CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-21 12:16:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: optimize rms_norm, and allow the work to spread across multiple SMs (#15281)

Browse Source 
* vulkan: optimize rms_norm, and allow the work to spread across multiple SMs

There are really two parts to this change:
(1) Some optimizations similar to what we have in soft_max, to unroll with
different numbers of iterations.
(2) A fusion optimization where we detect add followed by rms_norm, and make
the add shader atomically accumulate the values^2 into memory. Then the
rms_norm shader can just load that sum. This allows the rms_norm to be
parallelized across multiple workgroups, it just becomes a simple per-element
multiply.

The fusion optimization is currently only applied when the rms_norm is on a
single vector. This previously always ran on a single SM. It could apply more
broadly, but when there are other dimensions the work can already spread across
SMs, and there would be some complexity to tracking multiple atomic sums.

* Change add+rms_norm optimization to write out an array of partial sums
rather than using atomic add, to make it deterministic. The rms_norm
shader fetches a subgroup's worth in parallel and uses subgroupAdd to
add them up.

* complete rebase against fused adds - multi_add shader can also compute partial sums

* fix validation errors

* disable add_rms_fusion for Intel due to possible driver bug

* resolve against #15489, sync after clearing partial sums
This commit is contained in:
Jeff Bolz
2025-08-23 13:16:17 -05:00
committed by GitHub
parent b1afcab804
commit 611f419cff
7 changed files with 379 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files

60
ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
View File

@@ -10,9 +10,9 @@ layout (constant_id = 1) const bool do_multiply = false;
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
shared FLOAT_TYPE sum[BLOCK_SIZE];
shared FLOAT_TYPE sumsh[BLOCK_SIZE];
void main() {
void rms_norm(uint num_iters) {
const uint ncols = p.ne00;
const uint nrows = gl_NumWorkGroups.x;
const uint nchannels = gl_NumWorkGroups.y;
@@ -30,38 +30,76 @@ void main() {
uint32_t b_offset = src1_idx(0, row, channel, samp) + get_boffset();
uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
FLOAT_TYPE sum = FLOAT_TYPE(0.0f); // partial sum for thread in warp
[[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;
[[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
FLOAT_TYPE xi = FLOAT_TYPE(0);
if (col < ncols) {
xi = FLOAT_TYPE(data_a[a_offset + col]);
}
sum += xi * xi;
}
sumsh[tid] = sum;
// sum up partial sums and write back result
barrier();
[[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
if (tid < s) {
sum[tid] += sum[tid + s];
sum += sumsh[tid + s];
sumsh[tid] = sum;
}
barrier();
}
sum = sumsh[0];
const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(ncols);
const FLOAT_TYPE mean = sum / FLOAT_TYPE(ncols);
const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
if (do_multiply) {
if (ncols > p.ne10) {
[[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
[[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
if (col >= ncols) {
continue;
}
data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + fastmod(col, p.ne10)]));
}
} else {
[[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
[[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
if (col >= ncols) {
continue;
}
data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
}
}
} else {
[[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
[[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
if (col >= ncols) {
continue;
}
data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
}
}
}
void main() {
// instantiate the rms_norm function for several different
// dimensions, to allow loop unrolling
uint num_blocks = (p.ne00 + BLOCK_SIZE - 1) / BLOCK_SIZE;
if (num_blocks > 32) {
rms_norm(num_blocks);
} else if (num_blocks > 16) {
rms_norm(32);
} else if (num_blocks > 8) {
rms_norm(16);
} else if (num_blocks > 4) {
rms_norm(8);
} else if (num_blocks == 4) {
rms_norm(4);
} else if (num_blocks == 3) {
rms_norm(3);
} else if (num_blocks == 2) {
rms_norm(2);
} else if (num_blocks == 1) {
rms_norm(1);
}
}