metal : opt mul_mm_id map0

ggerganov · ggerganov · commit b534e39ba863 · 2025-08-24T18:37:54.000+03:00
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
@@ -3894,8 +3894,6 @@ static int ggml_metal_encode_node(
                     }
 
                     {
-                        const int nth = MIN(1024, ne10/4);
-
                         ggml_metal_kargs_mul_mm_id_map0 args = {
                             ne10,
                             ne11, // n_expert_used (bcast)
@@ -3916,7 +3914,7 @@ static int ggml_metal_encode_node(
                         [encoder setBuffer: h_tpe  offset:0            atIndex:2];
                         [encoder setBuffer: h_ids  offset:0            atIndex:3];
 
-                        [encoder dispatchThreadgroups:MTLSizeMake(ne02, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        [encoder dispatchThreadgroups:MTLSizeMake(ne02, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                     }
 
                     {
diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
@@ -7497,18 +7497,14 @@ kernel void kernel_mul_mm_id_map0(
                 continue;
             }
 
-            if (tpitg.x == 0) {
-                ids_i32[ide*args.ne21 + n_all] = i21*args.ne20 + i20;
-            }
+            ids_i32[ide*args.ne21 + n_all] = i21*args.ne20 + i20;
 
             ++n_all;
         }
     }
 
-    if (tpitg.x == 0) {
-        device int32_t * tpe_i32 = (device int32_t *) (htpe);
-        tpe_i32[ide] = n_all;
-    }
+    device int32_t * tpe_i32 = (device int32_t *) (htpe);
+    tpe_i32[ide] = n_all;
 }
 
 typedef decltype(kernel_mul_mm_id_map0<half4>) kernel_mul_mm_id_map0_t;

Original file line number	Diff line number	Diff line change
`@@ -3894,8 +3894,6 @@ static int ggml_metal_encode_node(`
`3894`	`3894`	`}`
`3895`	`3895`
`3896`	`3896`	`{`
`3897`		`- const int nth = MIN(1024, ne10/4);`
`3898`		`-`
`3899`	`3897`	`ggml_metal_kargs_mul_mm_id_map0 args = {`
`3900`	`3898`	`ne10,`
`3901`	`3899`	`ne11, // n_expert_used (bcast)`
`@@ -3916,7 +3914,7 @@ static int ggml_metal_encode_node(`
`3916`	`3914`	`[encoder setBuffer: h_tpe offset:0 atIndex:2];`
`3917`	`3915`	`[encoder setBuffer: h_ids offset:0 atIndex:3];`
`3918`	`3916`
`3919`		`- [encoder dispatchThreadgroups:MTLSizeMake(ne02, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];`
	`3917`	`+ [encoder dispatchThreadgroups:MTLSizeMake(ne02, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];`
`3920`	`3918`	`}`
`3921`	`3919`
`3922`	`3920`	`{`
Original file line number	Diff line number	Diff line change
`@@ -7497,18 +7497,14 @@ kernel void kernel_mul_mm_id_map0(`
`7497`	`7497`	`continue;`
`7498`	`7498`	`}`
`7499`	`7499`
`7500`		`- if (tpitg.x == 0) {`
`7501`		`- ids_i32[ideargs.ne21 + n_all] = i21args.ne20 + i20;`
`7502`		`- }`
	`7500`	`+ ids_i32[ideargs.ne21 + n_all] = i21args.ne20 + i20;`
`7503`	`7501`
`7504`	`7502`	`++n_all;`
`7505`	`7503`	`}`
`7506`	`7504`	`}`
`7507`	`7505`
`7508`		`- if (tpitg.x == 0) {`
`7509`		`- device int32_t * tpe_i32 = (device int32_t *) (htpe);`
`7510`		`- tpe_i32[ide] = n_all;`
`7511`		`- }`
	`7506`	`+ device int32_t * tpe_i32 = (device int32_t *) (htpe);`
	`7507`	`+ tpe_i32[ide] = n_all;`
`7512`	`7508`	`}`
`7513`	`7509`
`7514`	`7510`	`typedef decltype(kernel_mul_mm_id_map0<half4>) kernel_mul_mm_id_map0_t;`