apply more optimization

junhee-yoo · junhee-yoo · commit bd86c4c4df44 · 2024-10-23T11:58:58.000+09:00
- original IM2COL kernel + _ext with MIN()

Signed-off-by: Junhee Yoo &lt;junhee.yoo@navercorp.com&gt;
diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m
@@ -2625,7 +2625,7 @@ static void ggml_metal_encode_node(
                     [encoder setBytes:&KW       length:sizeof(int32_t) atIndex:15];
 
                     const int64_t D = N / M + (N % M > 0 ? 1 : 0);
-                    [encoder dispatchThreadgroups:MTLSizeMake(D * CHW, OH, OW) threadsPerThreadgroup:MTLSizeMake(M, 1, 1)];
+                    [encoder dispatchThreadgroups:MTLSizeMake(D * CHW, OH, OW) threadsPerThreadgroup:MTLSizeMake(MIN((uint64_t)N, M), 1, 1)];
                 } else {
                     [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];
                 }

Original file line number	Diff line number	Diff line change
`@@ -2625,7 +2625,7 @@ static void ggml_metal_encode_node(`
`2625`	`2625`	`[encoder setBytes:&KW length:sizeof(int32_t) atIndex:15];`
`2626`	`2626`
`2627`	`2627`	`const int64_t D = N / M + (N % M > 0 ? 1 : 0);`
`2628`		`- [encoder dispatchThreadgroups:MTLSizeMake(D * CHW, OH, OW) threadsPerThreadgroup:MTLSizeMake(M, 1, 1)];`
	`2628`	`+ [encoder dispatchThreadgroups:MTLSizeMake(D * CHW, OH, OW) threadsPerThreadgroup:MTLSizeMake(MIN((uint64_t)N, M), 1, 1)];`
`2629`	`2629`	`} else {`
`2630`	`2630`	`[encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];`
`2631`	`2631`	`}`