| 1 |
#version 310 es |
| 2 |
layout(local_size_x = 64) in; |
| 3 |
|
| 4 |
layout(set = 0, binding = 0, std430) readonly buffer SSBO0 |
| 5 |
{ |
| 6 |
float inputs[]; |
| 7 |
}; |
| 8 |
|
| 9 |
layout(set = 0, binding = 1, std430) writeonly buffer SSBO1 |
| 10 |
{ |
| 11 |
float outputs[]; |
| 12 |
}; |
| 13 |
|
| 14 |
shared float tmp[gl_WorkGroupSize.x]; |
| 15 |
|
| 16 |
void main() |
| 17 |
{ |
| 18 |
uint local = gl_LocalInvocationIndex; |
| 19 |
uint work_group = gl_WorkGroupID.x; |
| 20 |
|
| 21 |
// Does a trivial parallel reduction through shared memory. |
| 22 |
tmp[local] = inputs[work_group * gl_WorkGroupSize.x * 2u + local] + inputs[work_group * gl_WorkGroupSize.x * 2u + local + gl_WorkGroupSize.x]; |
| 23 |
memoryBarrierShared(); |
| 24 |
barrier(); |
| 25 |
|
| 26 |
for (uint limit = 32u; limit > 1u; limit >>= 1u) |
| 27 |
{ |
| 28 |
if (local < limit) |
| 29 |
tmp[local] = tmp[local] + tmp[local + limit]; |
| 30 |
memoryBarrierShared(); |
| 31 |
barrier(); |
| 32 |
} |
| 33 |
|
| 34 |
if (local == 0u) |
| 35 |
outputs[work_group] = tmp[0] + tmp[1]; |
| 36 |
} |