11/*********************************************************************/
22/* Copyright 2009, 2010 The University of Texas at Austin. */
3+ /* Copyright 2025-2026 The OpenBLAS Project. */
34/* All rights reserved. */
45/* */
56/* Redistribution and use in source and binary forms, with or */
@@ -334,6 +335,10 @@ int strmm_ilnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX
334335int strmm_ilnncopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , float * b );
335336int strmm_iltucopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , float * b );
336337int strmm_iltncopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , float * b );
338+ int ssymm_incopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , float * b );
339+ int ssymm_itcopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , float * b );
340+ int strmm_incopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , float * b );
341+ int strmm_itcopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , float * b );
337342int strmm_olnucopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , float * b );
338343int strmm_olnncopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , float * b );
339344int strmm_oltucopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , float * b );
@@ -351,6 +356,10 @@ int dtrmm_ilnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG pos
351356int dtrmm_ilnncopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , double * b );
352357int dtrmm_iltucopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , double * b );
353358int dtrmm_iltncopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , double * b );
359+ int dsymm_itcopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , double * b );
360+ int dsymm_incopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , double * b );
361+ int dtrmm_itcopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , double * b );
362+ int dtrmm_incopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , double * b );
354363int dtrmm_olnucopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , double * b );
355364int dtrmm_olnncopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , double * b );
356365int dtrmm_oltucopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , double * b );
@@ -385,6 +394,10 @@ int ctrmm_ilnucopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, BLASLONG posX
385394int ctrmm_ilnncopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , float * b );
386395int ctrmm_iltucopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , float * b );
387396int ctrmm_iltncopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , float * b );
397+ int csymm_itcopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , float * b );
398+ int csymm_incopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , float * b );
399+ int ctrmm_itcopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , float * b );
400+ int ctrmm_incopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , float * b );
388401int ctrmm_olnucopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , float * b );
389402int ctrmm_olnncopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , float * b );
390403int ctrmm_oltucopy (BLASLONG m , BLASLONG n , float * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , float * b );
@@ -402,6 +415,10 @@ int ztrmm_ilnucopy(BLASLONG m, BLASLONG n, double *a, BLASLONG lda, BLASLONG pos
402415int ztrmm_ilnncopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , double * b );
403416int ztrmm_iltucopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , double * b );
404417int ztrmm_iltncopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , double * b );
418+ int zsymm_itcopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , double * b );
419+ int zsymm_incopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , double * b );
420+ int ztrmm_itcopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , double * b );
421+ int ztrmm_incopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , double * b );
405422int ztrmm_olnucopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , double * b );
406423int ztrmm_olnncopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , double * b );
407424int ztrmm_oltucopy (BLASLONG m , BLASLONG n , double * a , BLASLONG lda , BLASLONG posX , BLASLONG posY , double * b );
@@ -629,6 +646,10 @@ int bgemm_kernel(BLASLONG, BLASLONG, BLASLONG, bfloat16, bfloat16 *, bfloat16 *
629646int sbgemm_kernel (BLASLONG , BLASLONG , BLASLONG , float , bfloat16 * , bfloat16 * , float * , BLASLONG );
630647int sgemm_kernel (BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG );
631648int dgemm_kernel (BLASLONG , BLASLONG , BLASLONG , double , double * , double * , double * , BLASLONG );
649+ int ssymm_kernel (BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG );
650+ int dsymm_kernel (BLASLONG , BLASLONG , BLASLONG , double , double * , double * , double * , BLASLONG );
651+ int strmm_gemm_kernel (BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG );
652+ int dtrmm_gemm_kernel (BLASLONG , BLASLONG , BLASLONG , double , double * , double * , double * , BLASLONG );
632653
633654#ifdef QUAD_PRECISION
634655int qgemm_kernel (BLASLONG , BLASLONG , BLASLONG , xidouble * , xidouble * , xidouble * , xdouble * , BLASLONG );
@@ -778,6 +799,26 @@ int cgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float
778799int zgemm3m_kernel (BLASLONG , BLASLONG , BLASLONG , double , double , double * , double * , double * , BLASLONG );
779800int xgemm3m_kernel (BLASLONG , BLASLONG , BLASLONG , xdouble , xdouble , xdouble * , xdouble * , xdouble * , BLASLONG );
780801
802+ int csymm_kernel_n (BLASLONG , BLASLONG , BLASLONG , float , float , float * , float * , float * , BLASLONG );
803+ int csymm_kernel_l (BLASLONG , BLASLONG , BLASLONG , float , float , float * , float * , float * , BLASLONG );
804+ int csymm_kernel_r (BLASLONG , BLASLONG , BLASLONG , float , float , float * , float * , float * , BLASLONG );
805+ int csymm_kernel_b (BLASLONG , BLASLONG , BLASLONG , float , float , float * , float * , float * , BLASLONG );
806+
807+ int ctrmm_gemm_kernel_n (BLASLONG , BLASLONG , BLASLONG , float , float , float * , float * , float * , BLASLONG );
808+ int ctrmm_gemm_kernel_l (BLASLONG , BLASLONG , BLASLONG , float , float , float * , float * , float * , BLASLONG );
809+ int ctrmm_gemm_kernel_r (BLASLONG , BLASLONG , BLASLONG , float , float , float * , float * , float * , BLASLONG );
810+ int ctrmm_gemm_kernel_b (BLASLONG , BLASLONG , BLASLONG , float , float , float * , float * , float * , BLASLONG );
811+
812+ int zsymm_kernel_n (BLASLONG , BLASLONG , BLASLONG , double , double , double * , double * , double * , BLASLONG );
813+ int zsymm_kernel_l (BLASLONG , BLASLONG , BLASLONG , double , double , double * , double * , double * , BLASLONG );
814+ int zsymm_kernel_r (BLASLONG , BLASLONG , BLASLONG , double , double , double * , double * , double * , BLASLONG );
815+ int zsymm_kernel_b (BLASLONG , BLASLONG , BLASLONG , double , double , double * , double * , double * , BLASLONG );
816+
817+ int ztrmm_gemm_kernel_n (BLASLONG , BLASLONG , BLASLONG , double , double , double * , double * , double * , BLASLONG );
818+ int ztrmm_gemm_kernel_l (BLASLONG , BLASLONG , BLASLONG , double , double , double * , double * , double * , BLASLONG );
819+ int ztrmm_gemm_kernel_r (BLASLONG , BLASLONG , BLASLONG , double , double , double * , double * , double * , BLASLONG );
820+ int ztrmm_gemm_kernel_b (BLASLONG , BLASLONG , BLASLONG , double , double , double * , double * , double * , BLASLONG );
821+
781822int shgemm_nn (blas_arg_t * , BLASLONG * , BLASLONG * , hfloat16 * , hfloat16 * , BLASLONG );
782823int shgemm_nt (blas_arg_t * , BLASLONG * , BLASLONG * , hfloat16 * , hfloat16 * , BLASLONG );
783824int shgemm_tn (blas_arg_t * , BLASLONG * , BLASLONG * , hfloat16 * , hfloat16 * , BLASLONG );
0 commit comments