metal : opt mul_mm_id map0

ggerganov · ggerganov · commit 8a3c0b3de5bb · 2025-08-26T12:15:05.000+03:00
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
@@ -3924,8 +3924,6 @@ 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)
@@ -3946,7 +3944,7 @@ 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
@@ -7514,18 +7514,14 @@ 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
`@@ -3924,8 +3924,6 @@ static int ggml_metal_encode_node(`
`3924`	`3924`	`}`
`3925`	`3925`
`3926`	`3926`	`{`
`3927`		`- const int nth = MIN(1024, ne10/4);`
`3928`		`-`
`3929`	`3927`	`ggml_metal_kargs_mul_mm_id_map0 args = {`
`3930`	`3928`	`ne10,`
`3931`	`3929`	`ne11, // n_expert_used (bcast)`
`@@ -3946,7 +3944,7 @@ static int ggml_metal_encode_node(`
`3946`	`3944`	`[encoder setBuffer: h_tpe offset:0 atIndex:2];`
`3947`	`3945`	`[encoder setBuffer: h_ids offset:0 atIndex:3];`
`3948`	`3946`
`3949`		`- [encoder dispatchThreadgroups:MTLSizeMake(ne02, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];`
	`3947`	`+ [encoder dispatchThreadgroups:MTLSizeMake(ne02, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];`
`3950`	`3948`	`}`
`3951`	`3949`
`3952`	`3950`	`{`
Original file line number	Diff line number	Diff line change
`@@ -7514,18 +7514,14 @@ kernel void kernel_mul_mm_id_map0(`
`7514`	`7514`	`continue;`
`7515`	`7515`	`}`
`7516`	`7516`
`7517`		`- if (tpitg.x == 0) {`
`7518`		`- ids_i32[ideargs.ne21 + n_all] = i21args.ne20 + i20;`
`7519`		`- }`
	`7517`	`+ ids_i32[ideargs.ne21 + n_all] = i21args.ne20 + i20;`
`7520`	`7518`
`7521`	`7519`	`++n_all;`
`7522`	`7520`	`}`
`7523`	`7521`	`}`
`7524`	`7522`
`7525`		`- if (tpitg.x == 0) {`
`7526`		`- device int32_t * tpe_i32 = (device int32_t *) (htpe);`
`7527`		`- tpe_i32[ide] = n_all;`
`7528`		`- }`
	`7523`	`+ device int32_t * tpe_i32 = (device int32_t *) (htpe);`
	`7524`	`+ tpe_i32[ide] = n_all;`
`7529`	`7525`	`}`
`7530`	`7526`
`7531`	`7527`	`typedef decltype(kernel_mul_mm_id_map0<half4>) kernel_mul_mm_id_map0_t;`