Segments: "compute" parameter is not checked always
- BiEllpack:
compute
seems to be checked correctly - ChunkedEllpack:
compute
seems to be checked correctly - Ellpack:
compute
is checked only in the general cases (1, 2), but not in the CUDA specializations (3, 4) - SlicedEllpack:
compute
is not checked at all: 1, 2 - CSR:
- Adaptive:
compute
is not checked: 1 - Hybrid:
compute
is checked in the multivector kernel, but not in the hybrid kernel - Light:
compute
is checked in the multivector kernel, but not in the other kernels - Scalar:
compute
seems to be checked correctly - Vector:
compute
is not checked: 1
- Adaptive:
Obviously we don't have any tests for this feature. But do we have some benchmark which proves that this optimization helps in some cases?