diff --git a/frame/1m/packm/bli_packm.h b/frame/1m/packm/bli_packm.h index 96542790a..3f68a5bb9 100644 --- a/frame/1m/packm/bli_packm.h +++ b/frame/1m/packm/bli_packm.h @@ -52,3 +52,13 @@ #include "bli_packm_cxk.h" #include "bli_packm_cxk_ri.h" #include "bli_packm_cxk_ri3.h" + +#include "bli_packm_ref_2xk.h" +#include "bli_packm_ref_4xk.h" +#include "bli_packm_ref_6xk.h" +#include "bli_packm_ref_8xk.h" +#include "bli_packm_ref_10xk.h" +#include "bli_packm_ref_12xk.h" +#include "bli_packm_ref_14xk.h" +#include "bli_packm_ref_16xk.h" + diff --git a/frame/1m/packm/bli_packm_blk_var3.c b/frame/1m/packm/bli_packm_blk_var3.c index cc5f567f1..f79b5a48e 100644 --- a/frame/1m/packm/bli_packm_blk_var3.c +++ b/frame/1m/packm/bli_packm_blk_var3.c @@ -314,8 +314,8 @@ void PASTEMAC(ch,varname)( \ \ p_begin = p_cast; \ \ - for ( ic = ic0, ip = ip0, it = 0; it < num_iter; \ - ic += ic_inc, ip += ip_inc, it += 1 ) \ + for ( ic = ic0, ip = ip0, it = 0; it < num_iter; \ + ic += ic_inc, ip += ip_inc, it += 1 ) \ { \ panel_dim_i = bli_min( panel_dim_max, iter_dim - ic ); \ \ @@ -387,6 +387,7 @@ void PASTEMAC(ch,varname)( \ } \ \ \ + /* NOTE: This value is usually LESS than (ps_p*3)/2. */ \ p_inc = ( ldp * panel_len_max_i * 3 ) / 2; \ \ /* @@ -423,7 +424,7 @@ void PASTEMAC(ch,varname)( \ p_begin, rs_p, cs_p ); \ } \ \ - /* NOTE: This value is equivalent to ps_p. */ \ + /* NOTE: This value is equivalent to (ps_p*3)/2. */ \ p_inc = ( ldp * panel_len_max_i * 3 ) / 2; \ } \ else \ @@ -450,7 +451,7 @@ void PASTEMAC(ch,varname)( \ p_begin, rs_p, cs_p ); \ } \ \ - /* NOTE: This value is equivalent to ps_p. */ \ + /* NOTE: This value is equivalent to (ps_p*3)/2. */ \ p_inc = ( ldp * panel_len_max_i * 3 ) / 2; \ \ } \ diff --git a/frame/1m/packm/bli_packm_blk_var4.c b/frame/1m/packm/bli_packm_blk_var4.c index 06d66a8bd..adcc92088 100644 --- a/frame/1m/packm/bli_packm_blk_var4.c +++ b/frame/1m/packm/bli_packm_blk_var4.c @@ -314,7 +314,7 @@ void PASTEMAC(ch,varname)( \ \ p_begin = p_cast; \ \ - for ( ic = ic0, ip = ip0, it = 0; it < num_iter; \ + for ( ic = ic0, ip = ip0, it = 0; it < num_iter; \ ic += ic_inc, ip += ip_inc, it += 1 ) \ { \ panel_dim_i = bli_min( panel_dim_max, iter_dim - ic ); \ @@ -386,6 +386,7 @@ void PASTEMAC(ch,varname)( \ p_use, rs_p, cs_p ); \ } \ \ + /* NOTE: This value is usually LESS than ps_p. */ \ p_inc = ldp * panel_len_max_i; \ \ /* diff --git a/frame/1m/packm/bli_packm_cxk.h b/frame/1m/packm/bli_packm_cxk.h index ce41fa936..38eb6c1bc 100644 --- a/frame/1m/packm/bli_packm_cxk.h +++ b/frame/1m/packm/bli_packm_cxk.h @@ -32,17 +32,6 @@ */ -// Include headers for various packm kernels. -#include "bli_packm_ref_2xk.h" -#include "bli_packm_ref_4xk.h" -#include "bli_packm_ref_6xk.h" -#include "bli_packm_ref_8xk.h" -#include "bli_packm_ref_10xk.h" -#include "bli_packm_ref_12xk.h" -#include "bli_packm_ref_14xk.h" -#include "bli_packm_ref_16xk.h" - - #undef GENTPROT #define GENTPROT( ctype, ch, varname ) \ \ diff --git a/frame/1m/packm/bli_packm_cxk_ri.h b/frame/1m/packm/bli_packm_cxk_ri.h index 25ace973a..200b2f532 100644 --- a/frame/1m/packm/bli_packm_cxk_ri.h +++ b/frame/1m/packm/bli_packm_cxk_ri.h @@ -32,17 +32,6 @@ */ -// Include headers for various packm kernels. -#include "bli_packm_ref_2xk.h" -#include "bli_packm_ref_4xk.h" -#include "bli_packm_ref_6xk.h" -#include "bli_packm_ref_8xk.h" -#include "bli_packm_ref_10xk.h" -#include "bli_packm_ref_12xk.h" -#include "bli_packm_ref_14xk.h" -#include "bli_packm_ref_16xk.h" - - #undef GENTPROTCO #define GENTPROTCO( ctype, ctype_r, ch, chr, varname ) \ \ diff --git a/frame/1m/packm/bli_packm_cxk_ri3.h b/frame/1m/packm/bli_packm_cxk_ri3.h index 73285fd09..d45508453 100644 --- a/frame/1m/packm/bli_packm_cxk_ri3.h +++ b/frame/1m/packm/bli_packm_cxk_ri3.h @@ -32,17 +32,6 @@ */ -// Include headers for various packm kernels. -#include "bli_packm_ref_2xk.h" -#include "bli_packm_ref_4xk.h" -#include "bli_packm_ref_6xk.h" -#include "bli_packm_ref_8xk.h" -#include "bli_packm_ref_10xk.h" -#include "bli_packm_ref_12xk.h" -#include "bli_packm_ref_14xk.h" -#include "bli_packm_ref_16xk.h" - - #undef GENTPROTCO #define GENTPROTCO( ctype, ctype_r, ch, chr, varname ) \ \