Skip to content

Commit 1ad932c

Browse files
committed
opencl: enforce subgroup size for soft_max
1 parent cd57f51 commit 1ad932c

File tree

1 file changed

+6
-0
lines changed

1 file changed

+6
-0
lines changed

ggml/src/ggml-opencl/kernels/ggml-opencl.cl

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -679,6 +679,9 @@ kernel void kernel_diag_mask_inf_8(
679679
//------------------------------------------------------------------------------
680680
// softmax
681681
//------------------------------------------------------------------------------
682+
#ifdef ADRENO_GPU
683+
REQD_SUBGROUP_SIZE_64
684+
#endif
682685
kernel void kernel_soft_max(
683686
global float * src0,
684687
ulong offset0,
@@ -811,6 +814,9 @@ kernel void kernel_soft_max_4(
811814
}
812815
}
813816

817+
#ifdef ADRENO_GPU
818+
REQD_SUBGROUP_SIZE_64
819+
#endif
814820
kernel void kernel_soft_max_f16(
815821
global float * src0,
816822
ulong offset0,

0 commit comments

Comments
 (0)