0.25.11
Added
-
New prelude function:
manifest
. For doing subtle things to memory. -
The GPU backends now handle up to 20 operators in a single fused
reduction. -
CUDA/HIP terminology for GPU concepts (e.g. "thread block") is now
used in all public interfaces. The OpenCL names are still supported
for backwards compatibility. -
More fusion across array slicing.
Fixed
- Compatibility with CUDA versions prior than 12.