Merge pull request #5071 from annop-w/sgemm_throttling

Add thread throttling profile for SGEMM on NEOVERSEV1

CONTRIBUTORS.md

@@ -232,3 +232,6 @@ In chronological order:
 
 * Aniket P. Garade <https://github.com/garadeaniket>   Sushil Pratap Singh <https://github.com/SushilPratap04>  Juliya James <https://github.com/Juliya32> 
   *  [2024-12-13] Optimized swap and rot  Level-1 BLAS routines with ARM SVE  
+
+* Annop Wongwathanarat <annop.wongwathanarat@arm.com>
+  * [2025-01-10] Add thread throttling profile for SGEMM on NEOVERSEV1

interface/gemm.c

@@ -1,5 +1,5 @@
 /*********************************************************************/
-/* Copyright 2024 The OpenBLAS Project                               */
+/* Copyright 2024, 2025 The OpenBLAS Project                         */
 /* Copyright 2009, 2010 The University of Texas at Austin.           */
 /* All rights reserved.                                              */
 /*                                                                   */
@@ -177,6 +177,49 @@ static int init_amxtile_permission() {
 }
 #endif
 
+#ifdef DYNAMIC_ARCH
+extern char* gotoblas_corename(void);
+#endif
+
+#if defined(DYNAMIC_ARCH) || defined(NEOVERSEV1)
+static inline int get_gemm_optimal_nthreads_neoversev1(double MNK, int ncpu) {
+  return
+      MNK < 262144L    ? 1
+    : MNK < 1124864L   ? MIN(ncpu, 6)
+    : MNK < 7880599L   ? MIN(ncpu, 12)
+    : MNK < 17173512L  ? MIN(ncpu, 16)
+    : MNK < 33386248L  ? MIN(ncpu, 20)
+    : MNK < 57066625L  ? MIN(ncpu, 24)
+    : MNK < 91733851L  ? MIN(ncpu, 32)
+    : MNK < 265847707L ? MIN(ncpu, 40)
+    : MNK < 458314011L ? MIN(ncpu, 48)
+    : MNK < 729000000L ? MIN(ncpu, 56)
+    : ncpu;
+}
+#endif
+
+static inline int get_gemm_optimal_nthreads(double MNK) {
+  int ncpu = num_cpu_avail(3);
+#if defined(NEOVERSEV1) && !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16)
+  return get_gemm_optimal_nthreads_neoversev1(MNK, ncpu);
+#elif defined(DYNAMIC_ARCH) && !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16)
+  if (strcmp(gotoblas_corename(), "neoversev1") == 0) {
+    return get_gemm_optimal_nthreads_neoversev1(MNK, ncpu);
+  }
+#endif
+  if ( MNK <= (SMP_THRESHOLD_MIN  * (double) GEMM_MULTITHREAD_THRESHOLD) ) {
+    return 1;
+  }
+  else {
+    if (MNK/ncpu < SMP_THRESHOLD_MIN*(double)GEMM_MULTITHREAD_THRESHOLD) {
+      return MNK/(SMP_THRESHOLD_MIN*(double)GEMM_MULTITHREAD_THRESHOLD);
+    }
+    else {
+      return ncpu;
+    }
+  }
+}
+
 #ifndef CBLAS
 
 void NAME(char *TRANSA, char *TRANSB,
@@ -310,7 +353,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
   FLOAT *beta  = (FLOAT*) vbeta;
   FLOAT *a = (FLOAT*) va;
   FLOAT *b = (FLOAT*) vb;
-  FLOAT *c = (FLOAT*) vc;	   
+  FLOAT *c = (FLOAT*) vc;
 #endif
 
   blas_arg_t args;
@@ -352,7 +395,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
 #if !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16) && defined(USE_SGEMM_KERNEL_DIRECT)
 #ifdef DYNAMIC_ARCH
  if (support_avx512() )
-#endif  
+#endif
   if (beta == 0 && alpha == 1.0 && order == CblasRowMajor && TransA == CblasNoTrans && TransB == CblasNoTrans && SGEMM_DIRECT_PERFORMANT(m,n,k)) {
 	SGEMM_DIRECT(m, n, k, a, lda, b, ldb, c, ldc);
 	return;
@@ -604,13 +647,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
 #endif
 
   MNK = (double) args.m * (double) args.n * (double) args.k;
-  if ( MNK <= (SMP_THRESHOLD_MIN  * (double) GEMM_MULTITHREAD_THRESHOLD)  )
-	args.nthreads = 1;
-  else {
-	args.nthreads = num_cpu_avail(3);
-	if (MNK/args.nthreads < SMP_THRESHOLD_MIN*(double)GEMM_MULTITHREAD_THRESHOLD)
-		args.nthreads = MNK/(SMP_THRESHOLD_MIN*(double)GEMM_MULTITHREAD_THRESHOLD);
-  }
+  args.nthreads = get_gemm_optimal_nthreads(MNK);
 
   args.common = NULL;