// SPDX-License-Identifier: AGPL-3.0-or-later
// deformable_conv2d.wgsl - Deformable Convolution 2D (f64 canonical)
//
// Convolution with learnable offsets for sampling positions
struct Params {
batch_size: u32,
in_channels: u32,
out_channels: u32,
in_height: u32,
in_width: u32,
out_height: u32,
out_width: u32,
kernel_size: u32,
stride: u32,
padding: u32,
dilation: u32,
deform_groups: u32,
}
@group(0) @binding(0) var<storage, read> input: array<f64>;
@group(0) @binding(1) var<storage, read> offset: array<f64>;
@group(0) @binding(2) var<storage, read> weight: array<f64>;
@group(0) @binding(3) var<storage, read> bias: array<f64>;
@group(0) @binding(4) var<storage, read_write> output: array<f64>;
@group(0) @binding(5) var<uniform> params: Params;
fn bilinear_sample(x: f64, y: f64, b: u32, c: u32) -> f64 {
let x0 = floor(x);
let x1 = x0 + 1.0;
let y0 = floor(y);
let y1 = y0 + 1.0;
if (x0 < 0.0 || x1 >= f64(params.in_width) || y0 < 0.0 || y1 >= f64(params.in_height)) {
return 0.0;
}
let ix0 = u32(x0);
let ix1 = u32(x1);
let iy0 = u32(y0);
let iy1 = u32(y1);
let wa = (x1 - x) * (y1 - y);
let wb = (x - x0) * (y1 - y);
let wc = (x1 - x) * (y - y0);
let wd = (x - x0) * (y - y0);
let idx_a = b * params.in_channels * params.in_height * params.in_width +
c * params.in_height * params.in_width +
iy0 * params.in_width + ix0;
let idx_b = b * params.in_channels * params.in_height * params.in_width +
c * params.in_height * params.in_width +
iy0 * params.in_width + ix1;
let idx_c = b * params.in_channels * params.in_height * params.in_width +
c * params.in_height * params.in_width +
iy1 * params.in_width + ix0;
let idx_d = b * params.in_channels * params.in_height * params.in_width +
c * params.in_height * params.in_width +
iy1 * params.in_width + ix1;
return wa * input[idx_a] + wb * input[idx_b] + wc * input[idx_c] + wd * input[idx_d];
}
@compute @workgroup_size(8, 8, 1)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
let b = global_id.z % params.batch_size;
let c_out = global_id.z / params.batch_size;
let oh = global_id.y;
let ow = global_id.x;
if (c_out >= params.out_channels || oh >= params.out_height || ow >= params.out_width) {
return;
}
var sum: f64 = 0.0;
for (var c_in: u32 = 0u; c_in < params.in_channels; c_in = c_in + 1u) {
var k_idx: u32 = 0u;
for (var kh: u32 = 0u; kh < params.kernel_size; kh = kh + 1u) {
for (var kw: u32 = 0u; kw < params.kernel_size; kw = kw + 1u) {
let ih_base = f64(oh * params.stride) - f64(params.padding) + f64(kh * params.dilation);
let iw_base = f64(ow * params.stride) - f64(params.padding) + f64(kw * params.dilation);
let offset_idx = b * 2u * params.kernel_size * params.kernel_size * params.out_height * params.out_width +
(2u * k_idx) * params.out_height * params.out_width +
oh * params.out_width +
ow;
let offset_y = offset[offset_idx];
let offset_x = offset[offset_idx + params.out_height * params.out_width];
let ih = ih_base + offset_y;
let iw = iw_base + offset_x;
let sampled_val = bilinear_sample(iw, ih, b, c_in);
let w_idx = c_out * params.in_channels * params.kernel_size * params.kernel_size +
c_in * params.kernel_size * params.kernel_size +
kh * params.kernel_size +
kw;
sum = sum + sampled_val * weight[w_idx];
k_idx = k_idx + 1u;
}
}
}
sum = sum + bias[c_out];
let out_idx = b * params.out_channels * params.out_height * params.out_width +
c_out * params.out_height * params.out_width +
oh * params.out_width +
ow;
output[out_idx] = sum;
}