#version 450
// One invocation copies one 2x2 NV12 block into optimal R8 and RG8 images. No transfer-source
// usage or TFU operation touches the producer allocation.
layout(local_size_x = 16, local_size_y = 16, local_size_z = 1) in;
layout(set = 0, binding = 0) uniform usamplerBuffer source_buffer;
layout(set = 0, binding = 1, r8) writeonly uniform image2D luma_image;
layout(set = 0, binding = 2, rg8) writeonly uniform image2D chroma_image;
layout(push_constant) uniform Parameters {
uint width;
uint height;
uint y_offset;
uint y_pitch;
uint uv_offset;
uint uv_pitch;
uint color_range;
uint x_chroma_offset;
uint y_chroma_offset;
} params;
uint load_byte(uint offset) {
uint word = texelFetch(source_buffer, int(offset >> 2u)).r;
return (word >> ((offset & 3u) * 8u)) & 255u;
}
uvec2 load_pair(uint offset) {
uint shift = (offset & 3u) * 8u;
uint word = texelFetch(source_buffer, int(offset >> 2u)).r;
if (shift <= 16u) {
return uvec2((word >> shift) & 255u, (word >> (shift + 8u)) & 255u);
}
uint next_word = texelFetch(source_buffer, int((offset >> 2u) + 1u)).r;
return uvec2((word >> 24u) & 255u, next_word & 255u);
}
void main() {
ivec2 block = ivec2(gl_GlobalInvocationID.xy);
ivec2 pixel = block * 2;
if (pixel.x >= int(params.width) || pixel.y >= int(params.height)) {
return;
}
uint y0_offset = params.y_offset + uint(pixel.y) * params.y_pitch + uint(pixel.x);
uvec2 y0 = load_pair(y0_offset);
uvec2 y1 = load_pair(y0_offset + params.y_pitch);
imageStore(luma_image, pixel, vec4(float(y0.x) / 255.0, 0.0, 0.0, 1.0));
imageStore(luma_image, pixel + ivec2(1, 0), vec4(float(y0.y) / 255.0, 0.0, 0.0, 1.0));
imageStore(luma_image, pixel + ivec2(0, 1), vec4(float(y1.x) / 255.0, 0.0, 0.0, 1.0));
imageStore(luma_image, pixel + ivec2(1, 1), vec4(float(y1.y) / 255.0, 0.0, 0.0, 1.0));
uint uv_offset = params.uv_offset + uint(block.y) * params.uv_pitch + uint(block.x) * 2u;
uvec2 uv = load_pair(uv_offset);
imageStore(
chroma_image,
block,
vec4(float(uv.x) / 255.0, float(uv.y) / 255.0, 0.0, 1.0)
);
}