EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases 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
9 changed files with 117 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
ggml/src/vulkan-shaders/acc.comp
View File

@@ -3,6 +3,8 @@
#include "types.comp"
#include "generic_binary_head.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = gl_GlobalInvocationID.x;
if (idx >= p.ne) {
@@ -15,10 +17,13 @@ void main() {
const uint oy = (src1_i - (oz * p.nb02)) / p.nb01;
const uint ox = src1_i % p.nb01;
uint i00, i01, i02, i03;
get_indices(idx, i00, i01, i02, i03);
if (ox < p.ne10 && oy < p.ne11 && oz < p.ne12) {
data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
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[ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
} else {
data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]));
data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]));
}
}