Skip to content

Commit a0dc47a

Browse files
ggerganovggerganov
committed
metal : print extra compute pipeline info
1 parent b34ab74 commit a0dc47a

File tree

1 file changed

+37
-36
lines changed

1 file changed

+37
-36
lines changed

‎ggml-metal.m

Lines changed: 37 additions & 36 deletions
Original file line numberDiff line numberDiff line change
@@ -167,7 +167,9 @@ @implementation GGMLMetalClass
167167
#define GGML_METAL_ADD_KERNEL(name) \
168168
ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
169169
ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
170-
fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name); \
170+
fprintf(stderr, "%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
171+
(int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
172+
(int) ctx->pipeline_##name.threadExecutionWidth); \
171173
if (error) { \
172174
fprintf(stderr, "%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
173175
return NULL; \
@@ -218,12 +220,12 @@ @implementation GGMLMetalClass
218220
#undef GGML_METAL_ADD_KERNEL
219221
}
220222

221-
fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
222-
fprintf(stderr, "%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
223+
fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
224+
fprintf(stderr, "%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
223225
if (ctx->device.maxTransferRate != 0) {
224-
fprintf(stderr, "%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
226+
fprintf(stderr, "%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
225227
} else {
226-
fprintf(stderr, "%s: maxTransferRate = built-in GPU\n", __func__);
228+
fprintf(stderr, "%s: maxTransferRate = built-in GPU\n", __func__);
227229
}
228230

229231
return ctx;
@@ -744,32 +746,31 @@ void ggml_metal_graph_compute(
744746
[ctx->device supportsFamily:MTLGPUFamilyApple7] &&
745747
ne00%32 == 0 &&
746748
ne11 > 1) {
747-
switch (src0->type) {
748-
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break;
749-
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
750-
case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
751-
case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
752-
case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
753-
case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
754-
case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
755-
case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
756-
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
757-
}
758-
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
759-
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
760-
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
761-
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
762-
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
763-
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5];
764-
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6];
765-
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
766-
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8];
767-
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9];
768-
[encoder setBytes:&gqa length:sizeof(gqa) atIndex:10];
769-
[encoder setThreadgroupMemoryLength:8192 atIndex:0];
770-
[encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
749+
switch (src0->type) {
750+
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break;
751+
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
752+
case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
753+
case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
754+
case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
755+
case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
756+
case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
757+
case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
758+
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
771759
}
772-
else {
760+
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
761+
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
762+
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
763+
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
764+
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
765+
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5];
766+
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6];
767+
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
768+
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8];
769+
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9];
770+
[encoder setBytes:&gqa length:sizeof(gqa) atIndex:10];
771+
[encoder setThreadgroupMemoryLength:8192 atIndex:0];
772+
[encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
773+
} else {
773774
int nth0 = 32;
774775
int nth1 = 1;
775776

@@ -868,24 +869,24 @@ void ggml_metal_graph_compute(
868869
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
869870
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:15];
870871
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:16];
871-
[encoder setBytes:&gqa length:sizeof(gqa) atIndex:17];
872+
[encoder setBytes:&gqa length:sizeof(gqa) atIndex:17];
872873

873874
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
874875
src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
875-
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
876+
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
876877
}
877878
else if (src0t == GGML_TYPE_Q3_K) {
878879
#ifdef GGML_QKK_64
879-
[encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
880+
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
880881
#else
881-
[encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
882+
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
882883
#endif
883884
}
884885
else if (src0t == GGML_TYPE_Q5_K) {
885-
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
886+
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
886887
}
887888
else if (src0t == GGML_TYPE_Q6_K) {
888-
[encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
889+
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
889890
} else {
890891
[encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
891892
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];

0 commit comments

Comments
 (0)