wehub-resource-sync
54 lines
1.4 KiB
Plaintext
54 lines
1.4 KiB
Plaintext
#version 440 core
|
|
layout(std430) buffer;
|
|
|
|
layout(set=0, binding=0) writeonly buffer pointoffsetSum{
|
|
highp uint data[];
|
|
} uPointoffsetSum;
|
|
|
|
|
|
layout(set=0, binding=1) readonly buffer pointoffset{
|
|
highp uint data[];
|
|
} uPointoffset;
|
|
|
|
layout(set=0, binding=2) uniform constBuffer {
|
|
ivec4 point; // point size
|
|
} uConstant;
|
|
|
|
shared uint local_sum[256];
|
|
|
|
#define UNIT 3
|
|
#define LOCAL_SIZE 8
|
|
layout (local_size_x = LOCAL_SIZE, local_size_y = 1, local_size_z = 1) in;
|
|
void main()
|
|
{
|
|
int tId = int(gl_LocalInvocationID.x);
|
|
int size = uConstant.point.x;
|
|
int curOffset = 0;
|
|
uint threadBuffer[UNIT];
|
|
while (curOffset < size) {
|
|
int sta = tId * UNIT + curOffset;
|
|
int fin = min(sta + UNIT, size);
|
|
for (int i=sta; i<fin; ++i) {
|
|
int lpos = i - sta;
|
|
threadBuffer[lpos] = uPointoffset.data[i];
|
|
}
|
|
for (int i=sta+1; i<fin; ++i) {
|
|
int lpos = i - sta;
|
|
threadBuffer[lpos] = threadBuffer[lpos] + threadBuffer[lpos-1];
|
|
}
|
|
local_sum[tId] = threadBuffer[fin-sta-1];
|
|
barrier();
|
|
if (fin > sta) {
|
|
uint sum = 0;
|
|
for (int i=0; i<tId; ++i) {
|
|
sum += local_sum[i];
|
|
}
|
|
for (int i=sta; i<fin; ++i) {
|
|
int lpos = i - sta;
|
|
uPointoffsetSum.data[i] = threadBuffer[lpos] + sum;
|
|
}
|
|
}
|
|
curOffset += LOCAL_SIZE * UNIT;
|
|
}
|
|
}
|