diff --git a/frame/1/bli_l1v_ker.h b/frame/1/bli_l1v_ker.h index a864e9390..b71e778a6 100644 --- a/frame/1/bli_l1v_ker.h +++ b/frame/1/bli_l1v_ker.h @@ -37,6 +37,11 @@ // Define template prototypes for level-1v kernels. // +// Note: Instead of defining function prototype macro templates and then +// instantiating those macros to define the individual function prototypes, +// we simply alias the official operations' prototypes as defined in +// bli_l1v_ker_prot.h. + #undef GENTPROT #define GENTPROT ADDV_KER_PROT diff --git a/frame/1/bli_l1v_ker_prot.h b/frame/1/bli_l1v_ker_prot.h index a79867f0a..b5594b1d3 100644 --- a/frame/1/bli_l1v_ker_prot.h +++ b/frame/1/bli_l1v_ker_prot.h @@ -41,11 +41,11 @@ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ); @@ -53,10 +53,10 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - dim_t* restrict index, \ - cntx_t* cntx \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + dim_t* restrict index, \ + cntx_t* restrict cntx \ ); \ @@ -64,13 +64,13 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict beta, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict beta, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ); \ @@ -78,12 +78,12 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ); \ @@ -91,11 +91,11 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ); @@ -103,13 +103,13 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjx, \ - conj_t conjy, \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - ctype* restrict rho, \ - cntx_t* cntx \ + conj_t conjx, \ + conj_t conjy, \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + ctype* restrict rho, \ + cntx_t* restrict cntx \ ); \ @@ -117,15 +117,15 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjx, \ - conj_t conjy, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - ctype* restrict beta, \ - ctype* restrict rho, \ - cntx_t* cntx \ + conj_t conjx, \ + conj_t conjy, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + ctype* restrict beta, \ + ctype* restrict rho, \ + cntx_t* restrict cntx \ ); \ @@ -133,9 +133,9 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - cntx_t* cntx \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + cntx_t* restrict cntx \ ); \ @@ -143,11 +143,11 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjalpha, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - cntx_t* cntx \ + conj_t conjalpha, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + cntx_t* restrict cntx \ ); \ @@ -155,12 +155,12 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ); \ @@ -168,11 +168,11 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjalpha, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - cntx_t* cntx \ + conj_t conjalpha, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + cntx_t* restrict cntx \ ); \ @@ -180,11 +180,11 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ); @@ -192,10 +192,10 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ); \ @@ -203,11 +203,11 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict beta, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict beta, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ); \ diff --git a/frame/1f/bli_l1f_ker.h b/frame/1f/bli_l1f_ker.h index 48eb84533..a3ed4fa4a 100644 --- a/frame/1f/bli_l1f_ker.h +++ b/frame/1f/bli_l1f_ker.h @@ -37,6 +37,11 @@ // Define template prototypes for level-1f kernels. // +// Note: Instead of defining function prototype macro templates and then +// instantiating those macros to define the individual function prototypes, +// we simply alias the official operations' prototypes as defined in +// bli_l1f_ker_prot.h. + #undef GENTPROT #define GENTPROT AXPY2V_KER_PROT diff --git a/frame/1f/bli_l1f_ker_prot.h b/frame/1f/bli_l1f_ker_prot.h index 17bd62320..3bbac2b89 100644 --- a/frame/1f/bli_l1f_ker_prot.h +++ b/frame/1f/bli_l1f_ker_prot.h @@ -41,15 +41,15 @@ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjx, \ - conj_t conjy, \ - dim_t n, \ - ctype* restrict alphax, \ - ctype* restrict alphay, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - ctype* restrict z, inc_t incz, \ - cntx_t* cntx \ + conj_t conjx, \ + conj_t conjy, \ + dim_t n, \ + ctype* restrict alphax, \ + ctype* restrict alphay, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + ctype* restrict z, inc_t incz, \ + cntx_t* restrict cntx \ ); @@ -57,15 +57,15 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conja, \ - conj_t conjx, \ - dim_t m, \ - dim_t b_n, \ - ctype* restrict alpha, \ - ctype* restrict a, inc_t inca, inc_t lda, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conja, \ + conj_t conjx, \ + dim_t m, \ + dim_t b_n, \ + ctype* restrict alpha, \ + ctype* restrict a, inc_t inca, inc_t lda, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ); @@ -73,16 +73,16 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjxt, \ - conj_t conjx, \ - conj_t conjy, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - ctype* restrict rho, \ - ctype* restrict z, inc_t incz, \ - cntx_t* cntx \ + conj_t conjxt, \ + conj_t conjx, \ + conj_t conjy, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + ctype* restrict rho, \ + ctype* restrict z, inc_t incz, \ + cntx_t* restrict cntx \ ); @@ -90,20 +90,20 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjat, \ - conj_t conja, \ - conj_t conjw, \ - conj_t conjx, \ - dim_t m, \ - dim_t b_n, \ - ctype* restrict alpha, \ - ctype* restrict a, inc_t inca, inc_t lda, \ - ctype* restrict w, inc_t incw, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict beta, \ - ctype* restrict y, inc_t incy, \ - ctype* restrict z, inc_t incz, \ - cntx_t* cntx \ + conj_t conjat, \ + conj_t conja, \ + conj_t conjw, \ + conj_t conjx, \ + dim_t m, \ + dim_t b_n, \ + ctype* restrict alpha, \ + ctype* restrict a, inc_t inca, inc_t lda, \ + ctype* restrict w, inc_t incw, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict beta, \ + ctype* restrict y, inc_t incy, \ + ctype* restrict z, inc_t incz, \ + cntx_t* restrict cntx \ ); @@ -111,15 +111,15 @@ void PASTEMAC(ch,opname) \ \ void PASTEMAC(ch,opname) \ ( \ - conj_t conjat, \ - conj_t conjx, \ - dim_t m, \ - dim_t b_n, \ - ctype* restrict alpha, \ - ctype* restrict a, inc_t inca, inc_t lda, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict beta, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjat, \ + conj_t conjx, \ + dim_t m, \ + dim_t b_n, \ + ctype* restrict alpha, \ + ctype* restrict a, inc_t inca, inc_t lda, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict beta, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ); diff --git a/frame/1m/bli_l1m_ker.h b/frame/1m/bli_l1m_ker.h index cf73cd390..78c65d365 100644 --- a/frame/1m/bli_l1m_ker.h +++ b/frame/1m/bli_l1m_ker.h @@ -37,6 +37,11 @@ // Define template prototypes for level-1m kernels. // +// Note: Instead of defining function prototype macro templates and then +// instantiating those macros to define the individual function prototypes, +// we simply alias the official operations' prototypes as defined in +// bli_l1m_ker_prot.h. + // native packm kernels #undef GENTPROT diff --git a/frame/3/bli_l3_ukr.h b/frame/3/bli_l3_ukr.h index 6de50f427..275f90049 100644 --- a/frame/3/bli_l3_ukr.h +++ b/frame/3/bli_l3_ukr.h @@ -36,6 +36,11 @@ // Define template prototypes for level-3 micro-kernels. // +// Note: Instead of defining function prototype macro templates and then +// instantiating those macros to define the individual function prototypes, +// we simply alias the official operations' prototypes as defined in +// bli_l3_ukr_prot.h. + #undef GENTPROT #define GENTPROT GEMM_UKR_PROT diff --git a/kernels/bgq/1/bli_axpyv_bgq_int.c b/kernels/bgq/1/bli_axpyv_bgq_int.c index 8885d9817..20b54cc00 100644 --- a/kernels/bgq/1/bli_axpyv_bgq_int.c +++ b/kernels/bgq/1/bli_axpyv_bgq_int.c @@ -41,7 +41,7 @@ void bli_daxpyv_bgq_int double* restrict alpha, double* restrict x, inc_t incx, double* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { if ( bli_zero_dim1( n ) ) return; diff --git a/kernels/bgq/1/bli_dotv_bgq_int.c b/kernels/bgq/1/bli_dotv_bgq_int.c index 96f2259aa..3fa21143d 100644 --- a/kernels/bgq/1/bli_dotv_bgq_int.c +++ b/kernels/bgq/1/bli_dotv_bgq_int.c @@ -42,7 +42,7 @@ void bli_ddotv_bgq_int double* restrict x, inc_t incx, double* restrict y, inc_t incy, double* restrict rho, - cntx_t* cntx + cntx_t* restrict cntx ) { bool_t use_ref = FALSE; diff --git a/kernels/bgq/1f/bli_axpyf_bgq_int.c b/kernels/bgq/1f/bli_axpyf_bgq_int.c index 9ea0fb674..b3b890d3e 100644 --- a/kernels/bgq/1f/bli_axpyf_bgq_int.c +++ b/kernels/bgq/1f/bli_axpyf_bgq_int.c @@ -45,7 +45,7 @@ void bli_daxpyf_bgq_int double* restrict a, inc_t inca, inc_t lda, double* restrict x, inc_t incx, double* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t fusefac = 8; diff --git a/kernels/penryn/1/bli_axpyv_penryn_int.c b/kernels/penryn/1/bli_axpyv_penryn_int.c index 69a98eb73..8a4f5c530 100644 --- a/kernels/penryn/1/bli_axpyv_penryn_int.c +++ b/kernels/penryn/1/bli_axpyv_penryn_int.c @@ -50,7 +50,7 @@ void bli_daxpyv_penryn_int double* restrict alpha, double* restrict x, inc_t incx, double* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { double* restrict alpha_cast = alpha; diff --git a/kernels/penryn/1/bli_dotv_penryn_int.c b/kernels/penryn/1/bli_dotv_penryn_int.c index 51c498f8b..493bb513d 100644 --- a/kernels/penryn/1/bli_dotv_penryn_int.c +++ b/kernels/penryn/1/bli_dotv_penryn_int.c @@ -51,7 +51,7 @@ void bli_ddotv_penryn_int double* restrict x, inc_t incx, double* restrict y, inc_t incy, double* restrict rho, - cntx_t* cntx + cntx_t* restrict cntx ) { double* restrict x_cast = x; diff --git a/kernels/penryn/1f/bli_axpy2v_penryn_int.c b/kernels/penryn/1f/bli_axpy2v_penryn_int.c index 8e43aa4c7..cc758ec37 100644 --- a/kernels/penryn/1f/bli_axpy2v_penryn_int.c +++ b/kernels/penryn/1f/bli_axpy2v_penryn_int.c @@ -53,7 +53,7 @@ void bli_daxpy2v_penryn_int double* restrict x, inc_t incx, double* restrict y, inc_t incy, double* restrict z, inc_t incz, - cntx_t* cntx + cntx_t* restrict cntx ) { double* restrict alpha_cast = alpha; diff --git a/kernels/penryn/1f/bli_axpyf_penryn_int.c b/kernels/penryn/1f/bli_axpyf_penryn_int.c index ce3eb99c8..d94d48927 100644 --- a/kernels/penryn/1f/bli_axpyf_penryn_int.c +++ b/kernels/penryn/1f/bli_axpyf_penryn_int.c @@ -53,7 +53,7 @@ void bli_daxpyf_penryn_int double* restrict a, inc_t inca, inc_t lda, double* restrict x, inc_t incx, double* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { double* restrict alpha_cast = alpha; diff --git a/kernels/penryn/1f/bli_dotaxpyv_penryn_int.c b/kernels/penryn/1f/bli_dotaxpyv_penryn_int.c index c2daeefdc..d86ac1338 100644 --- a/kernels/penryn/1f/bli_dotaxpyv_penryn_int.c +++ b/kernels/penryn/1f/bli_dotaxpyv_penryn_int.c @@ -54,7 +54,7 @@ void bli_ddotaxpyv_penryn_int double* restrict y, inc_t incy, double* restrict rho, double* restrict z, inc_t incz, - cntx_t* cntx + cntx_t* restrict cntx ) { double* restrict alpha_cast = alpha; diff --git a/kernels/penryn/1f/bli_dotxaxpyf_penryn_int.c b/kernels/penryn/1f/bli_dotxaxpyf_penryn_int.c index 13b461632..447a9a26a 100644 --- a/kernels/penryn/1f/bli_dotxaxpyf_penryn_int.c +++ b/kernels/penryn/1f/bli_dotxaxpyf_penryn_int.c @@ -58,7 +58,7 @@ void bli_ddotxaxpyf_penryn_int double* restrict beta, double* restrict y, inc_t incy, double* restrict z, inc_t incz, - cntx_t* cntx + cntx_t* restrict cntx ) { double* restrict alpha_cast = alpha; diff --git a/kernels/penryn/1f/bli_dotxf_penryn_int.c b/kernels/penryn/1f/bli_dotxf_penryn_int.c index 127bf1d6e..7c4e2cb11 100644 --- a/kernels/penryn/1f/bli_dotxf_penryn_int.c +++ b/kernels/penryn/1f/bli_dotxf_penryn_int.c @@ -54,7 +54,7 @@ void bli_ddotxf_penryn_int double* restrict x, inc_t incx, double* restrict beta, double* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { double* restrict alpha_cast = alpha; diff --git a/kernels/zen/1/bli_amaxv_zen_int.c b/kernels/zen/1/bli_amaxv_zen_int.c index bbdbcbac1..af001dad8 100644 --- a/kernels/zen/1/bli_amaxv_zen_int.c +++ b/kernels/zen/1/bli_amaxv_zen_int.c @@ -68,10 +68,10 @@ typedef union void bli_samaxv_zen_int ( - dim_t n, - float* x, inc_t incx, - dim_t* i_max, - cntx_t* cntx + dim_t n, + float* restrict x, inc_t incx, + dim_t* restrict i_max, + cntx_t* restrict cntx ) { float* minus_one = PASTEMAC(s,m1); @@ -232,10 +232,10 @@ void bli_samaxv_zen_int void bli_damaxv_zen_int ( - dim_t n, - double* x, inc_t incx, - dim_t* i_max, - cntx_t* cntx + dim_t n, + double* restrict x, inc_t incx, + dim_t* restrict i_max, + cntx_t* restrict cntx ) { double* minus_one = PASTEMAC(d,m1); diff --git a/kernels/zen/1/bli_axpyv_zen_int.c b/kernels/zen/1/bli_axpyv_zen_int.c index ff6018c6a..d84fed5af 100644 --- a/kernels/zen/1/bli_axpyv_zen_int.c +++ b/kernels/zen/1/bli_axpyv_zen_int.c @@ -61,7 +61,7 @@ void bli_saxpyv_zen_int float* restrict alpha, float* restrict x, inc_t incx, float* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 8; @@ -165,7 +165,7 @@ void bli_daxpyv_zen_int double* restrict alpha, double* restrict x, inc_t incx, double* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 4; diff --git a/kernels/zen/1/bli_axpyv_zen_int10.c b/kernels/zen/1/bli_axpyv_zen_int10.c index 02631bbbf..a6909ebcd 100644 --- a/kernels/zen/1/bli_axpyv_zen_int10.c +++ b/kernels/zen/1/bli_axpyv_zen_int10.c @@ -61,7 +61,7 @@ void bli_saxpyv_zen_int10 float* restrict alpha, float* restrict x, inc_t incx, float* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 8; @@ -267,7 +267,7 @@ void bli_daxpyv_zen_int10 double* restrict alpha, double* restrict x, inc_t incx, double* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 4; diff --git a/kernels/zen/1/bli_dotv_zen_int.c b/kernels/zen/1/bli_dotv_zen_int.c index a684c1727..71a7caa8c 100644 --- a/kernels/zen/1/bli_dotv_zen_int.c +++ b/kernels/zen/1/bli_dotv_zen_int.c @@ -61,7 +61,7 @@ void bli_sdotv_zen_int float* restrict x, inc_t incx, float* restrict y, inc_t incy, float* restrict rho, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 8; @@ -183,7 +183,7 @@ void bli_ddotv_zen_int double* restrict x, inc_t incx, double* restrict y, inc_t incy, double* restrict rho, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 4; diff --git a/kernels/zen/1/bli_dotv_zen_int10.c b/kernels/zen/1/bli_dotv_zen_int10.c index 5b414c441..0b8e7f729 100644 --- a/kernels/zen/1/bli_dotv_zen_int10.c +++ b/kernels/zen/1/bli_dotv_zen_int10.c @@ -62,7 +62,7 @@ void bli_sdotv_zen_int10 float* restrict x, inc_t incx, float* restrict y, inc_t incy, float* restrict rho, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 8; @@ -259,7 +259,7 @@ void bli_ddotv_zen_int10 double* restrict x, inc_t incx, double* restrict y, inc_t incy, double* restrict rho, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 4; diff --git a/kernels/zen/1/bli_dotxv_zen_int.c b/kernels/zen/1/bli_dotxv_zen_int.c index c09b34dce..cd8fa6816 100644 --- a/kernels/zen/1/bli_dotxv_zen_int.c +++ b/kernels/zen/1/bli_dotxv_zen_int.c @@ -63,7 +63,7 @@ void bli_sdotxv_zen_int float* restrict y, inc_t incy, float* restrict beta, float* restrict rho, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 8; @@ -191,7 +191,7 @@ void bli_ddotxv_zen_int double* restrict y, inc_t incy, double* restrict beta, double* restrict rho, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 4; diff --git a/kernels/zen/1/bli_scalv_zen_int.c b/kernels/zen/1/bli_scalv_zen_int.c index bdd52b079..29bcea478 100644 --- a/kernels/zen/1/bli_scalv_zen_int.c +++ b/kernels/zen/1/bli_scalv_zen_int.c @@ -60,7 +60,7 @@ void bli_sscalv_zen_int dim_t n, float* restrict alpha, float* restrict x, inc_t incx, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 8; @@ -166,7 +166,7 @@ void bli_dscalv_zen_int dim_t n, double* restrict alpha, double* restrict x, inc_t incx, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 4; diff --git a/kernels/zen/1/bli_scalv_zen_int10.c b/kernels/zen/1/bli_scalv_zen_int10.c index ec88a9015..4149309d9 100644 --- a/kernels/zen/1/bli_scalv_zen_int10.c +++ b/kernels/zen/1/bli_scalv_zen_int10.c @@ -60,7 +60,7 @@ void bli_sscalv_zen_int10 dim_t n, float* restrict alpha, float* restrict x, inc_t incx, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 8; @@ -257,7 +257,7 @@ void bli_dscalv_zen_int10 dim_t n, double* restrict alpha, double* restrict x, inc_t incx, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t n_elem_per_reg = 4; diff --git a/kernels/zen/1f/bli_axpyf_zen_int_8.c b/kernels/zen/1f/bli_axpyf_zen_int_8.c index fe8566b27..3e892f155 100644 --- a/kernels/zen/1f/bli_axpyf_zen_int_8.c +++ b/kernels/zen/1f/bli_axpyf_zen_int_8.c @@ -63,7 +63,7 @@ void bli_saxpyf_zen_int_8 float* restrict a, inc_t inca, inc_t lda, float* restrict x, inc_t incx, float* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t fuse_fac = 8; @@ -279,7 +279,7 @@ void bli_daxpyf_zen_int_8 double* restrict a, inc_t inca, inc_t lda, double* restrict x, inc_t incx, double* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t fuse_fac = 8; diff --git a/kernels/zen/1f/bli_dotxf_zen_int_8.c b/kernels/zen/1f/bli_dotxf_zen_int_8.c index 0147eaea6..7de80d24f 100644 --- a/kernels/zen/1f/bli_dotxf_zen_int_8.c +++ b/kernels/zen/1f/bli_dotxf_zen_int_8.c @@ -64,7 +64,7 @@ void bli_sdotxf_zen_int_8 float* restrict x, inc_t incx, float* restrict beta, float* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t fuse_fac = 8; @@ -394,7 +394,7 @@ void bli_ddotxf_zen_int_8 double* restrict x, inc_t incx, double* restrict beta, double* restrict y, inc_t incy, - cntx_t* cntx + cntx_t* restrict cntx ) { const dim_t fuse_fac = 8; diff --git a/ref_kernels/1/bli_addv_ref.c b/ref_kernels/1/bli_addv_ref.c index eb9239f8f..e8b7c3dca 100644 --- a/ref_kernels/1/bli_addv_ref.c +++ b/ref_kernels/1/bli_addv_ref.c @@ -39,11 +39,11 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1/bli_amaxv_ref.c b/ref_kernels/1/bli_amaxv_ref.c index 68c340d30..21ecd1966 100644 --- a/ref_kernels/1/bli_amaxv_ref.c +++ b/ref_kernels/1/bli_amaxv_ref.c @@ -43,10 +43,10 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - dim_t n, \ - ctype* x, inc_t incx, \ - dim_t* i_max, \ - cntx_t* cntx \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + dim_t* restrict i_max, \ + cntx_t* restrict cntx \ ) \ { \ ctype_r* minus_one = PASTEMAC(chr,m1); \ diff --git a/ref_kernels/1/bli_axpbyv_ref.c b/ref_kernels/1/bli_axpbyv_ref.c index ce503bea8..be3799d67 100644 --- a/ref_kernels/1/bli_axpbyv_ref.c +++ b/ref_kernels/1/bli_axpbyv_ref.c @@ -39,13 +39,13 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict beta, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict beta, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1/bli_axpyv_ref.c b/ref_kernels/1/bli_axpyv_ref.c index de5af407f..62662c337 100644 --- a/ref_kernels/1/bli_axpyv_ref.c +++ b/ref_kernels/1/bli_axpyv_ref.c @@ -39,12 +39,12 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1/bli_copyv_ref.c b/ref_kernels/1/bli_copyv_ref.c index 663287c5f..c809d9579 100644 --- a/ref_kernels/1/bli_copyv_ref.c +++ b/ref_kernels/1/bli_copyv_ref.c @@ -39,11 +39,11 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1/bli_dotv_ref.c b/ref_kernels/1/bli_dotv_ref.c index 145e61079..c52bae38d 100644 --- a/ref_kernels/1/bli_dotv_ref.c +++ b/ref_kernels/1/bli_dotv_ref.c @@ -39,13 +39,13 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjx, \ - conj_t conjy, \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - ctype* restrict rho, \ - cntx_t* cntx \ + conj_t conjx, \ + conj_t conjy, \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + ctype* restrict rho, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1/bli_dotxv_ref.c b/ref_kernels/1/bli_dotxv_ref.c index 2fef8abd0..ad7ccdfbb 100644 --- a/ref_kernels/1/bli_dotxv_ref.c +++ b/ref_kernels/1/bli_dotxv_ref.c @@ -39,15 +39,15 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjx, \ - conj_t conjy, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - ctype* restrict beta, \ - ctype* restrict rho, \ - cntx_t* cntx \ + conj_t conjx, \ + conj_t conjy, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + ctype* restrict beta, \ + ctype* restrict rho, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1/bli_invertv_ref.c b/ref_kernels/1/bli_invertv_ref.c index a503315d9..2fdbc306c 100644 --- a/ref_kernels/1/bli_invertv_ref.c +++ b/ref_kernels/1/bli_invertv_ref.c @@ -39,9 +39,9 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - cntx_t* cntx \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1/bli_scal2v_ref.c b/ref_kernels/1/bli_scal2v_ref.c index a5ee33449..39b1e20f7 100644 --- a/ref_kernels/1/bli_scal2v_ref.c +++ b/ref_kernels/1/bli_scal2v_ref.c @@ -39,12 +39,12 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1/bli_scalv_ref.c b/ref_kernels/1/bli_scalv_ref.c index 2a7ec3d59..b1433c4d6 100644 --- a/ref_kernels/1/bli_scalv_ref.c +++ b/ref_kernels/1/bli_scalv_ref.c @@ -39,11 +39,11 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjalpha, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - cntx_t* cntx \ + conj_t conjalpha, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1/bli_setv_ref.c b/ref_kernels/1/bli_setv_ref.c index 00894b13e..b624ddd7c 100644 --- a/ref_kernels/1/bli_setv_ref.c +++ b/ref_kernels/1/bli_setv_ref.c @@ -39,11 +39,11 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjalpha, \ - dim_t n, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - cntx_t* cntx \ + conj_t conjalpha, \ + dim_t n, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1/bli_subv_ref.c b/ref_kernels/1/bli_subv_ref.c index 4faba88ec..2f24fdcca 100644 --- a/ref_kernels/1/bli_subv_ref.c +++ b/ref_kernels/1/bli_subv_ref.c @@ -39,11 +39,11 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1/bli_swapv_ref.c b/ref_kernels/1/bli_swapv_ref.c index 06a02133b..599c83ed2 100644 --- a/ref_kernels/1/bli_swapv_ref.c +++ b/ref_kernels/1/bli_swapv_ref.c @@ -39,10 +39,10 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1/bli_xpbyv_ref.c b/ref_kernels/1/bli_xpbyv_ref.c index 48c9c5ec8..dc73e4116 100644 --- a/ref_kernels/1/bli_xpbyv_ref.c +++ b/ref_kernels/1/bli_xpbyv_ref.c @@ -39,12 +39,12 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjx, \ - dim_t n, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict beta, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjx, \ + dim_t n, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict beta, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ) \ { \ ctype* restrict chi1; \ diff --git a/ref_kernels/1f/bli_axpy2v_ref.c b/ref_kernels/1f/bli_axpy2v_ref.c index 6983b96f9..b6a8ca3ba 100644 --- a/ref_kernels/1f/bli_axpy2v_ref.c +++ b/ref_kernels/1f/bli_axpy2v_ref.c @@ -40,15 +40,15 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjx, \ - conj_t conjy, \ - dim_t n, \ - ctype* restrict alphax, \ - ctype* restrict alphay, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - ctype* restrict z, inc_t incz, \ - cntx_t* cntx \ + conj_t conjx, \ + conj_t conjy, \ + dim_t n, \ + ctype* restrict alphax, \ + ctype* restrict alphay, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + ctype* restrict z, inc_t incz, \ + cntx_t* restrict cntx \ ) \ { \ /* Query the context for the kernel function pointer. */ \ diff --git a/ref_kernels/1f/bli_axpyf_ref.c b/ref_kernels/1f/bli_axpyf_ref.c index 1d148deda..e0b91903d 100644 --- a/ref_kernels/1f/bli_axpyf_ref.c +++ b/ref_kernels/1f/bli_axpyf_ref.c @@ -40,15 +40,15 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conja, \ - conj_t conjx, \ - dim_t m, \ - dim_t b_n, \ - ctype* restrict alpha, \ - ctype* restrict a, inc_t inca, inc_t lda, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conja, \ + conj_t conjx, \ + dim_t m, \ + dim_t b_n, \ + ctype* restrict alpha, \ + ctype* restrict a, inc_t inca, inc_t lda, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ) \ { \ ctype* a1; \ diff --git a/ref_kernels/1f/bli_dotaxpyv_ref.c b/ref_kernels/1f/bli_dotaxpyv_ref.c index fde9cc476..472f78e87 100644 --- a/ref_kernels/1f/bli_dotaxpyv_ref.c +++ b/ref_kernels/1f/bli_dotaxpyv_ref.c @@ -40,16 +40,16 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjxt, \ - conj_t conjx, \ - conj_t conjy, \ - dim_t m, \ - ctype* restrict alpha, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict y, inc_t incy, \ - ctype* restrict rho, \ - ctype* restrict z, inc_t incz, \ - cntx_t* cntx \ + conj_t conjxt, \ + conj_t conjx, \ + conj_t conjy, \ + dim_t m, \ + ctype* restrict alpha, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict y, inc_t incy, \ + ctype* restrict rho, \ + ctype* restrict z, inc_t incz, \ + cntx_t* restrict cntx \ ) \ { \ ctype* one = PASTEMAC(ch,1); \ diff --git a/ref_kernels/1f/bli_dotxaxpyf_ref.c b/ref_kernels/1f/bli_dotxaxpyf_ref.c index 8fe184c4a..9db9fd732 100644 --- a/ref_kernels/1f/bli_dotxaxpyf_ref.c +++ b/ref_kernels/1f/bli_dotxaxpyf_ref.c @@ -40,20 +40,20 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjat, \ - conj_t conja, \ - conj_t conjw, \ - conj_t conjx, \ - dim_t m, \ - dim_t b_n, \ - ctype* restrict alpha, \ - ctype* restrict a, inc_t inca, inc_t lda, \ - ctype* restrict w, inc_t incw, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict beta, \ - ctype* restrict y, inc_t incy, \ - ctype* restrict z, inc_t incz, \ - cntx_t* cntx \ + conj_t conjat, \ + conj_t conja, \ + conj_t conjw, \ + conj_t conjx, \ + dim_t m, \ + dim_t b_n, \ + ctype* restrict alpha, \ + ctype* restrict a, inc_t inca, inc_t lda, \ + ctype* restrict w, inc_t incw, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict beta, \ + ctype* restrict y, inc_t incy, \ + ctype* restrict z, inc_t incz, \ + cntx_t* restrict cntx \ ) \ { \ /* A is m x n. */ \ diff --git a/ref_kernels/1f/bli_dotxf_ref.c b/ref_kernels/1f/bli_dotxf_ref.c index 2a64045b3..c82ce3d8f 100644 --- a/ref_kernels/1f/bli_dotxf_ref.c +++ b/ref_kernels/1f/bli_dotxf_ref.c @@ -40,16 +40,16 @@ \ void PASTEMAC3(ch,opname,arch,suf) \ ( \ - conj_t conjat, \ - conj_t conjx, \ - dim_t m, \ - dim_t b_n, \ - ctype* restrict alpha, \ - ctype* restrict a, inc_t inca, inc_t lda, \ - ctype* restrict x, inc_t incx, \ - ctype* restrict beta, \ - ctype* restrict y, inc_t incy, \ - cntx_t* cntx \ + conj_t conjat, \ + conj_t conjx, \ + dim_t m, \ + dim_t b_n, \ + ctype* restrict alpha, \ + ctype* restrict a, inc_t inca, inc_t lda, \ + ctype* restrict x, inc_t incx, \ + ctype* restrict beta, \ + ctype* restrict y, inc_t incy, \ + cntx_t* restrict cntx \ ) \ { \ ctype* a1; \