#pragma kernel ImageSample

Texture2D X_tex2D;
SamplerState LinearClampSampler;
RWStructuredBuffer<float> Optr;

uint O_width;
uint O_height;
uint O_channels;
uint X_width;
uint X_height;
float4x4 affineMatrix;

[numthreads(8, 8, 1)]
void ImageSample(uint3 dispatchThreadID : SV_DispatchThreadID)
{
    uint2 O_pos = dispatchThreadID.yx;

    if (O_pos.x >= O_width || O_pos.y >= O_height)
        return;

    float4 uv = mul(affineMatrix, float4(O_pos.x, O_pos.y, 1, 1)) / float4(X_width, X_height, 1, 1);

    bool mask = uv.x >= 0 && uv.x <= 1 && uv.y >= 0 && uv.y <= 1;
    float4 c = mask * X_tex2D.SampleLevel(LinearClampSampler, uv.xy, 0);
    // SRGB conversion
    bool3 maskRGB = c.rgb > 0.0031308f;
    c.rgb = lerp(12.92f * c.rgb, 1.055f * (pow(abs(c.rgb), 0.41666666666f) - 0.055f), maskRGB);

    Optr[(O_pos.y * O_width + O_pos.x) * O_channels + 0] = c.x;
    Optr[(O_pos.y * O_width + O_pos.x) * O_channels + 1] = c.y;
    Optr[(O_pos.y * O_width + O_pos.x) * O_channels + 2] = c.z;
}