Commit 4885c38a authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'transpose_opt' of https://github.com/ROCm/composable_kernel into rowwise_opt

parents cbf14ee1 7c8e92fa
if (GPU_TARGETS)
if (NOT GPU_TARGETS MATCHES "gfx94")
add_definitions(-DCK_SKIP_FLAKY_F8_TEST)
set(CK_SKIP_FLAKY_F8_TEST "ON")
endif()
else()
add_definitions(-DCK_SKIP_FLAKY_F8_TEST)
set(CK_SKIP_FLAKY_F8_TEST "ON")
endif()
if (USE_BITINT_EXTENSION_INT4)
add_gtest_executable(test_int4 test_int4.cpp)
if(result EQUAL 0)
......
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, 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::f8_convert_rne;
using ck::f8_convert_sr;
using ck::half_t;
using ck::type_convert;
......@@ -24,33 +25,36 @@ 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);
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_rne<bf8_t>(0.0f)), abs_tol);
// don't run the next test on gfx11 devices
#ifndef CK_SKIP_FLAKY_F8_TEST
// 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())),
type_convert<float>(f8_convert_rne<bf8_t>(std::numeric_limits<float>::min())),
abs_tol);
#endif
// 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);
ASSERT_NEAR(57344.0f, type_convert<float>(f8_convert_rne<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())),
type_convert<float>(f8_convert_rne<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()),
f8_convert_rne<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);
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<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);
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<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);
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<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);
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<bf8_t>(neg_float)), abs_tol);
}
TEST(BF8, ConvertFP32Stochastic)
......@@ -92,34 +96,34 @@ 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);
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_rne<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())),
type_convert<half_t>(f8_convert_rne<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);
half_t{57344.0}, type_convert<half_t>(f8_convert_rne<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())),
type_convert<half_t>(f8_convert_rne<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()),
f8_convert_rne<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);
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<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);
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<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);
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<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);
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<bf8_t>(neg_half)), abs_tol);
}
TEST(BF8, ConvertFP16Stochastic)
......
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, 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::f8_convert_rne;
using ck::f8_convert_sr;
using ck::f8_t;
using ck::half_t;
......@@ -24,33 +25,36 @@ TEST(FP8, ConvertFP32Nearest)
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to fp8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(type_convert<f8_t>(0.0f)), abs_tol);
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_rne<f8_t>(0.0f)), abs_tol);
// don't run the next test on gfx11 devices
#ifndef CK_SKIP_FLAKY_F8_TEST
// convert minimal float to fp8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(type_convert<f8_t>(std::numeric_limits<float>::min())),
type_convert<float>(f8_convert_rne<f8_t>(std::numeric_limits<float>::min())),
abs_tol);
#endif
// convert maximal f8_t to float and check if equal to 240.0
ASSERT_NEAR(240.0f, type_convert<float>(type_convert<f8_t>(240.0f)), abs_tol);
ASSERT_NEAR(240.0f, type_convert<float>(f8_convert_rne<f8_t>(240.0f)), abs_tol);
// convert maximal float to fp8 and back, check if clipped to 240.0
ASSERT_NEAR(240.0f,
type_convert<float>(type_convert<f8_t>(std::numeric_limits<float>::max())),
type_convert<float>(f8_convert_rne<f8_t>(std::numeric_limits<float>::max())),
abs_tol);
// convert inf float to f8_t and check if it is qNan
ASSERT_NEAR(type_convert<f8_t>(0x80),
type_convert<f8_t>(std::numeric_limits<float>::infinity()),
f8_convert_rne<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);
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<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);
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<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);
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<f8_t>(pos_float)), abs_tol);
// 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);
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<f8_t>(neg_float)), abs_tol);
}
TEST(FP8, ConvertFP32Stochastic)
......@@ -92,33 +96,33 @@ TEST(FP8, ConvertFP16Nearest)
// fix the tolerance value
float abs_tol = 1e-3;
// convert 0 fp16 to fp8 and back, check if holds
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(type_convert<f8_t>(half_t{0.0})), abs_tol);
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_rne<f8_t>(half_t{0.0})), abs_tol);
// convert minimal fp16 to fp8 and back, check if holds
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(type_convert<f8_t>(ck::NumericLimits<half_t>::Min())),
type_convert<half_t>(f8_convert_rne<f8_t>(ck::NumericLimits<half_t>::Min())),
abs_tol);
// convert maximal f8_t to fp16 and check if equal to 240.0
ASSERT_NEAR(half_t{240.0}, type_convert<half_t>(type_convert<f8_t>(half_t{240.0})), abs_tol);
ASSERT_NEAR(half_t{240.0}, type_convert<half_t>(f8_convert_rne<f8_t>(half_t{240.0})), abs_tol);
// convert maximal fp16 to fp8 and back, check if clipped to 240.0
ASSERT_NEAR(half_t{240.0},
type_convert<half_t>(type_convert<f8_t>(ck::NumericLimits<half_t>::Max())),
type_convert<half_t>(f8_convert_rne<f8_t>(ck::NumericLimits<half_t>::Max())),
abs_tol);
// convert QuietNaN fp16 to f8_t and check if it is QuietNaN
ASSERT_NEAR(type_convert<f8_t>(0x80),
type_convert<f8_t>(ck::NumericLimits<half_t>::QuietNaN()),
f8_convert_rne<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);
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<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);
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<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);
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<f8_t>(pos_half)), abs_tol);
// 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);
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<f8_t>(neg_half)), abs_tol);
}
TEST(FP8, ConvertFP16Stochastic)
......
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