Files
2026-07-13 13:33:03 +08:00

43 lines
1.4 KiB
Plaintext

#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 constBuffer {
ivec4 inputSize;
ivec4 outputSize;
ivec2 pad;
ivec2 kernelSize;
ivec2 stride;
} uConstant;
layout (local_size_x = 8, local_size_y = 8, local_size_z = 1) in;
void main()
{
ivec3 pos = ivec3(gl_GlobalInvocationID);
ivec3 outputSize = uConstant.outputSize.xyz;
ivec2 spos = pos.xy*uConstant.stride-uConstant.pad;
int batch = pos.z / uConstant.outputSize.z;
int depth = pos.z % uConstant.outputSize.z;
if (all(lessThan(ivec4(pos.x, pos.y, depth, batch), uConstant.outputSize)))
{
ivec2 inputSizeXY = uConstant.inputSize.xy;
vec4 color = vec4(-100000.0);
ivec2 sfxy = max(ivec2(0), -spos);
ivec2 efxy = min(uConstant.kernelSize, inputSizeXY-spos);
for (int fy=sfxy.y; fy<efxy.y; ++fy)
{
for (int fx=sfxy.x; fx<efxy.x; ++fx)
{
ivec2 spos_ = spos + ivec2(fx, fy);
color = max(texelFetch(uInput, ivec2(spos_.x + depth * uConstant.inputSize.x, spos_.y + batch * uConstant.inputSize.y), 0), color);
}
}
imageStore(uOutput, ivec2(pos.x + depth * uConstant.outputSize.x, pos.y + batch * uConstant.outputSize.y), color);
}
}