/******************************************************************************* * * MIT License * * Copyright (c) 2022-2025 Advanced Micro Devices, Inc. * * Permission is hereby granted, free of charge, to any person obtaining a copy * of this software and associated documentation files (the "Software"), to deal * in the Software without restriction, including without limitation the rights * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell * copies of the Software, and to permit persons to whom the Software is * furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice shall be included in * all copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. * *******************************************************************************/ /* ========================================Gtest Unit Check * ==================================================== */ /*! \brief gtest unit compare two matrices float/double/complex */ #pragma once #include "hipsparselt_math.hpp" #include "hipsparselt_test.hpp" #include "hipsparselt_vector.hpp" #include #include #ifndef GOOGLE_TEST #define UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, UNIT_ASSERT_EQ) #define UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, UNIT_ASSERT_EQ) #else #define UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, UNIT_ASSERT_EQ) \ do \ { \ for(size_t k = 0; k < batch_count; k++) \ for(size_t j = 0; j < N; j++) \ for(size_t i = 0; i < M; i++) \ if(hipsparselt_isnan(hCPU[i + j * size_t(lda) + k * strideA])) \ { \ ASSERT_TRUE(hipsparselt_isnan(hGPU[i + j * size_t(lda) + k * strideA])); \ } \ else \ { \ UNIT_ASSERT_EQ(hCPU[i + j * size_t(lda) + k * strideA], \ hGPU[i + j * size_t(lda) + k * strideA]); \ } \ } while(0) #define UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, UNIT_ASSERT_EQ) \ do \ { \ for(size_t k = 0; k < batch_count; k++) \ for(size_t j = 0; j < N; j++) \ for(size_t i = 0; i < M; i++) \ if(hipsparselt_isnan(hCPU[k][i + j * size_t(lda)])) \ { \ ASSERT_TRUE(hipsparselt_isnan(hGPU[k][i + j * size_t(lda)])); \ } \ else \ { \ UNIT_ASSERT_EQ(hCPU[k][i + j * size_t(lda)], \ hGPU[k][i + j * size_t(lda)]); \ } \ } while(0) //#define ASSERT_HALF_EQ(a, b) ASSERT_FLOAT_EQ(float(a), float(b)) //#define ASSERT_BF16_EQ(a, b) ASSERT_FLOAT_EQ(float(a), float(b)) #if defined(__HIP_PLATFORM_AMD__) #define ASSERT_HALF_EQ(a, b) \ do \ { \ union _HALF \ { \ uint16_t x; \ _Float16 data; \ }; \ _HALF a_half = {__half_raw(a).x}; \ _HALF b_half = {__half_raw(b).x}; \ auto absA = (a_half.data > 0) ? a_half.data : -a_half.data; \ auto absB = (b_half.data > 0) ? b_half.data : -b_half.data; \ auto absDiff = (a_half.data - b_half.data > 0) ? a_half.data - b_half.data \ : b_half.data - a_half.data; \ ASSERT_TRUE(absDiff / (absA + absB + 1) < 0.01); \ } while(0) #else #define ASSERT_HALF_EQ(a, b) \ do \ { \ auto zero = __half(0.0); \ auto absA = (a > zero) ? a : negate(a); \ auto absB = (b > zero) ? b : negate(b); \ auto absDiff = (a - b > zero) ? a - b : b - a; \ ASSERT_TRUE(absDiff / (absA + absB + __half(1.0)) < __half(0.01)); \ } while(0) #endif #define ASSERT_BF16_EQ(a, b) \ do \ { \ const hip_bfloat16 bf16A = static_cast(a); \ const hip_bfloat16 bf16B = static_cast(b); \ const hip_bfloat16 bf16Zero = static_cast(0.0f); \ const hip_bfloat16 bf16One = static_cast(1.0f); \ hip_bfloat16 absA = (bf16A > bf16Zero) ? bf16A : negate(bf16A); \ hip_bfloat16 absB = (bf16B > bf16Zero) ? bf16B : negate(bf16B); \ hip_bfloat16 absDiff = (bf16A - bf16B > bf16Zero) ? bf16A - bf16B : bf16B - bf16A; \ ASSERT_TRUE(absDiff / (absA + absB + bf16One) < static_cast(0.1f)); \ } while(0) #define ASSERT_FP8_EQ(a, b) \ do \ { \ const float f32A = static_cast(a); \ const float f32B = static_cast(b); \ float absA = (f32A > 0.0f) ? f32A : negate(f32A); \ float absB = (f32B > 0.0f) ? f32B : negate(f32B); \ float absDiff = (f32A - f32B > 0.0f) ? f32A - f32B : f32B - f32A; \ ASSERT_TRUE(absDiff / (absA + absB + 1.0f) < 0.125f); \ /*tolerance * eps = 2 * 0.0625; 2*eps needed for SR*/ \ } while(0) #define ASSERT_BF8_EQ(a, b) \ do \ { \ const float f32A = static_cast(a); \ const float f32B = static_cast(b); \ float absA = (f32A > 0.0f) ? f32A : negate(f32A); \ float absB = (f32B > 0.0f) ? f32B : negate(f32B); \ float absDiff = (f32A - f32B > 0.0f) ? f32A - f32B : f32B - f32A; \ ASSERT_TRUE(absDiff / (absA + absB + 1.0f) < 0.25f); \ /*tolerance * eps= 2 * 0.125; 2*eps needed for SR*/ \ } while(0) // Compare float to hip_bfloat16 // Allow the hip_bfloat16 to match the rounded or truncated value of float // Only call ASSERT_FLOAT_EQ with the rounded value if the truncated value does not match #include #define ASSERT_FLOAT_BF16_EQ(a, b) \ do \ { \ using testing::internal::FloatingPoint; \ if(!FloatingPoint(b).AlmostEquals( \ FloatingPoint(float_to_bfloat16_truncate(a)))) \ ASSERT_FLOAT_EQ(b, hip_bfloat16(a)); \ } while(0) #define ASSERT_FLOAT_COMPLEX_EQ(a, b) \ do \ { \ auto ta = (a), tb = (b); \ ASSERT_FLOAT_EQ(std::real(ta), std::real(tb)); \ ASSERT_FLOAT_EQ(std::imag(ta), std::imag(tb)); \ } while(0) #define ASSERT_DOUBLE_COMPLEX_EQ(a, b) \ do \ { \ auto ta = (a), tb = (b); \ ASSERT_DOUBLE_EQ(std::real(ta), std::real(tb)); \ ASSERT_DOUBLE_EQ(std::imag(ta), std::imag(tb)); \ } while(0) #endif // GOOGLE_TEST // TODO: Replace std::remove_cv_t with std::type_identity_t in C++20 // It is only used to make T_hpa non-deduced template void unit_check_general( int64_t M, int64_t N, int64_t lda, const std::remove_cv_t* hCPU, const T* hGPU); template <> inline void unit_check_general( int64_t M, int64_t N, int64_t lda, const hip_bfloat16* hCPU, const hip_bfloat16* hGPU) { UNIT_CHECK(M, N, lda, 0, hCPU, hGPU, 1, ASSERT_BF16_EQ); } template <> inline void unit_check_general( int64_t M, int64_t N, int64_t lda, const float* hCPU, const hip_bfloat16* hGPU) { UNIT_CHECK(M, N, lda, 0, hCPU, hGPU, 1, ASSERT_FLOAT_BF16_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const __half* hCPU, const __half* hGPU) { UNIT_CHECK(M, N, lda, 0, hCPU, hGPU, 1, ASSERT_HALF_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const float* hCPU, const float* hGPU) { UNIT_CHECK(M, N, lda, 0, hCPU, hGPU, 1, ASSERT_FLOAT_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const double* hCPU, const double* hGPU) { UNIT_CHECK(M, N, lda, 0, hCPU, hGPU, 1, ASSERT_DOUBLE_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const int64_t* hCPU, const int64_t* hGPU) { UNIT_CHECK(M, N, lda, 0, hCPU, hGPU, 1, ASSERT_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const int8_t* hCPU, const int8_t* hGPU) { UNIT_CHECK(M, N, lda, 0, hCPU, hGPU, 1, ASSERT_EQ); } #ifdef HIPSPARSELT_CLIENT_ENABLE_FP8_OCP template <> inline void unit_check_general( int64_t M, int64_t N, int64_t lda, const hipsparselt_fp8_e4m3* hCPU, const hipsparselt_fp8_e4m3* hGPU) { UNIT_CHECK(M, N, lda, 0, hCPU, hGPU, 1, ASSERT_FP8_EQ); } template <> inline void unit_check_general( int64_t M, int64_t N, int64_t lda, const hipsparselt_fp8_e5m2* hCPU, const hipsparselt_fp8_e5m2* hGPU) { UNIT_CHECK(M, N, lda, 0, hCPU, hGPU, 1, ASSERT_BF8_EQ); } #endif #ifdef HIPSPARSELT_CLIENT_ENABLE_FP8_FNUZ template <> inline void unit_check_general( int64_t M, int64_t N, int64_t lda, const hipsparselt_fp8_e4m3_fnuz* hCPU, const hipsparselt_fp8_e4m3_fnuz* hGPU) { UNIT_CHECK(M, N, lda, 0, hCPU, hGPU, 1, ASSERT_FP8_EQ); } template <> inline void unit_check_general( int64_t M, int64_t N, int64_t lda, const hipsparselt_fp8_e5m2_fnuz* hCPU, const hipsparselt_fp8_e5m2_fnuz* hGPU) { UNIT_CHECK(M, N, lda, 0, hCPU, hGPU, 1, ASSERT_BF8_EQ); } #endif template void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const std::remove_cv_t* hCPU, const T* hGPU, int64_t batch_count); template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const hip_bfloat16* hCPU, const hip_bfloat16* hGPU, int64_t batch_count) { UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, ASSERT_BF16_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const float* hCPU, const hip_bfloat16* hGPU, int64_t batch_count) { UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, ASSERT_FLOAT_BF16_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const __half* hCPU, const __half* hGPU, int64_t batch_count) { UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, ASSERT_HALF_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const float* hCPU, const float* hGPU, int64_t batch_count) { UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, ASSERT_FLOAT_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const double* hCPU, const double* hGPU, int64_t batch_count) { UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, ASSERT_DOUBLE_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const int64_t* hCPU, const int64_t* hGPU, int64_t batch_count) { UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, ASSERT_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const int8_t* hCPU, const int8_t* hGPU, int64_t batch_count) { UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, ASSERT_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const int32_t* hCPU, const int32_t* hGPU, int64_t batch_count) { UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, ASSERT_EQ); } #ifdef HIPSPARSELT_CLIENT_ENABLE_FP8_OCP template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const hipsparselt_fp8_e4m3* hCPU, const hipsparselt_fp8_e4m3* hGPU, int64_t batch_count) { UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, ASSERT_FP8_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const hipsparselt_fp8_e5m2* hCPU, const hipsparselt_fp8_e5m2* hGPU, int64_t batch_count) { UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, ASSERT_BF8_EQ); } #endif #ifdef HIPSPARSELT_CLIENT_ENABLE_FP8_FNUZ template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const hipsparselt_fp8_e4m3_fnuz* hCPU, const hipsparselt_fp8_e4m3_fnuz* hGPU, int64_t batch_count) { UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, ASSERT_FP8_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, int64_t strideA, const hipsparselt_fp8_e5m2_fnuz* hCPU, const hipsparselt_fp8_e5m2_fnuz* hGPU, int64_t batch_count) { UNIT_CHECK(M, N, lda, strideA, hCPU, hGPU, batch_count, ASSERT_BF8_EQ); } #endif template void unit_check_general(int64_t M, int64_t N, int64_t lda, const host_vector> hCPU[], const host_vector hGPU[], int64_t batch_count); template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const host_vector hCPU[], const host_vector hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_BF16_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const host_vector hCPU[], const host_vector hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_FLOAT_BF16_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const host_vector<__half> hCPU[], const host_vector<__half> hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_HALF_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const host_vector hCPU[], const host_vector hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const host_vector hCPU[], const host_vector hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const host_vector hCPU[], const host_vector hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_FLOAT_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const host_vector hCPU[], const host_vector hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_DOUBLE_EQ); } template void unit_check_general(int64_t M, int64_t N, int64_t lda, const std::remove_cv_t* const hCPU[], const T* const hGPU[], int64_t batch_count); template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const hip_bfloat16* const hCPU[], const hip_bfloat16* const hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_BF16_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const float* const hCPU[], const hip_bfloat16* const hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_FLOAT_BF16_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const __half* const hCPU[], const __half* const hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_HALF_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const int* const hCPU[], const int* const hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const int8_t* const hCPU[], const int8_t* const hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const float* const hCPU[], const float* const hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_FLOAT_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const double* const hCPU[], const double* const hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_DOUBLE_EQ); } #ifdef HIPSPARSELT_CLIENT_ENABLE_FP8_OCP template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const hipsparselt_fp8_e4m3* const hCPU[], const hipsparselt_fp8_e4m3* const hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_FP8_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const hipsparselt_fp8_e5m2* const hCPU[], const hipsparselt_fp8_e5m2* const hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_BF8_EQ); } #endif #ifdef HIPSPARSELT_CLIENT_ENABLE_FP8_FNUZ template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const hipsparselt_fp8_e4m3_fnuz* const hCPU[], const hipsparselt_fp8_e4m3_fnuz* const hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_FP8_EQ); } template <> inline void unit_check_general(int64_t M, int64_t N, int64_t lda, const hipsparselt_fp8_e5m2_fnuz* const hCPU[], const hipsparselt_fp8_e5m2_fnuz* const hGPU[], int64_t batch_count) { UNIT_CHECK_B(M, N, lda, hCPU, hGPU, batch_count, ASSERT_BF8_EQ); } #endif template inline void trsm_err_res_check(T max_error, int64_t M, T forward_tolerance, T eps) { #ifdef GOOGLE_TEST ASSERT_LE(max_error, forward_tolerance * eps * M); #endif } template constexpr double get_epsilon() { return std::numeric_limits::epsilon(); } template inline int64_t unit_check_diff( int64_t M, int64_t N, int64_t lda, int64_t stride, T* hCPU, T* hGPU, int64_t batch_count) { using c_type = std::conditional_t::value #ifdef HIPSPARSELT_CLIENT_ENABLE_FP8_OCP || std::is_same::value || std::is_same::value #endif #ifdef HIPSPARSELT_CLIENT_ENABLE_FP8_FNUZ || std::is_same::value || std::is_same::value #endif , float, T>; int64_t error = 0; do { for(size_t k = 0; k < batch_count; k++) for(size_t j = 0; j < N; j++) for(size_t i = 0; i < M; i++) if(hipsparselt_isnan(hCPU[i + j * size_t(lda) + k * stride])) { error += hipsparselt_isnan(hGPU[i + j * size_t(lda) + k * stride]) ? 0 : 1; } else { error += static_cast(hCPU[i + j * size_t(lda) + k * stride]) == static_cast( hGPU[i + j * size_t(lda) + k * stride]) ? 0 : 1; } } while(0); return error; }