root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-11-14 23:09:53 +00:00
Code Issues Actions 14 Packages Projects Releases Wiki Activity

vulkan : argsort barriers must be under uniform control flow (ggml/951)

Browse Source 
a return before a barrier (that happens only in some threads in
a workgroup) leads to UB.
While the old code actually works on some devices,
it fails on some others (i.e. "smaller" GPUs).

BTW, I think it would be better to set specialization constants
when the graph is built, in that way the local workgroup
could be sized appropriately.
But it would take a lot of work.

Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com>
This commit is contained in:
Salvatore Mesoraca 2024-09-26 08:59:42 +02:00 committed by Georgi Gerganov
parent 6084bfb261
commit 544f409b4b
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 4 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
ggml/src/vulkan-shaders/argsort.comp
View File

@ -29,20 +29,18 @@ void main() {
const int col = int(gl_LocalInvocationID.x); const int col = int(gl_LocalInvocationID.x);
const uint row = gl_WorkGroupID.y; const uint row = gl_WorkGroupID.y;
if (col >= p.ncols_pad) {
return;
}
const uint row_offset = row * p.ncols; const uint row_offset = row * p.ncols;
// initialize indices // initialize indices
if (col < p.ncols_pad) {
dst_row[col] = col; dst_row[col] = col;
}
barrier(); barrier();
for (uint k = 2; k <= p.ncols_pad; k *= 2) { for (uint k = 2; k <= p.ncols_pad; k *= 2) {
for (uint j = k / 2; j > 0; j /= 2) { for (uint j = k / 2; j > 0; j /= 2) {
const uint ixj = col ^ j; const uint ixj = col ^ j;
if (ixj > col) { if (col < p.ncols_pad && ixj > col) {
if ((col & k) == 0) { if ((col & k) == 0) {
if (dst_row[col] >= p.ncols || if (dst_row[col] >= p.ncols ||
(dst_row[ixj] < p.ncols && (p.order == ASC ? (dst_row[ixj] < p.ncols && (p.order == ASC ?