Skip to content

[ET-VK] Minor unroll tuning to improve conv2d pw perf. #11187

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] Minor unroll tuning to improve conv2d pw perf. #11187

Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
cb69a5b
[ET-VK] De vectorise conv2d pw shader to improve perf.
trivedivivek May 28, 2025
83ad3d3
[ET-VK] Remove the use of shared memory in conv2d pw to improve perf.
trivedivivek May 28, 2025
843f7d7
[ET-VK] Tuning conv 2d pw op tile size to improve perf.
trivedivivek May 28, 2025
c4780f3
[ET-VK] Minor tuning for conv2d pw op to improve performance.
trivedivivek May 28, 2025
8b3eba7
[ET-VK] De vectorise positions in conv2d pw shader to improve perf.
trivedivivek May 28, 2025
7348101
[ET-VK] Minor unroll tuning to improve conv2d pw perf.
trivedivivek May 28, 2025
1fcdaeb
Merge branch 'main' into gh/trivedivivek/94/orig
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
4 changes: 3 additions & 1 deletion 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 @@ -38,6 +38,8 @@ layout(push_constant) uniform restrict Block {

layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z_id = 2) in;

#extension GL_EXT_control_flow_attributes : require

/*
* Computes a 2D pointwise convolution of an NxN output tile. Calculating an
* output tile for pointwise convolution is more efficient because the kernel
Expand Down Expand Up @@ -105,7 +107,7 @@ void main() {
float kernel_values[4 * 4]; // 4 channels, 4 elements per channel

// Load kernel values from texels to array
for (int i = 0; i < 4; ++i) {
[[unroll]] for (int i = 0; i < 4; ++i) {
const vec4 k_tex = texelFetch(t_kernel, ivec2(z + i, gpos.z), 0);
kernel_values[i * 4 + 0] = k_tex.x;
kernel_values[i * 4 + 1] = k_tex.y;
Expand Down
Loading