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

metal : fix synchronization in new matrix multiplication kernel (#2686)

Browse source
This commit is contained in:
Shouzheng Liu 2023-08-21 06:59:29 -04:00 committed by GitHub
parent cb1c0727bd
commit dadbed99e6
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 2 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
ggml-metal.metal
View file

@ -1898,10 +1898,11 @@ kernel void kernel_mul_mm(device const uchar * src0,
threadgroup float *temp_str = ((threadgroup float *)shared_memory) \
+ 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
for (int i = 0; i < 8; i++) {
threadgroup_barrier(mem_flags::mem_device);
simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
}
threadgroup_barrier(mem_flags::mem_threadgroup);
threadgroup_barrier(mem_flags::mem_device);
device float *C = dst + BLOCK_SIZE_M * r0 + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
if (sgitg==0) {
for (int i = 0; i < n_rows; i++) {