Skip to content

[ET-VK] De vectorise all vectors in conv2d pw shader to improve perf. #11189

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 9 commits into from
May 29, 2025
Merged

[ET-VK] De vectorise all vectors in conv2d pw shader to improve perf. #11189

Changes from all commits
Commits
Show all changes
9 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
e275147
[ET-VK] Tuning local workgroup size calculation for conv2d pw to impr…
trivedivivek May 28, 2025
fab78a4
[ET-VK] De vectorise all vectors in conv2d pw shader to improve perf.
trivedivivek May 28, 2025
3a5cf6e
Merge branch 'main' into gh/trivedivivek/96/orig
trivedivivek May 29, 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
21 changes: 9 additions & 12 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 @@ -46,17 +46,14 @@ layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z_id = 2) in;
* size is only 1x1, making it easier to re-use loaded texels from t_kernel.
*/
void main() {
const ivec2 out_limits_scaled = (out_limits.xy + ivec2(TILE_SIZE_X - 1, TILE_SIZE_Y - 1)) / ivec2(TILE_SIZE_X, TILE_SIZE_Y);
const int out_limits_scaled[2] = {out_limits.x + (TILE_SIZE_X - 1) * TILE_SIZE_X, out_limits.y + (TILE_SIZE_Y - 1) * TILE_SIZE_Y};

const uint div_by_x = gl_GlobalInvocationID.x / out_limits_scaled.x;
const ivec3 gpos = ivec3(
gl_GlobalInvocationID.x % out_limits_scaled.x,
div_by_x,
gl_GlobalInvocationID.y);
const int div_by_x = int(gl_GlobalInvocationID.x / out_limits_scaled[0]);
const int out_pos[3] = {int(gl_GlobalInvocationID.x % out_limits_scaled[0]), div_by_x, int(gl_GlobalInvocationID.y)};

// If the top left position is out of bounds, then this invocation will have
// no work to do.
if (gpos.y >= out_limits_scaled.y || gpos.z >= out_limits.z) {
if (out_pos[1] >= out_limits_scaled[1] || out_pos[2] >= out_limits.z) {
return;
}

Expand All @@ -69,8 +66,8 @@ void main() {
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 * 2] = gpos.x * TILE_SIZE_X + x;
pos[i * 2 + 1] = gpos.y * TILE_SIZE_Y + y;
pos[i * 2] = out_pos[0] * TILE_SIZE_X + x;
pos[i * 2 + 1] = out_pos[1] * TILE_SIZE_Y + y;
i++;
}
}
Expand All @@ -88,7 +85,7 @@ void main() {
// Tuple of consecutive 4 elements represents a single output texel.
float sum[TILE_SIZE_X * TILE_SIZE_Y * 4];

const vec4 bias = texelFetch(t_bias, ivec2(gpos.z, 0), 0);
const vec4 bias = texelFetch(t_bias, ivec2(out_pos[2], 0), 0);

// Initialize the output array with the bias value
for (int i = 0; i < TILE_SIZE_X * TILE_SIZE_Y * 4; i += 4) {
Expand All @@ -108,7 +105,7 @@ void main() {

// Load kernel values from texels to array
[[unroll]] for (int i = 0; i < 4; ++i) {
const vec4 k_tex = texelFetch(t_kernel, ivec2(z + i, gpos.z), 0);
const vec4 k_tex = texelFetch(t_kernel, ivec2(z + i, out_pos[2]), 0);
kernel_values[i * 4 + 0] = k_tex.x;
kernel_values[i * 4 + 1] = k_tex.y;
kernel_values[i * 4 + 2] = k_tex.z;
Expand Down Expand Up @@ -167,7 +164,7 @@ void main() {
}

for (int i = 0; i < TILE_SIZE_X * TILE_SIZE_Y; ++i) {
const ivec3 pos_l = ivec3(pos[i * 2], pos[i * 2 + 1], gpos.z);
const ivec3 pos_l = ivec3(pos[i * 2], pos[i * 2 + 1], out_pos[2]);
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));
}
Expand Down
Loading