CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-05 09:36:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat: Use a secondary simd_sum instead of a for loop

Browse Source 
Branch: GraniteFourPerf

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
This commit is contained in:
Gabe Goodhart
2025-07-22 09:13:45 -06:00
parent 3866f766fe
commit 641276a816
1 changed files with 10 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
ggml/src/ggml-metal/ggml-metal.metal
View File

@@ -1840,15 +1840,18 @@ kernel void kernel_ssm_scan_f32_group(
// sum of the individual simd groups. // sum of the individual simd groups.
threadgroup_barrier(mem_flags::mem_threadgroup); threadgroup_barrier(mem_flags::mem_threadgroup);
// Sum the simd buckets => threadgroup sum // For simd group 0 at indices < num simd groups, extract the shared
// simd sum
sumf = 0.0f; sumf = 0.0f;
for (int64_t i0 = 0; i0 < sgptg; ++i0) { if (sgitg == 0) {
sumf += shared[i0]; if (tiisg < sgptg) {
sumf = shared[tiisg];
} }
sumf = simd_sum(sumf);
threadgroup_barrier(mem_flags::mem_threadgroup); if (tiisg == 0) {
y[0] = sumf; y[0] = sumf;
}
}
// recurse // recurse
s0 = s; s0 = s;