From cf31fcd02077840dccbe2ec707a688b0004af80e Mon Sep 17 00:00:00 2001
From: Vignesh Balasubramanian <vignbala@amd.com>
Date: Wed, 17 Aug 2022 17:17:43 +0530
Subject: [PATCH] Fine tuned threshold and aocl dynamic for zgemm for skinny
 matrices.

-Updated optimal threads in zgemm sup path for skinny matrices.
-Fine tuned the threshold values for small and sup paths
 to improve overall zgemm.
-Zgemm small is selected for inputs with transb as N.
-Redirection of input among small, sup and native path
 was fine tuned.

AMD-Internal : [CPUPL-1900]

Change-Id: Ide37c8255def770b4b74bc6e7c6edb5ee15d3b1f
---
 config/zen/bli_cntx_init_zen.c |  4 ++--
 frame/base/bli_rntm.c          | 19 +++++++++++++++++--
 frame/compat/bla_gemm_amd.c    |  6 ++++--
 3 files changed, 23 insertions(+), 6 deletions(-)

diff --git a/config/zen/bli_cntx_init_zen.c b/config/zen/bli_cntx_init_zen.c
index 3fea3ea8f..f527fe58d 100644
--- a/config/zen/bli_cntx_init_zen.c
+++ b/config/zen/bli_cntx_init_zen.c
@@ -231,9 +231,9 @@ void bli_cntx_init_zen( cntx_t* cntx )
 
     // Initialize sup thresholds with architecture-appropriate values.
     //                                          s     d     c     z
-    bli_blksz_init_easy( &thresh[ BLIS_MT ],   512,  256,   380,   110 );
+    bli_blksz_init_easy( &thresh[ BLIS_MT ],   512,  256,   380,   128 );
     bli_blksz_init_easy( &thresh[ BLIS_NT ],   512,  256,   256,   128 );
-    bli_blksz_init_easy( &thresh[ BLIS_KT ],   440,  220,   220,   110 );
+    bli_blksz_init_easy( &thresh[ BLIS_KT ],   440,  220,   220,   128 );
 
     // Initialize the context with the sup thresholds.
     bli_cntx_set_l3_sup_thresh
diff --git a/frame/base/bli_rntm.c b/frame/base/bli_rntm.c
index 5efba4f2f..d6712b060 100644
--- a/frame/base/bli_rntm.c
+++ b/frame/base/bli_rntm.c
@@ -624,14 +624,29 @@ void bli_nthreads_optimum(
 		dim_t n = bli_obj_width(c);
 		dim_t k = bli_obj_width_after_trans(a);
 
-		if((m<=128 || n<=128 || k<=128) && ((m+n+k) <= 400) )
+		if((m<=128 || n<=128 || k<=128) && ((m+n+k) <= 400))
 		{
 			n_threads_ideal = 8;
 		}
-		else if((m<=256 || n<=256 || k<=256) && ((m+n+k) <= 800) )
+		else if((m<=256 || n<=256 || k<=256) && ((m+n+k) <= 800))
 		{
 			n_threads_ideal = 16;
 		}
+		if((m<=48) || (n<=48) || (k<=48))
+		{
+			if((m+n+k) <= 840)
+			{
+				n_threads_ideal = 8;
+			}
+			else if((m+n+k) <= 1240)
+			{
+				n_threads_ideal = 16;
+			}
+			else if((m+n+k) <= 1540)
+			{
+				n_threads_ideal = 32;
+			}
+		}
 	}
 	else if( family == BLIS_SYRK && bli_obj_is_double(c))
 	{
diff --git a/frame/compat/bla_gemm_amd.c b/frame/compat/bla_gemm_amd.c
index adc83f073..2a9dcb99d 100644
--- a/frame/compat/bla_gemm_amd.c
+++ b/frame/compat/bla_gemm_amd.c
@@ -762,7 +762,7 @@ void zgemm_
     - For single thread, the API has no constraints before invoking.
     - For multiple threads, the constraint is that m and n should individually be less than 128.
     */
-    if((k0==1) && ((nt==0) || ((nt==1) && (m0 < 128) && (n0 < 128)))
+    if((k0 == 1) && ((nt == 0) || ((nt == 1) && (m0 < 128) && (n0 < 128)))
         && bli_is_notrans(blis_transa)
         && bli_is_notrans(blis_transb))
     {
@@ -853,9 +853,11 @@ void zgemm_
         }
 #endif
     }
+
 #ifdef BLIS_ENABLE_SMALL_MATRIX
 
-    if (((nt == 0) && (m0 <= 40) && (n0 <= 40) && (k0 <= 512)) ||
+    if (((nt == 0) && (((m0 <= 40) && (n0 <= 40)) || 
+         (m0 <= 128) && (n0 <= 128) && bli_is_notrans(blis_transb)) && (k0 <= 512)) ||
         ((nt == 1) && (((m0 <= 32) || (n0 <= 32) || (k0 <= 32)) && ((m0 + n0 + k0) <= 100))))
     {
         err_t status = BLIS_NOT_YET_IMPLEMENTED;