55 lines
2.0 KiB
WebGPU Shading Language
55 lines
2.0 KiB
WebGPU Shading Language
struct Params {
|
|
filterDim : u32;
|
|
blockDim : u32;
|
|
};
|
|
|
|
[[group(0), binding(0)]] var samp : sampler;
|
|
|
|
[[group(0), binding(1)]] var<uniform> params : Params;
|
|
|
|
[[group(1), binding(1)]] var inputTex : texture_2d<f32>;
|
|
|
|
[[group(1), binding(2)]] var outputTex : texture_storage_2d<rgba8unorm, write>;
|
|
|
|
struct Flip {
|
|
value : u32;
|
|
};
|
|
|
|
[[group(1), binding(3)]] var<uniform> flip : Flip;
|
|
|
|
var<workgroup> tile : array<array<vec3<f32>, 256>, 4>;
|
|
|
|
[[stage(compute), workgroup_size(64, 1, 1)]]
|
|
fn main([[builtin(workgroup_id)]] WorkGroupID : vec3<u32>, [[builtin(local_invocation_id)]] LocalInvocationID : vec3<u32>) {
|
|
let filterOffset : u32 = ((params.filterDim - 1u) / 2u);
|
|
let dims : vec2<i32> = textureDimensions(inputTex, 0);
|
|
let baseIndex = (vec2<i32>(((WorkGroupID.xy * vec2<u32>(params.blockDim, 4u)) + (LocalInvocationID.xy * vec2<u32>(4u, 1u)))) - vec2<i32>(i32(filterOffset), 0));
|
|
for(var r : u32 = 0u; (r < 4u); r = (r + 1u)) {
|
|
for(var c : u32 = 0u; (c < 4u); c = (c + 1u)) {
|
|
var loadIndex = (baseIndex + vec2<i32>(i32(c), i32(r)));
|
|
if ((flip.value != 0u)) {
|
|
loadIndex = loadIndex.yx;
|
|
}
|
|
tile[r][((4u * LocalInvocationID.x) + c)] = textureSampleLevel(inputTex, samp, ((vec2<f32>(loadIndex) + vec2<f32>(0.25, 0.25)) / vec2<f32>(dims)), 0.0).rgb;
|
|
}
|
|
}
|
|
workgroupBarrier();
|
|
for(var r : u32 = 0u; (r < 4u); r = (r + 1u)) {
|
|
for(var c : u32 = 0u; (c < 4u); c = (c + 1u)) {
|
|
var writeIndex = (baseIndex + vec2<i32>(i32(c), i32(r)));
|
|
if ((flip.value != 0u)) {
|
|
writeIndex = writeIndex.yx;
|
|
}
|
|
let center : u32 = ((4u * LocalInvocationID.x) + c);
|
|
if ((((center >= filterOffset) && (center < (256u - filterOffset))) && all((writeIndex < dims)))) {
|
|
var acc : vec3<f32> = vec3<f32>(0.0, 0.0, 0.0);
|
|
for(var f : u32 = 0u; (f < params.filterDim); f = (f + 1u)) {
|
|
var i : u32 = ((center + f) - filterOffset);
|
|
acc = (acc + ((1.0 / f32(params.filterDim)) * tile[r][i]));
|
|
}
|
|
textureStore(outputTex, writeIndex, vec4<f32>(acc, 1.0));
|
|
}
|
|
}
|
|
}
|
|
}
|