Files
gpgpu-two/shaders/A2Task1KernelDecompositionAtomic.comp
2026-01-03 22:55:08 +01:00

57 lines
1.0 KiB
Plaintext

#version 450
/* built in:
in uvec3 gl_NumWorkGroups;
in uvec3 gl_WorkGroupID;
in uvec3 gl_LocalInvocationID;
in uvec3 gl_GlobalInvocationID;
in uint gl_LocalInvocationIndex;
*/
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout(push_constant) uniform PushStruct
{
uint offset;
}
p;
layout(binding = 0) buffer inBuffer
{
uint v[];
};
layout(binding = 1) buffer outBuffer
{
uint g_v[];
};
const uint bufferSize = 256;
shared uint localBuffer;
void main()
{
uint tid = gl_LocalInvocationID.x;
uint gid = gl_WorkGroupID.x;
uint offset = gid * bufferSize;
uint idx1 = offset + tid;
uint idx2 = offset + tid + gl_WorkGroupSize.x;
uint val1 = 0;
uint val2 = 0;
if (idx1 < p.offset)
val1 = v[idx1];
if (idx2 < p.offset)
val2 = v[idx2];
if (tid == 0)
localBuffer = 0;
barrier();
uint partial = val1 + val2;
atomicAdd(localBuffer, partial);
barrier();
if (tid == 0)
g_v[gid] = localBuffer;
}