Skip to content

GitLab

Unverified Commit 3550cefe authored by rocking's avatar rocking Committed by GitHub
Browse files

Merge branch 'develop' into max-pool-bwd

parents 508d7a59 027e46ee
Hide whitespace changes
Inline Side-by-side
Showing with 608 additions and 41 deletions
Dockerfile
View file @ 3550cefe
......@@ -12,20 +12,20 @@ RUN useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins
RUN chmod 1777 /tmp
RUN apt-get update
RUN apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl
RUN --mount=type=ssh if [ "$ROCMVERSION" != "5.6" ]; then \
RUN if [ "$ROCMVERSION" != "5.6" ]; then \
wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.list.d/rocm.list"; \
elif [ "$ROCMVERSION" = "5.6" ] && [ "$compiler_version" = "" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amd-nonfree-radeon_20.04-1_all.deb" && \
apt update && apt-get install -y ./amd-nonfree-radeon_20.04-1_all.deb && \
amdgpu-repo --amdgpu-build=1567752 --rocm-build=compute-rocm-dkms-no-npi-hipclang/11914; \
elif [ "$ROCMVERSION" = "5.6" ] && [ "$compiler_version" = "rc3" ]; then \
amdgpu-repo --amdgpu-build=1567752 --rocm-build=compute-rocm-dkms-no-npi-hipclang/11914 && \
amdgpu-install -y --usecase=rocm --no-dkms; \
elif [ "$ROCMVERSION" = "5.6" ] && [ "$compiler_version" = "rc3" ] || [ "$compiler_version" = "amd-stg-open" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_5.6-20.04-1_all.deb" && \
apt update && apt-get install -y ./amdgpu-install-internal_5.6-20.04-1_all.deb && \
sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 5.6 rel-45 > /etc/apt/sources.list.d/rocm-build.list' && \
amdgpu-repo --amdgpu-build=1602498; \
amdgpu-repo --amdgpu-build=1602498 && amdgpu-install -y --usecase=rocm --no-dkms; \
fi
RUN amdgpu-install -y --usecase=rocm --no-dkms
RUN wget --no-check-certificate -qO - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | apt-key add -
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
......@@ -105,7 +105,7 @@ ENV compiler_commit=$compiler_commit
RUN sh -c "echo compiler version = '$compiler_version'"
RUN sh -c "echo compiler commit = '$compiler_commit'"
RUN --mount=type=ssh if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" = "" ]; then \
RUN if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" = "" ]; then \
git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \
cd llvm-project && mkdir build && cd build && \
cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \
......@@ -113,7 +113,7 @@ RUN --mount=type=ssh if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler
else echo "using the release compiler"; \
fi
RUN --mount=type=ssh if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" != "" ]; then \
RUN if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" != "" ]; then \
git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \
cd llvm-project && git checkout "$compiler_commit" && echo "checking out commit $compiler_commit" && mkdir build && cd build && \
cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \
......
Jenkinsfile
View file @ 3550cefe
......@@ -695,8 +695,8 @@ pipeline {
}
agent{ label rocmnode("gfx908 || gfx90a") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940" """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
}
steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
......
client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
View file @ 3550cefe
......@@ -11,3 +11,17 @@ target_link_libraries(client_gemm_fastgelu PRIVATE composable_kernel::device_ope
add_dependencies(client_gemm_fastgelu_examples client_gemm_add_add_fastgelu client_gemm_add_fastgelu
client_gemm_fastgelu)
add_custom_target(client_gemm_fastgelu_generic_examples)
add_executable(client_gemm_add_add_fastgelu_generic gemm_add_add_fastgelu_generic.cpp)
target_link_libraries(client_gemm_add_add_fastgelu_generic PRIVATE composable_kernel::device_operations)
add_executable(client_gemm_add_fastgelu_generic gemm_add_fastgelu_generic.cpp)
target_link_libraries(client_gemm_add_fastgelu_generic PRIVATE composable_kernel::device_operations)
add_executable(client_gemm_fastgelu_generic gemm_fastgelu_generic.cpp)
target_link_libraries(client_gemm_fastgelu_generic PRIVATE composable_kernel::device_operations)
add_dependencies(client_gemm_fastgelu_generic_examples client_gemm_add_add_fastgelu_generic
client_gemm_add_fastgelu_generic client_gemm_fastgelu_generic)
client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp 0 → 100644
View file @ 3550cefe
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include <stdexcept>
#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/gemm_add_add_fastgelu.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 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 constexpr(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{}));
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout, D1Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType, D1DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddAddFastGelu>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
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{};
// get generic instance
auto& op_ptr = op_ptrs[0];
std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
// run the generic instance
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});
}
else
{
throw std::runtime_error(
"Generic instance should be suitable for various input lengths/strides");
}
std::cout << "Done" << std::endl;
return 0;
}
client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp
View file @ 3550cefe
......@@ -76,7 +76,7 @@ int main(int argc, char* argv[])
StrideA = std::stoi(argv[4]);
StrideB = std::stoi(argv[5]);
StrideD0 = std::stoi(argv[6]);
StrideE = std::stoi(argv[8]);
StrideE = std::stoi(argv[7]);
}
else
{
......
client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp 0 → 100644
View file @ 3550cefe
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include <stdexcept>
#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/gemm_add_fastgelu.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 AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddFastGelu;
using ADataType = F16;
using BDataType = F16;
using D0DataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using D0Layout = 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 StrideE = 4096;
if(argc == 1)
{
// use default case
}
else if(argc == 8)
{
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]);
StrideE = std::stoi(argv[7]);
}
else
{
printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD0, 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 constexpr(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 e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddFastGelu>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
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{};
// get generic instance
auto& op_ptr = op_ptrs[0];
std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
// run the generic instance
auto argument_ptr =
op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD0},
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});
}
else
{
throw std::runtime_error(
"Generic instance should be suitable for various input lengths/strides");
}
std::cout << "Done" << std::endl;
return 0;
}
client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp
View file @ 3550cefe
......@@ -72,7 +72,7 @@ int main(int argc, char* argv[])
StrideA = std::stoi(argv[4]);
StrideB = std::stoi(argv[5]);
StrideE = std::stoi(argv[8]);
StrideE = std::stoi(argv[6]);
}
else
{
......
client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp 0 → 100644
View file @ 3550cefe
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include <stdexcept>
#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/gemm_fastgelu.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 FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = FastGelu;
using ADataType = F16;
using BDataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
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 StrideE = 4096;
if(argc == 1)
{
// use default case
}
else if(argc == 7)
{
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]);
StrideE = std::stoi(argv[6]);
}
else
{
printf("arg1 to 6: M, N, K, StrideA, StrideB, 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 constexpr(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 e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<>,
ELayout,
ADataType,
BDataType,
ck::Tuple<>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::FastGelu>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
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{};
// get generic instance
auto& op_ptr = op_ptrs[0];
std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
// run the generic instance
auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{},
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});
}
else
{
throw std::runtime_error(
"Generic instance should be suitable for various input lengths/strides");
}
std::cout << "Done" << std::endl;
return 0;
}
client_example/16_convnd_fwd/common.hpp
View file @ 3550cefe
......@@ -141,14 +141,10 @@ bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialD
std::next(rbegin(in_strides)),
std::next(rbegin(in_strides), NumDimSpatial + 1));
std::rotate(
std::next(rbegin(wei_lengths)), std::next(rbegin(wei_lengths), 2), rend(wei_lengths));
std::rotate(rbegin(wei_lengths),
std::next(rbegin(wei_lengths)),
std::next(rbegin(wei_lengths), NumDimSpatial + 1));
std::rotate(
std::next(rbegin(wei_strides)), std::next(rbegin(wei_strides), 2), rend(wei_strides));
std::rotate(rbegin(wei_strides),
std::next(rbegin(wei_strides)),
std::next(rbegin(wei_strides), NumDimSpatial + 1));
......
client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp
View file @ 3550cefe
......@@ -11,7 +11,7 @@ using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::KZYXGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
static constexpr ck::index_t NumDimSpatial = 3;
......@@ -38,7 +38,7 @@ int main()
InLayout,
WeiLayout,
OutLayout>(
{N, Di, Hi, Wi, G, C}, {K, Z, Y, X, G, C}, {N, Do, Ho, Wo, G, K})
{N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
? EXIT_SUCCESS
: EXIT_FAILURE;
}
client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp
View file @ 3550cefe
......@@ -11,7 +11,7 @@ using WeiDataType = float;
using OutDataType = float;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::KZYXGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
static constexpr ck::index_t NumDimSpatial = 3;
......@@ -38,7 +38,7 @@ int main()
InLayout,
WeiLayout,
OutLayout>(
{N, Di, Hi, Wi, G, C}, {K, Z, Y, X, G, C}, {N, Do, Ho, Wo, G, K})
{N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
? EXIT_SUCCESS
: EXIT_FAILURE;
}
example/02_gemm_bilinear/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
list(APPEND gpu_list1 gfx1100 gfx1101 gfx1102)
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp)
endif()
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
set(target 1)
endif()
endforeach()
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
endif()
set(target 1)
endif()
endforeach()
example/03_gemm_bias_relu/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_example_executable(example_gemm_bias_relu_xdl_fp16 gemm_bias_relu_xdl_fp16.cpp)
endif()
\ No newline at end of file
set(target 1)
endif()
endforeach()
example/04_gemm_add_add_fastgelu/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_gemm_add_add_fastgelu_xdl)
add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp)
......@@ -16,4 +19,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4)
endif(USE_BITINT_EXTENSION_INT4)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8)
endif()
\ No newline at end of file
set(target 1)
endif()
endforeach()
\ No newline at end of file
example/09_convnd_fwd/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
endif()
set(target 1)
endif()
endforeach()
add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp)
add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp)
add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp)
example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_convnd_fwd_reduce_xdl)
add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
......@@ -12,4 +15,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M
add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4)
endif(USE_BITINT_EXTENSION_INT4)
endif()
\ No newline at end of file
set(target 1)
endif()
endforeach()
\ No newline at end of file
example/14_gemm_quantization/CMakeLists.txt
View file @ 3550cefe
......@@ -2,7 +2,12 @@
add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
# xdlops
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp)
add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp)
endif()
\ No newline at end of file
set(target 1)
endif()
endforeach()
\ No newline at end of file
example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_gemm_reduce_xdl)
add_custom_target(example_gemm_reduce_xdl_max)
add_custom_target(example_gemm_reduce_xdl_mean_meansquare)
......@@ -39,4 +42,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M
add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
endif()
endif()
set(target 1)
endif()
endforeach()
example/17_convnd_bwd_data/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp)
target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)
endif()
set(target 1)
endif()
endforeach()
add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
example/18_batched_gemm_reduce/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_example_executable(example_batched_gemm_reduce_xdl_fp16 batched_gemm_reduce_xdl_fp16.cpp)
endif()
set(target 1)
endif()
endforeach()
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