nvmath-python 1.0.0__cp314-cp314t-win_amd64.whl

nvmath/bindings/cublas.pxd

@@ -0,0 +1,558 @@
+# Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# This code was automatically generated across versions from 11.0.3 to 13.2.0, generator version 0.3.1.dev1471+g7122059e9. Do not modify it directly.
+
+from libc.stdint cimport intptr_t
+
+from .cycublas cimport *
+
+
+###############################################################################
+# Types
+###############################################################################
+
+ctypedef cublasHandle_t Handle
+ctypedef cublasLogCallback Callback
+
+ctypedef cudaStream_t Stream
+ctypedef cudaDataType DataType
+ctypedef libraryPropertyType_t LibraryPropertyType
+
+
+###############################################################################
+# Enum
+###############################################################################
+
+ctypedef cublasStatus_t _Status
+ctypedef cublasFillMode_t _FillMode
+ctypedef cublasDiagType_t _DiagType
+ctypedef cublasSideMode_t _SideMode
+ctypedef cublasOperation_t _Operation
+ctypedef cublasPointerMode_t _PointerMode
+ctypedef cublasAtomicsMode_t _AtomicsMode
+ctypedef cublasGemmAlgo_t _GemmAlgo
+ctypedef cublasMath_t _Math
+ctypedef cublasComputeType_t _ComputeType
+ctypedef cublasEmulationStrategy_t _EmulationStrategy
+
+
+###############################################################################
+# Functions
+###############################################################################
+
+cpdef intptr_t create() except? 0
+cpdef destroy(intptr_t handle)
+cpdef int get_version(intptr_t handle) except? -1
+cpdef int get_property(int type) except? -1
+cpdef size_t get_cudart_version() except? 0
+cpdef set_workspace(intptr_t handle, intptr_t workspace, size_t workspace_size_in_bytes)
+cpdef set_stream(intptr_t handle, intptr_t stream_id)
+cpdef intptr_t get_stream(intptr_t handle) except? 0
+cpdef int get_pointer_mode(intptr_t handle) except? -1
+cpdef set_pointer_mode(intptr_t handle, int mode)
+cpdef int get_atomics_mode(intptr_t handle) except? -1
+cpdef set_atomics_mode(intptr_t handle, int mode)
+cpdef int get_math_mode(intptr_t handle) except? -1
+cpdef set_math_mode(intptr_t handle, int mode)
+cpdef logger_configure(int log_is_on, int log_to_std_out, int log_to_std_err, log_file_name)
+cpdef set_vector(int n, int elem_size, intptr_t x, int incx, intptr_t device_ptr, int incy)
+cpdef get_vector(int n, int elem_size, intptr_t x, int incx, intptr_t y, int incy)
+cpdef set_matrix(int rows, int cols, int elem_size, intptr_t a, int lda, intptr_t b, int ldb)
+cpdef get_matrix(int rows, int cols, int elem_size, intptr_t a, int lda, intptr_t b, int ldb)
+cpdef set_vector_async(int n, int elem_size, intptr_t host_ptr, int incx, intptr_t device_ptr, int incy, intptr_t stream)
+cpdef get_vector_async(int n, int elem_size, intptr_t device_ptr, int incx, intptr_t host_ptr, int incy, intptr_t stream)
+cpdef set_matrix_async(int rows, int cols, int elem_size, intptr_t a, int lda, intptr_t b, int ldb, intptr_t stream)
+cpdef get_matrix_async(int rows, int cols, int elem_size, intptr_t a, int lda, intptr_t b, int ldb, intptr_t stream)
+cpdef nrm2_ex(intptr_t handle, int n, intptr_t x, int x_type, int incx, intptr_t result, int result_type, int execution_type)
+cpdef snrm2(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef dnrm2(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef scnrm2(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef dznrm2(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef dot_ex(intptr_t handle, int n, intptr_t x, int x_type, int incx, intptr_t y, int y_type, int incy, intptr_t result, int result_type, int execution_type)
+cpdef dotc_ex(intptr_t handle, int n, intptr_t x, int x_type, int incx, intptr_t y, int y_type, int incy, intptr_t result, int result_type, int execution_type)
+cpdef sdot(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t result)
+cpdef ddot(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t result)
+cpdef cdotu(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t result)
+cpdef cdotc(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t result)
+cpdef zdotu(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t result)
+cpdef zdotc(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t result)
+cpdef scal_ex(intptr_t handle, int n, intptr_t alpha, int alpha_type, intptr_t x, int x_type, int incx, int execution_type)
+cpdef sscal(intptr_t handle, int n, intptr_t alpha, intptr_t x, int incx)
+cpdef dscal(intptr_t handle, int n, intptr_t alpha, intptr_t x, int incx)
+cpdef cscal(intptr_t handle, int n, intptr_t alpha, intptr_t x, int incx)
+cpdef csscal(intptr_t handle, int n, intptr_t alpha, intptr_t x, int incx)
+cpdef zscal(intptr_t handle, int n, intptr_t alpha, intptr_t x, int incx)
+cpdef zdscal(intptr_t handle, int n, intptr_t alpha, intptr_t x, int incx)
+cpdef axpy_ex(intptr_t handle, int n, intptr_t alpha, int alpha_type, intptr_t x, int x_type, int incx, intptr_t y, int y_type, int incy, int executiontype)
+cpdef saxpy(intptr_t handle, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy)
+cpdef daxpy(intptr_t handle, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy)
+cpdef caxpy(intptr_t handle, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy)
+cpdef zaxpy(intptr_t handle, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy)
+cpdef copy_ex(intptr_t handle, int n, intptr_t x, int x_type, int incx, intptr_t y, int y_type, int incy)
+cpdef scopy(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy)
+cpdef dcopy(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy)
+cpdef ccopy(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy)
+cpdef zcopy(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy)
+cpdef sswap(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy)
+cpdef dswap(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy)
+cpdef cswap(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy)
+cpdef zswap(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy)
+cpdef swap_ex(intptr_t handle, int n, intptr_t x, int x_type, int incx, intptr_t y, int y_type, int incy)
+cpdef isamax(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef idamax(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef icamax(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef izamax(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef iamax_ex(intptr_t handle, int n, intptr_t x, int x_type, int incx, intptr_t result)
+cpdef isamin(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef idamin(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef icamin(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef izamin(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef iamin_ex(intptr_t handle, int n, intptr_t x, int x_type, int incx, intptr_t result)
+cpdef asum_ex(intptr_t handle, int n, intptr_t x, int x_type, int incx, intptr_t result, int result_type, int executiontype)
+cpdef sasum(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef dasum(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef scasum(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef dzasum(intptr_t handle, int n, intptr_t x, int incx, intptr_t result)
+cpdef srot(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t c, intptr_t s)
+cpdef drot(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t c, intptr_t s)
+cpdef crot(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t c, intptr_t s)
+cpdef csrot(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t c, intptr_t s)
+cpdef zrot(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t c, intptr_t s)
+cpdef zdrot(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t c, intptr_t s)
+cpdef rot_ex(intptr_t handle, int n, intptr_t x, int x_type, int incx, intptr_t y, int y_type, int incy, intptr_t c, intptr_t s, int cs_type, int executiontype)
+cpdef srotg(intptr_t handle, intptr_t a, intptr_t b, intptr_t c, intptr_t s)
+cpdef drotg(intptr_t handle, intptr_t a, intptr_t b, intptr_t c, intptr_t s)
+cpdef crotg(intptr_t handle, intptr_t a, intptr_t b, intptr_t c, intptr_t s)
+cpdef zrotg(intptr_t handle, intptr_t a, intptr_t b, intptr_t c, intptr_t s)
+cpdef rotg_ex(intptr_t handle, intptr_t a, intptr_t b, int ab_type, intptr_t c, intptr_t s, int cs_type, int executiontype)
+cpdef srotm(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t param)
+cpdef drotm(intptr_t handle, int n, intptr_t x, int incx, intptr_t y, int incy, intptr_t param)
+cpdef rotm_ex(intptr_t handle, int n, intptr_t x, int x_type, int incx, intptr_t y, int y_type, int incy, intptr_t param, int param_type, int executiontype)
+cpdef srotmg(intptr_t handle, intptr_t d1, intptr_t d2, intptr_t x1, intptr_t y1, intptr_t param)
+cpdef drotmg(intptr_t handle, intptr_t d1, intptr_t d2, intptr_t x1, intptr_t y1, intptr_t param)
+cpdef rotmg_ex(intptr_t handle, intptr_t d1, int d1type, intptr_t d2, int d2type, intptr_t x1, int x1type, intptr_t y1, int y1type, intptr_t param, int param_type, int executiontype)
+cpdef sgemv(intptr_t handle, int trans, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef dgemv(intptr_t handle, int trans, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef cgemv(intptr_t handle, int trans, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef zgemv(intptr_t handle, int trans, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef sgbmv(intptr_t handle, int trans, int m, int n, int kl, int ku, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef dgbmv(intptr_t handle, int trans, int m, int n, int kl, int ku, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef cgbmv(intptr_t handle, int trans, int m, int n, int kl, int ku, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef zgbmv(intptr_t handle, int trans, int m, int n, int kl, int ku, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef strmv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t a, int lda, intptr_t x, int incx)
+cpdef dtrmv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t a, int lda, intptr_t x, int incx)
+cpdef ctrmv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t a, int lda, intptr_t x, int incx)
+cpdef ztrmv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t a, int lda, intptr_t x, int incx)
+cpdef stbmv(intptr_t handle, int uplo, int trans, int diag, int n, int k, intptr_t a, int lda, intptr_t x, int incx)
+cpdef dtbmv(intptr_t handle, int uplo, int trans, int diag, int n, int k, intptr_t a, int lda, intptr_t x, int incx)
+cpdef ctbmv(intptr_t handle, int uplo, int trans, int diag, int n, int k, intptr_t a, int lda, intptr_t x, int incx)
+cpdef ztbmv(intptr_t handle, int uplo, int trans, int diag, int n, int k, intptr_t a, int lda, intptr_t x, int incx)
+cpdef stpmv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t ap, intptr_t x, int incx)
+cpdef dtpmv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t ap, intptr_t x, int incx)
+cpdef ctpmv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t ap, intptr_t x, int incx)
+cpdef ztpmv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t ap, intptr_t x, int incx)
+cpdef strsv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t a, int lda, intptr_t x, int incx)
+cpdef dtrsv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t a, int lda, intptr_t x, int incx)
+cpdef ctrsv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t a, int lda, intptr_t x, int incx)
+cpdef ztrsv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t a, int lda, intptr_t x, int incx)
+cpdef stpsv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t ap, intptr_t x, int incx)
+cpdef dtpsv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t ap, intptr_t x, int incx)
+cpdef ctpsv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t ap, intptr_t x, int incx)
+cpdef ztpsv(intptr_t handle, int uplo, int trans, int diag, int n, intptr_t ap, intptr_t x, int incx)
+cpdef stbsv(intptr_t handle, int uplo, int trans, int diag, int n, int k, intptr_t a, int lda, intptr_t x, int incx)
+cpdef dtbsv(intptr_t handle, int uplo, int trans, int diag, int n, int k, intptr_t a, int lda, intptr_t x, int incx)
+cpdef ctbsv(intptr_t handle, int uplo, int trans, int diag, int n, int k, intptr_t a, int lda, intptr_t x, int incx)
+cpdef ztbsv(intptr_t handle, int uplo, int trans, int diag, int n, int k, intptr_t a, int lda, intptr_t x, int incx)
+cpdef ssymv(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef dsymv(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef csymv(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef zsymv(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef chemv(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef zhemv(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef ssbmv(intptr_t handle, int uplo, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef dsbmv(intptr_t handle, int uplo, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef chbmv(intptr_t handle, int uplo, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef zhbmv(intptr_t handle, int uplo, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef sspmv(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t ap, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef dspmv(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t ap, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef chpmv(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t ap, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef zhpmv(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t ap, intptr_t x, int incx, intptr_t beta, intptr_t y, int incy)
+cpdef sger(intptr_t handle, int m, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t a, int lda)
+cpdef dger(intptr_t handle, int m, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t a, int lda)
+cpdef cgeru(intptr_t handle, int m, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t a, int lda)
+cpdef cgerc(intptr_t handle, int m, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t a, int lda)
+cpdef zgeru(intptr_t handle, int m, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t a, int lda)
+cpdef zgerc(intptr_t handle, int m, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t a, int lda)
+cpdef ssyr(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t a, int lda)
+cpdef dsyr(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t a, int lda)
+cpdef csyr(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t a, int lda)
+cpdef zsyr(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t a, int lda)
+cpdef cher(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t a, int lda)
+cpdef zher(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t a, int lda)
+cpdef sspr(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t ap)
+cpdef dspr(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t ap)
+cpdef chpr(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t ap)
+cpdef zhpr(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t ap)
+cpdef ssyr2(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t a, int lda)
+cpdef dsyr2(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t a, int lda)
+cpdef csyr2(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t a, int lda)
+cpdef zsyr2(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t a, int lda)
+cpdef cher2(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t a, int lda)
+cpdef zher2(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t a, int lda)
+cpdef sspr2(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t ap)
+cpdef dspr2(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t ap)
+cpdef chpr2(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t ap)
+cpdef zhpr2(intptr_t handle, int uplo, int n, intptr_t alpha, intptr_t x, int incx, intptr_t y, int incy, intptr_t ap)
+cpdef sgemm(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef dgemm(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef cgemm(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef cgemm3m(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef cgemm3m_ex(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int atype, int lda, intptr_t b, int btype, int ldb, intptr_t beta, intptr_t c, int ctype, int ldc)
+cpdef zgemm(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef zgemm3m(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef sgemm_ex(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int atype, int lda, intptr_t b, int btype, int ldb, intptr_t beta, intptr_t c, int ctype, int ldc)
+cpdef gemm_ex(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int atype, int lda, intptr_t b, int btype, int ldb, intptr_t beta, intptr_t c, int ctype, int ldc, int compute_type, int algo)
+cpdef cgemm_ex(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int atype, int lda, intptr_t b, int btype, int ldb, intptr_t beta, intptr_t c, int ctype, int ldc)
+cpdef uint8gemm_bias(intptr_t handle, int transa, int transb, int transc, int m, int n, int k, intptr_t a, int a_bias, int lda, intptr_t b, int b_bias, int ldb, intptr_t c, int c_bias, int ldc, int c_mult, int c_shift)
+cpdef ssyrk(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t beta, intptr_t c, int ldc)
+cpdef dsyrk(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t beta, intptr_t c, int ldc)
+cpdef csyrk(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t beta, intptr_t c, int ldc)
+cpdef zsyrk(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t beta, intptr_t c, int ldc)
+cpdef csyrk_ex(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int atype, int lda, intptr_t beta, intptr_t c, int ctype, int ldc)
+cpdef csyrk3m_ex(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int atype, int lda, intptr_t beta, intptr_t c, int ctype, int ldc)
+cpdef cherk(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t beta, intptr_t c, int ldc)
+cpdef zherk(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t beta, intptr_t c, int ldc)
+cpdef cherk_ex(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int atype, int lda, intptr_t beta, intptr_t c, int ctype, int ldc)
+cpdef cherk3m_ex(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int atype, int lda, intptr_t beta, intptr_t c, int ctype, int ldc)
+cpdef ssyr2k(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef dsyr2k(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef csyr2k(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef zsyr2k(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef cher2k(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef zher2k(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef ssyrkx(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef dsyrkx(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef csyrkx(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef zsyrkx(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef cherkx(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef zherkx(intptr_t handle, int uplo, int trans, int n, int k, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef ssymm(intptr_t handle, int side, int uplo, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef dsymm(intptr_t handle, int side, int uplo, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef csymm(intptr_t handle, int side, int uplo, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef zsymm(intptr_t handle, int side, int uplo, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef chemm(intptr_t handle, int side, int uplo, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef zhemm(intptr_t handle, int side, int uplo, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t beta, intptr_t c, int ldc)
+cpdef strsm(intptr_t handle, int side, int uplo, int trans, int diag, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb)
+cpdef dtrsm(intptr_t handle, int side, int uplo, int trans, int diag, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb)
+cpdef ctrsm(intptr_t handle, int side, int uplo, int trans, int diag, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb)
+cpdef ztrsm(intptr_t handle, int side, int uplo, int trans, int diag, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb)
+cpdef strmm(intptr_t handle, int side, int uplo, int trans, int diag, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t c, int ldc)
+cpdef dtrmm(intptr_t handle, int side, int uplo, int trans, int diag, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t c, int ldc)
+cpdef ctrmm(intptr_t handle, int side, int uplo, int trans, int diag, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t c, int ldc)
+cpdef ztrmm(intptr_t handle, int side, int uplo, int trans, int diag, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, intptr_t c, int ldc)
+cpdef sgemm_batched(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t aarray, int lda, intptr_t barray, int ldb, intptr_t beta, intptr_t carray, int ldc, int batch_count)
+cpdef dgemm_batched(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t aarray, int lda, intptr_t barray, int ldb, intptr_t beta, intptr_t carray, int ldc, int batch_count)
+cpdef cgemm_batched(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t aarray, int lda, intptr_t barray, int ldb, intptr_t beta, intptr_t carray, int ldc, int batch_count)
+cpdef cgemm3m_batched(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t aarray, int lda, intptr_t barray, int ldb, intptr_t beta, intptr_t carray, int ldc, int batch_count)
+cpdef zgemm_batched(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t aarray, int lda, intptr_t barray, int ldb, intptr_t beta, intptr_t carray, int ldc, int batch_count)
+cpdef gemm_batched_ex(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t aarray, int atype, int lda, intptr_t barray, int btype, int ldb, intptr_t beta, intptr_t carray, int ctype, int ldc, int batch_count, int compute_type, int algo)
+cpdef gemm_strided_batched_ex(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int atype, int lda, long long int stride_a, intptr_t b, int btype, int ldb, long long int stride_b, intptr_t beta, intptr_t c, int ctype, int ldc, long long int stride_c, int batch_count, int compute_type, int algo)
+cpdef sgemm_strided_batched(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int lda, long long int stride_a, intptr_t b, int ldb, long long int stride_b, intptr_t beta, intptr_t c, int ldc, long long int stride_c, int batch_count)
+cpdef dgemm_strided_batched(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int lda, long long int stride_a, intptr_t b, int ldb, long long int stride_b, intptr_t beta, intptr_t c, int ldc, long long int stride_c, int batch_count)
+cpdef cgemm_strided_batched(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int lda, long long int stride_a, intptr_t b, int ldb, long long int stride_b, intptr_t beta, intptr_t c, int ldc, long long int stride_c, int batch_count)
+cpdef cgemm3m_strided_batched(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int lda, long long int stride_a, intptr_t b, int ldb, long long int stride_b, intptr_t beta, intptr_t c, int ldc, long long int stride_c, int batch_count)
+cpdef zgemm_strided_batched(intptr_t handle, int transa, int transb, int m, int n, int k, intptr_t alpha, intptr_t a, int lda, long long int stride_a, intptr_t b, int ldb, long long int stride_b, intptr_t beta, intptr_t c, int ldc, long long int stride_c, int batch_count)
+cpdef sgeam(intptr_t handle, int transa, int transb, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t beta, intptr_t b, int ldb, intptr_t c, int ldc)
+cpdef dgeam(intptr_t handle, int transa, int transb, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t beta, intptr_t b, int ldb, intptr_t c, int ldc)
+cpdef cgeam(intptr_t handle, int transa, int transb, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t beta, intptr_t b, int ldb, intptr_t c, int ldc)
+cpdef zgeam(intptr_t handle, int transa, int transb, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t beta, intptr_t b, int ldb, intptr_t c, int ldc)
+cpdef sgetrf_batched(intptr_t handle, int n, intptr_t a, int lda, intptr_t p, intptr_t info, int batch_size)
+cpdef dgetrf_batched(intptr_t handle, int n, intptr_t a, int lda, intptr_t p, intptr_t info, int batch_size)
+cpdef cgetrf_batched(intptr_t handle, int n, intptr_t a, int lda, intptr_t p, intptr_t info, int batch_size)
+cpdef zgetrf_batched(intptr_t handle, int n, intptr_t a, int lda, intptr_t p, intptr_t info, int batch_size)
+cpdef sgetri_batched(intptr_t handle, int n, intptr_t a, int lda, intptr_t p, intptr_t c, int ldc, intptr_t info, int batch_size)
+cpdef dgetri_batched(intptr_t handle, int n, intptr_t a, int lda, intptr_t p, intptr_t c, int ldc, intptr_t info, int batch_size)
+cpdef cgetri_batched(intptr_t handle, int n, intptr_t a, int lda, intptr_t p, intptr_t c, int ldc, intptr_t info, int batch_size)
+cpdef zgetri_batched(intptr_t handle, int n, intptr_t a, int lda, intptr_t p, intptr_t c, int ldc, intptr_t info, int batch_size)
+cpdef sgetrs_batched(intptr_t handle, int trans, int n, int nrhs, intptr_t aarray, int lda, intptr_t dev_ipiv, intptr_t barray, int ldb, intptr_t info, int batch_size)
+cpdef dgetrs_batched(intptr_t handle, int trans, int n, int nrhs, intptr_t aarray, int lda, intptr_t dev_ipiv, intptr_t barray, int ldb, intptr_t info, int batch_size)
+cpdef cgetrs_batched(intptr_t handle, int trans, int n, int nrhs, intptr_t aarray, int lda, intptr_t dev_ipiv, intptr_t barray, int ldb, intptr_t info, int batch_size)
+cpdef zgetrs_batched(intptr_t handle, int trans, int n, int nrhs, intptr_t aarray, int lda, intptr_t dev_ipiv, intptr_t barray, int ldb, intptr_t info, int batch_size)
+cpdef strsm_batched(intptr_t handle, int side, int uplo, int trans, int diag, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, int batch_count)
+cpdef dtrsm_batched(intptr_t handle, int side, int uplo, int trans, int diag, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, int batch_count)
+cpdef ctrsm_batched(intptr_t handle, int side, int uplo, int trans, int diag, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, int batch_count)
+cpdef ztrsm_batched(intptr_t handle, int side, int uplo, int trans, int diag, int m, int n, intptr_t alpha, intptr_t a, int lda, intptr_t b, int ldb, int batch_count)
+cpdef smatinv_batched(intptr_t handle, int n, intptr_t a, int lda, intptr_t ainv, int lda_inv, intptr_t info, int batch_size)
+cpdef dmatinv_batched(intptr_t handle, int n, intptr_t a, int lda, intptr_t ainv, int lda_inv, intptr_t info, int batch_size)
+cpdef cmatinv_batched(intptr_t handle, int n, intptr_t a, int lda, intptr_t ainv, int lda_inv, intptr_t info, int batch_size)
+cpdef zmatinv_batched(intptr_t handle, int n, intptr_t a, int lda, intptr_t ainv, int lda_inv, intptr_t info, int batch_size)
+cpdef sgeqrf_batched(intptr_t handle, int m, int n, intptr_t aarray, int lda, intptr_t tau_array, intptr_t info, int batch_size)
+cpdef dgeqrf_batched(intptr_t handle, int m, int n, intptr_t aarray, int lda, intptr_t tau_array, intptr_t info, int batch_size)
+cpdef cgeqrf_batched(intptr_t handle, int m, int n, intptr_t aarray, int lda, intptr_t tau_array, intptr_t info, int batch_size)
+cpdef zgeqrf_batched(intptr_t handle, int m, int n, intptr_t aarray, int lda, intptr_t tau_array, intptr_t info, int batch_size)
+cpdef sgels_batched(intptr_t handle, int trans, int m, int n, int nrhs, intptr_t aarray, int lda, intptr_t carray, int ldc, intptr_t info, intptr_t dev_info_array, int batch_size)
+cpdef dgels_batched(intptr_t handle, int trans, int m, int n, int nrhs, intptr_t aarray, int lda, intptr_t carray, int ldc, intptr_t info, intptr_t dev_info_array, int batch_size)
+cpdef cgels_batched(intptr_t handle, int trans, int m, int n, int nrhs, intptr_t aarray, int lda, intptr_t carray, int ldc, intptr_t info, intptr_t dev_info_array, int batch_size)
+cpdef zgels_batched(intptr_t handle, int trans, int m, int n, int nrhs, intptr_t aarray, int lda, intptr_t carray, int ldc, intptr_t info, intptr_t dev_info_array, int batch_size)
+cpdef sdgmm(intptr_t handle, int mode, int m, int n, intptr_t a, int lda, intptr_t x, int incx, intptr_t c, int ldc)
+cpdef ddgmm(intptr_t handle, int mode, int m, int n, intptr_t a, int lda, intptr_t x, int incx, intptr_t c, int ldc)
+cpdef cdgmm(intptr_t handle, int mode, int m, int n, intptr_t a, int lda, intptr_t x, int incx, intptr_t c, int ldc)
+cpdef zdgmm(intptr_t handle, int mode, int m, int n, intptr_t a, int lda, intptr_t x, int incx, intptr_t c, int ldc)
+cpdef stpttr(intptr_t handle, int uplo, int n, intptr_t ap, intptr_t a, int lda)
+cpdef dtpttr(intptr_t handle, int uplo, int n, intptr_t ap, intptr_t a, int lda)
+cpdef ctpttr(intptr_t handle, int uplo, int n, intptr_t ap, intptr_t a, int lda)
+cpdef ztpttr(intptr_t handle, int uplo, int n, intptr_t ap, intptr_t a, int lda)
+cpdef strttp(intptr_t handle, int uplo, int n, intptr_t a, int lda, intptr_t ap)
+cpdef dtrttp(intptr_t handle, int uplo, int n, intptr_t a, int lda, intptr_t ap)
+cpdef ctrttp(intptr_t handle, int uplo, int n, intptr_t a, int lda, intptr_t ap)
+cpdef ztrttp(intptr_t handle, int uplo, int n, intptr_t a, int lda, intptr_t ap)
+cpdef int get_sm_count_target(intptr_t handle) except? -1
+cpdef set_sm_count_target(intptr_t handle, int sm_count_target)
+cpdef str get_status_name(int status)
+cpdef str get_status_string(int status)
+cpdef sgemv_batched(intptr_t handle, int trans, int m, int n, intptr_t alpha, intptr_t aarray, int lda, intptr_t xarray, int incx, intptr_t beta, intptr_t yarray, int incy, int batch_count)
+cpdef dgemv_batched(intptr_t handle, int trans, int m, int n, intptr_t alpha, intptr_t aarray, int lda, intptr_t xarray, int incx, intptr_t beta, intptr_t yarray, int incy, int batch_count)
+cpdef cgemv_batched(intptr_t handle, int trans, int m, int n, intptr_t alpha, intptr_t aarray, int lda, intptr_t xarray, int incx, intptr_t beta, intptr_t yarray, int incy, int batch_count)
+cpdef zgemv_batched(intptr_t handle, int trans, int m, int n, intptr_t alpha, intptr_t aarray, int lda, intptr_t xarray, int incx, intptr_t beta, intptr_t yarray, int incy, int batch_count)
+cpdef sgemv_strided_batched(intptr_t handle, int trans, int m, int n, intptr_t alpha, intptr_t a, int lda, long long int stride_a, intptr_t x, int incx, long long int stridex, intptr_t beta, intptr_t y, int incy, long long int stridey, int batch_count)
+cpdef dgemv_strided_batched(intptr_t handle, int trans, int m, int n, intptr_t alpha, intptr_t a, int lda, long long int stride_a, intptr_t x, int incx, long long int stridex, intptr_t beta, intptr_t y, int incy, long long int stridey, int batch_count)
+cpdef cgemv_strided_batched(intptr_t handle, int trans, int m, int n, intptr_t alpha, intptr_t a, int lda, long long int stride_a, intptr_t x, int incx, long long int stridex, intptr_t beta, intptr_t y, int incy, long long int stridey, int batch_count)
+cpdef zgemv_strided_batched(intptr_t handle, int trans, int m, int n, intptr_t alpha, intptr_t a, int lda, long long int stride_a, intptr_t x, int incx, long long int stridex, intptr_t beta, intptr_t y, int incy, long long int stridey, int batch_count)
+cpdef set_vector_64(int64_t n, int64_t elem_size, intptr_t x, int64_t incx, intptr_t device_ptr, int64_t incy)
+cpdef get_vector_64(int64_t n, int64_t elem_size, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef set_matrix_64(int64_t rows, int64_t cols, int64_t elem_size, intptr_t a, int64_t lda, intptr_t b, int64_t ldb)
+cpdef get_matrix_64(int64_t rows, int64_t cols, int64_t elem_size, intptr_t a, int64_t lda, intptr_t b, int64_t ldb)
+cpdef set_vector_async_64(int64_t n, int64_t elem_size, intptr_t host_ptr, int64_t incx, intptr_t device_ptr, int64_t incy, intptr_t stream)
+cpdef get_vector_async_64(int64_t n, int64_t elem_size, intptr_t device_ptr, int64_t incx, intptr_t host_ptr, int64_t incy, intptr_t stream)
+cpdef set_matrix_async_64(int64_t rows, int64_t cols, int64_t elem_size, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t stream)
+cpdef get_matrix_async_64(int64_t rows, int64_t cols, int64_t elem_size, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t stream)
+cpdef nrm2ex_64(intptr_t handle, int64_t n, intptr_t x, int x_type, int64_t incx, intptr_t result, int result_type, int execution_type)
+cpdef snrm2_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef dnrm2_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef scnrm2_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef dznrm2_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef dot_ex_64(intptr_t handle, int64_t n, intptr_t x, int x_type, int64_t incx, intptr_t y, int y_type, int64_t incy, intptr_t result, int result_type, int execution_type)
+cpdef dotc_ex_64(intptr_t handle, int64_t n, intptr_t x, int x_type, int64_t incx, intptr_t y, int y_type, int64_t incy, intptr_t result, int result_type, int execution_type)
+cpdef sdot_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t result)
+cpdef ddot_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t result)
+cpdef cdotu_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t result)
+cpdef cdotc_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t result)
+cpdef zdotu_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t result)
+cpdef zdotc_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t result)
+cpdef scal_ex_64(intptr_t handle, int64_t n, intptr_t alpha, int alpha_type, intptr_t x, int x_type, int64_t incx, int execution_type)
+cpdef sscal_64(intptr_t handle, int64_t n, intptr_t alpha, intptr_t x, int64_t incx)
+cpdef dscal_64(intptr_t handle, int64_t n, intptr_t alpha, intptr_t x, int64_t incx)
+cpdef cscal_64(intptr_t handle, int64_t n, intptr_t alpha, intptr_t x, int64_t incx)
+cpdef csscal_64(intptr_t handle, int64_t n, intptr_t alpha, intptr_t x, int64_t incx)
+cpdef zscal_64(intptr_t handle, int64_t n, intptr_t alpha, intptr_t x, int64_t incx)
+cpdef zdscal_64(intptr_t handle, int64_t n, intptr_t alpha, intptr_t x, int64_t incx)
+cpdef axpy_ex_64(intptr_t handle, int64_t n, intptr_t alpha, int alpha_type, intptr_t x, int x_type, int64_t incx, intptr_t y, int y_type, int64_t incy, int executiontype)
+cpdef saxpy_64(intptr_t handle, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef daxpy_64(intptr_t handle, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef caxpy_64(intptr_t handle, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef zaxpy_64(intptr_t handle, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef copy_ex_64(intptr_t handle, int64_t n, intptr_t x, int x_type, int64_t incx, intptr_t y, int y_type, int64_t incy)
+cpdef scopy_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef dcopy_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef ccopy_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef zcopy_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef sswap_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef dswap_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef cswap_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef zswap_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy)
+cpdef swap_ex_64(intptr_t handle, int64_t n, intptr_t x, int x_type, int64_t incx, intptr_t y, int y_type, int64_t incy)
+cpdef isamax_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef idamax_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef icamax_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef izamax_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef iamax_ex_64(intptr_t handle, int64_t n, intptr_t x, int x_type, int64_t incx, intptr_t result)
+cpdef isamin_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef idamin_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef icamin_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef izamin_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef iamin_ex_64(intptr_t handle, int64_t n, intptr_t x, int x_type, int64_t incx, intptr_t result)
+cpdef asum_ex_64(intptr_t handle, int64_t n, intptr_t x, int x_type, int64_t incx, intptr_t result, int result_type, int executiontype)
+cpdef sasum_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef dasum_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef scasum_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef dzasum_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t result)
+cpdef srot_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t c, intptr_t s)
+cpdef drot_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t c, intptr_t s)
+cpdef crot_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t c, intptr_t s)
+cpdef csrot_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t c, intptr_t s)
+cpdef zrot_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t c, intptr_t s)
+cpdef zdrot_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t c, intptr_t s)
+cpdef rot_ex_64(intptr_t handle, int64_t n, intptr_t x, int x_type, int64_t incx, intptr_t y, int y_type, int64_t incy, intptr_t c, intptr_t s, int cs_type, int executiontype)
+cpdef srotm_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t param)
+cpdef drotm_64(intptr_t handle, int64_t n, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t param)
+cpdef rotm_ex_64(intptr_t handle, int64_t n, intptr_t x, int x_type, int64_t incx, intptr_t y, int y_type, int64_t incy, intptr_t param, int param_type, int executiontype)
+cpdef sgemv_64(intptr_t handle, int trans, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef dgemv_64(intptr_t handle, int trans, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef cgemv_64(intptr_t handle, int trans, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef zgemv_64(intptr_t handle, int trans, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef sgbmv_64(intptr_t handle, int trans, int64_t m, int64_t n, int64_t kl, int64_t ku, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef dgbmv_64(intptr_t handle, int trans, int64_t m, int64_t n, int64_t kl, int64_t ku, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef cgbmv_64(intptr_t handle, int trans, int64_t m, int64_t n, int64_t kl, int64_t ku, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef zgbmv_64(intptr_t handle, int trans, int64_t m, int64_t n, int64_t kl, int64_t ku, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef strmv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef dtrmv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef ctrmv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef ztrmv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef stbmv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, int64_t k, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef dtbmv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, int64_t k, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef ctbmv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, int64_t k, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef ztbmv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, int64_t k, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef stpmv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t ap, intptr_t x, int64_t incx)
+cpdef dtpmv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t ap, intptr_t x, int64_t incx)
+cpdef ctpmv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t ap, intptr_t x, int64_t incx)
+cpdef ztpmv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t ap, intptr_t x, int64_t incx)
+cpdef strsv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef dtrsv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef ctrsv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef ztrsv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef stpsv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t ap, intptr_t x, int64_t incx)
+cpdef dtpsv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t ap, intptr_t x, int64_t incx)
+cpdef ctpsv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t ap, intptr_t x, int64_t incx)
+cpdef ztpsv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, intptr_t ap, intptr_t x, int64_t incx)
+cpdef stbsv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, int64_t k, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef dtbsv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, int64_t k, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef ctbsv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, int64_t k, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef ztbsv_64(intptr_t handle, int uplo, int trans, int diag, int64_t n, int64_t k, intptr_t a, int64_t lda, intptr_t x, int64_t incx)
+cpdef ssymv_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef dsymv_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef csymv_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef zsymv_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef chemv_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef zhemv_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef ssbmv_64(intptr_t handle, int uplo, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef dsbmv_64(intptr_t handle, int uplo, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef chbmv_64(intptr_t handle, int uplo, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef zhbmv_64(intptr_t handle, int uplo, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef sspmv_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t ap, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef dspmv_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t ap, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef chpmv_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t ap, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef zhpmv_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t ap, intptr_t x, int64_t incx, intptr_t beta, intptr_t y, int64_t incy)
+cpdef sger_64(intptr_t handle, int64_t m, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t a, int64_t lda)
+cpdef dger_64(intptr_t handle, int64_t m, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t a, int64_t lda)
+cpdef cgeru_64(intptr_t handle, int64_t m, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t a, int64_t lda)
+cpdef cgerc_64(intptr_t handle, int64_t m, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t a, int64_t lda)
+cpdef zgeru_64(intptr_t handle, int64_t m, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t a, int64_t lda)
+cpdef zgerc_64(intptr_t handle, int64_t m, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t a, int64_t lda)
+cpdef ssyr_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t a, int64_t lda)
+cpdef dsyr_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t a, int64_t lda)
+cpdef csyr_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t a, int64_t lda)
+cpdef zsyr_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t a, int64_t lda)
+cpdef cher_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t a, int64_t lda)
+cpdef zher_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t a, int64_t lda)
+cpdef sspr_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t ap)
+cpdef dspr_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t ap)
+cpdef chpr_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t ap)
+cpdef zhpr_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t ap)
+cpdef ssyr2_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t a, int64_t lda)
+cpdef dsyr2_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t a, int64_t lda)
+cpdef csyr2_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t a, int64_t lda)
+cpdef zsyr2_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t a, int64_t lda)
+cpdef cher2_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t a, int64_t lda)
+cpdef zher2_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t a, int64_t lda)
+cpdef sspr2_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t ap)
+cpdef dspr2_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t ap)
+cpdef chpr2_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t ap)
+cpdef zhpr2_64(intptr_t handle, int uplo, int64_t n, intptr_t alpha, intptr_t x, int64_t incx, intptr_t y, int64_t incy, intptr_t ap)
+cpdef sgemv_batched_64(intptr_t handle, int trans, int64_t m, int64_t n, intptr_t alpha, intptr_t aarray, int64_t lda, intptr_t xarray, int64_t incx, intptr_t beta, intptr_t yarray, int64_t incy, int64_t batch_count)
+cpdef dgemv_batched_64(intptr_t handle, int trans, int64_t m, int64_t n, intptr_t alpha, intptr_t aarray, int64_t lda, intptr_t xarray, int64_t incx, intptr_t beta, intptr_t yarray, int64_t incy, int64_t batch_count)
+cpdef cgemv_batched_64(intptr_t handle, int trans, int64_t m, int64_t n, intptr_t alpha, intptr_t aarray, int64_t lda, intptr_t xarray, int64_t incx, intptr_t beta, intptr_t yarray, int64_t incy, int64_t batch_count)
+cpdef zgemv_batched_64(intptr_t handle, int trans, int64_t m, int64_t n, intptr_t alpha, intptr_t aarray, int64_t lda, intptr_t xarray, int64_t incx, intptr_t beta, intptr_t yarray, int64_t incy, int64_t batch_count)
+cpdef sgemv_strided_batched_64(intptr_t handle, int trans, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, long long int stride_a, intptr_t x, int64_t incx, long long int stridex, intptr_t beta, intptr_t y, int64_t incy, long long int stridey, int64_t batch_count)
+cpdef dgemv_strided_batched_64(intptr_t handle, int trans, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, long long int stride_a, intptr_t x, int64_t incx, long long int stridex, intptr_t beta, intptr_t y, int64_t incy, long long int stridey, int64_t batch_count)
+cpdef cgemv_strided_batched_64(intptr_t handle, int trans, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, long long int stride_a, intptr_t x, int64_t incx, long long int stridex, intptr_t beta, intptr_t y, int64_t incy, long long int stridey, int64_t batch_count)
+cpdef zgemv_strided_batched_64(intptr_t handle, int trans, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, long long int stride_a, intptr_t x, int64_t incx, long long int stridex, intptr_t beta, intptr_t y, int64_t incy, long long int stridey, int64_t batch_count)
+cpdef sgemm_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef dgemm_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef cgemm_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef cgemm3m_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef cgemm3m_ex_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int atype, int64_t lda, intptr_t b, int btype, int64_t ldb, intptr_t beta, intptr_t c, int ctype, int64_t ldc)
+cpdef zgemm_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef zgemm3m_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef sgemm_ex_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int atype, int64_t lda, intptr_t b, int btype, int64_t ldb, intptr_t beta, intptr_t c, int ctype, int64_t ldc)
+cpdef gemm_ex_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int atype, int64_t lda, intptr_t b, int btype, int64_t ldb, intptr_t beta, intptr_t c, int ctype, int64_t ldc, int compute_type, int algo)
+cpdef cgemm_ex_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int atype, int64_t lda, intptr_t b, int btype, int64_t ldb, intptr_t beta, intptr_t c, int ctype, int64_t ldc)
+cpdef ssyrk_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef dsyrk_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef csyrk_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef zsyrk_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef csyrk_ex_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int atype, int64_t lda, intptr_t beta, intptr_t c, int ctype, int64_t ldc)
+cpdef csyrk3m_ex_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int atype, int64_t lda, intptr_t beta, intptr_t c, int ctype, int64_t ldc)
+cpdef cherk_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef zherk_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef cherk_ex_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int atype, int64_t lda, intptr_t beta, intptr_t c, int ctype, int64_t ldc)
+cpdef cherk3m_ex_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int atype, int64_t lda, intptr_t beta, intptr_t c, int ctype, int64_t ldc)
+cpdef ssyr2k_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef dsyr2k_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef csyr2k_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef zsyr2k_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef cher2k_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef zher2k_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef ssyrkx_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef dsyrkx_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef csyrkx_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef zsyrkx_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef cherkx_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef zherkx_64(intptr_t handle, int uplo, int trans, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef ssymm_64(intptr_t handle, int side, int uplo, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef dsymm_64(intptr_t handle, int side, int uplo, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef csymm_64(intptr_t handle, int side, int uplo, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef zsymm_64(intptr_t handle, int side, int uplo, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef chemm_64(intptr_t handle, int side, int uplo, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef zhemm_64(intptr_t handle, int side, int uplo, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t beta, intptr_t c, int64_t ldc)
+cpdef strsm_64(intptr_t handle, int side, int uplo, int trans, int diag, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb)
+cpdef dtrsm_64(intptr_t handle, int side, int uplo, int trans, int diag, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb)
+cpdef ctrsm_64(intptr_t handle, int side, int uplo, int trans, int diag, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb)
+cpdef ztrsm_64(intptr_t handle, int side, int uplo, int trans, int diag, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb)
+cpdef strmm_64(intptr_t handle, int side, int uplo, int trans, int diag, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t c, int64_t ldc)
+cpdef dtrmm_64(intptr_t handle, int side, int uplo, int trans, int diag, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t c, int64_t ldc)
+cpdef ctrmm_64(intptr_t handle, int side, int uplo, int trans, int diag, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t c, int64_t ldc)
+cpdef ztrmm_64(intptr_t handle, int side, int uplo, int trans, int diag, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, intptr_t c, int64_t ldc)
+cpdef sgemm_batched_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t aarray, int64_t lda, intptr_t barray, int64_t ldb, intptr_t beta, intptr_t carray, int64_t ldc, int64_t batch_count)
+cpdef dgemm_batched_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t aarray, int64_t lda, intptr_t barray, int64_t ldb, intptr_t beta, intptr_t carray, int64_t ldc, int64_t batch_count)
+cpdef cgemm_batched_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t aarray, int64_t lda, intptr_t barray, int64_t ldb, intptr_t beta, intptr_t carray, int64_t ldc, int64_t batch_count)
+cpdef cgemm3m_batched_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t aarray, int64_t lda, intptr_t barray, int64_t ldb, intptr_t beta, intptr_t carray, int64_t ldc, int64_t batch_count)
+cpdef zgemm_batched_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t aarray, int64_t lda, intptr_t barray, int64_t ldb, intptr_t beta, intptr_t carray, int64_t ldc, int64_t batch_count)
+cpdef sgemm_strided_batched_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, long long int stride_a, intptr_t b, int64_t ldb, long long int stride_b, intptr_t beta, intptr_t c, int64_t ldc, long long int stride_c, int64_t batch_count)
+cpdef dgemm_strided_batched_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, long long int stride_a, intptr_t b, int64_t ldb, long long int stride_b, intptr_t beta, intptr_t c, int64_t ldc, long long int stride_c, int64_t batch_count)
+cpdef cgemm_strided_batched_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, long long int stride_a, intptr_t b, int64_t ldb, long long int stride_b, intptr_t beta, intptr_t c, int64_t ldc, long long int stride_c, int64_t batch_count)
+cpdef cgemm3m_strided_batched_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, long long int stride_a, intptr_t b, int64_t ldb, long long int stride_b, intptr_t beta, intptr_t c, int64_t ldc, long long int stride_c, int64_t batch_count)
+cpdef zgemm_strided_batched_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int64_t lda, long long int stride_a, intptr_t b, int64_t ldb, long long int stride_b, intptr_t beta, intptr_t c, int64_t ldc, long long int stride_c, int64_t batch_count)
+cpdef gemm_batched_ex_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t aarray, int atype, int64_t lda, intptr_t barray, int btype, int64_t ldb, intptr_t beta, intptr_t carray, int ctype, int64_t ldc, int64_t batch_count, int compute_type, int algo)
+cpdef gemm_strided_batched_ex_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, int64_t k, intptr_t alpha, intptr_t a, int atype, int64_t lda, long long int stride_a, intptr_t b, int btype, int64_t ldb, long long int stride_b, intptr_t beta, intptr_t c, int ctype, int64_t ldc, long long int stride_c, int64_t batch_count, int compute_type, int algo)
+cpdef sgeam_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t beta, intptr_t b, int64_t ldb, intptr_t c, int64_t ldc)
+cpdef dgeam_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t beta, intptr_t b, int64_t ldb, intptr_t c, int64_t ldc)
+cpdef cgeam_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t beta, intptr_t b, int64_t ldb, intptr_t c, int64_t ldc)
+cpdef zgeam_64(intptr_t handle, int transa, int transb, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t beta, intptr_t b, int64_t ldb, intptr_t c, int64_t ldc)
+cpdef strsm_batched_64(intptr_t handle, int side, int uplo, int trans, int diag, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, int64_t batch_count)
+cpdef dtrsm_batched_64(intptr_t handle, int side, int uplo, int trans, int diag, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, int64_t batch_count)
+cpdef ctrsm_batched_64(intptr_t handle, int side, int uplo, int trans, int diag, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, int64_t batch_count)
+cpdef ztrsm_batched_64(intptr_t handle, int side, int uplo, int trans, int diag, int64_t m, int64_t n, intptr_t alpha, intptr_t a, int64_t lda, intptr_t b, int64_t ldb, int64_t batch_count)
+cpdef sdgmm_64(intptr_t handle, int mode, int64_t m, int64_t n, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t c, int64_t ldc)
+cpdef ddgmm_64(intptr_t handle, int mode, int64_t m, int64_t n, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t c, int64_t ldc)
+cpdef cdgmm_64(intptr_t handle, int mode, int64_t m, int64_t n, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t c, int64_t ldc)
+cpdef zdgmm_64(intptr_t handle, int mode, int64_t m, int64_t n, intptr_t a, int64_t lda, intptr_t x, int64_t incx, intptr_t c, int64_t ldc)
+cpdef sgemm_grouped_batched(intptr_t handle, transa_array, transb_array, m_array, n_array, k_array, alpha_array, intptr_t aarray, lda_array, intptr_t barray, ldb_array, beta_array, intptr_t carray, ldc_array, int group_count, group_size)
+cpdef sgemm_grouped_batched_64(intptr_t handle, transa_array, transb_array, m_array, n_array, k_array, alpha_array, intptr_t aarray, lda_array, intptr_t barray, ldb_array, beta_array, intptr_t carray, ldc_array, int64_t group_count, group_size)
+cpdef dgemm_grouped_batched(intptr_t handle, transa_array, transb_array, m_array, n_array, k_array, alpha_array, intptr_t aarray, lda_array, intptr_t barray, ldb_array, beta_array, intptr_t carray, ldc_array, int group_count, group_size)
+cpdef dgemm_grouped_batched_64(intptr_t handle, transa_array, transb_array, m_array, n_array, k_array, alpha_array, intptr_t aarray, lda_array, intptr_t barray, ldb_array, beta_array, intptr_t carray, ldc_array, int64_t group_count, group_size)
+cpdef gemm_grouped_batched_ex(intptr_t handle, transa_array, transb_array, m_array, n_array, k_array, intptr_t alpha_array, intptr_t aarray, int atype, lda_array, intptr_t barray, int btype, ldb_array, intptr_t beta_array, intptr_t carray, int ctype, ldc_array, int group_count, group_size, int compute_type)
+cpdef gemm_grouped_batched_ex_64(intptr_t handle, transa_array, transb_array, m_array, n_array, k_array, intptr_t alpha_array, intptr_t aarray, int atype, lda_array, intptr_t barray, int btype, ldb_array, intptr_t beta_array, intptr_t carray, int ctype, ldc_array, int64_t group_count, group_size, int compute_type)
+cpdef int get_emulation_strategy(intptr_t handle) except? -1
+cpdef set_emulation_strategy(intptr_t handle, int emulation_strategy)
+cpdef int get_emulation_special_values_support(intptr_t handle) except? -1
+cpdef set_emulation_special_values_support(intptr_t handle, int mask)
+cpdef int get_fixed_point_emulation_mantissa_control(intptr_t handle) except? -1
+cpdef set_fixed_point_emulation_mantissa_control(intptr_t handle, int mantissa_control)
+cpdef int get_fixed_point_emulation_max_mantissa_bit_count(intptr_t handle) except? -1
+cpdef set_fixed_point_emulation_max_mantissa_bit_count(intptr_t handle, int max_mantissa_bit_count)
+cpdef int get_fixed_point_emulation_mantissa_bit_offset(intptr_t handle) except? -1
+cpdef set_fixed_point_emulation_mantissa_bit_offset(intptr_t handle, int mantissa_bit_offset)
+cpdef intptr_t get_fixed_point_emulation_mantissa_bit_count_pointer(intptr_t handle) except? -1
+cpdef set_fixed_point_emulation_mantissa_bit_count_pointer(intptr_t handle, intptr_t mantissa_bit_count)