Skip to content

GitLab

Commit 11279540 authored by Astha Rai's avatar Astha Rai
Browse files

Merge branch 'transpose_5d' of github.com:ROCmSoftwarePlatform/composable_kernel into transpose_5d

parents 14daa201 33e78b9a
Hide whitespace changes
Inline Side-by-side
Showing with 406 additions and 198 deletions
profiler/src/profile_grouped_conv_bwd_weight.cpp
View file @ 11279540
......@@ -20,10 +20,11 @@ enum struct ConvLayout
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_F32_BF16, // 2
F16_F16_F16_BF8_F8 // 3
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_F32_BF16, // 2
F16_F16_F16_BF8_F8, // 3
I8_I8_I8 // 4
};
#define OP_NAME "grouped_conv_bwd_weight"
......@@ -35,7 +36,8 @@ static void print_helper_msg()
<< "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
<< " 1: Input fp16, Weight fp16, Output fp16\n"
<< " 2: Input bf16, Weight fp32, Output bf16\n"
<< " 3: Input fp16, Weight fp16, Output fp16, Gemm bf8@fp8)\n"
<< " 3: Input fp16, Weight fp16, Output fp16, Gemm bf8@fp8\n"
<< " 4: Input int8, Weight int8, Output int8)\n"
<< "arg3: tensor layout (0: Input[G, N, C, Hi, Wi], Weight[G, K, C, Y, X], Output[G, "
"N, K, Ho, Wo]\n"
<< " 1: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, "
......@@ -84,12 +86,8 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
#ifdef CK_ENABLE_FP8
using F8 = ck::f8_t;
#endif
#ifdef CK_ENABLE_BF8
using BF8 = ck::bf8_t;
#endif
using F8 = ck::f8_t;
using BF8 = ck::bf8_t;
using namespace ck::tensor_layout::convolution;
......@@ -139,83 +137,93 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
{
return profile(I1, GNWC{}, GKXC{}, GNWK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I1, GNWC{}, GKXC{}, GNWK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I1, GNWC{}, GKXC{}, GNWK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
if(num_dim_spatial == 2 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::I8_I8_I8)
{
return profile(
I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, int8_t{}, int8_t{}, int8_t{}, int8_t{}, int8_t{});
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F16_F16_F16_BF8_F8)
if(data_type == ConvDataType::F16_F16_F16_BF8_F8)
{
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F16{}, F16{}, F16{}, BF8{}, F8{});
}
else if(data_type == ConvDataType::I8_I8_I8)
{
return profile(
I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, int8_t{}, int8_t{}, int8_t{}, int8_t{}, int8_t{});
}
}
std::cout << "this data_type & layout is not implemented" << std::endl;
......
profiler/src/profile_grouped_gemm.cpp
View file @ 11279540
......@@ -27,6 +27,8 @@ enum struct GemmDataType
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
F8_F16_F16, // 4
F16_F8_F16, // 5
};
#define OP_NAME "grouped_gemm"
......@@ -56,7 +58,7 @@ int profile_grouped_gemm(int argc, char* argv[])
{
std::cout
<< "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
<< "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: fp8@fp6; 5: f16@f8)\n"
<< "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"
<< " 1: A[m, k] * B[n, k] = C[m, n];\n"
<< " 2: A[k, m] * B[k, n] = C[m, n];\n"
......@@ -169,6 +171,46 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideCs,
kbatch);
}
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::f8_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
time_kernel,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs,
kbatch);
}
else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::half_t,
ck::f8_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
time_kernel,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs,
kbatch);
}
else
{
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
......
profiler/src/profile_groupnorm.cpp
View file @ 11279540
......@@ -93,12 +93,12 @@ int profile_groupnorm(int argc, char* argv[])
if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_groupnorm_impl<F32, F32, F32, F32, F32>(
ck::profiler::profile_groupnorm_impl<F32, F32, F32, F32, F32, F32, false>(
do_verification, init_method, do_log, time_kernel, length);
}
else if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16>(
ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16, F32, false>(
do_verification, init_method, do_log, time_kernel, length);
}
else
......
profiler/src/profile_layernorm.cpp
View file @ 11279540
......@@ -82,12 +82,12 @@ int profile_layernorm(int argc, char* argv[])
if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, rank>(
ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, F32, false, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_layernorm_impl<F32, F32, F32, F32, F32, rank>(
ck::profiler::profile_layernorm_impl<F32, F32, F32, F32, F32, F32, false, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else
......
script/cmake-ck-dev.sh
View file @ 11279540
......@@ -8,8 +8,7 @@ MY_PROJECT_SOURCE=$1
cmake \
-D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_CXX_FLAGS="-std=c++17 -O3 -ftemplate-backtrace-limit=0 -fPIE -Wno-gnu-line-marker \
-save-temps=$PWD" \
-D CMAKE_CXX_FLAGS="-std=c++17 -O3 -ftemplate-backtrace-limit=0 -fPIE -Wno-gnu-line-marker" \
-D CMAKE_BUILD_TYPE=Release \
-D BUILD_DEV=ON \
-D GPU_TARGETS="gfx908;gfx90a;gfx940" \
......
script/redis-cli.conf 0 → 100644
View file @ 11279540
fips = no
setuid = root
setgid = root
pid = /var/run/stunnel.pid
debug = 7
options = NO_SSLv2
options = NO_SSLv3
[redis-cli]
client = yes
accept = 127.0.0.1:6379
script/sccache_wrapper.sh 0 → 100755
View file @ 11279540
#!/bin/bash
set -e
COMPILERS_HASH_DIR=${COMPILERS_HASH_DIR:-"/tmp/.sccache"}
SCCACHE_EXTRAFILES=${SCCACHE_EXTRAFILES:-"${COMPILERS_HASH_DIR}/rocm_compilers_hash_file"}
SCCACHE_BIN=${SCCACHE_BIN:-"${SCCACHE_INSTALL_LOCATION}/sccache"}
ENFORCE_REDIS="false"
while [ "$1" != "" ];
do
case $1 in
--enforce_redis )
shift; ENFORCE_REDIS="true" ;;
--no-hipcc )
shift ;;
*)
break ;;
esac
done
setup_rocm_compilers_hash_file() {
mkdir -p "$COMPILERS_HASH_DIR"
HIPCC_MD5="$(md5sum "${ROCM_PATH}/bin/hipcc")"
pushd "${ROCM_PATH}/amdgcn/bitcode"
DEVICELIBS_BITCODES_MD5="$(find . -type f -exec md5sum {} \; | sort | md5sum)"
popd
HIPCC_HASH_VALUE="${HIPCC_MD5%% *}"
DEVICELIBS_BITCODES_HASH_VALUE="${DEVICELIBS_BITCODES_MD5%% *}"
# MD5 checksums of clang and clang-offload-bundler cannot be used since they will keep changing
# if the ROCM_PATH changes, ie; for every mainline build.
# This is because ROCM_PATH gets encoded into the clang/clang-offload-bundler binaries as part
# of RPATH.
# The versions themselves contain the commit hash of the compiler repo at the time of building.
# Hence, this should be a viable alternative to using the binary checksum itself.
CLANG_VERSION="$("${ROCM_PATH}/llvm/bin/clang" --version | head -n 1)"
CLANG_OFFLOAD_BUNDLER_VERSION="$("${ROCM_PATH}/llvm/bin/clang-offload-bundler" --version | head -n 1)"
printf '%s: %s\n' 'clang version' "${CLANG_VERSION}" | tee -a "$SCCACHE_EXTRAFILES"
printf '%s: %s\n' 'clang-offload-bundler version' "${CLANG_OFFLOAD_BUNDLER_VERSION}" | tee -a "$SCCACHE_EXTRAFILES"
printf '%s: %s\n' 'hipcc md5sum' "${HIPCC_HASH_VALUE}" | tee -a "$SCCACHE_EXTRAFILES"
printf '%s: %s\n' 'devicelibs bitcode md5sum' "${DEVICELIBS_BITCODES_HASH_VALUE}" | tee -a "$SCCACHE_EXTRAFILES"
echo "sccache-wrapper: compilers hash file set up at ${SCCACHE_EXTRAFILES}"
cat "$SCCACHE_EXTRAFILES"
}
if [ "${ENFORCE_REDIS}" == "true" ]; then
if [ -z "${SCCACHE_REDIS}" ]; then
echo "SCCACHE_REDIS not set. Not wrapping compilers with sccache."
exit 10
else
response=$(redis-cli -u ${SCCACHE_REDIS} ping) || true
if [ "${response}" != "PONG" ]; then
echo "Redis server unreachable. Not wrapping compilers with sccache."
exit 20
fi
fi
fi
setup_rocm_compilers_hash_file
$SCCACHE_BIN --version
$SCCACHE_BIN --start-server
test/CMakeLists.txt
View file @ 11279540
......@@ -11,40 +11,36 @@ function(add_test_executable TEST_NAME)
message("adding test ${TEST_NAME}")
set(result 1)
if(DEFINED DTYPES)
foreach(source IN LISTS ARGN)
set(test 0)
foreach(type IN LISTS DTYPES)
if(type MATCHES "fp16")
set(type1 "_f16")
elseif(type MATCHES "fp32")
set(type1 "_f32")
elseif(type MATCHES "fp8")
set(type1 "_f8")
elseif(type MATCHES "bf16")
set(type1 "_b16")
elseif(type MATCHES "fp64")
set(type1 "_f64")
elseif(type MATCHES "int8")
set(type1 "_i8")
endif()
if("${source}" MATCHES "${type}" OR "${source}" MATCHES "${type1}")
#if filename matches any selected type, exit type loop and do no exclude the file from the list
set(test 0)
break()
elseif((source MATCHES "fp8" OR source MATCHES "fp32" OR source MATCHES "fp64" OR source MATCHES "bf16" OR source MATCHES "int8" OR source MATCHES "fp16" OR
source MATCHES "_f8" OR source MATCHES "_f32" OR source MATCHES "_f64" OR source MATCHES "_i8" OR source MATCHES "_f16" OR source MATCHES "_b16") AND
NOT(source MATCHES type OR source MATCHES type1))
#if filename contains a type which doesn't match any selected type, mark it for removal
set(test 1)
foreach(source IN LISTS ARGN)
set(test 0)
if((source MATCHES "_fp16" OR source MATCHES "_f16") AND NOT "fp16" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp32" OR source MATCHES "_f32") AND NOT "fp32" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp64" OR source MATCHES "_f64") AND NOT "fp64" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp8" OR source MATCHES "_f8") AND NOT "fp8" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_bf8" OR source MATCHES "_bf8") AND NOT "bf8" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_bf16" OR source MATCHES "_b16") AND NOT "bf16" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_int8" OR source MATCHES "_i8") AND NOT "int8" IN_LIST DTYPES)
set(test 1)
endif()
if(test EQUAL 1)
message("removing test ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
if(test EQUAL 1)
message("removing test ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
endif()
foreach(source IN LISTS ARGN)
endif()
foreach(source IN LISTS ARGN)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
message("removing dl test ${source} ")
list(REMOVE_ITEM ARGN "${source}")
......@@ -70,38 +66,34 @@ function(add_gtest_executable TEST_NAME)
message("adding gtest ${TEST_NAME}")
set(result 1)
if(DEFINED DTYPES)
foreach(source IN LISTS ARGN)
set(test 0)
foreach(type IN LISTS DTYPES)
if(type MATCHES "fp16")
set(type1 "_f16")
elseif(type MATCHES "fp32")
set(type1 "_f32")
elseif(type MATCHES "fp8")
set(type1 "_f8")
elseif(type MATCHES "bf16")
set(type1 "_b16")
elseif(type MATCHES "fp64")
set(type1 "_f64")
elseif(type MATCHES "int8")
set(type1 "_i8")
endif()
if("${source}" MATCHES "${type}" OR "${source}" MATCHES "${type1}")
#if filename matches any selected type, exit type loop and do no exclude the file from the list
set(test 0)
break()
elseif((source MATCHES "fp8" OR source MATCHES "fp32" OR source MATCHES "fp64" OR source MATCHES "bf16" OR source MATCHES "int8" OR source MATCHES "fp16" OR
source MATCHES "_f8" OR source MATCHES "_f32" OR source MATCHES "_f64" OR source MATCHES "_i8" OR source MATCHES "_f16" OR source MATCHES "_b16") AND
NOT(source MATCHES type OR source MATCHES type1))
#if filename contains a type which doesn't match any selected type, mark it for removal
set(test 1)
foreach(source IN LISTS ARGN)
set(test 0)
if((source MATCHES "_fp16" OR source MATCHES "_f16") AND NOT "fp16" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp32" OR source MATCHES "_f32") AND NOT "fp32" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp64" OR source MATCHES "_f64") AND NOT "fp64" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp8" OR source MATCHES "_f8") AND NOT "fp8" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_bf8" OR source MATCHES "_bf8") AND NOT "bf8" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_bf16" OR source MATCHES "_b16") AND NOT "bf16" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_int8" OR source MATCHES "_i8") AND NOT "int8" IN_LIST DTYPES)
set(test 1)
endif()
if(test EQUAL 1)
message("removing gtest ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
if(test EQUAL 1)
message("removing gtest ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
endif()
foreach(source IN LISTS ARGN)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
......
test/contraction/CMakeLists.txt
View file @ 11279540
add_gtest_executable(test_contraction test_contraction.cpp)
target_link_libraries(test_contraction PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance)
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_gtest_executable(test_contraction_interface test_contraction_interface.cpp)
target_link_libraries(test_contraction_interface PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance)
set(target 1)
endif()
if(gpu IN_LIST gpu_list AND target EQUAL 0)
if((DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64") OR NOT DEFINED DTYPES)
add_gtest_executable(test_contraction test_contraction.cpp)
target_link_libraries(test_contraction PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance)
add_gtest_executable(test_contraction_interface test_contraction_interface.cpp)
target_link_libraries(test_contraction_interface PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance)
set(target 1)
endif()
endif()
endforeach()
test/contraction/test_contraction.cpp
View file @ 11279540
......@@ -10,9 +10,12 @@
#include <gtest/gtest.h>
#include "profiler/profile_contraction_impl.hpp"
#include "profiler/profile_contraction_utils.hpp"
using F32 = float;
using F64 = double;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using F64 = double;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -20,49 +23,49 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using Scale = ck::tensor_operation::element_wise::Scale;
struct MemoryParams
struct Dimensions
{
std::vector<ck::index_t> M;
std::vector<ck::index_t> N;
std::vector<ck::index_t> K;
std::vector<ck::index_t> StridesA;
std::vector<ck::index_t> StridesB;
std::vector<ck::index_t> StridesC;
std::vector<ck::index_t> StridesD;
};
template <typename Tuple>
class TestContraction : public ::testing::Test
{
protected:
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using CDLayout = std::tuple_element_t<2, Tuple>;
using DataType = std::tuple_element_t<3, Tuple>;
using DTupleDataType = std::tuple_element_t<4, Tuple>;
using CDElementOp = std::tuple_element_t<5, Tuple>;
std::vector<MemoryParams> list_of_memory_params = {{{32, 32},
{32, 32},
{32, 32},
{32768, 1024, 32, 1},
{32768, 1024, 32, 1},
{32768, 1024, 32, 1},
{32768, 1024, 32, 1}},
{{16, 16},
{32, 32},
{16, 16},
{4096, 256, 16, 1},
{16, 1, 8192, 256},
{16384, 1024, 32, 1},
{16384, 1024, 32, 1}}};
std::vector<ck::index_t> init_methods = {0, 1, 2};
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using CDLayout = std::tuple_element_t<2, Tuple>;
using DataType = std::tuple_element_t<3, Tuple>;
using DTupleDataType = std::tuple_element_t<4, Tuple>;
using ComputeDataType = std::tuple_element_t<5, Tuple>;
using CDElementOp = std::tuple_element_t<6, Tuple>;
std::vector<Dimensions> dimension_list = {{{32, 32}, {32, 32}, {32, 32}},
{{16, 16}, {32, 32}, {16, 16}}};
std::vector<ck::index_t> init_methods = {1, 2};
std::unique_ptr<CDElementOp> p_cd_element_op;
void Run()
{
for(auto& memory_params : list_of_memory_params)
for(auto& dimension_params : dimension_list)
{
std::vector<ck::index_t> StridesA;
std::vector<ck::index_t> StridesB;
std::vector<ck::index_t> StridesC;
std::vector<ck::index_t> StridesD;
const auto& M = dimension_params.M;
const auto& N = dimension_params.N;
const auto& K = dimension_params.K;
assign_default_strides(ALayout{}, StridesA, {M[0], M[1], K[0], K[1]});
assign_default_strides(BLayout{}, StridesB, {N[0], N[1], K[0], K[1]});
assign_default_strides(CDLayout{}, StridesC, {M[0], M[1], N[0], N[1]});
assign_default_strides(CDLayout{}, StridesD, {M[0], M[1], N[0], N[1]});
for(const ck::index_t init_method : init_methods)
{
bool pass =
......@@ -70,19 +73,20 @@ class TestContraction : public ::testing::Test
BLayout,
CDLayout,
DataType,
ComputeDataType,
DTupleDataType,
CDElementOp>(true /*do_verification*/,
init_method,
false /*do_logs*/,
false /*time_kernel*/,
*p_cd_element_op,
memory_params.M,
memory_params.N,
memory_params.K,
memory_params.StridesA,
memory_params.StridesB,
memory_params.StridesC,
memory_params.StridesD);
dimension_params.M,
dimension_params.N,
dimension_params.K,
StridesA,
StridesB,
StridesC,
StridesD);
EXPECT_TRUE(pass);
}
}
......@@ -99,24 +103,18 @@ class TestContractionBilinear : public TestContraction<Tuple>
{
};
#define ALL_LAYOUT_COMBINATIONS(dt, tuple_dt, compute_dt, op) \
std::tuple<Row, Row, Row, dt, tuple_dt, compute_dt, op>, \
std::tuple<Row, Col, Row, dt, tuple_dt, compute_dt, op>, \
std::tuple<Col, Row, Row, dt, tuple_dt, compute_dt, op>, \
std::tuple<Col, Col, Row, dt, tuple_dt, compute_dt, op>
using BilinearKernelTypes =
::testing::Types<std::tuple<Row, Row, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Row, Col, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Row, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Col, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Row, Row, Row, F64, ck::Tuple<F32>, Bilinear>,
std::tuple<Row, Col, Row, F64, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Row, Row, F64, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Col, Row, F64, ck::Tuple<F32>, Bilinear>>;
using ScaleKernelTypes = ::testing::Types<std::tuple<Row, Row, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Row, Col, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Col, Row, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Col, Col, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Row, Row, Row, F64, ck::Tuple<>, Scale>,
std::tuple<Row, Col, Row, F64, ck::Tuple<>, Scale>,
std::tuple<Col, Row, Row, F64, ck::Tuple<>, Scale>,
std::tuple<Col, Col, Row, F64, ck::Tuple<>, Scale>>;
::testing::Types<ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<F32>, F32, Bilinear),
ALL_LAYOUT_COMBINATIONS(F64, ck::Tuple<F64>, F64, Bilinear)>;
using ScaleKernelTypes = ::testing::Types<ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<>, F32, Scale),
ALL_LAYOUT_COMBINATIONS(F64, ck::Tuple<>, F64, Scale)>;
TYPED_TEST_SUITE(TestContractionBilinear, BilinearKernelTypes);
TYPED_TEST_SUITE(TestContractionScale, ScaleKernelTypes);
......@@ -136,3 +134,46 @@ TYPED_TEST(TestContractionScale, scale)
this->p_cd_element_op = std::make_unique<Scale>(0.5f);
this->Run();
}
template <typename Tuple>
class TestContractionScaleMixedPrecision : public TestContraction<Tuple>
{
};
template <typename Tuple>
class TestContractionBilinearMixedPrecision : public TestContraction<Tuple>
{
};
using BilinearKernelTypesMixedPrecision =
::testing::Types<ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<F32>, F16, Bilinear),
ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<F32>, BF16, Bilinear),
ALL_LAYOUT_COMBINATIONS(F64, ck::Tuple<F64>, F32, Bilinear),
ALL_LAYOUT_COMBINATIONS(F16, ck::Tuple<F16>, F32, Bilinear),
ALL_LAYOUT_COMBINATIONS(BF16, ck::Tuple<BF16>, F32, Bilinear)>;
using ScaleKernelTypesMixedPrecision =
::testing::Types<ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<>, F16, Scale),
ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<>, BF16, Scale),
ALL_LAYOUT_COMBINATIONS(F64, ck::Tuple<>, F32, Scale),
ALL_LAYOUT_COMBINATIONS(F16, ck::Tuple<>, F32, Scale),
ALL_LAYOUT_COMBINATIONS(BF16, ck::Tuple<>, F32, Scale)>;
TYPED_TEST_SUITE(TestContractionBilinearMixedPrecision, BilinearKernelTypesMixedPrecision);
TYPED_TEST_SUITE(TestContractionScaleMixedPrecision, ScaleKernelTypesMixedPrecision);
TYPED_TEST(TestContractionBilinearMixedPrecision, bilinear)
{
this->p_cd_element_op = std::make_unique<Bilinear>(1.f, 1.f);
this->Run();
this->p_cd_element_op = std::make_unique<Bilinear>(-0.5f, 0.5f);
this->Run();
}
TYPED_TEST(TestContractionScaleMixedPrecision, scale)
{
this->p_cd_element_op = std::make_unique<Scale>(1.f);
this->Run();
this->p_cd_element_op = std::make_unique<Scale>(0.5f);
this->Run();
}
test/contraction/test_contraction_interface.cpp
View file @ 11279540
......@@ -34,11 +34,11 @@ class ContractionInstanceWrapper
static constexpr ck::index_t NumDim = 2;
// clang-format off
using ContractionDeviceInstance = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| 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|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//#####################################| | | | | | | | | | 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_NWaveNPerXdl| _NWaveNPerXdl|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDim, NumDim, NumDim, F32, F32, F32, F32, ck::Tuple<F32>, F32, Pass, Pass, Bilinear, GemmSpec, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, ABlockTransferSrcVectorDim, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, BBlockTransferSrcVectorDim, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, CDEBlockTransferScalarPerVector>;
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| 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| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | 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_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDim, NumDim, NumDim, F32, F32, F32, F32, ck::Tuple<F32>, F32, Pass, Pass, Bilinear, GemmSpec, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, ABlockTransferSrcVectorDim, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, BBlockTransferSrcVectorDim, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, CDEBlockTransferScalarPerVector, F32>;
// clang-format on
bool isSupported(std::vector<ck::index_t>& ADims,
......
test/conv_tensor_rearrange/CMakeLists.txt
View file @ 11279540
add_gtest_executable(test_conv_tensor_rearrange test_conv_tensor_rearrange.cpp)
target_link_libraries(test_conv_tensor_rearrange PRIVATE utility device_image_to_column_instance device_column_to_image_instance)
add_gtest_executable(test_conv_tensor_rearrange_interface test_conv_tensor_rearrange_interface.cpp)
target_link_libraries(test_conv_tensor_rearrange_interface PRIVATE utility)
test/conv_tensor_rearrange/test_conv_tensor_rearrange.cpp
View file @ 11279540
......@@ -45,14 +45,20 @@ class TestConvTensorRearrange : public ::testing::Test
using namespace ck::tensor_layout::convolution;
using namespace ck::conv_tensor_rearrange_op;
using KernelTypes1d =
::testing::Types<std::tuple<GNWC, ImageToColumn>, std::tuple<GNWC, ColumnToImage>>;
using KernelTypes1d = ::testing::Types<std::tuple<GNWC, ImageToColumn>,
std::tuple<GNWC, ColumnToImage>,
std::tuple<NWGC, ImageToColumn>,
std::tuple<NWGC, ColumnToImage>>;
using KernelTypes2d =
::testing::Types<std::tuple<GNHWC, ImageToColumn>, std::tuple<GNHWC, ColumnToImage>>;
using KernelTypes2d = ::testing::Types<std::tuple<GNHWC, ImageToColumn>,
std::tuple<GNHWC, ColumnToImage>,
std::tuple<NHWGC, ImageToColumn>,
std::tuple<NHWGC, ColumnToImage>>;
using KernelTypes3d =
::testing::Types<std::tuple<GNDHWC, ImageToColumn>, std::tuple<GNDHWC, ColumnToImage>>;
using KernelTypes3d = ::testing::Types<std::tuple<GNDHWC, ImageToColumn>,
std::tuple<GNDHWC, ColumnToImage>,
std::tuple<NDHWGC, ImageToColumn>,
std::tuple<NDHWGC, ColumnToImage>>;
template <typename Tuple>
class TestConvTensorRearrange1d : public TestConvTensorRearrange<Tuple>
......@@ -77,16 +83,16 @@ TYPED_TEST(TestConvTensorRearrange1d, Test1D)
{
this->conv_params.clear();
this->conv_params.push_back({1, 1, 4, 1, 192, {3}, {28}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 1, 64, 1, 64, {3}, {14}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 1, 64, 1, 64, {1}, {7}, {3}, {1}, {0}, {0}});
this->conv_params.push_back({1, 1, 64, 1, 64, {1}, {3}, {1}, {1}, {0}, {0}});
this->conv_params.push_back({1, 2, 4, 1, 192, {3}, {28}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 2, 64, 1, 64, {3}, {14}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 2, 64, 1, 64, {1}, {7}, {3}, {1}, {0}, {0}});
this->conv_params.push_back({1, 2, 64, 1, 64, {1}, {3}, {1}, {1}, {0}, {0}});
// ScalarPerVector should be 1
this->conv_params.push_back({1, 1, 4, 1, 1, {3}, {28}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 2, 4, 1, 1, {3}, {28}, {1}, {1}, {1}, {1}});
// stride != 1
this->conv_params.push_back({1, 1, 1, 1, 4, {3}, {28}, {2}, {1}, {1}, {1}});
this->conv_params.push_back({1, 2, 1, 1, 4, {3}, {28}, {2}, {1}, {1}, {1}});
// dilation != 1
this->conv_params.push_back({1, 1, 1, 1, 4, {3}, {28}, {1}, {2}, {1}, {1}});
this->conv_params.push_back({1, 2, 1, 1, 4, {3}, {28}, {1}, {2}, {1}, {1}});
#ifdef CK_ENABLE_FP32
this->template Run<1, float, float>();
#endif
......@@ -106,13 +112,13 @@ TYPED_TEST(TestConvTensorRearrange2d, Test2D)
this->conv_params.clear();
this->conv_params.push_back(
{2, 1, 4, 1, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
{2, 2, 4, 1, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back(
{2, 1, 64, 1, 64, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
{2, 2, 64, 1, 64, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 64, 1, 64, {1, 1}, {7, 7}, {3, 3}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back({2, 1, 64, 1, 64, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back(
{2, 1, 64, 1, 64, {3, 3}, {28, 28}, {2, 2}, {2, 2}, {1, 1}, {1, 1}});
{2, 2, 64, 1, 64, {3, 3}, {28, 28}, {2, 2}, {2, 2}, {1, 1}, {1, 1}});
#ifdef CK_ENABLE_FP32
this->template Run<2, float, float>();
#endif
......@@ -131,13 +137,13 @@ TYPED_TEST(TestConvTensorRearrange3d, Test3D)
{
this->conv_params.clear();
this->conv_params.push_back(
{3, 1, 16, 1, 64, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {3, 3, 3}, {0, 0, 0}, {0, 0, 0}});
{3, 2, 16, 1, 64, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {3, 3, 3}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 1, 2, 1, 64, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
{3, 2, 2, 1, 64, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 32, 1, 64, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
{3, 2, 32, 1, 64, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 1, 64, 1, 64, {3, 3, 3}, {14, 14, 14}, {2, 2, 2}, {2, 2, 2}, {1, 1, 1}, {1, 1, 1}});
{3, 2, 64, 1, 64, {3, 3, 3}, {14, 14, 14}, {2, 2, 2}, {2, 2, 2}, {1, 1, 1}, {1, 1, 1}});
#ifdef CK_ENABLE_FP32
this->template Run<3, float, float>();
#endif
......
test/conv_tensor_rearrange/test_conv_tensor_rearrange_interface.cpp
View file @ 11279540
......@@ -53,7 +53,7 @@ class TestConvTensorRearrangeInterface : public ::testing::Test
template <typename ConvTensorRearrangeOp>
bool Run()
{
const auto G = conv_param.G_;
const auto N = conv_param.N_;
const auto C = conv_param.C_;
const auto FakeC =
......@@ -71,13 +71,13 @@ class TestConvTensorRearrangeInterface : public ::testing::Test
const auto image_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ImLayout>(
conv_param);
const auto gemm_desc = HostTensorDescriptor({NDoHoWo, CZYX});
const auto gemm_desc = HostTensorDescriptor({G, NDoHoWo, CZYX});
std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
std::array<ck::index_t, NDimSpatial + 3> input_g_n_c_wis_strides{};
std::array<ck::index_t, 2> output_m_k_strides{};
std::array<ck::index_t, 3> output_g_m_k_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
......@@ -89,7 +89,7 @@ class TestConvTensorRearrangeInterface : public ::testing::Test
copy(conv_param.filter_spatial_lengths_, filter_spatial_lengths);
copy(conv_param.output_spatial_lengths_, output_spatial_lengths);
copy(image_desc.GetStrides(), input_g_n_c_wis_strides);
copy(gemm_desc.GetStrides(), output_m_k_strides);
copy(gemm_desc.GetStrides(), output_g_m_k_strides);
copy(conv_param.conv_filter_strides_, conv_filter_strides);
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
copy(conv_param.input_left_pads_, input_left_pads);
......@@ -100,13 +100,14 @@ class TestConvTensorRearrangeInterface : public ::testing::Test
auto img2col = DeviceImgToColInstance{};
auto argument = img2col.MakeArgument(nullptr,
nullptr,
G,
N,
IsCPacked ? C : FakeC,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
input_g_n_c_wis_strides,
output_m_k_strides,
output_g_m_k_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......@@ -119,13 +120,14 @@ class TestConvTensorRearrangeInterface : public ::testing::Test
auto col2img = DeviceColToimgInstance{};
auto argument = col2img.MakeArgument(nullptr,
nullptr,
G,
N,
IsCPacked ? C : FakeC,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
input_g_n_c_wis_strides,
output_m_k_strides,
output_g_m_k_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......
test/data_type/CMakeLists.txt
View file @ 11279540
if (USE_BITINT_EXTENSION_INT4)
add_gtest_executable(test_int4 int4.cpp)
add_gtest_executable(test_int4 test_int4.cpp)
if(result EQUAL 0)
target_link_libraries(test_int4 PRIVATE utility)
endif()
endif()
add_gtest_executable(test_fp8 fp8.cpp)
add_gtest_executable(test_fp8 test_fp8.cpp)
if(result EQUAL 0)
target_link_libraries(test_fp8 PRIVATE utility)
endif()
add_gtest_executable(test_bf8 bf8.cpp)
add_gtest_executable(test_bf8 test_bf8.cpp)
if(result EQUAL 0)
target_link_libraries(test_bf8 PRIVATE utility)
endif()
......
test/grouped_convnd_bwd_weight/CMakeLists.txt
View file @ 11279540
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list_xdl gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list_wmma gfx1100 gfx1101 gfx1102)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
if(gpu IN_LIST gpu_list_xdl AND target EQUAL 0)
add_gtest_executable(test_grouped_convnd_bwd_weight test_grouped_convnd_bwd_weight.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance)
add_gtest_executable(test_grouped_convnd_bwd_weight_interface test_grouped_convnd_bwd_weight_interface.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight_interface PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance)
add_gtest_executable(test_grouped_convnd_bwd_weight_interface test_grouped_convnd_bwd_weight_interface_xdl.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight_interface PRIVATE utility)
set(target 1)
endif()
if(gpu IN_LIST gpu_list_wmma AND target EQUAL 0)
add_gtest_executable(test_grouped_convnd_bwd_weight test_grouped_convnd_bwd_weight.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance)
add_gtest_executable(test_grouped_convnd_bwd_weight_interface test_grouped_convnd_bwd_weight_interface_wmma.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight_interface PRIVATE utility)
set(target 1)
endif()
endforeach()
\ No newline at end of file
test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp
View file @ 11279540
......@@ -11,6 +11,7 @@
#include "ck/utility/common_header.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "profiler/profile_grouped_conv_bwd_weight_impl.hpp"
......@@ -33,8 +34,9 @@ class TestGroupedConvndBwdWeight : public ::testing::Test
bool skip_case(const ck::utils::conv::ConvParam& params, const ck::index_t split_k)
{
// Odd K or C values are supported only by DL kernel (only applies to fp16)
// DL kernel currently supports only `split_k=1`
// Odd K or C values are supported only by DL and WMMA
// kernels (only applies to fp16)
// DL and WMMA kernels currently support only `split_k=1`
if constexpr(std::is_same_v<InDataType, ck::half_t>)
{
if(split_k != 1 && (params.K_ % 2 != 0 || params.C_ % 2 != 0))
......@@ -53,6 +55,42 @@ class TestGroupedConvndBwdWeight : public ::testing::Test
}
}
const bool is_navi3x = ck::get_device_name() == "gfx1100" ||
ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102";
if(is_navi3x)
{
// on navi3x only support for 3d is implemented
if constexpr(NDimSpatial{} != 3)
{
return true;
}
// on navi3x only support for i8 and fp16 is implemented
if constexpr(!((std::is_same_v<InDataType, int8_t> &&
std::is_same_v<WeiDataType, int8_t> &&
std::is_same_v<OutDataType, int8_t>) ||
(std::is_same_v<InDataType, ck::half_t> &&
std::is_same_v<WeiDataType, ck::half_t> &&
std::is_same_v<OutDataType, ck::half_t>)))
{
return true;
}
// WMMA kernel is only supported for split_k=1
if(split_k != 1)
{
return true;
}
}
else
{
// support for i8 is only implemented on navi3x
if constexpr(std::is_same_v<InDataType, int8_t> &&
std::is_same_v<WeiDataType, int8_t> && std::is_same_v<OutDataType, int8_t>)
{
return true;
}
}
return false;
}
......@@ -120,9 +158,11 @@ using KernelTypes3d = ::testing::Types<
std::tuple<float, float, float, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
std::tuple<ck::bhalf_t, float, ck::bhalf_t, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
std::tuple<int8_t, int8_t, int8_t, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
std::tuple<float, float, float, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>,
std::tuple<ck::bhalf_t, float, ck::bhalf_t, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>>;
std::tuple<ck::bhalf_t, float, ck::bhalf_t, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>,
std::tuple<int8_t, int8_t, int8_t, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>>;
TYPED_TEST_SUITE(TestGroupedConvndBwdWeight1d, KernelTypes1d);
TYPED_TEST_SUITE(TestGroupedConvndBwdWeight2d, KernelTypes2d);
......
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