Structural shape of the generated compute kernel.
Determines thread dispatch math, binding structure, and loop shape.
| Enumerator |
|---|
| Elementwise | f(x[i]) -> y[i]; one thread per element
|
| Reduction | f(x[0..n]) -> scalar; shared-memory tree reduction
|
| Stencil | f(x[i-k..i+k]) -> y[i]; neighbourhood reads, radius in PC
|
| GeometryEmit | Writes into vertex SSBO with atomic counter.
|
| BitonicSort | Bitonic sort network; one thread per element.
|
| Convolve2D | 2D separable or non-separable convolution; kernel weights in SSBO, radius in PC
|
| Scan | Inclusive prefix scan over one InOut SSBO, double-buffered Hillis-Steele in shared memory.
Op selects the combine function.
|