[SYCL] Fix DMMV dequantization (#9279)

OuadiElfarouki · web-flow · commit 5910ea942772 · 2024-09-04T16:26:33.000+01:00
Fixed dmmv dequant for ncols== GGML_SYCL_DMMV_X
diff --git a/ggml/src/ggml-sycl/dmmv.cpp b/ggml/src/ggml-sycl/dmmv.cpp
@@ -76,8 +76,8 @@ static void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat *
     }
 
     // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    const int mask_start = ncols > GGML_SYCL_DMMV_X ? WARP_SIZE >> 1 : WARP_SIZE >> 2;
+    for (int mask = mask_start; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }

Original file line number	Diff line number	Diff line change
`@@ -76,8 +76,8 @@ static void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat *`
`76`	`76`	`}`
`77`	`77`
`78`	`78`	`// sum up partial sums and write back result`
`79`		`-#pragma unroll`
`80`		`- for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {`
	`79`	`+ const int mask_start = ncols > GGML_SYCL_DMMV_X ? WARP_SIZE >> 1 : WARP_SIZE >> 2;`
	`80`	`+ for (int mask = mask_start; mask > 0; mask >>= 1) {`
`81`	`81`	`tmp +=`
`82`	`82`	`dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);`
`83`	`83`	`}`