From 096b366ddcfe386f44419ef84d8df8be13825f86 Mon Sep 17 00:00:00 2001 From: "Field G. Van Zee" Date: Thu, 25 Apr 2013 16:43:43 -0500 Subject: [PATCH] Use cntl trees that block in n dimension. Details: - Updated _cntl.c files for each level-3 operation to induce blocked algorithms that first paritition in the n dimension with a blocksize of NC. Typically this is not an issue since only very large problems exceed that of NC. But developers often run very large problems, and so this extra blocking should be the default. - Removed some recently introduced but now unused macros from bli_param_macro_defs.h. --- frame/3/gemm/bli_gemm_cntl.c | 4 ++-- frame/3/hemm/bli_hemm_cntl.c | 4 ++-- frame/3/her2k/bli_her2k_cntl.c | 4 ++-- frame/3/herk/bli_herk_cntl.c | 4 ++-- frame/3/trmm/bli_trmm_cntl.c | 4 ++-- frame/3/trmm3/bli_trmm3_cntl.c | 4 ++-- frame/3/trsm/bli_trsm_cntl.c | 4 ++-- frame/include/bli_param_macro_defs.h | 8 -------- 8 files changed, 14 insertions(+), 22 deletions(-) diff --git a/frame/3/gemm/bli_gemm_cntl.c b/frame/3/gemm/bli_gemm_cntl.c index 8ed26fb18..55970b57e 100644 --- a/frame/3/gemm/bli_gemm_cntl.c +++ b/frame/3/gemm/bli_gemm_cntl.c @@ -219,8 +219,8 @@ void bli_gemm_cntl_init() NULL ); // Alias the "master" gemm control tree to a shorter name. - gemm_cntl = gemm_cntl_mm_op; - + //gemm_cntl = gemm_cntl_mm_op; + gemm_cntl = gemm_cntl_vl_mm; } void bli_gemm_cntl_finalize() diff --git a/frame/3/hemm/bli_hemm_cntl.c b/frame/3/hemm/bli_hemm_cntl.c index 2a1767649..e6d94ec7a 100644 --- a/frame/3/hemm/bli_hemm_cntl.c +++ b/frame/3/hemm/bli_hemm_cntl.c @@ -219,8 +219,8 @@ void bli_hemm_cntl_init() NULL ); // Alias the "master" hemm control tree to a shorter name. - hemm_cntl = hemm_cntl_mm_op; - + //hemm_cntl = hemm_cntl_mm_op; + hemm_cntl = hemm_cntl_vl_mm; } void bli_hemm_cntl_finalize() diff --git a/frame/3/her2k/bli_her2k_cntl.c b/frame/3/her2k/bli_her2k_cntl.c index 186eebf4a..adee9eb48 100644 --- a/frame/3/her2k/bli_her2k_cntl.c +++ b/frame/3/her2k/bli_her2k_cntl.c @@ -222,8 +222,8 @@ void bli_her2k_cntl_init() NULL ); // Alias the "master" her2k control tree to a shorter name. - her2k_cntl = her2k_cntl_mm_op; - + //her2k_cntl = her2k_cntl_mm_op; + her2k_cntl = her2k_cntl_vl_mm; } void bli_her2k_cntl_finalize() diff --git a/frame/3/herk/bli_herk_cntl.c b/frame/3/herk/bli_herk_cntl.c index 3ff2ae63c..19c897ffa 100644 --- a/frame/3/herk/bli_herk_cntl.c +++ b/frame/3/herk/bli_herk_cntl.c @@ -219,8 +219,8 @@ void bli_herk_cntl_init() NULL ); // Alias the "master" herk control tree to a shorter name. - herk_cntl = herk_cntl_mm_op; - + //herk_cntl = herk_cntl_mm_op; + herk_cntl = herk_cntl_vl_mm; } void bli_herk_cntl_finalize() diff --git a/frame/3/trmm/bli_trmm_cntl.c b/frame/3/trmm/bli_trmm_cntl.c index 002df204f..a4a18320f 100644 --- a/frame/3/trmm/bli_trmm_cntl.c +++ b/frame/3/trmm/bli_trmm_cntl.c @@ -227,8 +227,8 @@ void bli_trmm_cntl_init() NULL ); // Alias the "master" trmm control tree to a shorter name. - trmm_cntl = trmm_cntl_mm_op; - + //trmm_cntl = trmm_cntl_mm_op; + trmm_cntl = trmm_cntl_vl_mm; } void bli_trmm_cntl_finalize() diff --git a/frame/3/trmm3/bli_trmm3_cntl.c b/frame/3/trmm3/bli_trmm3_cntl.c index 5b80c0f3d..17c41f8e5 100644 --- a/frame/3/trmm3/bli_trmm3_cntl.c +++ b/frame/3/trmm3/bli_trmm3_cntl.c @@ -227,8 +227,8 @@ void bli_trmm3_cntl_init() NULL ); // Alias the "master" trmm3 control tree to a shorter name. - trmm3_cntl = trmm3_cntl_mm_op; - + //trmm3_cntl = trmm3_cntl_mm_op; + trmm3_cntl = trmm3_cntl_vl_mm; } void bli_trmm3_cntl_finalize() diff --git a/frame/3/trsm/bli_trsm_cntl.c b/frame/3/trsm/bli_trsm_cntl.c index 0f7fe847c..87e2dca13 100644 --- a/frame/3/trsm/bli_trsm_cntl.c +++ b/frame/3/trsm/bli_trsm_cntl.c @@ -227,8 +227,8 @@ void bli_trsm_cntl_init() NULL ); // Alias the "master" trsm control tree to a shorter name. - trsm_cntl = trsm_cntl_mm_op; - + //trsm_cntl = trsm_cntl_mm_op; + trsm_cntl = trsm_cntl_vl_mm; } void bli_trsm_cntl_finalize() diff --git a/frame/include/bli_param_macro_defs.h b/frame/include/bli_param_macro_defs.h index ab9012d52..3e24a12c5 100644 --- a/frame/include/bli_param_macro_defs.h +++ b/frame/include/bli_param_macro_defs.h @@ -424,10 +424,6 @@ \ ( i1 != iter - 1 || left == 0 ) -#define bli_is_last_iter_f( i1, iter, left ) \ -\ - ( i1 == iter - 1 && left == 0 ) - #define bli_is_edge_b( i1, iter, left ) \ \ ( i1 == 0 && left != 0 ) @@ -436,10 +432,6 @@ \ ( i1 != 0 || left == 0 ) -#define bli_is_last_iter_b( i1, iter, left ) \ -\ - ( i1 == 0 && left == 0 ) - // packbuf_t-related