Skip to content

[ET-VK] De vectorise positions in conv2d pw shader to improve perf. #11122

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 7 commits into from
May 28, 2025
Merged

[ET-VK] De vectorise positions in conv2d pw shader to improve perf. #11122

Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
9c19980
[ET-VK] Fully De vectorise conv2d pw shader to improve perf.
trivedivivek May 24, 2025
6e8d8f9
Update on "[ET-VK] Fully De vectorise conv2d pw shader to improve perf."
trivedivivek May 27, 2025
d8413a9
Update on "[ET-VK] Fully De vectorise conv2d pw shader to improve perf."
trivedivivek May 27, 2025
01610d8
Update on "[ET-VK] Fully De vectorise conv2d pw shader to improve perf."
trivedivivek May 27, 2025
fd580b6
Update on "[ET-VK] Fully De vectorise conv2d pw shader to improve perf."
trivedivivek May 28, 2025
b6f02c1
Update on "[ET-VK] Fully De vectorise conv2d pw shader to improve perf."
trivedivivek May 28, 2025
305d686
Update on "[ET-VK] Fully De vectorise conv2d pw shader to improve perf."
trivedivivek May 28, 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
17 changes: 10 additions & 7 deletions backends/vulkan/runtime/graph/ops/glsl/conv2d_pw.glsl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,20 +64,22 @@ void main() {
// +--------+--------+
// | pos[2] | pos[3] |
// +--------+--------+
ivec3 pos[TILE_SIZE_X * TILE_SIZE_Y];
int pos[TILE_SIZE_X * TILE_SIZE_Y * 2];
for (int y = 0, i = 0; y < TILE_SIZE_Y; ++y) {
for (int x = 0; x < TILE_SIZE_X; ++x) {
pos[i] = ivec3(gpos.x * TILE_SIZE_X + x, gpos.y * TILE_SIZE_Y + y, gpos.z);
pos[i * 2] = gpos.x * TILE_SIZE_X + x;
pos[i * 2 + 1] = gpos.y * TILE_SIZE_Y + y;
i++;
}
}

// Compute the index of the input texture that needs to be loaded for each
// output position. Note that negative indices can be produced indicating that
// the top-left element is in a region added by padding.
ivec2 ipos[TILE_SIZE_X * TILE_SIZE_Y];
int ipos[TILE_SIZE_X * TILE_SIZE_Y * 2];
for (int i = 0; i < TILE_SIZE_X * TILE_SIZE_Y; ++i) {
ipos[i] = pos[i].xy * stride - padding;
ipos[i * 2] = pos[i * 2] * stride.x - padding.x;
ipos[i * 2 + 1] = pos[i * 2 + 1] * stride.y - padding.y;
}

// Final output array where each element is a tensor value.
Expand Down Expand Up @@ -112,7 +114,7 @@ void main() {
}

for (int i = 0; i < TILE_SIZE_X * TILE_SIZE_Y; ++i) {
const vec4 in_tex = texelFetch(t_in, ivec3(ipos[i], z4), 0);
const vec4 in_tex = texelFetch(t_in, ivec3(ipos[i * 2], ipos[i * 2 + 1], z4), 0);
// Load the input texel into an array
float tex_values[4];
tex_values[0] = in_tex.x;
Expand Down Expand Up @@ -163,8 +165,9 @@ void main() {
}

for (int i = 0; i < TILE_SIZE_X * TILE_SIZE_Y; ++i) {
if (all(lessThan(pos[i], out_limits.xyz))) {
imageStore(t_out, pos[i], op(vec4(sum[i * 4], sum[i * 4 + 1], sum[i * 4 + 2], sum[i * 4 + 3]), out_min, out_max));
const ivec3 pos_l = ivec3(pos[i * 2], pos[i * 2 + 1], gpos.z);
if (all(lessThan(pos_l, out_limits.xyz))) {
imageStore(t_out, pos_l, op(vec4(sum[i * 4], sum[i * 4 + 1], sum[i * 4 + 2], sum[i * 4 + 3]), out_min, out_max));
}
}
}
Loading