use fma instead of dot to fix Nvidia and Apple performance issues

0cc4m · 0cc4m · commit e604b63d1d9b · 2025-09-19T16:01:56.000Z
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
@@ -357,7 +357,7 @@ void main() {
                     [[unroll]] for (uint cc = 0; cc < TN; cc++) {
                         [[unroll]] for (uint cr = 0; cr < TM; cr++) {
                             const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
-                            sums[sums_idx] += dot(ACC_TYPE_VEC2(cache_a[wsir * TM + cr]), ACC_TYPE_VEC2(cache_b[cc]));
+                            sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr].x), ACC_TYPE(cache_b[cc].x), fma(ACC_TYPE(cache_a[wsir * TM + cr].y), ACC_TYPE(cache_b[cc].y), sums[sums_idx]));
                         }
                     }
                 }

Original file line number	Diff line number	Diff line change
`@@ -357,7 +357,7 @@ void main() {`
`357`	`357`	`[[unroll]] for (uint cc = 0; cc < TN; cc++) {`
`358`	`358`	`[[unroll]] for (uint cr = 0; cr < TM; cr++) {`
`359`	`359`	`const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;`
`360`		`- sums[sums_idx] += dot(ACC_TYPE_VEC2(cache_a[wsir * TM + cr]), ACC_TYPE_VEC2(cache_b[cc]));`
	`360`	`+ sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr].x), ACC_TYPE(cache_b[cc].x), fma(ACC_TYPE(cache_a[wsir * TM + cr].y), ACC_TYPE(cache_b[cc].y), sums[sums_idx]));`
`361`	`361`	`}`
`362`	`362`	`}`
`363`	`363`	`}`