Expand description
Per-op recorders that implement super::GraphOp.
Each submodule owns the op type for one CUDA library or mechanism. New ops added in later phases drop in as additional submodules here without editing any central enum.
Modulesยง
- cudnn
GraphOpRecordimpls forcrate::kernel::CudnnActorrequests.- cusparse
GraphOpRecordimpls forcrate::kernel::SparseActorrequests.- fft_r2c
- 1-D R2C FFT op for
super::super::GraphOp. - memcpy
- Device-to-device memcpy op for
super::super::GraphOp. - rng_
fill_ uniform - Uniform RNG fill op for
super::super::GraphOp. - sgemm
- SGEMM op for
super::super::GraphOp.