Merge pull request #5141 from michalowski-arm/fork-throttle

Add throttling profile for SGEMM and SGEMV on `NEOVERSEV2`

Author: martin-frbg

CONTRIBUTORS.md

@@ -237,8 +237,10 @@ In chronological order:
   * [2025-01-10] Add thread throttling profile for SGEMM on NEOVERSEV1
   * [2025-01-21] Optimize gemv_t_sve_v1x3 kernel
 
-* Marek Michalowski <https://github.com/michalowski-arm>
+* Marek Michalowski <marek.michalowski@arm.com>
   * [2025-01-21] Add thread throttling profile for SGEMV on `NEOVERSEV1`
+  * [2025-02-18] Add thread throttling profile for SGEMM on `NEOVERSEV2`
+  * [2025-02-19] Add thread throttling profile for SGEMV on `NEOVERSEV2`
 
 * Ye Tao <ye.tao@arm.com>
   * [2025-02-03] Optimize SBGEMM kernel on NEOVERSEV1

interface/gemm.c

@@ -198,14 +198,37 @@ static inline int get_gemm_optimal_nthreads_neoversev1(double MNK, int ncpu) {
 }
 #endif
 
+#if defined(DYNAMIC_ARCH) || defined(NEOVERSEV2)
+static inline int get_gemm_optimal_nthreads_neoversev2(double MNK, int ncpu) {
+  return
+      MNK < 125000L     ? 1
+    : MNK < 1092727L    ? MIN(ncpu, 6)
+    : MNK < 2628072L    ? MIN(ncpu, 8)
+    : MNK < 8000000L    ? MIN(ncpu, 12)
+    : MNK < 20346417L   ? MIN(ncpu, 16)
+    : MNK < 57066625L   ? MIN(ncpu, 24)
+    : MNK < 91125000L   ? MIN(ncpu, 28)
+    : MNK < 238328000L  ? MIN(ncpu, 40)
+    : MNK < 454756609L  ? MIN(ncpu, 48)
+    : MNK < 857375000L  ? MIN(ncpu, 56)
+    : MNK < 1073741824L ? MIN(ncpu, 64)
+    : 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(NEOVERSEV2) && !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16)
+  return get_gemm_optimal_nthreads_neoversev2(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);
   }
+  if (strcmp(gotoblas_corename(), "neoversev2") == 0) {
+    return get_gemm_optimal_nthreads_neoversev2(MNK, ncpu);
+  }
 #endif
   if ( MNK <= (SMP_THRESHOLD_MIN  * (double) GEMM_MULTITHREAD_THRESHOLD) ) {
     return 1;

interface/gemv.c

@@ -77,14 +77,30 @@ static inline int get_gemv_optimal_nthreads_neoversev1(BLASLONG MN, int ncpu) {
 }
 #endif
 
+#if defined(DYNAMIC_ARCH) || defined(NEOVERSEV2)
+static inline int get_gemv_optimal_nthreads_neoversev2(BLASLONG MN, int ncpu) {
+  return
+      MN < 24964L    ? 1
+    : MN < 65536L    ? MIN(ncpu, 8)
+    : MN < 262144L   ? MIN(ncpu, 32)
+    : MN < 1638400L  ? MIN(ncpu, 64)
+    : 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(NEOVERSEV2) && !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16)
+  return get_gemv_optimal_nthreads_neoversev2(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);
   }
+  if (strcmp(gotoblas_corename(), "neoversev2") == 0) {
+    return get_gemv_optimal_nthreads_neoversev2(MN, ncpu);
+  }
 #endif
 
   if ( MN < 115200L * GEMM_MULTITHREAD_THRESHOLD )