From: Cordell Bloor Date: Sun, 9 Apr 2023 02:33:46 -0600 Subject: make openmp optional Applied-Upstream: https://github.com/ROCmSoftwarePlatform/rocBLAS/commit/45bef63fc4988d11f25c9cc84b12513c7b8abc81 --- clients/common/blis_interface.cpp | 2 + clients/common/cblas_interface.cpp | 12 +++++ clients/include/rocblas_init.hpp | 52 ++++++++++++++++++++++ clients/samples/example_openmp.cpp | 2 + tensile/HostLibraryTests/CachingLibrary_test.cpp | 2 + .../HostLibraryTests/testlib/include/TestUtils.hpp | 7 ++- 6 files changed, 76 insertions(+), 1 deletion(-) diff --git a/clients/common/blis_interface.cpp b/clients/common/blis_interface.cpp index da7aef3..b302d57 100644 --- a/clients/common/blis_interface.cpp +++ b/clients/common/blis_interface.cpp @@ -21,7 +21,9 @@ * ************************************************************************ */ #include "blis.h" +#ifdef _OPENMP #include "omp.h" +#endif void setup_blis() { diff --git a/clients/common/cblas_interface.cpp b/clients/common/cblas_interface.cpp index 2eb50ab..86b10cd 100644 --- a/clients/common/cblas_interface.cpp +++ b/clients/common/cblas_interface.cpp @@ -23,7 +23,9 @@ #include "rocblas_vector.hpp" #include "utility.hpp" #include +#ifdef _OPENMP #include +#endif /* * =========================================================================== @@ -462,7 +464,9 @@ void cblas_geam_helper(rocblas_operation transA, rocblas_int inc1_B = transB == rocblas_operation_none ? 1 : ldb; rocblas_int inc2_B = transB == rocblas_operation_none ? ldb : 1; +#ifdef _OPENMP #pragma omp parallel for +#endif for(rocblas_int i = 0; i < M; i++) { for(rocblas_int j = 0; j < N; j++) @@ -971,7 +975,9 @@ void cblas_herkx(rocblas_fill uplo, { if(uplo == rocblas_fill_upper) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(int j = 0; j < n; ++j) { for(int i = 0; i <= j; i++) @@ -993,7 +999,9 @@ void cblas_herkx(rocblas_fill uplo, } else // lower { +#ifdef _OPENMP #pragma omp parallel for +#endif for(int j = 0; j < n; ++j) { for(int i = j; i < n; i++) @@ -1018,7 +1026,9 @@ void cblas_herkx(rocblas_fill uplo, { if(uplo == rocblas_fill_upper) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(int j = 0; j < n; ++j) { for(int i = 0; i <= j; i++) @@ -1042,7 +1052,9 @@ void cblas_herkx(rocblas_fill uplo, } else // lower { +#ifdef _OPENMP #pragma omp parallel for +#endif for(int j = 0; j < n; ++j) { for(int i = j; i < n; i++) diff --git a/clients/include/rocblas_init.hpp b/clients/include/rocblas_init.hpp index e971b72..4a07fb9 100644 --- a/clients/include/rocblas_init.hpp +++ b/clients/include/rocblas_init.hpp @@ -29,7 +29,9 @@ #include "rocblas_random.hpp" #include #include +#ifdef _OPENMP #include +#endif #include //! @@ -70,7 +72,9 @@ void rocblas_init_matrix_alternating_sign(rocblas_check_matrix_type matrix_type, if(matrix_type == rocblas_client_general_matrix) { for(size_t b = 0; b < batch_count; b++) +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < M; ++i) for(size_t j = 0; j < N; ++j) { @@ -81,7 +85,9 @@ void rocblas_init_matrix_alternating_sign(rocblas_check_matrix_type matrix_type, else if(matrix_type == rocblas_client_triangular_matrix) { for(size_t b = 0; b < batch_count; b++) +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < M; ++i) for(size_t j = 0; j < N; ++j) { @@ -107,7 +113,9 @@ void rocblas_init_matrix_alternating_sign(rocblas_check_matrix_type matrix_type, if(matrix_type == rocblas_client_general_matrix) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < M; ++i) for(size_t j = 0; j < N; ++j) { @@ -117,7 +125,9 @@ void rocblas_init_matrix_alternating_sign(rocblas_check_matrix_type matrix_type, } else if(matrix_type == rocblas_client_triangular_matrix) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < M; ++i) for(size_t j = 0; j < N; ++j) { @@ -136,7 +146,9 @@ void rocblas_init_vector_alternating_sign(T rand_gen(), T* x, rocblas_int N, roc if(incx < 0) x -= (N - 1) * incx; +#ifdef _OPENMP #pragma omp parallel for +#endif for(rocblas_int j = 0; j < N; ++j) { auto value = rand_gen(); @@ -162,7 +174,9 @@ void rocblas_init_matrix(rocblas_check_matrix_type matrix_type, if(matrix_type == rocblas_client_general_matrix) { for(size_t b = 0; b < batch_count; b++) +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < M; ++i) for(size_t j = 0; j < N; ++j) A[i + j * lda + b * stride] = rand_gen(); @@ -170,7 +184,9 @@ void rocblas_init_matrix(rocblas_check_matrix_type matrix_type, else if(matrix_type == rocblas_client_hermitian_matrix) { for(size_t b = 0; b < batch_count; ++b) +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < N; ++i) for(size_t j = 0; j <= i; ++j) { @@ -197,7 +213,9 @@ void rocblas_init_matrix(rocblas_check_matrix_type matrix_type, else if(matrix_type == rocblas_client_symmetric_matrix) { for(size_t b = 0; b < batch_count; ++b) +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < N; ++i) for(size_t j = 0; j <= i; ++j) { @@ -224,7 +242,9 @@ void rocblas_init_matrix(rocblas_check_matrix_type matrix_type, else if(matrix_type == rocblas_client_triangular_matrix) { for(size_t b = 0; b < batch_count; b++) +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < M; ++i) for(size_t j = 0; j < N; ++j) { @@ -249,14 +269,18 @@ void rocblas_init_matrix(rocblas_check_matrix_type matrix_type, auto lda = hA.lda(); if(matrix_type == rocblas_client_general_matrix) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < M; ++i) for(size_t j = 0; j < N; ++j) A[i + j * lda] = rand_gen(); } else if(matrix_type == rocblas_client_hermitian_matrix) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < N; ++i) for(size_t j = 0; j <= i; ++j) { @@ -282,7 +306,9 @@ void rocblas_init_matrix(rocblas_check_matrix_type matrix_type, } else if(matrix_type == rocblas_client_symmetric_matrix) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < N; ++i) for(size_t j = 0; j <= i; ++j) { @@ -308,7 +334,9 @@ void rocblas_init_matrix(rocblas_check_matrix_type matrix_type, } else if(matrix_type == rocblas_client_triangular_matrix) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < M; ++i) for(size_t j = 0; j < N; ++j) { @@ -329,7 +357,9 @@ void rocblas_init_vector(T rand_gen(), T* x, rocblas_int N, rocblas_stride incx) if(incx < 0) x -= (N - 1) * incx; +#ifdef _OPENMP #pragma omp parallel for +#endif for(rocblas_int j = 0; j < N; ++j) x[j * incx] = rand_gen(); } @@ -352,7 +382,9 @@ void rocblas_init_matrix_trig(rocblas_check_matrix_type matrix_type, if(matrix_type == rocblas_client_general_matrix) { for(size_t b = 0; b < batch_count; b++) +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < M; ++i) for(size_t j = 0; j < N; ++j) A[i + j * lda + b * stride] = T(seedReset ? cos(i + j * lda + b * stride) @@ -361,7 +393,9 @@ void rocblas_init_matrix_trig(rocblas_check_matrix_type matrix_type, else if(matrix_type == rocblas_client_hermitian_matrix) { for(size_t b = 0; b < batch_count; ++b) +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < N; ++i) for(size_t j = 0; j <= i; ++j) { @@ -390,7 +424,9 @@ void rocblas_init_matrix_trig(rocblas_check_matrix_type matrix_type, else if(matrix_type == rocblas_client_symmetric_matrix) { for(size_t b = 0; b < batch_count; ++b) +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < N; ++i) for(size_t j = 0; j <= i; ++j) { @@ -418,7 +454,9 @@ void rocblas_init_matrix_trig(rocblas_check_matrix_type matrix_type, else if(matrix_type == rocblas_client_triangular_matrix) { for(size_t b = 0; b < batch_count; b++) +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < M; ++i) for(size_t j = 0; j < N; ++j) { @@ -449,14 +487,18 @@ void rocblas_init_matrix_trig(rocblas_check_matrix_type matrix_type, if(matrix_type == rocblas_client_general_matrix) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < M; ++i) for(size_t j = 0; j < N; ++j) A[i + j * lda] = T(seedReset ? cos(i + j * lda) : sin(i + j * lda)); } else if(matrix_type == rocblas_client_hermitian_matrix) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < N; ++i) for(size_t j = 0; j <= i; ++j) { @@ -483,7 +525,9 @@ void rocblas_init_matrix_trig(rocblas_check_matrix_type matrix_type, } else if(matrix_type == rocblas_client_symmetric_matrix) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < N; ++i) for(size_t j = 0; j <= i; ++j) { @@ -509,7 +553,9 @@ void rocblas_init_matrix_trig(rocblas_check_matrix_type matrix_type, } else if(matrix_type == rocblas_client_triangular_matrix) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t i = 0; i < M; ++i) for(size_t j = 0; j < N; ++j) { @@ -533,7 +579,9 @@ void rocblas_init_vector_trig(T* x, rocblas_int N, rocblas_stride incx, bool see if(incx < 0) x -= (N - 1) * incx; +#ifdef _OPENMP #pragma omp parallel for +#endif for(rocblas_int j = 0; j < N; ++j) x[j * incx] = T(seedReset ? cos(j * incx) : sin(j * incx)); } @@ -809,7 +857,9 @@ void rocblas_copy_matrix(const T* A, { size_t stride_offset_a = i_batch * stridea; size_t stride_offset_b = i_batch * strideb; +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t j = 0; j < N; ++j) { size_t offset_a = stride_offset_a + j * lda; @@ -828,7 +878,9 @@ void rocblas_copy_matrix( for(size_t i_batch = 0; i_batch < batch_count; i_batch++) { +#ifdef _OPENMP #pragma omp parallel for +#endif for(size_t j = 0; j < N; ++j) { size_t offset_a = j * lda; diff --git a/clients/samples/example_openmp.cpp b/clients/samples/example_openmp.cpp index f62dae6..f844480 100644 --- a/clients/samples/example_openmp.cpp +++ b/clients/samples/example_openmp.cpp @@ -42,7 +42,9 @@ #include #include #include +#ifdef _OPENMP #include +#endif #include #define NUM_THREADS 4 diff --git a/tensile/HostLibraryTests/CachingLibrary_test.cpp b/tensile/HostLibraryTests/CachingLibrary_test.cpp index 2868530..2bdf118 100644 --- a/tensile/HostLibraryTests/CachingLibrary_test.cpp +++ b/tensile/HostLibraryTests/CachingLibrary_test.cpp @@ -63,7 +63,9 @@ TEST(Cache, Threaded) using namespace Tensile; CacheMap cache(-1); +#ifdef _OPENMP #pragma omp parallel num_threads(32) +#endif { int seed = 0; #ifdef _OPENMP diff --git a/tensile/HostLibraryTests/testlib/include/TestUtils.hpp b/tensile/HostLibraryTests/testlib/include/TestUtils.hpp index f146c58..10475ed 100644 --- a/tensile/HostLibraryTests/testlib/include/TestUtils.hpp +++ b/tensile/HostLibraryTests/testlib/include/TestUtils.hpp @@ -27,7 +27,9 @@ #pragma once #include +#ifdef _OPENMP #include +#endif #include #include @@ -99,8 +101,9 @@ namespace Tensile throw std::runtime_error("Fix this function to work with dimensions != 3"); auto seed_base = rng(); - +#ifdef _OPENMP #pragma omp parallel num_threads(32) +#endif { RNG myrng = rng; auto seed = seed_base; @@ -111,7 +114,9 @@ namespace Tensile std::vector index3{0, 0, 0}; +#ifdef _OPENMP #pragma omp for schedule(static) collapse(2) +#endif for(size_t i = 0; i < desc.sizes()[2]; i++) { for(size_t j = 0; j < desc.sizes()[1]; j++)