/external/cblas/examples/ |
H A D | cblas_example1.c | 10 enum CBLAS_TRANSPOSE transa; local 17 transa = CblasNoTrans; 60 cblas_dgemv( order, transa, m, n, alpha, a, lda, x, incx, beta,
|
/external/tensorflow/tensorflow/contrib/rnn/kernels/ |
H A D | blas_gemm.h | 30 void operator()(OpKernelContext* ctx, bool transa, bool transb, uint64 m, 40 static void compute(OpKernelContext* ctx, const Device& d, bool transa, argument 46 int64 k = transa ? a.dimensions()[0] : a.dimensions()[1]; 48 TensorCuBlasGemm<T>()(ctx, transb, transa, n, m, k, alpha, b.data(), 49 transb ? k : n, a.data(), transa ? m : k, beta, 56 static void compute(OpKernelContext* ctx, const Device& d, bool transa, argument 62 Eigen::IndexPair<Eigen::DenseIndex>(transa == false, transb == true);
|
H A D | blas_gemm.cc | 39 void TensorCuBlasGemm<T>::operator()(OpKernelContext* ctx, bool transa, argument 55 ->ThenBlasGemm(trans[transa], trans[transb], m, n, k, alpha, a_ptr,
|
/external/cblas/testing/ |
H A D | c_dblas3.c | 20 enum CBLAS_TRANSPOSE transa, transb; local 22 get_transpose_type(transpa, &transa); 26 if (transa == CblasNoTrans) { 60 cblas_dgemm( CblasRowMajor, transa, transb, *m, *n, *k, *alpha, A, LDA, 70 cblas_dgemm( CblasColMajor, transa, transb, *m, *n, *k, *alpha, a, *lda, 73 cblas_dgemm( UNDEFINED, transa, transb, *m, *n, *k, *alpha, a, *lda,
|
H A D | c_sblas3.c | 18 enum CBLAS_TRANSPOSE transa, transb; local 20 get_transpose_type(transpa, &transa); 24 if (transa == CblasNoTrans) { 57 cblas_sgemm( CblasRowMajor, transa, transb, *m, *n, *k, *alpha, A, LDA, 67 cblas_sgemm( CblasColMajor, transa, transb, *m, *n, *k, *alpha, a, *lda, 70 cblas_sgemm( UNDEFINED, transa, transb, *m, *n, *k, *alpha, a, *lda,
|
H A D | c_cblas3.c | 21 enum CBLAS_TRANSPOSE transa, transb; local 23 get_transpose_type(transpa, &transa); 27 if (transa == CblasNoTrans) { 72 cblas_cgemm( CblasRowMajor, transa, transb, *m, *n, *k, alpha, A, LDA, 84 cblas_cgemm( CblasColMajor, transa, transb, *m, *n, *k, alpha, a, *lda, 87 cblas_cgemm( UNDEFINED, transa, transb, *m, *n, *k, alpha, a, *lda,
|
H A D | c_zblas3.c | 21 enum CBLAS_TRANSPOSE transa, transb; local 23 get_transpose_type(transpa, &transa); 27 if (transa == CblasNoTrans) { 72 cblas_zgemm( CblasRowMajor, transa, transb, *m, *n, *k, alpha, A, LDA, 84 cblas_zgemm( CblasColMajor, transa, transb, *m, *n, *k, alpha, a, *lda, 87 cblas_zgemm( UNDEFINED, transa, transb, *m, *n, *k, alpha, a, *lda,
|
/external/tensorflow/tensorflow/core/kernels/ |
H A D | matmul_op.h | 61 MatmulParameters(bool transa, bool transb, uint64 m, uint64 n, uint64 k, argument 63 : transa_(transa), 70 hash_code_ = transa;
|
H A D | mkl_matmul_op.cc | 109 // @param transa Specifies the form of op(a) used in MatMul. If transa is 127 // depending on transa parameter. Since TF uses row-major 129 // lda = max(1,k) when transa is false, otherwise lda = max(1,m) 144 void MklBlasGemm(bool transa, bool transb, const int m, const int n, argument 153 cblas_sgemm(CblasRowMajor, transa ? CblasTrans : CblasNoTrans, 160 void MklBlasGemm(bool transa, bool transb, const int m, const int n, argument 165 cblas_dgemm(CblasRowMajor, transa ? CblasTrans : CblasNoTrans, 172 void MklBlasGemm(bool transa, bool transb, const int m, const int n, argument 178 cblas_cgemm(CblasRowMajor, transa 188 MklBlasGemm(bool transa, bool transb, const int m, const int n, const int k, const complex128* a, const int lda, const complex128* b, const int ldb, complex128* c, const int ldc) argument [all...] |
H A D | cuda_solvers.cc | 326 cublasOperation_t transa, 334 TF_RETURN_IF_CUBLAS_ERROR(solver(cublas_handle, transa, transb, m, n, 346 cublasOperation_t transa, cublasOperation_t transb, int m, int n, \ 351 return GeamImpl(BLAS_SOLVER_FN(geam, type_prefix), cublas_handle_, transa, \ 325 GeamImpl(SolverFnT solver, cublasHandle_t cublas_handle, cublasOperation_t transa, cublasOperation_t transb, int m, int n, const Scalar* alpha, const Scalar* A, int lda, const Scalar* beta, const Scalar* B, int ldb, Scalar* C, int ldc) argument
|
/external/opencv/cxcore/src/ |
H A D | cxmatmul.cpp | 1041 const char* transa = flags & CV_GEMM_A_T ? "t" : "n"; local 1070 blas_func( transb, transa, &d_size.width, &d_size.height, &len, 1088 blas_func( transb, transa, &d_size.width, &d_size.height, &len,
|
/external/tensorflow/tensorflow/stream_executor/cuda/ |
H A D | cuda_blas.cc | 1815 Stream *stream, blas::Transpose transa, 1825 static_cast<int>(transa), static_cast<int>(transb), m, n, k, alpha, 1827 if (transa == blas::Transpose::kNoTranspose) { 1864 true /* = err_on_failure= */, use_tensor_ops, CUDABlasTranspose(transa), 1876 bool CUDABlas::DoBlasGemm(Stream *stream, blas::Transpose transa, argument 1885 static_cast<int>(transa), static_cast<int>(transb), m, n, k, alpha, 1887 if (transa == blas::Transpose::kNoTranspose) { 1911 CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha, 1915 bool CUDABlas::DoBlasGemm(Stream *stream, blas::Transpose transa, argument 1922 CUDABlasTranspose(transa), CUDABlasTranspos 1814 DoBlasGemm( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, float alpha, const DeviceMemory<Eigen::half> &a, int lda, const DeviceMemory<Eigen::half> &b, int ldb, float beta, DeviceMemory<Eigen::half> *c, int ldc) argument 1926 DoBlasGemm(Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<float> alpha, const DeviceMemory<std::complex<float>> &a, int lda, const DeviceMemory<std::complex<float>> &b, int ldb, std::complex<float> beta, DeviceMemory<std::complex<float>> *c, int ldc) argument 1941 DoBlasGemm(Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<double> alpha, const DeviceMemory<std::complex<double>> &a, int lda, const DeviceMemory<std::complex<double>> &b, int ldb, std::complex<double> beta, DeviceMemory<std::complex<double>> *c, int ldc) argument 1998 DoBlasGemmWithProfiling( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, float alpha, const DeviceMemory<Eigen::half> &a, int lda, const DeviceMemory<Eigen::half> &b, int ldb, float beta, DeviceMemory<Eigen::half> *c, int ldc, blas::ProfileResult *output_profile_result) argument 2009 DoBlasGemmWithProfiling( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, float alpha, const DeviceMemory<float> &a, int lda, const DeviceMemory<float> &b, int ldb, float beta, DeviceMemory<float> *c, int ldc, blas::ProfileResult *output_profile_result) argument 2019 DoBlasGemmWithProfiling( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, double alpha, const DeviceMemory<double> &a, int lda, const DeviceMemory<double> &b, int ldb, double beta, DeviceMemory<double> *c, int ldc, blas::ProfileResult *output_profile_result) argument 2030 DoBlasGemmWithProfiling( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<float> alpha, const DeviceMemory<std::complex<float>> &a, int lda, const DeviceMemory<std::complex<float>> &b, int ldb, std::complex<float> beta, DeviceMemory<std::complex<float>> *c, int ldc, blas::ProfileResult *output_profile_result) argument 2042 DoBlasGemmWithProfiling( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<double> alpha, const DeviceMemory<std::complex<double>> &a, int lda, const DeviceMemory<std::complex<double>> &b, int ldb, std::complex<double> beta, DeviceMemory<std::complex<double>> *c, int ldc, blas::ProfileResult *output_profile_result) argument 2093 DoBlasGemmWithProfilingImpl( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, const ParamType &alpha, const DeviceMemory<T> &a, int lda, const DeviceMemory<T> &b, int ldb, const ParamType &beta, DeviceMemory<T> *c, int ldc, blas::ProfileResult *output_profile_result) argument 2151 DoBlasGemmWithAlgorithmImpl( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, const CompT &alpha, const DeviceMemory<InT> &a, int lda, const DeviceMemory<InT> &b, int ldb, const CompT &beta, DeviceMemory<OutT> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 2240 DoBlasGemmWithAlgorithm( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, int alpha, const DeviceMemory<int8> &a, int lda, const DeviceMemory<int8> &b, int ldb, int beta, DeviceMemory<int> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 2251 DoBlasGemmWithAlgorithm( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, const Eigen::half &alpha, const DeviceMemory<Eigen::half> &a, int lda, const DeviceMemory<Eigen::half> &b, int ldb, const Eigen::half &beta, DeviceMemory<Eigen::half> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 2264 DoBlasGemmWithAlgorithm( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, float alpha, const DeviceMemory<float> &a, int lda, const DeviceMemory<float> &b, int ldb, float beta, DeviceMemory<float> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 2275 DoBlasGemmWithAlgorithm( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, double alpha, const DeviceMemory<double> &a, int lda, const DeviceMemory<double> &b, int ldb, double beta, DeviceMemory<double> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 2286 DoBlasGemmWithAlgorithm( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<float> alpha, const DeviceMemory<std::complex<float>> &a, int lda, const DeviceMemory<std::complex<float>> &b, int ldb, std::complex<float> beta, DeviceMemory<std::complex<float>> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 2299 DoBlasGemmWithAlgorithm( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<double> alpha, const DeviceMemory<std::complex<double>> &a, int lda, const DeviceMemory<std::complex<double>> &b, int ldb, std::complex<double> beta, DeviceMemory<std::complex<double>> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 2313 DoBlasGemmBatchedInternal( FuncT cublas_func, Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, T alpha, const port::ArraySlice<DeviceMemory<T> *> &a_ptrs_to_wrappers, int lda, const port::ArraySlice<DeviceMemory<T> *> &b_ptrs_to_wrappers, int ldb, T beta, const port::ArraySlice<DeviceMemory<T> *> &c_ptrs_to_wrappers, int ldc, int batch_count, ScratchAllocator *scratch_allocator) argument 2393 DoBlasGemmBatched( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, float alpha, const port::ArraySlice<DeviceMemory<float> *> &a_array, int lda, const port::ArraySlice<DeviceMemory<float> *> &b_array, int ldb, float beta, const port::ArraySlice<DeviceMemory<float> *> &c_array, int ldc, int batch_count, ScratchAllocator *scratch_allocator) argument 2409 DoBlasGemmBatched( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, double alpha, const port::ArraySlice<DeviceMemory<double> *> &a_array, int lda, const port::ArraySlice<DeviceMemory<double> *> &b_array, int ldb, double beta, const port::ArraySlice<DeviceMemory<double> *> &c_array, int ldc, int batch_count, ScratchAllocator *scratch_allocator) argument 2425 DoBlasGemmBatched( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<float> alpha, const port::ArraySlice<DeviceMemory<std::complex<float>> *> &a_array, int lda, const port::ArraySlice<DeviceMemory<std::complex<float>> *> &b_array, int ldb, std::complex<float> beta, const port::ArraySlice<DeviceMemory<std::complex<float>> *> &c_array, int ldc, int batch_count, ScratchAllocator *scratch_allocator) argument 2443 DoBlasGemmBatched( Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<double> alpha, const port::ArraySlice<DeviceMemory<std::complex<double>> *> &a_array, int lda, const port::ArraySlice<DeviceMemory<std::complex<double>> *> &b_array, int ldb, std::complex<double> beta, const port::ArraySlice<DeviceMemory<std::complex<double>> *> &c_array, int ldc, int batch_count, ScratchAllocator *scratch_allocator) argument 2693 DoBlasTrmm(Stream *stream, blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, float alpha, const DeviceMemory<float> &a, int lda, DeviceMemory<float> *b, int ldb) argument 2705 DoBlasTrmm(Stream *stream, blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, double alpha, const DeviceMemory<double> &a, int lda, DeviceMemory<double> *b, int ldb) argument 2717 DoBlasTrmm(Stream *stream, blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, std::complex<float> alpha, const DeviceMemory<std::complex<float>> &a, int lda, DeviceMemory<std::complex<float>> *b, int ldb) argument 2731 DoBlasTrmm(Stream *stream, blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, std::complex<double> alpha, const DeviceMemory<std::complex<double>> &a, int lda, DeviceMemory<std::complex<double>> *b, int ldb) argument 2745 DoBlasTrsm(Stream *stream, blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, float alpha, const DeviceMemory<float> &a, int lda, DeviceMemory<float> *b, int ldb) argument 2757 DoBlasTrsm(Stream *stream, blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, double alpha, const DeviceMemory<double> &a, int lda, DeviceMemory<double> *b, int ldb) argument 2769 DoBlasTrsm(Stream *stream, blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, std::complex<float> alpha, const DeviceMemory<std::complex<float>> &a, int lda, DeviceMemory<std::complex<float>> *b, int ldb) argument 2782 DoBlasTrsm(Stream *stream, blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, std::complex<double> alpha, const DeviceMemory<std::complex<double>> &a, int lda, DeviceMemory<std::complex<double>> *b, int ldb) argument [all...] |
/external/tensorflow/tensorflow/stream_executor/ |
H A D | stream.cc | 3496 Stream &Stream::ThenBlasGemm(blas::Transpose transa, blas::Transpose transb, argument 3502 VLOG_CALL(PARAM(transa), PARAM(transb), PARAM(m), PARAM(n), PARAM(k), 3510 return impl(this, &blas::BlasSupport::DoBlasGemm, transa, transb, m, n, k, 3514 Stream &Stream::ThenBlasGemm(blas::Transpose transa, blas::Transpose transb, argument 3519 VLOG_CALL(PARAM(transa), PARAM(transb), PARAM(m), PARAM(n), PARAM(k), 3526 return impl(this, &blas::BlasSupport::DoBlasGemm, transa, transb, m, n, k, 3530 Stream &Stream::ThenBlasGemm(blas::Transpose transa, blas::Transpose transb, argument 3535 VLOG_CALL(PARAM(transa), PARAM(transb), PARAM(m), PARAM(n), PARAM(k), 3542 return impl(this, &blas::BlasSupport::DoBlasGemm, transa, transb, m, n, k, 3546 Stream &Stream::ThenBlasGemm(blas::Transpose transa, bla argument 3567 ThenBlasGemm(blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<double> alpha, const DeviceMemory<std::complex<double>> &a, int lda, const DeviceMemory<std::complex<double>> &b, int ldb, std::complex<double> beta, DeviceMemory<std::complex<double>> *c, int ldc) argument 3681 ThenBlasGemmWithProfiling( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, float alpha, const DeviceMemory<Eigen::half> &a, int lda, const DeviceMemory<Eigen::half> &b, int ldb, float beta, DeviceMemory<Eigen::half> *c, int ldc, blas::ProfileResult *output_profile_result) argument 3701 ThenBlasGemmWithProfiling( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, float alpha, const DeviceMemory<float> &a, int lda, const DeviceMemory<float> &b, int ldb, float beta, DeviceMemory<float> *c, int ldc, blas::ProfileResult *output_profile_result) argument 3720 ThenBlasGemmWithProfiling( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, double alpha, const DeviceMemory<double> &a, int lda, const DeviceMemory<double> &b, int ldb, double beta, DeviceMemory<double> *c, int ldc, blas::ProfileResult *output_profile_result) argument 3740 ThenBlasGemmWithProfiling( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<float> alpha, const DeviceMemory<std::complex<float>> &a, int lda, const DeviceMemory<std::complex<float>> &b, int ldb, std::complex<float> beta, DeviceMemory<std::complex<float>> *c, int ldc, blas::ProfileResult *output_profile_result) argument 3762 ThenBlasGemmWithProfiling( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<double> alpha, const DeviceMemory<std::complex<double>> &a, int lda, const DeviceMemory<std::complex<double>> &b, int ldb, std::complex<double> beta, DeviceMemory<std::complex<double>> *c, int ldc, blas::ProfileResult *output_profile_result) argument 3784 ThenBlasGemmWithAlgorithm( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, const Eigen::half &alpha, const DeviceMemory<Eigen::half> &a, int lda, const DeviceMemory<Eigen::half> &b, int ldb, const Eigen::half &beta, DeviceMemory<Eigen::half> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 3808 ThenBlasGemmWithAlgorithm( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, int alpha, const DeviceMemory<int8> &a, int lda, const DeviceMemory<int8> &b, int ldb, int beta, DeviceMemory<int> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 3829 ThenBlasGemmWithAlgorithm( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, float alpha, const DeviceMemory<float> &a, int lda, const DeviceMemory<float> &b, int ldb, float beta, DeviceMemory<float> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 3850 ThenBlasGemmWithAlgorithm( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, double alpha, const DeviceMemory<double> &a, int lda, const DeviceMemory<double> &b, int ldb, double beta, DeviceMemory<double> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 3872 ThenBlasGemmWithAlgorithm( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<float> alpha, const DeviceMemory<std::complex<float>> &a, int lda, const DeviceMemory<std::complex<float>> &b, int ldb, std::complex<float> beta, DeviceMemory<std::complex<float>> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 3897 ThenBlasGemmWithAlgorithm( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<double> alpha, const DeviceMemory<std::complex<double>> &a, int lda, const DeviceMemory<std::complex<double>> &b, int ldb, std::complex<double> beta, DeviceMemory<std::complex<double>> *c, int ldc, blas::ComputationType computation_type, blas::AlgorithmType algorithm, blas::ProfileResult *output_profile_result) argument 4234 ThenBlasTrmm(blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, float alpha, const DeviceMemory<float> &a, int lda, DeviceMemory<float> *b, int ldb) argument 4249 ThenBlasTrmm(blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, double alpha, const DeviceMemory<double> &a, int lda, DeviceMemory<double> *b, int ldb) argument 4264 ThenBlasTrmm(blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, std::complex<float> alpha, const DeviceMemory<std::complex<float>> &a, int lda, DeviceMemory<std::complex<float>> *b, int ldb) argument 4281 ThenBlasTrmm(blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, std::complex<double> alpha, const DeviceMemory<std::complex<double>> &a, int lda, DeviceMemory<std::complex<double>> *b, int ldb) argument 4298 ThenBlasTrsm(blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, float alpha, const DeviceMemory<float> &a, int lda, DeviceMemory<float> *b, int ldb) argument 4313 ThenBlasTrsm(blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, double alpha, const DeviceMemory<double> &a, int lda, DeviceMemory<double> *b, int ldb) argument 4328 ThenBlasTrsm(blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, std::complex<float> alpha, const DeviceMemory<std::complex<float>> &a, int lda, DeviceMemory<std::complex<float>> *b, int ldb) argument 4345 ThenBlasTrsm(blas::Side side, blas::UpperLower uplo, blas::Transpose transa, blas::Diagonal diag, uint64 m, uint64 n, std::complex<double> alpha, const DeviceMemory<std::complex<double>> &a, int lda, DeviceMemory<std::complex<double>> *b, int ldb) argument 4362 ThenBlasGemmBatched( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, float alpha, const port::ArraySlice<DeviceMemory<float> *> &a, int lda, const port::ArraySlice<DeviceMemory<float> *> &b, int ldb, float beta, const port::ArraySlice<DeviceMemory<float> *> &c, int ldc, int batch_count) argument 4373 ThenBlasGemmBatchedWithScratch( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, float alpha, const port::ArraySlice<DeviceMemory<float> *> &a, int lda, const port::ArraySlice<DeviceMemory<float> *> &b, int ldb, float beta, const port::ArraySlice<DeviceMemory<float> *> &c, int ldc, int batch_count, ScratchAllocator *scratch_allocator) argument 4394 ThenBlasGemmBatched( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, double alpha, const port::ArraySlice<DeviceMemory<double> *> &a, int lda, const port::ArraySlice<DeviceMemory<double> *> &b, int ldb, double beta, const port::ArraySlice<DeviceMemory<double> *> &c, int ldc, int batch_count) argument 4405 ThenBlasGemmBatchedWithScratch( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, double alpha, const port::ArraySlice<DeviceMemory<double> *> &a, int lda, const port::ArraySlice<DeviceMemory<double> *> &b, int ldb, double beta, const port::ArraySlice<DeviceMemory<double> *> &c, int ldc, int batch_count, ScratchAllocator *scratch_allocator) argument 4426 ThenBlasGemmBatched( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<float> alpha, const port::ArraySlice<DeviceMemory<std::complex<float>> *> &a, int lda, const port::ArraySlice<DeviceMemory<std::complex<float>> *> &b, int ldb, std::complex<float> beta, const port::ArraySlice<DeviceMemory<std::complex<float>> *> &c, int ldc, int batch_count) argument 4439 ThenBlasGemmBatchedWithScratch( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<float> alpha, const port::ArraySlice<DeviceMemory<std::complex<float>> *> &a, int lda, const port::ArraySlice<DeviceMemory<std::complex<float>> *> &b, int ldb, std::complex<float> beta, const port::ArraySlice<DeviceMemory<std::complex<float>> *> &c, int ldc, int batch_count, ScratchAllocator *scratch_allocator) argument 4465 ThenBlasGemmBatched( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<double> alpha, const port::ArraySlice<DeviceMemory<std::complex<double>> *> &a, int lda, const port::ArraySlice<DeviceMemory<std::complex<double>> *> &b, int ldb, std::complex<double> beta, const port::ArraySlice<DeviceMemory<std::complex<double>> *> &c, int ldc, int batch_count) argument 4478 ThenBlasGemmBatchedWithScratch( blas::Transpose transa, blas::Transpose transb, uint64 m, uint64 n, uint64 k, std::complex<double> alpha, const port::ArraySlice<DeviceMemory<std::complex<double>> *> &a, int lda, const port::ArraySlice<DeviceMemory<std::complex<double>> *> &b, int ldb, std::complex<double> beta, const port::ArraySlice<DeviceMemory<std::complex<double>> *> &c, int ldc, int batch_count, ScratchAllocator *scratch_allocator) argument [all...] |