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

metal : add missing barriers for mul-mat (#2699)

Browse source
This commit is contained in:
Shouzheng Liu 2023-08-22 02:18:40 -04:00 committed by GitHub
parent 226255b44e
commit 14b1d7e6f7
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 3 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

5
ggml-metal.metal
View file

@ -1850,6 +1850,7 @@ kernel void kernel_mul_mm(device const uchar * src0,
//load data and store to threadgroup memory
half4x4 temp_a;
dequantize_func(x, il, temp_a);
threadgroup_barrier(mem_flags::mem_threadgroup);
#pragma unroll(16)
for (int i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
@ -1895,14 +1896,14 @@ kernel void kernel_mul_mm(device const uchar * src0,
}
} else {
// block is smaller than 64x32, we should avoid writing data outside of the matrix
threadgroup_barrier(mem_flags::mem_threadgroup);
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_device);
threadgroup_barrier(mem_flags::mem_threadgroup);
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++) {