aboutsummaryrefslogtreecommitdiff
path: root/ggml-metal.m
diff options
context:
space:
mode:
Diffstat (limited to 'ggml-metal.m')
-rw-r--r--ggml-metal.m15
1 files changed, 9 insertions, 6 deletions
diff --git a/ggml-metal.m b/ggml-metal.m
index 135bda9..2810fa2 100644
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -685,8 +685,8 @@ void ggml_metal_graph_compute(
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
- nth0 = 4;
- nth1 = 16;
+ nth0 = 2;
+ nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_K_f32];
} break;
case GGML_TYPE_Q4_K:
@@ -743,15 +743,18 @@ void ggml_metal_graph_compute(
src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
+ else if (src0t == GGML_TYPE_Q3_K) {
+#ifdef GGML_QKK_64
+ [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+#else
+ [encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+#endif
+ }
else if (src0t == GGML_TYPE_Q5_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q6_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
- }
- else if (src0t == GGML_TYPE_Q3_K) {
- [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
- [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else {
[encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];