Merge master code till 2016_11_25 to amd-staging

config/haswell/bli_kernel.h

@@ -64,12 +64,14 @@
 #endif
 
 #if 0
-#define BLIS_SGEMM_UKERNEL         bli_sgemm_asm_24x4
-#define BLIS_DEFAULT_MC_S          264
-#define BLIS_DEFAULT_KC_S          128
+#define BLIS_SGEMM_UKERNEL         bli_sgemm_asm_4x24
+#define BLIS_DEFAULT_MC_S          256
+#define BLIS_DEFAULT_KC_S          256
 #define BLIS_DEFAULT_NC_S          4080
-#define BLIS_DEFAULT_MR_S          24
-#define BLIS_DEFAULT_NR_S          4
+#define BLIS_DEFAULT_MR_S          4
+#define BLIS_DEFAULT_NR_S          24
+
+#define BLIS_SGEMM_UKERNEL_PREFERS_CONTIG_ROWS
 #endif
 
 #if 0
@@ -95,12 +97,14 @@
 #endif
 
 #if 0
-#define BLIS_DGEMM_UKERNEL         bli_dgemm_asm_12x4
-#define BLIS_DEFAULT_MC_D          96
-#define BLIS_DEFAULT_KC_D          192
+#define BLIS_DGEMM_UKERNEL         bli_dgemm_asm_4x12
+#define BLIS_DEFAULT_MC_D          152
+#define BLIS_DEFAULT_KC_D          160
 #define BLIS_DEFAULT_NC_D          4080
-#define BLIS_DEFAULT_MR_D          12
-#define BLIS_DEFAULT_NR_D          4
+#define BLIS_DEFAULT_MR_D          4
+#define BLIS_DEFAULT_NR_D          12
+
+#define BLIS_DGEMM_UKERNEL_PREFERS_CONTIG_ROWS
 #endif
 
 #if 0

frame/3/trsm/bli_trsm_front.c

@@ -85,7 +85,7 @@ void bli_trsm_front
 		bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, a_local );
 	}
 
-#if 0
+#if 1
 
 	// If A is being solved against from the right, transpose all operands
 	// so that we can perform the computation as if A were being solved

frame/include/bli_config_macro_defs.h

@@ -101,6 +101,14 @@
 
 // -- MISCELLANEOUS OPTIONS ----------------------------------------------------
 
+// Do NOT require the cross-blocksize constraints. That is, do not enforce
+// MC % NR = 0 and NC % MR = 0 in bli_kernel_macro_defs.h. These are ONLY
+// needed when implementing trsm_r by allowing the right-hand matrix B to
+// be triangular.
+#ifndef BLIS_RELAX_MCNR_NCMR_CONSTRAINTS
+  #define BLIS_RELAX_MCNR_NCMR_CONSTRAINTS
+#endif
+
 // Stay initialized after auto-initialization, unless and until the user
 // explicitly calls bli_finalize().
 #ifdef BLIS_DISABLE_STAY_AUTO_INITIALIZED