From d6fcfe734517a1a53fb0fa38d9a650841c9f09b0 Mon Sep 17 00:00:00 2001
From: Madan mohan Manokar <madanmohan.manokar@amd.com>
Date: Fri, 17 Sep 2021 15:32:47 +0530
Subject: [PATCH] gemmt SUP limitThread count for small sizes

1. Max thread cap added for small dimension based on product(n*k).

AMD-Internal: [CPUPL-1388]

Change-Id: I34412a1374bb58a9c4b3fd8e40949a69006cf057
---
 frame/3/bli_l3_sup.c  |  9 ++++++++-
 frame/base/bli_rntm.c | 16 ++++++++++++++++
 2 files changed, 24 insertions(+), 1 deletion(-)

diff --git a/frame/3/bli_l3_sup.c b/frame/3/bli_l3_sup.c
index 163a828f8..a7d7a7874 100644
--- a/frame/3/bli_l3_sup.c
+++ b/frame/3/bli_l3_sup.c
@@ -158,7 +158,6 @@ printf( "dims: %d %d %d (threshs: %d %d %d)\n",
     AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_2);
 }
 
-
 err_t bli_gemmtsup
      (
        obj_t*  alpha,
@@ -243,6 +242,14 @@ err_t bli_gemmtsup
     if ( rntm == NULL ) { bli_rntm_init_from_global( &rntm_l ); rntm = &rntm_l; }
     else                { rntm_l = *rntm;                       rntm = &rntm_l; }
 
+#ifdef AOCL_DYNAMIC
+	// If dynamic-threading is enabled, calculate optimum number
+	// of threads and update in rntm
+
+    // Limit the number of thread for smaller sizes.
+    bli_nthreads_optimum( a, b, c, BLIS_GEMMT, rntm );
+#endif
+
 #if 0
 const num_t dt = bli_obj_dt( c );
 const dim_t m  = bli_obj_length( c );
diff --git a/frame/base/bli_rntm.c b/frame/base/bli_rntm.c
index ba878ac6d..6a100bbe8 100644
--- a/frame/base/bli_rntm.c
+++ b/frame/base/bli_rntm.c
@@ -605,6 +605,22 @@ void bli_nthreads_optimum(
 	  if(m<=512 && n<=512)
 	    n_threads_ideal = 4;
 	}
+	else if( family == BLIS_GEMMT && bli_obj_is_double(c)  )
+	{
+		dim_t n = bli_obj_length(c);
+		dim_t k = bli_obj_width_after_trans(a);
+		dim_t product = (n*k)>>4; /* product is derived based on n and k */
+		//    Limit the number thread for smaller sizes:
+		if(product <= 346)
+		{
+			n_threads_ideal = 1;
+		}
+		/* finer threshold needs to set for max_thread cap of 2,3,4,5,6..32 */
+		else
+		{
+			n_threads_ideal = n_threads;
+		}
+	}
 
 	dim_t n_threads_opt = bli_min(n_threads, n_threads_ideal);