CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

metal : simplify soft max kernel

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2023-12-01 10:31:21 +02:00
parent c4db59230d
commit d9c8fa3bce
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
ggml-metal.metal
View File

@@ -203,9 +203,9 @@ kernel void kernel_soft_max(
device float * pdst = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
// parallel max
float lmax = (tpitg < ne00) ? (psrc0[tpitg]*scale + (pmask ? pmask[tpitg] : 0.0f)) : -INFINITY;
float lmax = -INFINITY;
for (int i00 = tpitg + ntg; i00 < ne00; i00 += ntg) {
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f));
}
@@ -284,9 +284,9 @@ kernel void kernel_soft_max_4(
device float4 * pdst4 = (device float4 *)(dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
// parallel max
float4 lmax4 = tpitg < ne00/4 ? (psrc4[tpitg]*scale + (pmask ? pmask[tpitg] : 0.0f)) : -INFINITY;
float4 lmax4 = -INFINITY;
for (int i00 = tpitg + ntg; i00 < ne00/4; i00 += ntg) {
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f));
}