Skip to content

[ET-VK] Using shared memory to save position in conv2d dw output op. #7818

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

Open
wants to merge 2 commits into
base: gh/trivedivivek/50/base
Choose a base branch
from
Open

[ET-VK] Using shared memory to save position in conv2d dw output op. #7818

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
83c67ef
[ET-VK] Using shared memory to save position in conv2d dw output op.
trivedivivek Jan 21, 2025
c5e59ac
Update on "[ET-VK] Using shared memory to save position in conv2d dw …
trivedivivek Jan 22, 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
13 changes: 10 additions & 3 deletions backends/vulkan/runtime/graph/ops/glsl/conv2d_dw_output_tile.glsl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,6 +38,10 @@ ${layout_declare_ubo(8, "float", "out_min", "float", "out_max")}

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

// shared memory to hold calculated positions, this would reduce register usage thus improving performance.
// 64 is the number of threads in the local wg
shared ivec3 pos_shared[64];

/*
* Computes a depthwise convolution. Each shader invocation calculates the
* output at a single output location.
Expand All @@ -63,6 +67,8 @@ void main() {
return;
}

pos_shared[gl_LocalInvocationIndex] = pos;

// Compute the index of the top-left element of the overlay region. Negative
// indices indicate that the top-left element is in a region added by padding.
const ivec2 ipos = pos.xy * stride - padding;
Expand Down Expand Up @@ -109,18 +115,19 @@ void main() {
for (int j = 0; j < TILE_SIZE; j++, kx++) {
prev_kernel_line[j] = texelFetch(t_kernel, ivec2(kx, pos.z), 0);
for (int s = 0; s < BATCH_SIZE_X; s++) {
sum[0][s] = fma(in_texels[j + s], prev_kernel_line[j], sum[0][s]);
sum[0][s] = fma(in_texels[j + s], prev_kernel_line[j], sum[0][s]);
}
}
}
}

const ivec3 out_pos = pos_shared[gl_LocalInvocationIndex];
for (int y = 0; y < BATCH_SIZE_Y; y++) {
for (int x = 0; x < BATCH_SIZE_X; x++) {
if (any(greaterThanEqual(ivec3(pos.x + x, pos.y + y, pos.z), out_limits))) {
if (any(greaterThanEqual(ivec3(out_pos.x + x, out_pos.y + y, out_pos.z), out_limits))) {
continue;
}
imageStore(t_out, ivec3(pos.x + x, pos.y + y, pos.z), op(sum[y][x], out_min, out_max));
imageStore(t_out, ivec3(out_pos.x + x, out_pos.y + y, out_pos.z), op(sum[y][x], out_min, out_max));
}
}
}
Loading