Skip to content

GitLab

Commit 63b152d6 authored by danyao12's avatar danyao12
Browse files

Merge branch 'develop' into ck_tile/fa_bwd_v3

parents ae2d7d2b 14c3cfb1
Hide whitespace changes
Inline Side-by-side
Showing with 1319 additions and 49 deletions
library/include/ck/library/reference_tensor_operation/gpu/reference_gemm.hpp 0 → 100644
View file @ 63b152d6
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename ComputeTypeA,
typename ComputeTypeB>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
naive_gemm_kernel(const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
CDataType* __restrict__ p_c_grid,
index_t m,
index_t n,
index_t k,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation c_element_op)
{
using RowMajor = ck::tensor_layout::gemm::RowMajor;
const int row_idx = blockIdx.x * blockDim.x + threadIdx.x;
const int col_idx = blockIdx.y * blockDim.y + threadIdx.y;
if(row_idx < m && col_idx < n)
{
AccDataType v_acc = static_cast<AccDataType>(0.0);
ComputeTypeA v_a = static_cast<ComputeTypeA>(0.0);
ComputeTypeB v_b = static_cast<ComputeTypeB>(0.0);
CDataType v_c = static_cast<CDataType>(0.0);
for(int k_idx = 0; k_idx < k; ++k_idx)
{
// check input matrices layout
int element_idx_a = 0;
int element_idx_b = 0;
if constexpr(std::is_same_v<ALayout, RowMajor>)
{
element_idx_a = row_idx * k + k_idx;
}
else
{
element_idx_a = row_idx + m * k_idx;
}
if constexpr(std::is_same_v<BLayout, RowMajor>)
{
element_idx_b = k_idx * n + col_idx;
}
else
{
element_idx_b = k_idx + k * col_idx;
}
// apply a_element_op
a_element_op(v_a, p_a_grid[element_idx_a]);
// apply b_element_op
b_element_op(v_b, p_b_grid[element_idx_b]);
// multiply and accumulate
v_acc += static_cast<AccDataType>(v_a) * static_cast<AccDataType>(v_b);
}
// apply c_element_op
c_element_op(v_c, v_acc);
// check output matrix layout
int element_idx_c = 0;
if constexpr(std::is_same_v<CLayout, RowMajor>)
{
element_idx_c = row_idx * n + col_idx;
}
else
{
element_idx_c = row_idx + m * col_idx;
}
// prepare output
p_c_grid[element_idx_c] = v_c;
}
}
} // namespace ck
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename ComputeTypeA = CDataType,
typename ComputeTypeB = ComputeTypeA>
struct ReferenceGemm : public device::BaseOperator
{
// Argument
struct Argument : public device::BaseArgument
{
Argument(const void* p_a_grid,
const void* p_b_grid,
void* p_c_grid,
index_t m,
index_t n,
index_t k,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
: p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
p_c_grid_{static_cast<CDataType*>(p_c_grid)},
m_{m},
n_{n},
k_{k},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
c_element_op_{c_element_op}
{
}
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
index_t m_;
index_t n_;
index_t k_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
};
// Invoker
struct Invoker : public device::BaseInvoker
{
using Argument = ReferenceGemm::Argument;
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
int block_size = 16;
dim3 block_dim(block_size, block_size, 1);
dim3 grid_dim(
(arg.m_ + block_size - 1) / block_size, (arg.n_ + block_size - 1) / block_size, 1);
auto launch_kernel = [&]() {
const auto kernel = naive_gemm_kernel<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
ComputeTypeA,
ComputeTypeB>;
return launch_and_time_kernel(stream_config,
kernel,
grid_dim,
block_dim,
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.m_,
arg.n_,
arg.k_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_);
};
return launch_kernel();
}
float Run(const device::BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
static auto MakeArgument(const void* p_a_grid,
const void* p_b_grid,
void* p_c_grid,
index_t m,
index_t n,
index_t k,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
return Argument{
p_a_grid, p_b_grid, p_c_grid, m, n, k, a_element_op, b_element_op, c_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "Device Reference Gemm"
<< std::endl;
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ 63b152d6
......@@ -37,11 +37,7 @@ function(add_instance_library INSTANCE_NAME)
endforeach()
endif()
if(INSTANCES_ONLY)
set(INST_TARGETS ${DEFAULT_GPU_TARGETS})
else()
set(INST_TARGETS ${GPU_TARGETS})
endif()
set(INST_TARGETS ${SUPPORTED_GPU_TARGETS})
# Do not build DL instances if DL_KERNELS macro is not set
foreach(source IN LISTS ARGN)
......@@ -64,9 +60,9 @@ function(add_instance_library INSTANCE_NAME)
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
# Do not build mha instances if gfx94 targets are not on the target list
# Do not build mha instances if gfx94 or gfx90a targets are not on the target list
foreach(source IN LISTS ARGN)
if(NOT INST_TARGETS MATCHES "gfx94" AND source MATCHES "mha")
if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx90a" AND source MATCHES "mha")
message("removing mha instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
......@@ -75,17 +71,13 @@ function(add_instance_library INSTANCE_NAME)
if(ARGN)
set(INST_OBJ)
foreach(source IN LISTS ARGN)
if(INSTANCES_ONLY)
set(INST_TARGETS ${DEFAULT_GPU_TARGETS})
else()
set(INST_TARGETS ${GPU_TARGETS})
endif()
set(INST_TARGETS ${SUPPORTED_GPU_TARGETS})
if(source MATCHES "_xdl")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
elseif(ARGN MATCHES "_wmma")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
elseif(ARGN MATCHES "mha")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
endif()
set(offload_targets)
foreach(target IN LISTS INST_TARGETS)
......@@ -102,12 +94,14 @@ function(add_instance_library INSTANCE_NAME)
set(FMHA_FWD_FAST_EXP2 true)
endif()
if(FMHA_FWD_FAST_EXP2)
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=1 -fgpu-flush-denormals-to-zero)
list(APPEND FMHA_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=1 -fgpu-flush-denormals-to-zero)
else()
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=0)
list(APPEND FMHA_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=0)
endif()
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-float-equal)
target_compile_options(device_mha_instance PRIVATE ${EXAMPLE_FMHA_FWD_COMPILE_OPTIONS})
list(APPEND FMHA_COMPILE_OPTIONS -Wno-float-equal)
list(APPEND FMHA_COMPILE_OPTIONS -DCK_TILE_FMHA_FWD_SPLITKV_API=1)
list(APPEND FMHA_COMPILE_OPTIONS -DCK_TILE_FMHA_FWD_APPENDKV_API=1)
target_compile_options(device_mha_instance PRIVATE ${FMHA_COMPILE_OPTIONS})
endif()
target_compile_features(${INSTANCE_NAME} PUBLIC)
......@@ -189,12 +183,7 @@ FOREACH(subdir_path ${dir_list})
set(add_inst 1)
endif()
if(INSTANCES_ONLY)
set(INST_TARGETS ${DEFAULT_GPU_TARGETS})
else()
set(INST_TARGETS ${GPU_TARGETS})
endif()
set(INST_TARGETS ${SUPPORTED_GPU_TARGETS})
if(("${cmake_instance}" MATCHES "quantization") AND (DEFINED DTYPES) AND (NOT DTYPES MATCHES "int8"))
message("quantization instances will not be built!")
......@@ -318,8 +307,7 @@ if(CK_DEVICE_CONV_INSTANCES)
endif()
if(CK_DEVICE_MHA_INSTANCES)
set(gpu_list ${INST_TARGETS})
list(FILTER gpu_list INCLUDE REGEX "^gfx94")
if(gpu_list)
if(gpu_list MATCHES "gfx94" OR gpu_list MATCHES "gfx90a")
add_library(device_mha_operations STATIC ${CK_DEVICE_MHA_INSTANCES})
add_library(composablekernels::device_mha_operations ALIAS device_mha_operations)
target_compile_features(device_mha_operations PUBLIC)
......
library/src/tensor_operation_instance/gpu/mha/CMakeLists.txt
View file @ 63b152d6
......@@ -32,23 +32,33 @@ if(EXISTS ${FMHA_CPP_FOLDER}/blob_list.txt)
file(REMOVE ${FMHA_CPP_FOLDER}/blob_list.txt)
endif()
set(FMHA_KNOWN_APIS "fwd,fwd_splitkv,fwd_appendkv,bwd")
# generate a list of kernels, but not actually emit files at config stage
# Note: The receipt 3 arg filters the generated backwards instances to reduce compilation time.
# With receipt 3 set, we are generating instances for datatype == {fp16 || bfp16}, bias == {no || alibi}, deterministic == off, and dpad == dvpad.
execute_process(
COMMAND ${PYTHON_EXECUTABLE} ${CMAKE_SOURCE_DIR}/example/ck_tile/01_fmha/generate.py
COMMAND ${PYTHON_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
--list_blobs ${FMHA_CPP_FOLDER}/blob_list.txt
--api ${FMHA_KNOWN_APIS}
--receipt 3
RESULT_VARIABLE ret
)
if(ret AND NOT ret EQUAL 0)
message( FATAL_ERROR "CK Tile MHA FAILED to genrate a list of kernels via Python.")
else()
file(STRINGS ${FMHA_CPP_FOLDER}/blob_list.txt FMHA_FWD_GEN_BLOBS)
file(STRINGS ${FMHA_CPP_FOLDER}/blob_list.txt FMHA_GEN_BLOBS)
endif()
# actually generate the kernel content now
# Note: The receipt 3 arg filters the generated backwards instances to reduce compilation time.
# With receipt 3 set, we are generating instances for datatype == {fp16 || bfp16}, bias == {no || alibi}, deterministic == off, and dpad == dvpad.
add_custom_command(
OUTPUT ${FMHA_FWD_GEN_BLOBS}
COMMAND ${PYTHON_EXECUTABLE} ${CMAKE_SOURCE_DIR}/example/ck_tile/01_fmha/generate.py
OUTPUT ${FMHA_GEN_BLOBS}
COMMAND ${PYTHON_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
--output_dir ${FMHA_CPP_FOLDER}
--api ${FMHA_KNOWN_APIS}
--receipt 3
COMMENT "Generating mha kernel (cpp) files now ..."
VERBATIM
)
......@@ -57,12 +67,12 @@ add_custom_command(
# have filename. Since, it was cauing the cmake
# to throw "File name too long"
set(device_files)
foreach(filepath IN LISTS FMHA_FWD_GEN_BLOBS)
foreach(filepath IN LISTS FMHA_GEN_BLOBS)
get_filename_component(filename ${filepath} NAME)
# Append the filename to the device_files list
list(APPEND device_files ${filename})
endforeach()
add_custom_target(generate_cpp_files DEPENDS ${FMHA_FWD_GEN_BLOBS})
add_custom_target(generate_cpp_files DEPENDS ${FMHA_GEN_BLOBS})
add_instance_library(device_mha_instance ${device_files})
......
profiler/src/CMakeLists.txt
View file @ 63b152d6
......@@ -24,7 +24,7 @@ set(PROFILER_SOURCES
profile_permute_scale.cpp
)
if(GPU_TARGETS MATCHES "gfx9")
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
......@@ -49,7 +49,7 @@ if(GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
endif()
list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
if(GPU_TARGETS MATCHES "gfx94")
if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
endif()
......@@ -69,7 +69,7 @@ if(GPU_TARGETS MATCHES "gfx9")
endif()
if(GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12" OR GPU_TARGETS MATCHES "gfx9")
if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
endif()
......@@ -111,7 +111,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_inst
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
if(GPU_TARGETS MATCHES "gfx9")
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
......@@ -135,7 +135,7 @@ if(GPU_TARGETS MATCHES "gfx9")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
if(GPU_TARGETS MATCHES "gfx94")
if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
endif()
......@@ -159,7 +159,7 @@ if(GPU_TARGETS MATCHES "gfx9")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance)
endif()
if(GPU_TARGETS MATCHES "gfx9" OR GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12")
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
endif()
......
script/cmake-ck-dev.sh
View file @ 63b152d6
......@@ -7,8 +7,11 @@ MY_PROJECT_SOURCE=$1
if [ $# -ge 2 ] ; then
GPU_TARGETS=$2
shift 2
REST_ARGS=$@
else
GPU_TARGETS="gfx908;gfx90a;gfx940"
REST_ARGS=
fi
cmake \
......@@ -20,4 +23,5 @@ cmake
-D GPU_TARGETS=$GPU_TARGETS \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D USE_BITINT_EXTENSION_INT4=OFF \
$REST_ARGS \
${MY_PROJECT_SOURCE}
script/cmake-ck-release.sh
View file @ 63b152d6
......@@ -7,8 +7,11 @@ MY_PROJECT_SOURCE=$1
if [ $# -ge 2 ] ; then
GPU_TARGETS=$2
shift 2
REST_ARGS=$@
else
GPU_TARGETS="gfx908;gfx90a;gfx940"
REST_ARGS=
fi
cmake \
......@@ -20,5 +23,6 @@ cmake
-D GPU_TARGETS=$GPU_TARGETS \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D USE_BITINT_EXTENSION_INT4=OFF \
$REST_ARGS \
${MY_PROJECT_SOURCE}
test/CMakeLists.txt
View file @ 63b152d6
......@@ -41,11 +41,7 @@ function(add_test_executable TEST_NAME)
endforeach()
endif()
if(INSTANCES_ONLY)
set(TEST_TARGETS ${DEFAULT_GPU_TARGETS})
else()
set(TEST_TARGETS ${GPU_TARGETS})
endif()
set(TEST_TARGETS ${SUPPORTED_GPU_TARGETS})
foreach(source IN LISTS ARGN)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
......@@ -122,11 +118,7 @@ function(add_gtest_executable TEST_NAME)
endforeach()
endif()
if(INSTANCES_ONLY)
set(TEST_TARGETS ${DEFAULT_GPU_TARGETS})
else()
set(TEST_TARGETS ${GPU_TARGETS})
endif()
set(TEST_TARGETS ${SUPPORTED_GPU_TARGETS})
foreach(source IN LISTS ARGN)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
......@@ -173,6 +165,7 @@ function(add_gtest_executable TEST_NAME)
endfunction()
add_compile_options(-Wno-c++20-extensions)
add_subdirectory(ck_tile)
add_subdirectory(magic_number_division)
add_subdirectory(space_filling_curve)
add_subdirectory(conv_util)
......@@ -210,10 +203,10 @@ add_subdirectory(conv_tensor_rearrange)
add_subdirectory(transpose)
add_subdirectory(permute_scale)
add_subdirectory(wrapper)
if(GPU_TARGETS MATCHES "gfx11")
if(SUPPORTED_GPU_TARGETS MATCHES "gfx11")
add_subdirectory(wmma_op)
endif()
if(GPU_TARGETS MATCHES "gfx942" AND CK_HIP_VERSION_MAJOR GREATER_EQUAL 6 AND CK_HIP_VERSION_MINOR GREATER_EQUAL 2) # smfmac needs ROCm6.2
if(SUPPORTED_GPU_TARGETS MATCHES "gfx942" AND CK_HIP_VERSION_MAJOR GREATER_EQUAL 6 AND CK_HIP_VERSION_MINOR GREATER_EQUAL 2) # smfmac needs ROCm6.2
add_subdirectory(smfmac_op)
endif()
add_subdirectory(position_embedding)
test/ck_tile/CMakeLists.txt 0 → 100644
View file @ 63b152d6
add_subdirectory(image_to_column)
test/ck_tile/image_to_column/CMakeLists.txt 0 → 100644
View file @ 63b152d6
# Currently ck_tile is only built on gfx9
if(GPU_TARGETS MATCHES "gfx9")
add_gtest_executable(test_tile_image_to_column test_tile_image_to_column.cpp)
endif()
test/ck_tile/image_to_column/test_tile_image_to_column.cpp 0 → 100644
View file @ 63b152d6
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <gtest/gtest.h>
#include "ck_tile/host.hpp"
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/image_to_column.hpp"
// Host API implementation
template <typename DataType>
class TestCkTileImageToColumn : public ::testing::Test
{
static constexpr ck_tile::index_t VectorSize = 1;
static constexpr ck_tile::index_t NDimSpatial = 2;
protected:
void Run(const ck_tile::conv::ConvParam conv_params)
{
using ImLayout = ck_tile::tensor_layout::convolution::NHWGC;
const auto G = conv_params.G_;
const auto N = conv_params.N_;
const auto C = conv_params.C_;
const ck_tile::long_index_t NDoHoWo =
N * std::accumulate(conv_params.output_spatial_lengths_.begin(),
std::next(conv_params.output_spatial_lengths_.begin(), NDimSpatial),
1,
std::multiplies<>());
const ck_tile::long_index_t CZYX =
C * std::accumulate(conv_params.filter_spatial_lengths_.begin(),
std::next(conv_params.filter_spatial_lengths_.begin(), NDimSpatial),
1,
std::multiplies<>());
const auto in_desc =
ck_tile::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ImLayout>(
conv_params);
const auto out_desc = ck_tile::HostTensorDescriptor({G, NDoHoWo, CZYX});
// host verify
ck_tile::HostTensor<DataType> in(in_desc);
ck_tile::HostTensor<DataType> out_device(out_desc);
ck_tile::HostTensor<DataType> out_host(out_desc);
std::cout << "input: " << in.mDesc << std::endl;
std::cout << "output: " << out_device.mDesc << std::endl;
ck_tile::FillUniformDistributionIntegerValue<DataType>{-5.f, 5.f}(in);
ck_tile::DeviceMem in_device_buf(in.get_element_space_size_in_bytes());
ck_tile::DeviceMem out_device_buf(out_device.get_element_space_size_in_bytes());
in_device_buf.ToDevice(in.data());
using thread_tile = ck_tile::sequence<4, 4>;
using warp_tile = ck_tile::sequence<8, 128>;
using block_tile = ck_tile::sequence<32, 128>;
using Shape = ck_tile::TileImageToColumnShape<thread_tile, warp_tile, block_tile>;
using PipelineProblem = ck_tile::BlockImageToColumnProblem<DataType,
DataType,
Shape,
NDimSpatial,
VectorSize,
VectorSize>;
using Kernel = ck_tile::ImageToColumn<PipelineProblem>;
auto kargs = Kernel::MakeKargs(
in_device_buf.GetDeviceBuffer(),
out_device_buf.GetDeviceBuffer(),
G,
N,
C,
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(
conv_params.input_spatial_lengths_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(
conv_params.filter_spatial_lengths_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(
conv_params.output_spatial_lengths_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial + 3>(in_desc.get_strides()),
ck_tile::to_array<ck_tile::long_index_t, 3>(out_desc.get_strides()),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.conv_filter_strides_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(
conv_params.conv_filter_dilations_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.input_left_pads_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.input_right_pads_));
const dim3 grids = Kernel::GridSize(
kargs.N * kargs.output_spatial_lengths[0] * kargs.output_spatial_lengths[1],
kargs.filter_spatial_lengths[0] * kargs.filter_spatial_lengths[1] * kargs.C,
kargs.G);
constexpr dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 2;
ck_tile::launch_kernel(
ck_tile::stream_config{},
ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
// reference
ck_tile::reference_im2col<DataType, DataType, NDimSpatial>(in, out_host, conv_params);
out_device_buf.FromDevice(out_device.data());
bool pass = ck_tile::check_err(out_device, out_host);
EXPECT_TRUE(pass);
}
};
class TestCkTileImageToColumnFloat : public TestCkTileImageToColumn<float>
{
};
class TestCkTileImageToColumnHalf : public TestCkTileImageToColumn<ck_tile::half_t>
{
};
TEST_F(TestCkTileImageToColumnFloat, TestCorrectness)
{
this->Run({2, 2, 4, 1, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->Run({2, 2, 64, 1, 64, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->Run({2, 1, 64, 1, 64, {1, 1}, {7, 7}, {3, 3}, {1, 1}, {0, 0}, {0, 0}});
this->Run({2, 1, 64, 1, 64, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->Run({2, 2, 64, 1, 64, {3, 3}, {28, 28}, {2, 2}, {2, 2}, {1, 1}, {1, 1}});
}
TEST_F(TestCkTileImageToColumnHalf, TestCorrectness)
{
this->Run({2, 2, 4, 1, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->Run({2, 2, 64, 1, 64, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->Run({2, 1, 64, 1, 64, {1, 1}, {7, 7}, {3, 3}, {1, 1}, {0, 0}, {0, 0}});
this->Run({2, 1, 64, 1, 64, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->Run({2, 2, 64, 1, 64, {3, 3}, {28, 28}, {2, 2}, {2, 2}, {1, 1}, {1, 1}});
}
test/data_type/CMakeLists.txt
View file @ 63b152d6
......@@ -18,4 +18,9 @@ if(result EQUAL 0)
target_link_libraries(test_bf8 PRIVATE utility)
endif()
add_gtest_executable(test_custom_type test_custom_type.cpp)
if(result EQUAL 0)
target_link_libraries(test_custom_type PRIVATE utility)
endif()
add_gtest_executable(test_type_convert_const type_convert_const.cpp)
test/data_type/test_custom_type.cpp 0 → 100644
View file @ 63b152d6
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/utility/data_type.hpp"
#include "ck/utility/type_convert.hpp"
using ck::bf8_t;
using ck::bhalf_t;
using ck::f8_t;
using ck::half_t;
using ck::Number;
using ck::type_convert;
using ck::vector_type;
TEST(Custom_bool, TestSize)
{
struct custom_bool_t
{
bool data;
};
ASSERT_EQ(sizeof(custom_bool_t), sizeof(bool));
ASSERT_EQ(sizeof(vector_type<custom_bool_t, 2>), sizeof(vector_type<bool, 2>));
ASSERT_EQ(sizeof(vector_type<custom_bool_t, 4>), sizeof(vector_type<bool, 4>));
ASSERT_EQ(sizeof(vector_type<custom_bool_t, 8>), sizeof(vector_type<bool, 8>));
ASSERT_EQ(sizeof(vector_type<custom_bool_t, 16>), sizeof(vector_type<bool, 16>));
ASSERT_EQ(sizeof(vector_type<custom_bool_t, 32>), sizeof(vector_type<bool, 32>));
ASSERT_EQ(sizeof(vector_type<custom_bool_t, 64>), sizeof(vector_type<bool, 64>));
}
TEST(Custom_bool, TestAsType)
{
struct custom_bool_t
{
using type = bool;
type data;
custom_bool_t() : data{type{}} {}
custom_bool_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<bool> test_vec = {false, true, false, true};
// reference vector
vector_type<custom_bool_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_bool_t>()(Number<i>{}).data, false);
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_bool_t>()(Number<i>{}) = custom_bool_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_bool_t, size> left_vec{right_vec};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_bool_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_bool, TestAsTypeReshape)
{
struct custom_bool_t
{
using type = bool;
type data;
custom_bool_t() : data{type{}} {}
custom_bool_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<bool> test_vec = {false, true, false, true};
// reference vector
vector_type<custom_bool_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_bool_t>()(Number<i>{}).data, false);
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_bool_t>()(Number<i>{}) = custom_bool_t{test_vec.at(i)};
});
// copy the first half of a vector
vector_type<custom_bool_t, size / 2> left_vec{
right_vec.template AsType<vector_type<custom_bool_t, size / 2>::type>()(Number<0>{})};
// check if values were copied correctly
ck::static_for<0, size / 2, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_bool_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_int8, TestSize)
{
struct custom_int8_t
{
int8_t data;
};
ASSERT_EQ(sizeof(custom_int8_t), sizeof(int8_t));
ASSERT_EQ(sizeof(vector_type<custom_int8_t, 2>), sizeof(vector_type<int8_t, 2>));
ASSERT_EQ(sizeof(vector_type<custom_int8_t, 4>), sizeof(vector_type<int8_t, 4>));
ASSERT_EQ(sizeof(vector_type<custom_int8_t, 8>), sizeof(vector_type<int8_t, 8>));
ASSERT_EQ(sizeof(vector_type<custom_int8_t, 16>), sizeof(vector_type<int8_t, 16>));
ASSERT_EQ(sizeof(vector_type<custom_int8_t, 32>), sizeof(vector_type<int8_t, 32>));
ASSERT_EQ(sizeof(vector_type<custom_int8_t, 64>), sizeof(vector_type<int8_t, 64>));
}
TEST(Custom_int8, TestAsType)
{
struct custom_int8_t
{
using type = int8_t;
type data;
custom_int8_t() : data{type{}} {}
custom_int8_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<int8_t> test_vec = {3, -6, 8, -2};
// reference vector
vector_type<custom_int8_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_int8_t>()(Number<i>{}).data, 0);
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_int8_t>()(Number<i>{}) = custom_int8_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_int8_t, size> left_vec{right_vec};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_int8_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_int8, TestAsTypeReshape)
{
struct custom_int8_t
{
using type = int8_t;
type data;
custom_int8_t() : data{type{}} {}
custom_int8_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<int8_t> test_vec = {3, -6, 8, -2};
// reference vector
vector_type<custom_int8_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_int8_t>()(Number<i>{}).data, 0);
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_int8_t>()(Number<i>{}) = custom_int8_t{test_vec.at(i)};
});
// copy the first half of a vector
vector_type<custom_int8_t, size / 2> left_vec{
right_vec.template AsType<vector_type<custom_int8_t, size / 2>::type>()(Number<0>{})};
// check if values were copied correctly
ck::static_for<0, size / 2, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_int8_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_uint8, TestSize)
{
struct custom_uint8_t
{
uint8_t data;
};
ASSERT_EQ(sizeof(custom_uint8_t), sizeof(uint8_t));
ASSERT_EQ(sizeof(vector_type<custom_uint8_t, 2>), sizeof(vector_type<uint8_t, 2>));
ASSERT_EQ(sizeof(vector_type<custom_uint8_t, 4>), sizeof(vector_type<uint8_t, 4>));
ASSERT_EQ(sizeof(vector_type<custom_uint8_t, 8>), sizeof(vector_type<uint8_t, 8>));
ASSERT_EQ(sizeof(vector_type<custom_uint8_t, 16>), sizeof(vector_type<uint8_t, 16>));
ASSERT_EQ(sizeof(vector_type<custom_uint8_t, 32>), sizeof(vector_type<uint8_t, 32>));
ASSERT_EQ(sizeof(vector_type<custom_uint8_t, 64>), sizeof(vector_type<uint8_t, 64>));
}
TEST(Custom_uint8, TestAsType)
{
struct custom_uint8_t
{
using type = uint8_t;
type data;
custom_uint8_t() : data{type{}} {}
custom_uint8_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<uint8_t> test_vec = {3, 6, 8, 2};
// reference vector
vector_type<custom_uint8_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_uint8_t>()(Number<i>{}).data, 0);
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_uint8_t>()(Number<i>{}) = custom_uint8_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_uint8_t, size> left_vec{right_vec};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_uint8_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_uint8, TestAsTypeReshape)
{
struct custom_uint8_t
{
using type = uint8_t;
type data;
custom_uint8_t() : data{type{}} {}
custom_uint8_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<uint8_t> test_vec = {3, 6, 8, 2};
// reference vector
vector_type<custom_uint8_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_uint8_t>()(Number<i>{}).data, 0);
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_uint8_t>()(Number<i>{}) = custom_uint8_t{test_vec.at(i)};
});
// copy the first half of a vector
vector_type<custom_uint8_t, size / 2> left_vec{
right_vec.template AsType<vector_type<custom_uint8_t, size / 2>::type>()(Number<0>{})};
// check if values were copied correctly
ck::static_for<0, size / 2, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_uint8_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_f8, TestSize)
{
struct custom_f8_t
{
_BitInt(8) data;
};
ASSERT_EQ(sizeof(custom_f8_t), sizeof(_BitInt(8)));
ASSERT_EQ(sizeof(vector_type<custom_f8_t, 2>), sizeof(vector_type<_BitInt(8), 2>));
ASSERT_EQ(sizeof(vector_type<custom_f8_t, 4>), sizeof(vector_type<_BitInt(8), 4>));
ASSERT_EQ(sizeof(vector_type<custom_f8_t, 8>), sizeof(vector_type<_BitInt(8), 8>));
ASSERT_EQ(sizeof(vector_type<custom_f8_t, 16>), sizeof(vector_type<_BitInt(8), 16>));
ASSERT_EQ(sizeof(vector_type<custom_f8_t, 32>), sizeof(vector_type<_BitInt(8), 32>));
ASSERT_EQ(sizeof(vector_type<custom_f8_t, 64>), sizeof(vector_type<_BitInt(8), 64>));
}
TEST(Custom_f8, TestAsType)
{
struct custom_f8_t
{
using type = _BitInt(8);
type data;
custom_f8_t() : data{type{}} {}
custom_f8_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<_BitInt(8)> test_vec = {type_convert<_BitInt(8)>(0.3f),
type_convert<_BitInt(8)>(-0.6f),
type_convert<_BitInt(8)>(0.8f),
type_convert<_BitInt(8)>(-0.2f)};
// reference vector
vector_type<custom_f8_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}(
[&](auto i) { ASSERT_EQ(right_vec.template AsType<custom_f8_t>()(Number<i>{}).data, 0); });
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_f8_t>()(Number<i>{}) = custom_f8_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_f8_t, size> left_vec{right_vec};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_f8_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_f8, TestAsTypeReshape)
{
struct custom_f8_t
{
using type = _BitInt(8);
type data;
custom_f8_t() : data{type{}} {}
custom_f8_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<_BitInt(8)> test_vec = {type_convert<_BitInt(8)>(0.3f),
type_convert<_BitInt(8)>(-0.6f),
type_convert<_BitInt(8)>(0.8f),
type_convert<_BitInt(8)>(-0.2f)};
// reference vector
vector_type<custom_f8_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}(
[&](auto i) { ASSERT_EQ(right_vec.template AsType<custom_f8_t>()(Number<i>{}).data, 0); });
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_f8_t>()(Number<i>{}) = custom_f8_t{test_vec.at(i)};
});
// copy the first half of a vector
vector_type<custom_f8_t, size / 2> left_vec{
right_vec.template AsType<vector_type<custom_f8_t, size / 2>::type>()(Number<0>{})};
// check if values were copied correctly
ck::static_for<0, size / 2, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_f8_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_bf8, TestSize)
{
struct custom_bf8_t
{
unsigned _BitInt(8) data;
};
ASSERT_EQ(sizeof(custom_bf8_t), sizeof(unsigned _BitInt(8)));
ASSERT_EQ(sizeof(vector_type<custom_bf8_t, 2>), sizeof(vector_type<unsigned _BitInt(8), 2>));
ASSERT_EQ(sizeof(vector_type<custom_bf8_t, 4>), sizeof(vector_type<unsigned _BitInt(8), 4>));
ASSERT_EQ(sizeof(vector_type<custom_bf8_t, 8>), sizeof(vector_type<unsigned _BitInt(8), 8>));
ASSERT_EQ(sizeof(vector_type<custom_bf8_t, 16>), sizeof(vector_type<unsigned _BitInt(8), 16>));
ASSERT_EQ(sizeof(vector_type<custom_bf8_t, 32>), sizeof(vector_type<unsigned _BitInt(8), 32>));
ASSERT_EQ(sizeof(vector_type<custom_bf8_t, 64>), sizeof(vector_type<unsigned _BitInt(8), 64>));
}
TEST(Custom_bf8, TestAsType)
{
struct custom_bf8_t
{
using type = unsigned _BitInt(8);
type data;
custom_bf8_t() : data{type{}} {}
custom_bf8_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<unsigned _BitInt(8)> test_vec = {type_convert<unsigned _BitInt(8)>(0.3f),
type_convert<unsigned _BitInt(8)>(-0.6f),
type_convert<unsigned _BitInt(8)>(0.8f),
type_convert<unsigned _BitInt(8)>(-0.2f)};
// reference vector
vector_type<custom_bf8_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}(
[&](auto i) { ASSERT_EQ(right_vec.template AsType<custom_bf8_t>()(Number<i>{}).data, 0); });
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_bf8_t>()(Number<i>{}) = custom_bf8_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_bf8_t, size> left_vec{right_vec};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_bf8_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_bf8, TestAsTypeReshape)
{
struct custom_bf8_t
{
using type = unsigned _BitInt(8);
type data;
custom_bf8_t() : data{type{}} {}
custom_bf8_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<unsigned _BitInt(8)> test_vec = {type_convert<unsigned _BitInt(8)>(0.3f),
type_convert<unsigned _BitInt(8)>(-0.6f),
type_convert<unsigned _BitInt(8)>(0.8f),
type_convert<unsigned _BitInt(8)>(-0.2f)};
// reference vector
vector_type<custom_bf8_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}(
[&](auto i) { ASSERT_EQ(right_vec.template AsType<custom_bf8_t>()(Number<i>{}).data, 0); });
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_bf8_t>()(Number<i>{}) = custom_bf8_t{test_vec.at(i)};
});
// copy the first half of a vector
vector_type<custom_bf8_t, size / 2> left_vec{
right_vec.template AsType<vector_type<custom_bf8_t, size / 2>::type>()(Number<0>{})};
// check if values were copied correctly
ck::static_for<0, size / 2, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_bf8_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_half, TestSize)
{
struct custom_half_t
{
half_t data;
};
ASSERT_EQ(sizeof(custom_half_t), sizeof(half_t));
ASSERT_EQ(sizeof(vector_type<custom_half_t, 2>), sizeof(vector_type<half_t, 2>));
ASSERT_EQ(sizeof(vector_type<custom_half_t, 4>), sizeof(vector_type<half_t, 4>));
ASSERT_EQ(sizeof(vector_type<custom_half_t, 8>), sizeof(vector_type<half_t, 8>));
ASSERT_EQ(sizeof(vector_type<custom_half_t, 16>), sizeof(vector_type<half_t, 16>));
ASSERT_EQ(sizeof(vector_type<custom_half_t, 32>), sizeof(vector_type<half_t, 32>));
ASSERT_EQ(sizeof(vector_type<custom_half_t, 64>), sizeof(vector_type<half_t, 64>));
}
TEST(Custom_half, TestAsType)
{
struct custom_half_t
{
using type = half_t;
type data;
custom_half_t() : data{type{}} {}
custom_half_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<half_t> test_vec = {half_t{0.3f}, half_t{-0.6f}, half_t{0.8f}, half_t{-0.2f}};
// reference vector
vector_type<custom_half_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_half_t>()(Number<i>{}).data,
type_convert<half_t>(0.0f));
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_half_t>()(Number<i>{}) = custom_half_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_half_t, size> left_vec{right_vec};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_half_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_half, TestAsTypeReshape)
{
struct custom_half_t
{
using type = half_t;
type data;
custom_half_t() : data{type{}} {}
custom_half_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<half_t> test_vec = {half_t{0.3f}, half_t{-0.6f}, half_t{0.8f}, half_t{-0.2f}};
// reference vector
vector_type<custom_half_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_half_t>()(Number<i>{}).data,
type_convert<half_t>(0.0f));
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_half_t>()(Number<i>{}) = custom_half_t{test_vec.at(i)};
});
// copy the first half of a vector
vector_type<custom_half_t, size / 2> left_vec{
right_vec.template AsType<vector_type<custom_half_t, size / 2>::type>()(Number<0>{})};
// check if values were copied correctly
ck::static_for<0, size / 2, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_half_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_bhalf, TestSize)
{
struct custom_bhalf_t
{
bhalf_t data;
};
ASSERT_EQ(sizeof(custom_bhalf_t), sizeof(bhalf_t));
ASSERT_EQ(sizeof(vector_type<custom_bhalf_t, 2>), sizeof(vector_type<bhalf_t, 2>));
ASSERT_EQ(sizeof(vector_type<custom_bhalf_t, 4>), sizeof(vector_type<bhalf_t, 4>));
ASSERT_EQ(sizeof(vector_type<custom_bhalf_t, 8>), sizeof(vector_type<bhalf_t, 8>));
ASSERT_EQ(sizeof(vector_type<custom_bhalf_t, 16>), sizeof(vector_type<bhalf_t, 16>));
ASSERT_EQ(sizeof(vector_type<custom_bhalf_t, 32>), sizeof(vector_type<bhalf_t, 32>));
ASSERT_EQ(sizeof(vector_type<custom_bhalf_t, 64>), sizeof(vector_type<bhalf_t, 64>));
}
TEST(Custom_bhalf, TestAsType)
{
struct custom_bhalf_t
{
using type = bhalf_t;
type data;
custom_bhalf_t() : data{type{}} {}
custom_bhalf_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<bhalf_t> test_vec = {type_convert<bhalf_t>(0.3f),
type_convert<bhalf_t>(-0.6f),
type_convert<bhalf_t>(0.8f),
type_convert<bhalf_t>(-0.2f)};
// reference vector
vector_type<custom_bhalf_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_bhalf_t>()(Number<i>{}).data,
type_convert<bhalf_t>(0.0f));
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_bhalf_t>()(Number<i>{}) = custom_bhalf_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_bhalf_t, size> left_vec{right_vec};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_bhalf_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_bhalf, TestAsTypeReshape)
{
struct custom_bhalf_t
{
using type = bhalf_t;
type data;
custom_bhalf_t() : data{type{}} {}
custom_bhalf_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<bhalf_t> test_vec = {type_convert<bhalf_t>(0.3f),
type_convert<bhalf_t>(-0.6f),
type_convert<bhalf_t>(0.8f),
type_convert<bhalf_t>(-0.2f)};
// reference vector
vector_type<custom_bhalf_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_bhalf_t>()(Number<i>{}).data,
type_convert<bhalf_t>(0.0f));
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_bhalf_t>()(Number<i>{}) = custom_bhalf_t{test_vec.at(i)};
});
// copy the first half of a vector
vector_type<custom_bhalf_t, size / 2> left_vec{
right_vec.template AsType<vector_type<custom_bhalf_t, size / 2>::type>()(Number<0>{})};
// check if values were copied correctly
ck::static_for<0, size / 2, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_bhalf_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_float, TestSize)
{
struct custom_float_t
{
float data;
};
ASSERT_EQ(sizeof(custom_float_t), sizeof(float));
ASSERT_EQ(sizeof(vector_type<custom_float_t, 2>), sizeof(vector_type<float, 2>));
ASSERT_EQ(sizeof(vector_type<custom_float_t, 4>), sizeof(vector_type<float, 4>));
ASSERT_EQ(sizeof(vector_type<custom_float_t, 8>), sizeof(vector_type<float, 8>));
ASSERT_EQ(sizeof(vector_type<custom_float_t, 16>), sizeof(vector_type<float, 16>));
ASSERT_EQ(sizeof(vector_type<custom_float_t, 32>), sizeof(vector_type<float, 32>));
ASSERT_EQ(sizeof(vector_type<custom_float_t, 64>), sizeof(vector_type<float, 64>));
}
TEST(Custom_float, TestAsType)
{
struct custom_float_t
{
using type = float;
type data;
custom_float_t() : data{type{}} {}
custom_float_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<float> test_vec = {0.3f, -0.6f, 0.8f, -0.2f};
// reference vector
vector_type<custom_float_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_float_t>()(Number<i>{}).data, 0.0f);
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_float_t>()(Number<i>{}) = custom_float_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_float_t, size> left_vec{right_vec};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_float_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_float, TestAsTypeReshape)
{
struct custom_float_t
{
using type = float;
type data;
custom_float_t() : data{type{}} {}
custom_float_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<float> test_vec = {0.3f, -0.6f, 0.8f, -0.2f};
// reference vector
vector_type<custom_float_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_float_t>()(Number<i>{}).data, 0.0f);
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_float_t>()(Number<i>{}) = custom_float_t{test_vec.at(i)};
});
// copy the first half of a vector
vector_type<custom_float_t, size / 2> left_vec{
right_vec.template AsType<vector_type<custom_float_t, size / 2>::type>()(Number<0>{})};
// check if values were copied correctly
ck::static_for<0, size / 2, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_float_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_double, TestSize)
{
struct custom_double_t
{
double data;
};
ASSERT_EQ(sizeof(custom_double_t), sizeof(double));
ASSERT_EQ(sizeof(vector_type<custom_double_t, 2>), sizeof(vector_type<double, 2>));
ASSERT_EQ(sizeof(vector_type<custom_double_t, 4>), sizeof(vector_type<double, 4>));
ASSERT_EQ(sizeof(vector_type<custom_double_t, 8>), sizeof(vector_type<double, 8>));
ASSERT_EQ(sizeof(vector_type<custom_double_t, 16>), sizeof(vector_type<double, 16>));
ASSERT_EQ(sizeof(vector_type<custom_double_t, 32>), sizeof(vector_type<double, 32>));
ASSERT_EQ(sizeof(vector_type<custom_double_t, 64>), sizeof(vector_type<double, 64>));
}
TEST(Custom_double, TestAsType)
{
struct custom_double_t
{
using type = double;
type data;
custom_double_t() : data{type{}} {}
custom_double_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<double> test_vec = {0.3, 0.6, 0.8, 0.2};
// reference vector
vector_type<custom_double_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_double_t>()(Number<i>{}).data, 0.0);
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_double_t>()(Number<i>{}) = custom_double_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_double_t, size> left_vec{right_vec};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_double_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Custom_double, TestAsTypeReshape)
{
struct custom_double_t
{
using type = double;
type data;
custom_double_t() : data{type{}} {}
custom_double_t(type init) : data{init} {}
};
// test size
const int size = 4;
std::vector<double> test_vec = {0.3, 0.6, 0.8, 0.2};
// reference vector
vector_type<custom_double_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<custom_double_t>()(Number<i>{}).data, 0.0);
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_double_t>()(Number<i>{}) = custom_double_t{test_vec.at(i)};
});
// copy the first half of a vector
vector_type<custom_double_t, size / 2> left_vec{
right_vec.template AsType<vector_type<custom_double_t, size / 2>::type>()(Number<0>{})};
// check if values were copied correctly
ck::static_for<0, size / 2, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_double_t>()(Number<i>{}).data, test_vec.at(i));
});
}
TEST(Complex_half, TestSize)
{
struct complex_half_t
{
half_t real;
half_t img;
};
ASSERT_EQ(sizeof(complex_half_t), sizeof(half_t) + sizeof(half_t));
ASSERT_EQ(sizeof(vector_type<complex_half_t, 2>),
sizeof(vector_type<half_t, 2>) + sizeof(vector_type<half_t, 2>));
ASSERT_EQ(sizeof(vector_type<complex_half_t, 4>),
sizeof(vector_type<half_t, 4>) + sizeof(vector_type<half_t, 4>));
ASSERT_EQ(sizeof(vector_type<complex_half_t, 8>),
sizeof(vector_type<half_t, 8>) + sizeof(vector_type<half_t, 8>));
ASSERT_EQ(sizeof(vector_type<complex_half_t, 16>),
sizeof(vector_type<half_t, 16>) + sizeof(vector_type<half_t, 16>));
ASSERT_EQ(sizeof(vector_type<complex_half_t, 32>),
sizeof(vector_type<half_t, 32>) + sizeof(vector_type<half_t, 32>));
ASSERT_EQ(sizeof(vector_type<complex_half_t, 64>),
sizeof(vector_type<half_t, 64>) + sizeof(vector_type<half_t, 64>));
}
TEST(Complex_half, TestAlignment)
{
struct complex_half_t
{
half_t real;
half_t img;
};
ASSERT_EQ(alignof(vector_type<complex_half_t, 2>),
alignof(vector_type<half_t, 2>) + alignof(vector_type<half_t, 2>));
ASSERT_EQ(alignof(vector_type<complex_half_t, 4>),
alignof(vector_type<half_t, 4>) + alignof(vector_type<half_t, 4>));
ASSERT_EQ(alignof(vector_type<complex_half_t, 8>),
alignof(vector_type<half_t, 8>) + alignof(vector_type<half_t, 8>));
ASSERT_EQ(alignof(vector_type<complex_half_t, 16>),
alignof(vector_type<half_t, 16>) + alignof(vector_type<half_t, 16>));
ASSERT_EQ(alignof(vector_type<complex_half_t, 32>),
alignof(vector_type<half_t, 32>) + alignof(vector_type<half_t, 32>));
ASSERT_EQ(alignof(vector_type<complex_half_t, 64>),
alignof(vector_type<half_t, 64>) + alignof(vector_type<half_t, 64>));
}
TEST(Complex_half, TestAsType)
{
struct complex_half_t
{
using type = half_t;
type real;
type img;
complex_half_t() : real{type{}}, img{type{}} {}
complex_half_t(type real_init, type img_init) : real{real_init}, img{img_init} {}
};
// test size
const int size = 4;
// custom type number of elements
const int num_elem = sizeof(complex_half_t) / sizeof(complex_half_t::type);
std::vector<half_t> test_vec = {half_t{0.3f},
half_t{-0.6f},
half_t{0.8f},
half_t{-0.2f},
half_t{0.5f},
half_t{-0.7f},
half_t{0.9f},
half_t{-0.3f}};
// reference vector
vector_type<complex_half_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<complex_half_t>()(Number<i>{}).real,
type_convert<half_t>(0.0f));
ASSERT_EQ(right_vec.template AsType<complex_half_t>()(Number<i>{}).img,
type_convert<half_t>(0.0f));
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<complex_half_t>()(Number<i>{}) =
complex_half_t{test_vec.at(num_elem * i), test_vec.at(num_elem * i + 1)};
});
// copy the vector
vector_type<complex_half_t, size> left_vec{right_vec};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<complex_half_t>()(Number<i>{}).real,
test_vec.at(num_elem * i));
ASSERT_EQ(left_vec.template AsType<complex_half_t>()(Number<i>{}).img,
test_vec.at(num_elem * i + 1));
});
}
TEST(Complex_half, TestAsTypeReshape)
{
struct complex_half_t
{
using type = half_t;
type real;
type img;
complex_half_t() : real{type{}}, img{type{}} {}
complex_half_t(type real_init, type img_init) : real{real_init}, img{img_init} {}
};
// test size
const int size = 4;
// custom type number of elements
const int num_elem = sizeof(complex_half_t) / sizeof(complex_half_t::type);
std::vector<half_t> test_vec = {half_t{0.3f},
half_t{-0.6f},
half_t{0.8f},
half_t{-0.2f},
half_t{0.5f},
half_t{-0.7f},
half_t{0.9f},
half_t{-0.3f}};
// reference vector
vector_type<complex_half_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(right_vec.template AsType<complex_half_t>()(Number<i>{}).real,
type_convert<half_t>(0.0f));
ASSERT_EQ(right_vec.template AsType<complex_half_t>()(Number<i>{}).img,
type_convert<half_t>(0.0f));
});
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<complex_half_t>()(Number<i>{}) =
complex_half_t{test_vec.at(num_elem * i), test_vec.at(num_elem * i + 1)};
});
// copy the first half of a vector
vector_type<complex_half_t, size / 2> left_vec{
right_vec.template AsType<vector_type<complex_half_t, size / 2>::type>()(Number<0>{})};
// check if values were copied correctly
ck::static_for<0, size / 2, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<complex_half_t>()(Number<i>{}).real,
test_vec.at(num_elem * i));
ASSERT_EQ(left_vec.template AsType<complex_half_t>()(Number<i>{}).img,
test_vec.at(num_elem * i + 1));
});
}
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