Skip to content

GitLab

Commit 84dcf5d0 authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'develop' into amd-develop

parents 705d5a08 c9553832
Hide whitespace changes
Inline Side-by-side
Showing with 428 additions and 130 deletions
library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
View file @ 84dcf5d0
set(DEVICE_SOFTMAX_INSTANCES)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_SOFTMAX_INSTANCES device_softmax_f16_f16_instance_rank3_reduce1.cpp
list(APPEND DEVICE_SOFTMAX_INSTANCES
device_softmax_f16_f16_instance_rank3_reduce1.cpp
device_softmax_f16_f16_instance_rank3_reduce2.cpp
device_softmax_f16_f16_instance_rank3_reduce3.cpp
device_softmax_f16_f16_instance_rank4_reduce1.cpp
device_softmax_f16_f16_instance_rank4_reduce2.cpp
device_softmax_f16_f16_instance_rank4_reduce3.cpp
device_softmax_f16_f16_instance_rank4_reduce4.cpp)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_SOFTMAX_INSTANCES device_softmax_f32_f32_instance_rank3_reduce1.cpp
device_softmax_f16_f16_instance_rank4_reduce4.cpp
device_softmax_f32_f32_instance_rank3_reduce1.cpp
device_softmax_f32_f32_instance_rank3_reduce2.cpp
device_softmax_f32_f32_instance_rank3_reduce3.cpp
device_softmax_f32_f32_instance_rank4_reduce1.cpp
device_softmax_f32_f32_instance_rank4_reduce2.cpp
device_softmax_f32_f32_instance_rank4_reduce3.cpp
device_softmax_f32_f32_instance_rank4_reduce4.cpp)
endif()
add_instance_library(device_softmax_instance ${DEVICE_SOFTMAX_INSTANCES})
profiler/include/profiler/profile_gemm_multiply_add_impl.hpp
View file @ 84dcf5d0
......@@ -83,8 +83,8 @@ bool profile_gemm_multiply_add_impl(int do_verification,
d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-1, 1});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 0.2});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.1, 0.1});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
}
......
profiler/include/profiler/profile_gemm_splitk_impl.hpp
View file @ 84dcf5d0
......@@ -214,6 +214,7 @@ bool profile_gemm_splitk_impl(int do_verification,
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
<< kbatch_curr << std::endl;
#if defined CK_ENABLE_FP8
// set softer tolerances for fp8
if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
is_same_v<CDataType, f8_t>)
......@@ -226,8 +227,11 @@ bool profile_gemm_splitk_impl(int do_verification,
}
else
{
#endif
pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
#if defined CK_ENABLE_FP8
}
#endif
if(tflops > best_tflops)
{
......
profiler/src/profile_gemm_multiply_add.cpp
View file @ 84dcf5d0
......@@ -59,9 +59,11 @@ int profile_gemm_multiply_add(int argc, char* argv[])
const int StrideD1 = std::stoi(argv[14]);
const int StrideE = std::stoi(argv[15]);
using F8 = ck::f8_t;
using F16 = ck::half_t;
using F32 = float;
#if defined CK_ENABLE_FP8
using F8 = ck::f8_t;
#endif
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -132,6 +134,7 @@ int profile_gemm_multiply_add(int argc, char* argv[])
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
}
#if defined CK_ENABLE_FP8
else if(data_type == MatrixDataType::F16_F8_F32_F32_F16 &&
layout == MatrixLayout::MK_KN_MN_MN_MN)
{
......@@ -142,6 +145,7 @@ int profile_gemm_multiply_add(int argc, char* argv[])
{
return profile(F16{}, F8{}, F32{}, F32{}, F32{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
}
#endif
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
......
profiler/src/profile_gemm_splitk.cpp
View file @ 84dcf5d0
......@@ -67,7 +67,9 @@ int profile_gemm_splitk(int argc, char* argv[])
using F32 = float;
using F16 = ck::half_t;
using F8 = ck::f8_t;
#if defined CK_ENABLE_FP8
using F8 = ck::f8_t;
#endif
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -146,6 +148,7 @@ int profile_gemm_splitk(int argc, char* argv[])
{
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
}
#if defined CK_ENABLE_FP8
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
......@@ -178,6 +181,7 @@ int profile_gemm_splitk(int argc, char* argv[])
{
return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Col{}, Row{});
}
#endif
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
......
test/CMakeLists.txt
View file @ 84dcf5d0
......@@ -9,26 +9,121 @@ add_custom_target(tests)
function(add_test_executable TEST_NAME)
message("adding test ${TEST_NAME}")
add_executable(${TEST_NAME} ${ARGN})
add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}>)
add_dependencies(tests ${TEST_NAME})
add_dependencies(check ${TEST_NAME})
rocm_install(TARGETS ${TEST_NAME} COMPONENT tests)
set(result 1)
if(DEFINED DTYPES)
foreach(source IN LISTS ARGN)
set(test 0)
foreach(type IN LISTS DTYPES)
if(type MATCHES "fp16")
set(type1 "_f16")
elseif(type MATCHES "fp32")
set(type1 "_f32")
elseif(type MATCHES "fp8")
set(type1 "_f8")
elseif(type MATCHES "bf16")
set(type1 "_b16")
elseif(type MATCHES "fp64")
set(type1 "_f64")
elseif(type MATCHES "int8")
set(type1 "_i8")
endif()
if("${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)
endif()
endforeach()
if(test EQUAL 1)
message("removing test ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
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}")
endif()
endforeach()
#only continue if there are some source files left on the list
if(ARGN)
add_executable(${TEST_NAME} ${ARGN})
add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}>)
add_dependencies(tests ${TEST_NAME})
add_dependencies(check ${TEST_NAME})
rocm_install(TARGETS ${TEST_NAME} COMPONENT tests)
set(result 0)
endif()
#message("add_test returns ${result}")
return(PROPAGATE result)
endfunction(add_test_executable TEST_NAME)
include(GoogleTest)
function(add_gtest_executable TEST_NAME)
message("adding gtest ${TEST_NAME}")
add_executable(${TEST_NAME} ${ARGN})
add_dependencies(tests ${TEST_NAME})
add_dependencies(check ${TEST_NAME})
set(result 1)
if(DEFINED DTYPES)
foreach(source IN LISTS ARGN)
set(test 0)
foreach(type IN LISTS DTYPES)
if(type MATCHES "fp16")
set(type1 "_f16")
elseif(type MATCHES "fp32")
set(type1 "_f32")
elseif(type MATCHES "fp8")
set(type1 "_f8")
elseif(type MATCHES "bf16")
set(type1 "_b16")
elseif(type MATCHES "fp64")
set(type1 "_f64")
elseif(type MATCHES "int8")
set(type1 "_i8")
endif()
if("${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)
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")
message("removing dl test ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
#only continue if there are some source files left on the list
if(ARGN)
add_executable(${TEST_NAME} ${ARGN})
add_dependencies(tests ${TEST_NAME})
add_dependencies(check ${TEST_NAME})
# suppress gtest warnings
target_compile_options(${TEST_NAME} PRIVATE -Wno-global-constructors -Wno-undef)
target_link_libraries(${TEST_NAME} PRIVATE gtest_main)
add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}>)
rocm_install(TARGETS ${TEST_NAME} COMPONENT tests)
# suppress gtest warnings
target_compile_options(${TEST_NAME} PRIVATE -Wno-global-constructors -Wno-undef)
target_link_libraries(${TEST_NAME} PRIVATE gtest_main)
add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}>)
rocm_install(TARGETS ${TEST_NAME} COMPONENT tests)
set(result 0)
endif()
#message("add_gtest returns ${result}")
return(PROPAGATE result)
endfunction(add_gtest_executable TEST_NAME)
add_subdirectory(magic_number_division)
......
test/batched_gemm/CMakeLists.txt
View file @ 84dcf5d0
......@@ -2,25 +2,21 @@ 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)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_test_executable(test_batched_gemm_fp16 batched_gemm_fp16.cpp)
target_link_libraries(test_batched_gemm_fp16 PRIVATE utility)
target_link_libraries(test_batched_gemm_fp16 PRIVATE device_batched_gemm_instance)
add_test_executable(test_batched_gemm_fp16 batched_gemm_fp16.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_fp16 PRIVATE utility device_batched_gemm_instance)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_test_executable(test_batched_gemm_fp32 batched_gemm_fp32.cpp)
target_link_libraries(test_batched_gemm_fp32 PRIVATE utility)
target_link_libraries(test_batched_gemm_fp32 PRIVATE device_batched_gemm_instance)
add_test_executable(test_batched_gemm_fp32 batched_gemm_fp32.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_fp32 PRIVATE utility device_batched_gemm_instance)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_test_executable(test_batched_gemm_bf16 batched_gemm_bf16.cpp)
target_link_libraries(test_batched_gemm_bf16 PRIVATE utility)
target_link_libraries(test_batched_gemm_bf16 PRIVATE device_batched_gemm_instance)
add_test_executable(test_batched_gemm_bf16 batched_gemm_bf16.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_bf16 PRIVATE utility device_batched_gemm_instance)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_test_executable(test_batched_gemm_int8 batched_gemm_int8.cpp)
target_link_libraries(test_batched_gemm_int8 PRIVATE utility)
target_link_libraries(test_batched_gemm_int8 PRIVATE device_batched_gemm_instance)
add_test_executable(test_batched_gemm_int8 batched_gemm_int8.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_int8 PRIVATE utility device_batched_gemm_instance)
endif()
set(target 1)
endif()
......
test/batched_gemm_gemm/CMakeLists.txt
View file @ 84dcf5d0
......@@ -2,12 +2,12 @@ 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)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_custom_target(test_batched_gemm_gemm)
add_gtest_executable(test_batched_gemm_gemm_fp16 test_batched_gemm_gemm_fp16.cpp)
add_custom_target(test_batched_gemm_gemm)
add_gtest_executable(test_batched_gemm_gemm_fp16 test_batched_gemm_gemm_fp16.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_gemm_fp16 PRIVATE utility device_batched_gemm_gemm_instance)
add_dependencies(test_batched_gemm_gemm test_batched_gemm_gemm_fp16)
set(target 1)
endif()
endif()
endif()
endforeach()
\ No newline at end of file
test/batched_gemm_multi_d/CMakeLists.txt
View file @ 84dcf5d0
if(DL_KERNELS)
add_gtest_executable(test_batched_gemm_multi_d test_batched_gemm_multi_d.cpp)
add_gtest_executable(test_batched_gemm_multi_d test_batched_gemm_multi_d_dl.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_multi_d PRIVATE utility device_batched_gemm_multi_d_instance)
endif()
test/batched_gemm_reduce/CMakeLists.txt
View file @ 84dcf5d0
......@@ -2,10 +2,9 @@ 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)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_test_executable(test_batched_gemm_reduce_fp16 batched_gemm_reduce_fp16.cpp)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE utility)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE device_batched_gemm_reduce_instance)
add_test_executable(test_batched_gemm_reduce_fp16 batched_gemm_reduce_fp16.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE utility device_batched_gemm_reduce_instance)
set(target 1)
endif()
endif()
......
test/batched_gemm_softmax_gemm/CMakeLists.txt
View file @ 84dcf5d0
......@@ -2,12 +2,12 @@ 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)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_custom_target(test_batched_gemm_softmax_gemm)
add_gtest_executable(test_batched_gemm_softmax_gemm_fp16 test_batched_gemm_softmax_gemm_fp16.cpp)
target_link_libraries(test_batched_gemm_softmax_gemm_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_instance)
add_dependencies(test_batched_gemm_softmax_gemm test_batched_gemm_softmax_gemm_fp16)
set(target 1)
endif()
add_custom_target(test_batched_gemm_softmax_gemm)
add_gtest_executable(test_batched_gemm_softmax_gemm_fp16 test_batched_gemm_softmax_gemm_fp16.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_softmax_gemm_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_instance)
add_dependencies(test_batched_gemm_softmax_gemm test_batched_gemm_softmax_gemm_fp16)
set(target 1)
endif()
endif()
endforeach()
\ No newline at end of file
test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
View file @ 84dcf5d0
......@@ -2,25 +2,28 @@ 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)
if(DTYPES MATCHES "fp16" OR DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_custom_target(test_batched_gemm_softmax_gemm_permute)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_gtest_executable(test_batched_gemm_softmax_gemm_permute_fp16 test_batched_gemm_softmax_gemm_permute_fp16.cpp)
add_gtest_executable(test_batched_gemm_bias_softmax_gemm_permute_fp16 test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp)
target_link_libraries(test_batched_gemm_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_fp16)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_fp16)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_gtest_executable(test_batched_gemm_softmax_gemm_permute_bf16 test_batched_gemm_softmax_gemm_permute_bf16.cpp)
add_gtest_executable(test_batched_gemm_bias_softmax_gemm_permute_bf16 test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp)
target_link_libraries(test_batched_gemm_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_bf16)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_bf16)
endif()
add_custom_target(test_batched_gemm_softmax_gemm_permute)
add_gtest_executable(test_batched_gemm_softmax_gemm_permute_fp16 test_batched_gemm_softmax_gemm_permute_fp16.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_fp16)
endif()
add_gtest_executable(test_batched_gemm_bias_softmax_gemm_permute_fp16 test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_fp16)
endif()
add_gtest_executable(test_batched_gemm_softmax_gemm_permute_bf16 test_batched_gemm_softmax_gemm_permute_bf16.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_bf16)
endif()
add_gtest_executable(test_batched_gemm_bias_softmax_gemm_permute_bf16 test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_bf16)
endif()
set(target 1)
endif()
endforeach()
\ No newline at end of file
test/data_type/CMakeLists.txt
View file @ 84dcf5d0
if (USE_BITINT_EXTENSION_INT4)
add_gtest_executable(test_int4 int4.cpp)
target_link_libraries(test_int4 PRIVATE utility)
if(result EQUAL 0)
target_link_libraries(test_int4 PRIVATE utility)
endif()
endif()
add_gtest_executable(test_fp8 fp8.cpp)
target_link_libraries(test_fp8 PRIVATE utility)
if(result EQUAL 0)
target_link_libraries(test_fp8 PRIVATE utility)
endif()
add_gtest_executable(test_bf8 bf8.cpp)
if(result EQUAL 0)
target_link_libraries(test_bf8 PRIVATE utility)
endif()
test/data_type/bf8.cpp 0 → 100644
View file @ 84dcf5d0
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, 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::f8_convert_sr;
using ck::half_t;
using ck::type_convert;
TEST(BF8, NumericLimits)
{
// constants given for negative zero nan mode
EXPECT_EQ(ck::NumericLimits<bf8_t>::Min(), type_convert<bf8_t>(0x04));
EXPECT_EQ(ck::NumericLimits<bf8_t>::Max(), type_convert<bf8_t>(0x7F));
EXPECT_EQ(ck::NumericLimits<bf8_t>::Lowest(), type_convert<bf8_t>(0xFF));
EXPECT_EQ(ck::NumericLimits<bf8_t>::QuietNaN(), type_convert<bf8_t>(0x80));
}
TEST(BF8, ConvertFP32Nearest)
{
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to bf8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(type_convert<bf8_t>(0.0f)), abs_tol);
// convert minimal float to bf8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(type_convert<bf8_t>(std::numeric_limits<float>::min())),
abs_tol);
// convert maximal bf8_t to float and check if equal to 57344.0
ASSERT_NEAR(57344.0f, type_convert<float>(type_convert<bf8_t>(57344.0f)), abs_tol);
// convert maximal float to bf8 and back, check if clipped to 57344.0
ASSERT_NEAR(57344.0f,
type_convert<float>(type_convert<bf8_t>(std::numeric_limits<float>::max())),
abs_tol);
// convert inf float to bf8_t and check if it is qNan
ASSERT_NEAR(type_convert<bf8_t>(0x80),
type_convert<bf8_t>(std::numeric_limits<float>::infinity()),
abs_tol);
// positive norm float value to bf8 and back, check if holds
float pos_float = 0.0000762939f;
ASSERT_NEAR(pos_float, type_convert<float>(type_convert<bf8_t>(pos_float)), abs_tol);
// negative norm float value to bf8 and back, check if holds
float neg_float = -0.0000610351f;
ASSERT_NEAR(neg_float, type_convert<float>(type_convert<bf8_t>(neg_float)), abs_tol);
// positive subnorm float value to bf8 and back, check if holds
pos_float = 0.0000305175f;
ASSERT_NEAR(pos_float, type_convert<float>(type_convert<bf8_t>(pos_float)), abs_tol);
// negative subnorm float value to bf8 and back, check if holds
neg_float = -0.0000152587f;
ASSERT_NEAR(neg_float, type_convert<float>(type_convert<bf8_t>(neg_float)), abs_tol);
}
TEST(BF8, ConvertFP32Stochastic)
{
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to bf8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<bf8_t>(0.0f)), abs_tol);
// convert minimal float to bf8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(f8_convert_sr<bf8_t>(std::numeric_limits<float>::min())),
abs_tol);
// convert maximal bf8_t to float and check if equal to 57344.0
ASSERT_NEAR(57344.0f, type_convert<float>(f8_convert_sr<bf8_t>(57344.0f)), abs_tol);
// convert maximal float to bf8 and back, check if clipped to 57344.0
ASSERT_NEAR(57344.0f,
type_convert<float>(f8_convert_sr<bf8_t>(std::numeric_limits<float>::max())),
abs_tol);
// convert inf float to bf8_t and check if it is qNan
ASSERT_NEAR(type_convert<bf8_t>(0x80),
f8_convert_sr<bf8_t>(std::numeric_limits<float>::infinity()),
abs_tol);
// positive norm float value to bf8 and back, check if holds
float pos_float = 0.0000762939f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_t>(pos_float)), abs_tol);
// negative norm float value to bf8 and back, check if holds
float neg_float = -0.0000610351f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<bf8_t>(neg_float)), abs_tol);
// positive subnorm float value to bf8 and back, check if holds
pos_float = 0.0000305175f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_t>(pos_float)), abs_tol);
// negative subnorm float value to bf8 and back, check if holds
neg_float = -0.0000152587f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<bf8_t>(neg_float)), abs_tol);
}
TEST(BF8, ConvertFP16Nearest)
{
// fix the tolerance value
float abs_tol = 1e-3;
// convert 0 fp16 to bf8 and back, check if holds
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(type_convert<bf8_t>(half_t{0.0})), abs_tol);
// convert minimal fp16 to bf8 and back, check if holds
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(type_convert<bf8_t>(ck::NumericLimits<half_t>::Min())),
abs_tol);
// convert maximal bf8_t to fp16 and check if equal to 57344.0
ASSERT_NEAR(
half_t{57344.0}, type_convert<half_t>(type_convert<bf8_t>(half_t{57344.0})), abs_tol);
// convert maximal fp16 to bf8 and back, check if clipped to 57344.0
ASSERT_NEAR(half_t{57344.0},
type_convert<half_t>(type_convert<bf8_t>(ck::NumericLimits<half_t>::Max())),
abs_tol);
// convert QuietNaN fp16 to bf8_t and check if it is QuietNaN
ASSERT_NEAR(type_convert<bf8_t>(0x80),
type_convert<bf8_t>(ck::NumericLimits<half_t>::QuietNaN()),
abs_tol);
// positive norm fp16 value to bf8 and back, check if holds
half_t pos_half = half_t{0.0000762939};
ASSERT_NEAR(pos_half, type_convert<half_t>(type_convert<bf8_t>(pos_half)), abs_tol);
// negative norm fp16 value to bf8 and back, check if holds
half_t neg_half = half_t{-0.0000610351};
ASSERT_NEAR(neg_half, type_convert<half_t>(type_convert<bf8_t>(neg_half)), abs_tol);
// positive subnorm fp16 value to bf8 and back, check if holds
pos_half = half_t{0.0000305175};
ASSERT_NEAR(pos_half, type_convert<half_t>(type_convert<bf8_t>(pos_half)), abs_tol);
// negative subnorm fp16 value to bf8 and back, check if holds
neg_half = half_t{-0.0000152587};
ASSERT_NEAR(neg_half, type_convert<half_t>(type_convert<bf8_t>(neg_half)), abs_tol);
}
TEST(BF8, ConvertFP16Stochastic)
{
// fix the tolerance value
float abs_tol = 1e-3;
// convert 0 fp16 to bf8 and back, check if holds
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_sr<bf8_t>(half_t{0.0})), abs_tol);
// convert minimal fp16 to bf8 and back, check if holds
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(f8_convert_sr<bf8_t>(ck::NumericLimits<half_t>::Min())),
abs_tol);
// convert maximal bf8_t to fp16 and check if equal to 57344.0
ASSERT_NEAR(
half_t{57344.0}, type_convert<half_t>(f8_convert_sr<bf8_t>(half_t{57344.0})), abs_tol);
// convert maximal fp16 to bf8 and back, check if clipped to 57344.0
ASSERT_NEAR(half_t{57344.0},
type_convert<half_t>(f8_convert_sr<bf8_t>(ck::NumericLimits<half_t>::Max())),
abs_tol);
// convert QuietNaN fp16 to bf8_t and check if it is QuietNaN
ASSERT_NEAR(type_convert<bf8_t>(0x80),
f8_convert_sr<bf8_t>(ck::NumericLimits<half_t>::QuietNaN()),
abs_tol);
// positive norm fp16 value to bf8 and back, check if holds
half_t pos_half = half_t{0.0000762939};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<bf8_t>(pos_half)), abs_tol);
// negative norm fp16 value to bf8 and back, check if holds
half_t neg_half = half_t{-0.0000610351};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<bf8_t>(neg_half)), abs_tol);
// positive subnorm fp16 value to bf8 and back, check if holds
pos_half = half_t{0.0000305175};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<bf8_t>(pos_half)), abs_tol);
// negative subnorm fp16 value to bf8 and back, check if holds
neg_half = half_t{-0.0000152587};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<bf8_t>(neg_half)), abs_tol);
}
test/data_type/fp8.cpp
View file @ 84dcf5d0
......@@ -12,10 +12,11 @@ using ck::type_convert;
TEST(FP8, NumericLimits)
{
EXPECT_EQ(ck::NumericLimits<f8_t>::Min(), 0x08);
EXPECT_EQ(ck::NumericLimits<f8_t>::Max(), 0x77);
EXPECT_EQ(ck::NumericLimits<f8_t>::Lowest(), 0xF7);
EXPECT_EQ(ck::NumericLimits<f8_t>::QuietNaN(), 0x80);
// constants given for negative zero nan mode
EXPECT_EQ(ck::NumericLimits<f8_t>::Min(), type_convert<f8_t>(0x08));
EXPECT_EQ(ck::NumericLimits<f8_t>::Max(), type_convert<f8_t>(0x7F));
EXPECT_EQ(ck::NumericLimits<f8_t>::Lowest(), type_convert<f8_t>(0xFF));
EXPECT_EQ(ck::NumericLimits<f8_t>::QuietNaN(), type_convert<f8_t>(0x80));
}
TEST(FP8, ConvertFP32Nearest)
......@@ -35,12 +36,20 @@ TEST(FP8, ConvertFP32Nearest)
type_convert<float>(type_convert<f8_t>(std::numeric_limits<float>::max())),
abs_tol);
// convert inf float to f8_t and check if it is qNan
ASSERT_NEAR(0x80, type_convert<f8_t>(std::numeric_limits<float>::infinity()), abs_tol);
// positive float value to fp8 and back, check if holds
float pos_float = 0.0078125f;
ASSERT_NEAR(type_convert<f8_t>(0x80),
type_convert<f8_t>(std::numeric_limits<float>::infinity()),
abs_tol);
// positive norm float value to fp8 and back, check if holds
float pos_float = 0.017578125f;
ASSERT_NEAR(pos_float, type_convert<float>(type_convert<f8_t>(pos_float)), abs_tol);
// negative norm float value to fp8 and back, check if holds
float neg_float = -0.015625f;
ASSERT_NEAR(neg_float, type_convert<float>(type_convert<f8_t>(neg_float)), abs_tol);
// positive subnorm float value to fp8 and back, check if holds
pos_float = 0.00390625f;
ASSERT_NEAR(pos_float, type_convert<float>(type_convert<f8_t>(pos_float)), abs_tol);
// negative float value to fp8 and back, check if holds
float neg_float = -0.0156250f;
// negative subnorm float value to fp8 and back, check if holds
neg_float = -0.001953125f;
ASSERT_NEAR(neg_float, type_convert<float>(type_convert<f8_t>(neg_float)), abs_tol);
}
......@@ -61,12 +70,20 @@ TEST(FP8, ConvertFP32Stochastic)
type_convert<float>(f8_convert_sr<f8_t>(std::numeric_limits<float>::max())),
abs_tol);
// convert inf float to f8_t and check if it is qNan
ASSERT_NEAR(0x80, f8_convert_sr<f8_t>(std::numeric_limits<float>::infinity()), abs_tol);
// positive float value to fp8 and back, check if holds
float pos_float = 0.0078125f;
ASSERT_NEAR(type_convert<f8_t>(0x80),
f8_convert_sr<f8_t>(std::numeric_limits<float>::infinity()),
abs_tol);
// positive norm float value to fp8 and back, check if holds
float pos_float = 0.017578125f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<f8_t>(pos_float)), abs_tol);
// negative norm float value to fp8 and back, check if holds
float neg_float = -0.015625f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<f8_t>(neg_float)), abs_tol);
// positive subnorm float value to fp8 and back, check if holds
pos_float = 0.00390625f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<f8_t>(pos_float)), abs_tol);
// negative float value to fp8 and back, check if holds
float neg_float = -0.0156250f;
// negative subnorm float value to fp8 and back, check if holds
neg_float = -0.001953125f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<f8_t>(neg_float)), abs_tol);
}
......@@ -87,12 +104,20 @@ TEST(FP8, ConvertFP16Nearest)
type_convert<half_t>(type_convert<f8_t>(ck::NumericLimits<half_t>::Max())),
abs_tol);
// convert QuietNaN fp16 to f8_t and check if it is QuietNaN
ASSERT_NEAR(0x80, type_convert<f8_t>(ck::NumericLimits<half_t>::QuietNaN()), abs_tol);
// positive fp16 value to fp8 and back, check if holds
half_t pos_half = half_t{0.0078125};
ASSERT_NEAR(type_convert<f8_t>(0x80),
type_convert<f8_t>(ck::NumericLimits<half_t>::QuietNaN()),
abs_tol);
// positive norm fp16 value to fp8 and back, check if holds
half_t pos_half = half_t{0.017578125};
ASSERT_NEAR(pos_half, type_convert<half_t>(type_convert<f8_t>(pos_half)), abs_tol);
// negative norm fp16 value to fp8 and back, check if holds
half_t neg_half = half_t{-0.015625};
ASSERT_NEAR(neg_half, type_convert<half_t>(type_convert<f8_t>(neg_half)), abs_tol);
// positive subnorm fp16 value to fp8 and back, check if holds
pos_half = half_t{0.00390625};
ASSERT_NEAR(pos_half, type_convert<half_t>(type_convert<f8_t>(pos_half)), abs_tol);
// negative fp16 value to fp8 and back, check if holds
half_t neg_half = half_t{-0.0156250};
// negative subnorm fp16 value to fp8 and back, check if holds
neg_half = half_t{-0.001953125};
ASSERT_NEAR(neg_half, type_convert<half_t>(type_convert<f8_t>(neg_half)), abs_tol);
}
......@@ -113,11 +138,19 @@ TEST(FP8, ConvertFP16Stochastic)
type_convert<half_t>(f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::Max())),
abs_tol);
// convert QuietNaN fp16 to f8_t and check if it is QuietNaN
ASSERT_NEAR(0x80, f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::QuietNaN()), abs_tol);
// positive fp16 value to fp8 and back, check if holds
half_t pos_half = half_t{0.0078125};
ASSERT_NEAR(type_convert<f8_t>(0x80),
f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::QuietNaN()),
abs_tol);
// positive norm fp16 value to fp8 and back, check if holds
half_t pos_half = half_t{0.017578125};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<f8_t>(pos_half)), abs_tol);
// negative norm fp16 value to fp8 and back, check if holds
half_t neg_half = half_t{-0.015625};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<f8_t>(neg_half)), abs_tol);
// positive subnorm fp16 value to fp8 and back, check if holds
pos_half = half_t{0.00390625};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<f8_t>(pos_half)), abs_tol);
// negative fp16 value to fp8 and back, check if holds
half_t neg_half = half_t{-0.0156250};
// negative subnorm fp16 value to fp8 and back, check if holds
neg_half = half_t{-0.001953125};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<f8_t>(neg_half)), abs_tol);
}
test/elementwise_normalization/CMakeLists.txt
View file @ 84dcf5d0
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_custom_target(test_elementwise_normalization)
add_gtest_executable(test_elementwise_layernorm_fp16 test_elementwise_layernorm_fp16.cpp)
add_custom_target(test_elementwise_normalization)
add_gtest_executable(test_elementwise_layernorm_fp16 test_elementwise_layernorm_fp16.cpp)
if(result EQUAL 0)
target_link_libraries(test_elementwise_layernorm_fp16 PRIVATE utility device_elementwise_normalization_instance)
add_dependencies(test_elementwise_normalization test_elementwise_layernorm_fp16)
endif()
\ No newline at end of file
test/gemm/CMakeLists.txt
View file @ 84dcf5d0
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_test_executable(test_gemm_fp32 gemm_fp32.cpp)
target_link_libraries(test_gemm_fp32 PRIVATE utility)
target_link_libraries(test_gemm_fp32 PRIVATE device_gemm_instance)
if(result EQUAL 0)
target_link_libraries(test_gemm_fp32 PRIVATE utility device_gemm_instance)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_test_executable(test_gemm_fp16 gemm_fp16.cpp)
target_link_libraries(test_gemm_fp16 PRIVATE utility)
target_link_libraries(test_gemm_fp16 PRIVATE device_gemm_instance)
add_library(gemm_standalone_xdl_fp16_instances STATIC
if(result EQUAL 0)
target_link_libraries(test_gemm_fp16 PRIVATE utility device_gemm_instance)
add_library(gemm_standalone_xdl_fp16_instances STATIC
instance/gemm_f16_nn_instance.cpp
instance/gemm_f16_nt_instance.cpp
instance/gemm_f16_tn_instance.cpp
instance/gemm_wavelet_f16_tn_instance.cpp
instance/gemm_f16_tt_instance.cpp
)
)
endif()
add_test_executable(test_gemm_standalone_xdl_fp16 gemm_standalone_xdl_fp16.cpp)
target_link_libraries(test_gemm_standalone_xdl_fp16 PRIVATE gemm_standalone_xdl_fp16_instances utility)
target_include_directories(test_gemm_standalone_xdl_fp16 PRIVATE instance/)
if(result EQUAL 0)
target_link_libraries(test_gemm_standalone_xdl_fp16 PRIVATE gemm_standalone_xdl_fp16_instances utility)
target_include_directories(test_gemm_standalone_xdl_fp16 PRIVATE instance/)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_test_executable(test_gemm_bf16 gemm_bf16.cpp)
target_link_libraries(test_gemm_bf16 PRIVATE utility)
target_link_libraries(test_gemm_bf16 PRIVATE device_gemm_instance)
if(result EQUAL 0)
target_link_libraries(test_gemm_bf16 PRIVATE utility device_gemm_instance)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_test_executable(test_gemm_int8 gemm_int8.cpp)
target_link_libraries(test_gemm_int8 PRIVATE utility)
target_link_libraries(test_gemm_int8 PRIVATE device_gemm_instance)
if(result EQUAL 0)
target_link_libraries(test_gemm_int8 PRIVATE utility device_gemm_instance)
endif()
\ No newline at end of file
test/gemm_layernorm/CMakeLists.txt
View file @ 84dcf5d0
......@@ -2,12 +2,12 @@ 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)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_custom_target(test_gemm_layernorm)
add_gtest_executable(test_gemm_add_relu_add_layernorm_fp16 test_gemm_add_relu_add_layernorm_fp16.cpp)
target_link_libraries(test_gemm_add_relu_add_layernorm_fp16 PRIVATE utility device_gemm_add_relu_add_layernorm_instance)
add_dependencies(test_gemm_layernorm test_gemm_add_relu_add_layernorm_fp16)
set(target 1)
endif()
if(result EQUAL 0)
target_link_libraries(test_gemm_add_relu_add_layernorm_fp16 PRIVATE utility device_gemm_add_relu_add_layernorm_instance)
add_dependencies(test_gemm_layernorm test_gemm_add_relu_add_layernorm_fp16)
set(target 1)
endif()
endif()
endforeach()
test/gemm_reduce/CMakeLists.txt
View file @ 84dcf5d0
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_test_executable(test_gemm_reduce_fp16 gemm_reduce_fp16.cpp)
target_link_libraries(test_gemm_reduce_fp16 PRIVATE utility)
target_link_libraries(test_gemm_reduce_fp16 PRIVATE device_gemm_reduce_instance)
add_test_executable(test_gemm_reduce_fp16 gemm_reduce_fp16.cpp)
if(result EQUAL 0)
target_link_libraries(test_gemm_reduce_fp16 PRIVATE utility device_gemm_reduce_instance)
endif()
\ No newline at end of file
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