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

vulkan: Optimize binary ops (#10270)

Browse source 
Reuse the index calculations across all of src0/src1/dst. Add a shader
variant for when src0/src1 are the same dimensions and additional modulus
for src1 aren't needed. Div/mod are slow, so add "fast" div/mod that
have a fast path when the calculation isn't needed or can be done more
cheaply.
This commit is contained in:
Jeff Bolz 2024-11-13 23:22:55 -06:00 • committed by GitHub
parent 66798e42fb
commit af148c9386
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
9 changed files with 117 additions and 52 deletions
Download patch file Download diff file Expand all files Collapse all files

25
ggml/src/vulkan-shaders/add.comp
View file

@ -1,14 +1,29 @@
#version 450
#extension GL_EXT_shader_16bit_storage : require
#include "types.comp"
#include "generic_binary_head.comp"
const uint num_threads = 256;
layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
uint idx = get_idx();
if (idx >= p.ne) {
return;
// num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
const uint num_iter = 2;
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
if (idx >= p.ne) {
continue;
}
uint i00, i01, i02, i03;
get_indices(idx, i00, i01, i02, i03);
data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
idx += num_threads;
}
data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[src1_idx(idx)]));
}