vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

metal : slight speed-up for add and mul kernels (#2917)

Browse source
This commit is contained in:
Georgi Gerganov 2023-09-01 13:42:41 +03:00 • committed by GitHub
parent 4dcd47d71d
commit 13268c5331
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 32 additions and 20 deletions
Download patch file Download diff file Expand all files Collapse all files

20
ggml-metal.m
View file

@ -680,6 +680,12 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_ADD:
{
GGML_ASSERT(ggml_is_contiguous(src0));
// utilize float4
GGML_ASSERT(ne00 % 4 == 0);
const int64_t nb = ne00/4;
if (ggml_nelements(src1) == ne10) {
// src1 is a row
[encoder setComputePipelineState:ctx->pipeline_add_row];
@ -689,14 +695,20 @@ void ggml_metal_graph_compute(
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&nb length:sizeof(nb) atIndex:3];
const int64_t n = ggml_nelements(dst);
const int64_t n = ggml_nelements(dst)/4;
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_MUL:
{
GGML_ASSERT(ggml_is_contiguous(src0));
// utilize float4
GGML_ASSERT(ne00 % 4 == 0);
const int64_t nb = ne00/4;
if (ggml_nelements(src1) == ne10) {
// src1 is a row
[encoder setComputePipelineState:ctx->pipeline_mul_row];
@ -706,9 +718,9 @@ void ggml_metal_graph_compute(
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&nb length:sizeof(nb) atIndex:3];
const int64_t n = ggml_nelements(dst);
const int64_t n = ggml_nelements(dst)/4;
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;