Skip to content

opencl: fix rms_norm_mul #17250

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Nov 16, 2025
+26 −11
Merged

opencl: fix rms_norm_mul #17250

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5bca5cc
opencl: use subgrroup reduce for reduction in rms_norm_mul
lhez Nov 11, 2025
696343e
opencl: add comment about workgroup size
lhez Nov 13, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion ggml/src/ggml-opencl/ggml-opencl.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5682,7 +5682,7 @@ static void ggml_opencl_op_rms_norm_fused(ggml_backend_t backend, ggml_tensor *
CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb2));
CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &nb3));
CL_CHECK(clSetKernelArg(kernel, 23, sizeof(float), &eps));
CL_CHECK(clSetKernelArg(kernel, 24, sizeof(float)*nth/sgs, NULL));
CL_CHECK(clSetKernelArg(kernel, 24, sizeof(float)*sgs, NULL));

backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
}
Expand Down
35 changes: 25 additions & 10 deletions ggml/src/ggml-opencl/kernels/rms_norm.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -134,6 +134,15 @@ kernel void kernel_rms_norm_mul(
src1 = src1 + offset1;
dst = dst + offsetd;

// The size of sum is sizeof(float)*subgroup_size.
// Each subgroup writes its partial sum to this array.
// So the number of subgroups per workgroup for this kernel cannot exceed the subgroup size.
// This is generally true -
// for subgroup size 64, workgroup size should be less than 4096 (the max is usually 1024).
if (get_sub_group_id() == 0) {
sum[get_sub_group_local_id()] = 0.0f;
}

int i03 = get_group_id(2);
int i02 = get_group_id(1);
int i01 = get_group_id(0);
Expand All @@ -148,24 +157,30 @@ kernel void kernel_rms_norm_mul(
sumf += dot(x[i00], x[i00]);
}
sumf = sub_group_reduce_add(sumf);

barrier(CLK_LOCAL_MEM_FENCE);

if (get_sub_group_local_id() == 0) {
sum[get_sub_group_id()] = sumf;
}

barrier(CLK_LOCAL_MEM_FENCE);

for (uint i = get_local_size(0) / get_max_sub_group_size() / 2; i > 0; i /= 2) {
if (get_local_id(0) < i) {
sum[get_local_id(0)] += sum[get_local_id(0) + i];
}
}
if (get_local_id(0) == 0) {
sum[0] /= ne00;
}
//for (uint i = get_local_size(0) / get_max_sub_group_size() / 2; i > 0; i /= 2) {
// if (get_local_id(0) < i) {
// sum[get_local_id(0)] += sum[get_local_id(0) + i];
// }
//}
//if (get_local_id(0) == 0) {
// sum[0] /= ne00;
//}

barrier(CLK_LOCAL_MEM_FENCE);
//barrier(CLK_LOCAL_MEM_FENCE);

sumf = sum[get_sub_group_local_id()];
sumf = sub_group_reduce_add(sumf);

float mean = sum[0];
float mean = sumf / ne00;
float scale = 1.0f/sqrt(mean + eps);

global float4 * y = (global float4 *) (dst + i03*nb3 + i02*nb2 + i01*nb1);
Expand Down
Loading