EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

metal : batch rows copy in a single threadgroup (#14384)

Browse Source 
* metal : batch rows copy in a single threadgroup

ggml-ci

* metal : handle some edge cases when threadgroup size is not a power of 2

ggml-ci
This commit is contained in:
Georgi Gerganov
2025-06-26 15:50:15 +03:00
committed by GitHub
parent bf5bcd0b85
commit 5783ae4359
2 changed files with 45 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
ggml/src/ggml-metal/ggml-metal.metal
View File

@@ -4306,11 +4306,16 @@ kernel void kernel_cpy(
device const char * src0,
device char * dst,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
ushort3 tpitg[[thread_position_in_threadgroup]],
ushort3 ntg[[threads_per_threadgroup]]) {
ushort3 tptg[[threads_per_threadgroup]]) {
const int i03 = tgpig[2];
const int i02 = tgpig[1];
const int i01 = tgpig[0];
const int i01 = tgpig[0]*tptg.y + tiitg/tptg.x;
if (i01 >= args.ne01) {
return;
}
const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
@@ -4321,7 +4326,7 @@ kernel void kernel_cpy(
device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
for (int64_t i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
for (int64_t i00 = tiitg%tptg.x; i00 < args.ne00; i00 += tptg.x) {
device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
dst_data[i00] = (T1) src[0];
}