#version 440 core
layout(std430) buffer;
layout(set=0, binding=0) writeonly uniform image2D uOutput;
layout(set=0, binding=1) uniform sampler2D uInput;
layout(set=0, binding=2) uniform offsetBuffer {
    ivec4 size;//w/4, h/4, hLimit, w/4*h/4
    mat4 transform;
} uOffset;
layout (local_size_x = 256, local_size_y = 1, local_size_z = 1) in;

#define TRANSPOSE_POS(x, y) vec4(x, y, 0, 1) * uOffset.transform;
void main()
{
    int posX = ivec3(gl_GlobalInvocationID).x;

    if (posX < uOffset.size.w)
    {
        ivec2 pos;
        pos.x = posX % uOffset.size.x;// L
        pos.y = posX / uOffset.size.x;// E / H
        ivec2 texPos = ivec2((vec4(pos.x, pos.y, 0, 1) * uOffset.transform).xy);
        vec4 d0 = texelFetch(uInput, texPos + ivec2(0, 0 ), 0);
        vec4 d1 = vec4(0);
        if (texPos.y + 1 < uOffset.size.z) {
            d1 = texelFetch(uInput, texPos + ivec2(0, 1 ), 0);
        }
        vec4 d2 = vec4(0);
        if (texPos.y + 2 < uOffset.size.z) {
            d2 = texelFetch(uInput, texPos + ivec2(0, 2 ), 0);
        }
        vec4 d3 = vec4(0);
        if (texPos.y + 3 < uOffset.size.z) {
            d3 = texelFetch(uInput, texPos + ivec2(0, 3 ), 0);
        }
#ifdef TRANSPOSE
        vec4 c0 = vec4(d0.x, d1.x, d2.x, d3.x);
        vec4 c1 = vec4(d0.y, d1.y, d2.y, d3.y);
        vec4 c2 = vec4(d0.z, d1.z, d2.z, d3.z);
        vec4 c3 = vec4(d0.w, d1.w, d2.w, d3.w);
#else
        vec4 c0 = d0;
        vec4 c1 = d1;
        vec4 c2 = d2;
        vec4 c3 = d3;
#endif
        imageStore(uOutput, ivec2(4*pos.x+0, pos.y), c0);
        imageStore(uOutput, ivec2(4*pos.x+1, pos.y), c1);
        imageStore(uOutput, ivec2(4*pos.x+2, pos.y), c2);
        imageStore(uOutput, ivec2(4*pos.x+3, pos.y), c3);
    }
}