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

100 lines
2.1 KiB
Plaintext

#version 450
#define FLOAT float
#define MAX_FLOAT (3.402823466e+38)
layout(std430) buffer;
layout(set=0, binding=0) writeonly buffer destBuffer{
FLOAT data[];
} uOutput;
layout(set=0, binding=1) readonly buffer sourceBuffer0{
FLOAT data[];
} uInput;
layout(set=0, binding=2) uniform constBuffer {
int w;//inside
int h;//axis
int c;//outside
float k;//For mean
}uConst;
shared float sharedValues[128];
FLOAT initValue() {
FLOAT result;
#ifdef VMAX
result = -1 * MAX_FLOAT;
#endif
#ifdef VMIN
result = MAX_FLOAT;
#endif
#ifdef SUM
result = 0.0f;
#endif
#ifdef MEAN
result = 0.0f;
#endif
#ifdef PROD
result = 1.0f;
#endif
return result;
}
FLOAT reduceFunc(FLOAT a, FLOAT b) {
FLOAT result;
#ifdef VMAX
result = max(a, b);
#endif
#ifdef VMIN
result = min(a, b);
#endif
#ifdef SUM
result = a + b;
#endif
#ifdef MEAN
result = a + b;
#endif
#ifdef PROD
result = a * b;
#endif
return result;
}
layout(local_size_x = 128, local_size_y = 1, local_size_z = 1) in;
void main() {
uint axisIndex = gl_GlobalInvocationID.x;
uint outputIndex = gl_WorkGroupID.y;
uint insideIndex = outputIndex % uConst.w;
uint outsideIndex = outputIndex / uConst.w;
uint inputOffsetBase = insideIndex + uConst.w * (uConst.h * outsideIndex);
FLOAT beforeSharedResult = initValue();
for (uint i = axisIndex; i < uConst.h; i += gl_WorkGroupSize.x) {
beforeSharedResult = reduceFunc(beforeSharedResult, uInput.data[inputOffsetBase + i * uConst.w]);
}
sharedValues[axisIndex] = beforeSharedResult;
barrier();
for (uint stride = gl_WorkGroupSize.x >> 1; stride > 0; stride = stride >> 1) {
if (axisIndex < stride) {
sharedValues[axisIndex] = reduceFunc(sharedValues[axisIndex + stride], sharedValues[axisIndex]);
}
barrier();
}
if (axisIndex == 0) {
FLOAT result = sharedValues[0];
#ifdef MEAN
uOutput.data[outputIndex] = result * uConst.k;
#else
uOutput.data[outputIndex] = result;
#endif
}
}