ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

vulkan: support noncontiguous rms_norm (#13031)

Browse source
This commit is contained in:
Jeff Bolz 2025-04-20 03:50:02 -05:00 • committed by GitHub
parent 4ba9d711ba
commit 66168204be
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 39 additions and 15 deletions
Download patch file Download diff file Expand all files Collapse all files

32
ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
View file

@ -1,6 +1,6 @@
#version 450
#include "generic_head.comp"
#include "generic_unary_head.comp"
#include "types.comp"
#extension GL_EXT_control_flow_attributes : enable
@ -8,19 +8,29 @@
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
shared FLOAT_TYPE sum[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
const uint tid = gl_LocalInvocationID.x;
const uint ncols = p.ne00;
const uint nrows = gl_NumWorkGroups.x;
const uint nchannels = gl_NumWorkGroups.y;
const uint row = gl_WorkGroupID.x;
const uint channel = gl_WorkGroupID.y;
const uint samp = gl_WorkGroupID.z;
const uint tid = gl_LocalInvocationID.x;
const uint stride_row = p.nb01;
const uint stride_channel = p.nb02;
const uint stride_sample = p.nb03;
uint32_t a_offset = samp*stride_sample + channel*stride_channel + row*stride_row + get_aoffset();
uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]);
[[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
const FLOAT_TYPE xi = FLOAT_TYPE(data_a[a_offset + col]);
sum[tid] += xi * xi;
}
@ -33,10 +43,10 @@ void main() {
barrier();
}
const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(p.KX);
const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(ncols);
const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col]));
[[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
}
}