Merge pull request #4920 from martin-frbg/issue4917

martin-frbg · web-flow · commit bc0691a55602 · 2024-10-07T08:26:03.000+02:00
Fix potential inaccuracy in multithreaded level3 related to SWITCH_RATIO
diff --git a/driver/level3/level3_thread.c b/driver/level3/level3_thread.c
@@ -742,7 +742,7 @@ static int gemm_driver(blas_arg_t *args, BLASLONG *range_m, BLASLONG
     num_parts  = 0;
     while (n > 0){
       width = blas_quickdivide(n + nthreads - num_parts - 1, nthreads - num_parts);
-      if (width < switch_ratio) {
+      if (width < switch_ratio && width > 1) {
         width = switch_ratio;
       }
       width = round_up(n, width, GEMM_PREFERED_SIZE);

Original file line number	Diff line number	Diff line change
`@@ -742,7 +742,7 @@ static int gemm_driver(blas_arg_t args, BLASLONG range_m, BLASLONG`
`742`	`742`	`num_parts = 0;`
`743`	`743`	`while (n > 0){`
`744`	`744`	`width = blas_quickdivide(n + nthreads - num_parts - 1, nthreads - num_parts);`
`745`		`- if (width < switch_ratio) {`
	`745`	`+ if (width < switch_ratio && width > 1) {`
`746`	`746`	`width = switch_ratio;`
`747`	`747`	`}`
`748`	`748`	`width = round_up(n, width, GEMM_PREFERED_SIZE);`