wehub-resource-sync
105 lines
3.8 KiB
Plaintext
105 lines
3.8 KiB
Plaintext
#version 440 core
|
|
layout(std430) buffer;
|
|
layout(std430) uniform;
|
|
layout(set=0, binding=0, rgba8ui) writeonly restrict highp uniform uimage2D uOutput;
|
|
layout(set=0, binding=1) uniform highp usampler2D uInput;
|
|
layout(set=0, binding=2) uniform highp usampler2D uKernel;
|
|
layout(set=0, binding=3) uniform highp isampler2D uBias;
|
|
|
|
layout(set=0, binding=4) readonly restrict uniform constBuffer {
|
|
ivec4 outputSize;
|
|
int kernelZero;
|
|
int inputZero;
|
|
int outputShiftBefore;
|
|
int outputMulti;
|
|
int outputShiftAfter;
|
|
int outputZero;
|
|
int minValue;
|
|
int maxValue;
|
|
int multiLength;
|
|
}uConst;
|
|
|
|
layout (local_size_x = 8, local_size_y = 8, local_size_z = 1) in;
|
|
|
|
void main()
|
|
{
|
|
ivec3 pos = ivec3(gl_GlobalInvocationID);
|
|
if (pos.x < uConst.outputSize.x && pos.y < uConst.outputSize.y)
|
|
{
|
|
int kenerlY = pos.y + pos.z * uConst.outputSize.y;
|
|
int srcY = pos.x + pos.z * uConst.outputSize.x;
|
|
highp ivec4 bias = texelFetch(uBias, ivec2(pos.y, 0), 0);
|
|
highp ivec4 inputZero = ivec4(uConst.inputZero);
|
|
highp ivec4 kernelZero = ivec4(uConst.kernelZero);
|
|
highp ivec4 outputZero = ivec4(uConst.outputZero);
|
|
highp ivec4 outputMin = ivec4(uConst.minValue);
|
|
highp ivec4 outputMax = ivec4(uConst.maxValue);
|
|
highp ivec4 outputMulti = ivec4(uConst.outputMulti);
|
|
highp ivec4 o0 = bias;
|
|
highp ivec4 o1 = bias;
|
|
highp ivec4 o2 = bias;
|
|
highp ivec4 o3 = bias;
|
|
|
|
for (int k=0; k<uConst.multiLength; ++k)
|
|
{
|
|
ivec4 k0 = ivec4(texelFetch(uKernel, ivec2(4*k+0, kenerlY), 0)) - kernelZero;
|
|
ivec4 k1 = ivec4(texelFetch(uKernel, ivec2(4*k+1, kenerlY), 0)) - kernelZero;
|
|
ivec4 k2 = ivec4(texelFetch(uKernel, ivec2(4*k+2, kenerlY), 0)) - kernelZero;
|
|
ivec4 k3 = ivec4(texelFetch(uKernel, ivec2(4*k+3, kenerlY), 0)) - kernelZero;
|
|
|
|
ivec4 s0 = ivec4(texelFetch(uInput, ivec2(4*k+0, srcY), 0)) - inputZero;
|
|
ivec4 s1 = ivec4(texelFetch(uInput, ivec2(4*k+1, srcY), 0)) - inputZero;
|
|
ivec4 s2 = ivec4(texelFetch(uInput, ivec2(4*k+2, srcY), 0)) - inputZero;
|
|
ivec4 s3 = ivec4(texelFetch(uInput, ivec2(4*k+3, srcY), 0)) - inputZero;
|
|
|
|
o0 += k0 * s0.x;
|
|
o0 += k1 * s0.y;
|
|
o0 += k2 * s0.z;
|
|
o0 += k3 * s0.w;
|
|
|
|
o1 += k0 * s1.x;
|
|
o1 += k1 * s1.y;
|
|
o1 += k2 * s1.z;
|
|
o1 += k3 * s1.w;
|
|
|
|
o2 += k0 * s2.x;
|
|
o2 += k1 * s2.y;
|
|
o2 += k2 * s2.z;
|
|
o2 += k3 * s2.w;
|
|
|
|
o3 += k0 * s3.x;
|
|
o3 += k1 * s3.y;
|
|
o3 += k2 * s3.z;
|
|
o3 += k3 * s3.w;
|
|
}
|
|
o0 = o0 << uConst.outputShiftBefore;
|
|
o1 = o1 << uConst.outputShiftBefore;
|
|
o2 = o2 << uConst.outputShiftBefore;
|
|
o3 = o3 << uConst.outputShiftBefore;
|
|
ivec4 t0, t1, t2, t3;
|
|
|
|
imulExtended(o0, outputMulti, o0, t0);
|
|
o0 = o0 >> uConst.outputShiftAfter;
|
|
imulExtended(o1, outputMulti, o1, t1);
|
|
o1 = o1 >> uConst.outputShiftAfter;
|
|
imulExtended(o2, outputMulti, o2, t2);
|
|
o2 = o2 >> uConst.outputShiftAfter;
|
|
imulExtended(o3, outputMulti, o3, t3);
|
|
o3 = o3 >> uConst.outputShiftAfter;
|
|
|
|
o0 += outputZero;
|
|
o1 += outputZero;
|
|
o2 += outputZero;
|
|
o3 += outputZero;
|
|
o0 = clamp(o0, outputMin, outputMax);
|
|
o1 = clamp(o1, outputMin, outputMax);
|
|
o2 = clamp(o2, outputMin, outputMax);
|
|
o3 = clamp(o3, outputMin, outputMax);
|
|
|
|
imageStore(uOutput, ivec2(srcY, 4*pos.y+0), uvec4(o0));
|
|
imageStore(uOutput, ivec2(srcY, 4*pos.y+1), uvec4(o1));
|
|
imageStore(uOutput, ivec2(srcY, 4*pos.y+2), uvec4(o2));
|
|
imageStore(uOutput, ivec2(srcY, 4*pos.y+3), uvec4(o3));
|
|
}
|
|
}
|