
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
nullptr,
{},
c_device_buf.GetDeviceBuffer(),
p_reduces,
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
dxs_in_element_op,
dxs_out_element_op);
{},
gemm_element_ops,
{},
reduce_in_element_ops,
reduce_out_element_ops);
if(!gemm.IsSupportedArgument(argument))
{
......@@ -232,9 +241,9 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
// init reducetion buffer to 0
reduce0_device_buf.SetZero();
reduce1_device_buf.SetZero();
// if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
// will not be correct. need to set time_kernel = false for correctness test
......@@ -244,8 +253,8 @@ int main(int argc, char* argv[])
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
d0_device_buf.FromDevice(d0_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_m_device_result.mData.data());
reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -255,42 +264,40 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
auto d0_reduce_op = D0ReduceOp{};
auto d1_reduce_op = D1ReduceOp{};
auto reduce0_op = ReduceOp0{};
auto reduce1_op = ReduceOp1{};
for(int m = 0; m < M; ++m)
{
auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
auto c_val = ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
ReduceAccDataType d0_val;
ReduceAccDataType d1_val;
ReduceAccDataType square_c_val;
square(square_c_val, c_val);
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
reduce0_op(reduce0_acc, c_val);
reduce1_op(reduce1_acc, square_c_val);
}
dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
div(reduce0_acc, reduce0_acc);
div(reduce1_acc, reduce1_acc);
reduce0_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce0_acc);
reduce1_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce1_acc);
}
pass = ck::utils::check_err(c_m_n_device_result.mData,
c_m_n_host_result.mData,
"Error: Incorrect results c") &&
ck::utils::check_err(d0_m_device_result.mData,
d0_m_host_result.mData,
ck::utils::check_err(reduce0_m_device_result.mData,
reduce0_m_host_result.mData,
"Error: Incorrect results d0",
1e-4,
1e-5) &&
ck::utils::check_err(d1_m_device_result.mData,
d1_m_host_result.mData,
ck::utils::check_err(reduce1_m_device_result.mData,
reduce1_m_host_result.mData,
"Error: Incorrect results d1",
1e-3,
1e-5);
......@@ -300,7 +307,7 @@ int main(int argc, char* argv[])
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, DDataType>(ave_time, M, N, K);
DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, ReduceDataType>(ave_time, M, N, K);
}
return pass ? 0 : 1;
......
example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
View file @ 2b27d5fc
......@@ -31,26 +31,26 @@ using ADataType = F16;
using BDataType = F16;
using CDataType = F16;
using ReduceAccDataType = F32;
using DDataType = F32;
using DPtrsGlobal = ck::Tuple<DDataType*, DDataType*>;
using ReduceDataType = F32;
using ReducePtrsGlobal = ck::Tuple<ReduceDataType*, ReduceDataType*>;
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;
using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add;
using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using ReduceOp0 = ck::reduce::Add;
using ReduceOp1 = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceOp0, ReduceOp1>;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
using ReduceInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using ReduceOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
using DGlobalMemOp =
using ReduceGlobalMemOps =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
......@@ -63,7 +63,7 @@ using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatc
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOps, DxsOutElementOps, DGlobalMemOp, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
< Row, Col, Row, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, AElementOp, BElementOp, CElementOp, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceGlobalMemOps, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
// clang-format on
using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
......@@ -143,16 +143,16 @@ int main(int argc, char* argv[])
Tensor<CDataType> c_g_m_n_host_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
Tensor<ReduceDataType> d0_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
Tensor<ReduceDataType> d1_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
Tensor<ReduceDataType> d0_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
Tensor<ReduceDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
......@@ -177,38 +177,48 @@ int main(int argc, char* argv[])
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(DDataType) * d0_g_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_g_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
d0_g_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
d1_g_m_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
auto passthrough = UnaryIdenticElementOp{};
auto square = UnarySquareElementOp{};
std::array<void*, 2> reduce_in_element_ops = {&passthrough, &square};
std::array<void*, 2> reduce_out_element_ops = {&passthrough, &passthrough};
std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
reduce1_device_buf.GetDeviceBuffer()};
// do GEMM
auto batched_gemm = DeviceBatchedGemmReduceInstance{};
auto invoker = batched_gemm.MakeInvoker();
auto argument =
batched_gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
dxs_global,
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
DxsInElementOps{},
DxsOutElementOps{},
BatchCount);
auto argument = batched_gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
nullptr,
{},
c_device_buf.GetDeviceBuffer(),
p_reduces,
M,
N,
K,
StrideA,
StrideB,
StrideC,
{},
gemm_element_ops,
{},
reduce_in_element_ops,
reduce_out_element_ops,
BatchCount);
if(!batched_gemm.IsSupportedArgument(argument))
{
......@@ -218,8 +228,8 @@ int main(int argc, char* argv[])
}
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
reduce0_device_buf.SetZero();
reduce1_device_buf.SetZero();
// if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
// will not be correct. need to set time_kernel = false for correctness test
......@@ -241,8 +251,8 @@ int main(int argc, char* argv[])
if(do_verification)
{
c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
d0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
reduce0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
reduce1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
......@@ -252,15 +262,15 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
auto d0_reduce_op = D0ReduceOp{};
auto d1_reduce_op = D1ReduceOp{};
auto reduce0_op = ReduceOp0{};
auto reduce1_op = ReduceOp1{};
for(int batch = 0; batch < BatchCount; ++batch)
{
for(int m = 0; m < M; ++m)
{
auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
......@@ -271,12 +281,12 @@ int main(int argc, char* argv[])
UnaryIdenticElementOp{}(d0_val, c_val);
UnarySquareElementOp{}(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
reduce0_op(reduce0_acc, d0_val);
reduce1_op(reduce1_acc, d1_val);
}
d0_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d0_acc);
d1_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d1_acc);
d0_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce0_acc);
d1_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce1_acc);
}
}
......
example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
View file @ 2b27d5fc
......@@ -99,15 +99,17 @@ int main()
a_m_n_device_buf.ToDevice(a_m_n.mData.data());
b_n_device_buf.ToDevice(b_n.mData.data());
std::array<const void*, 2> input = {a_m_n_device_buf.GetDeviceBuffer(),
b_n_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {c_m_n_device_buf.GetDeviceBuffer()};
std::vector<ck::index_t> a_strides = {Stride, 1};
std::vector<ck::index_t> b_strides = {0, 1};
std::vector<ck::index_t> c_strides = {Stride, 1};
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(a_m_n_device_buf.GetDeviceBuffer(),
b_n_device_buf.GetDeviceBuffer(),
c_m_n_device_buf.GetDeviceBuffer(),
{M, N},
{Stride, 1},
{0, 1}, // broadcast in first dimension
{Stride, 1},
Add{});
auto argument = broadcastAdd.MakeArgumentPointer(
input, output, {M, N}, {a_strides, b_strides}, {c_strides}, Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{
......
example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
View file @ 2b27d5fc
......@@ -81,18 +81,24 @@ int main()
a_m_device_buf.ToDevice(a_m.mData.data());
b_m_n_k_device_buf.ToDevice(b_m_n_k.mData.data());
std::array<const void*, 2> input = {a_m_device_buf.GetDeviceBuffer(),
b_m_n_k_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {c_m_n_k_device_buf.GetDeviceBuffer()};
std::vector<ck::index_t> a_strides = {1, 0, 0};
std::vector<ck::index_t> b_strides{b_m_n_k.mDesc.GetStrides().begin(),
b_m_n_k.mDesc.GetStrides().end()};
std::vector<ck::index_t> c_strides{c_m_n_k.mDesc.GetStrides().begin(),
c_m_n_k.mDesc.GetStrides().end()};
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(
a_m_device_buf.GetDeviceBuffer(),
b_m_n_k_device_buf.GetDeviceBuffer(),
c_m_n_k_device_buf.GetDeviceBuffer(),
std::vector<ck::index_t>{mnk.begin(), mnk.end()},
{1, 0, 0}, // broadcast A on second and third dimension
std::vector<ck::index_t>{b_m_n_k.mDesc.GetStrides().begin(),
b_m_n_k.mDesc.GetStrides().end()},
std::vector<ck::index_t>{c_m_n_k.mDesc.GetStrides().begin(),
c_m_n_k.mDesc.GetStrides().end()},
Add{});
auto argument =
broadcastAdd.MakeArgumentPointer(input,
output,
std::vector<ck::index_t>{mnk.begin(), mnk.end()},
{a_strides, b_strides},
{c_strides},
Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{
......
example/19_binary_elementwise/elementwise_add_1d.cpp
View file @ 2b27d5fc
......@@ -79,15 +79,17 @@ int main()
a_m_device_buf.ToDevice(a_m.mData.data());
b_m_device_buf.ToDevice(b_m.mData.data());
std::array<const void*, 2> input = {a_m_device_buf.GetDeviceBuffer(),
b_m_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {c_m_device_buf.GetDeviceBuffer()};
std::vector<ck::index_t> a_strides = {1};
std::vector<ck::index_t> b_strides = {1};
std::vector<ck::index_t> c_strides = {1};
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(a_m_device_buf.GetDeviceBuffer(),
b_m_device_buf.GetDeviceBuffer(),
c_m_device_buf.GetDeviceBuffer(),
{M},
{1},
{1},
{1},
Add{});
auto argument = broadcastAdd.MakeArgumentPointer(
input, output, {M}, {{a_strides}, b_strides}, {c_strides}, Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{
......
example/19_binary_elementwise/elementwise_add_4d.cpp
View file @ 2b27d5fc
......@@ -81,16 +81,22 @@ int main()
a_device_buf.ToDevice(a.mData.data());
b_device_buf.ToDevice(b.mData.data());
std::array<const void*, 2> input = {a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {c_device_buf.GetDeviceBuffer()};
std::vector<ck::index_t> a_strides{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()};
std::vector<ck::index_t> b_strides{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()};
std::vector<ck::index_t> c_strides{c.mDesc.GetStrides().begin(), c.mDesc.GetStrides().end()};
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
c_device_buf.GetDeviceBuffer(),
std::vector<ck::index_t>{nchw.begin(), nchw.end()},
std::vector<ck::index_t>{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()},
std::vector<ck::index_t>{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()},
std::vector<ck::index_t>{c.mDesc.GetStrides().begin(), c.mDesc.GetStrides().end()},
Add{});
auto argument =
broadcastAdd.MakeArgumentPointer(input,
output,
std::vector<ck::index_t>{nchw.begin(), nchw.end()},
{{a_strides}, b_strides},
{c_strides},
Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{
......
example/23_softmax/softmax_blockwise.cpp
View file @ 2b27d5fc
......@@ -150,6 +150,9 @@ int main(int argc, char* argv[])
AccDataType alpha = args.scales[0];
AccDataType beta = args.scales[1];
std::cout << "in: " << in.mDesc << std::endl;
std::cout << "out: " << out.mDesc << std::endl;
std::size_t num_thread = 1;
if(args.do_verification)
......@@ -195,7 +198,7 @@ int main(int argc, char* argv[])
using ReferenceInstance =
tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>;
ReferenceInstance ref;
auto ref_arg = ref.MakeArgument(in, out_ref, alpha, beta, Rank, reduceDims);
auto ref_arg = ref.MakeArgument(in, out_ref, alpha, beta, reduceDims);
auto invoker = ref.MakeInvoker();
invoker.Run(ref_arg);
// LogRangeAsType<float>(std::cout << "tensor out_ref: ", out_ref.mData, ",") << std::endl;
......@@ -212,8 +215,8 @@ int main(int argc, char* argv[])
auto argument_ptr = device_instance.MakeArgumentPointer(i_inLengths,
i_inStrides,
reduceDims,
alpha,
beta,
&alpha,
&beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer());
......
example/25_gemm_bias_c_permute/CMakeLists.txt 0 → 100644
View file @ 2b27d5fc
add_example_executable(example_gemm_bias_c_permute_xdl_fp16 gemm_bias_c_permute_xdl_fp16.cpp)
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