diff options
Diffstat (limited to 'graphics/asymptote/base/shaders/partialsum.glsl')
-rw-r--r-- | graphics/asymptote/base/shaders/partialsum.glsl | 24 |
1 files changed, 9 insertions, 15 deletions
diff --git a/graphics/asymptote/base/shaders/partialsum.glsl b/graphics/asymptote/base/shaders/partialsum.glsl index 0506965c23..4b04a01dae 100644 --- a/graphics/asymptote/base/shaders/partialsum.glsl +++ b/graphics/asymptote/base/shaders/partialsum.glsl @@ -1,7 +1,6 @@ layout(local_size_x=PROCESSORS) in; uniform uint elements; -uniform uint steps; layout(binding=0, std430) buffer sumBuffer { @@ -15,11 +14,6 @@ layout(binding=1, std430) buffer offsetBuffer shared uint sharedData[PROCESSORS]; -uint ceilquotient(uint a, uint b) -{ - return (a+b-1u)/b; -} - void main(void) { uint id=gl_LocalInvocationID.x; @@ -30,25 +24,25 @@ void main(void) uint index=id << 1u; sharedData[index+1u] += sharedData[index]; barrier(); - uint step; - for(step=1u; step < steps-1u; step++) { + for(uint step=1u; step < STEPSM1; step++) { uint mask=(1u << step)-1u; uint index=((id >> step) << (step+1u))+mask; uint windex=index+(id&mask)+1u; sharedData[windex] += sharedData[index]; barrier(); } - uint mask=(1u << step)-1u; - index=((id >> step) << steps)+mask; + uint mask=(1u << STEPSM1)-1u; + index=((id >> STEPSM1) << (STEPSM1+1u))+mask; uint windex=index+(id&mask)+1u; if(windex < PROCESSORS) sharedData[windex] += sharedData[index]; barrier(); - uint m=ceilquotient(elements,PROCESSORS); - - if(id+1u < PROCESSORS) - offset[m*(id+1u)] += sharedData[id]; - else + uint id1=id+1u; + if(id1 < PROCESSORS) { + uint m=elements/PROCESSORS; + uint row=m*id1+min(id1,elements-m*PROCESSORS); + offset[row] += sharedData[id]; + } else sum[0]=sharedData[id]; // Store fragment size in sum[0] } |