MirrorHub/construct
0
0
Fork 0
mirror of https://github.com/matrix-construct/construct synced 2024-12-26 15:33:54 +01:00
Code Issues Releases Wiki Activity

ircd::gpt::gpu: Add global fence between attn and ffnn accumulations (gfx1011).

Browse source
This commit is contained in:
Jason Volk 2022-10-05 17:33:44 +00:00
parent c4cceb425c
commit 33afa8a4fc
1 changed files with 3 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
ircd/gpt_gpu.cl
View file

@ -528,7 +528,9 @@ ircd_gpt_coil(__global struct ircd_gpt_ctrl *const ctrl,
accum[wi].elem[li] = result;
}
// Backend mlp; layer-norm acquires any pending writes, no fence required.
barrier(CLK_GLOBAL_MEM_FENCE);
// Backend mlp
ircd_gpt_ffnn
(
ctrl,