Merge pull request #5088 from michalowski-arm/develop

Add thread throttling profile for SGEMV on `NEOVERSEV1`

CONTRIBUTORS.md

@@ -231,7 +231,10 @@ In chronological order:
   * [2024-01-24] Optimize GEMV forwarding on ARM64 systems
 
 * 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  
+  *  [2024-12-13] Optimized swap and rot  Level-1 BLAS routines with ARM SVE
 
 * Annop Wongwathanarat <[email protected]>
-  * [2025-01-10] Add thread throttling profile for SGEMM on NEOVERSEV1
+  * [2025-01-10] Add thread throttling profile for SGEMM on NEOVERSEV1
+
+* Marek Michalowski <https://github.com/michalowski-arm>
+  * [2025-01-21] Add thread throttling profile for SGEMV on `NEOVERSEV1`

interface/gemv.c

@@ -63,6 +63,36 @@ static int (*gemv_thread[])(BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG,  FLOAT
 };
 #endif
 
+#ifdef DYNAMIC_ARCH
+ extern char* gotoblas_corename(void);
+#endif
+
+#if defined(DYNAMIC_ARCH) || defined(NEOVERSEV1)
+static inline int get_gemv_optimal_nthreads_neoversev1(BLASLONG MN, int ncpu) {
+  return
+      MN < 25600L    ? 1
+    : MN < 63001L    ? MIN(ncpu, 4)
+    : MN < 459684L   ? MIN(ncpu, 16)
+    : ncpu;
+}
+#endif
+
+static inline int get_gemv_optimal_nthreads(BLASLONG MN) {
+  int ncpu = num_cpu_avail(3);
+#if defined(NEOVERSEV1) && !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16)
+  return get_gemv_optimal_nthreads_neoversev1(MN, ncpu);
+#elif defined(DYNAMIC_ARCH) && !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16)
+  if (strcmp(gotoblas_corename(), "neoversev1") == 0) {
+    return get_gemv_optimal_nthreads_neoversev1(MN, ncpu);
+  }
+#endif
+
+  if ( MN < 115200L * GEMM_MULTITHREAD_THRESHOLD )
+    return 1;
+  else
+    return num_cpu_avail(2);
+}
+
 #ifndef CBLAS
 
 void NAME(char *TRANS, blasint *M, blasint *N,
@@ -225,11 +255,7 @@ void CNAME(enum CBLAS_ORDER order,
   STACK_ALLOC(buffer_size, FLOAT, buffer);
 
 #ifdef SMP
-
-  if ( 1L * m * n < 115200L * GEMM_MULTITHREAD_THRESHOLD )
-    nthreads = 1;
-  else
-    nthreads = num_cpu_avail(2);
+  nthreads = get_gemv_optimal_nthreads(1L * m * n);
 
   if (nthreads == 1) {
 #endif