|
FZGPUModules 2.0
GPU-accelerated modular compression pipelines
|
#include <nvrtc_warp_fusion.h>
Public Attributes | |
| std::string | predictor_ti |
The warp-chain fingerprint the codegen composes: the predictor device-policy type name (defined in warp_fusion.cuh), any register→register transform policies between predictor and coder, the coder policy (the variable-length sink), and ElemsPerLane (= block_size / 32, a compile-time template arg of the kernels). Defaults spell cuSZp2.
| std::string fz::fused::WarpFusionSpec::predictor_ti |
Thread-independent (TI) variant names, copied straight from the predictor's and coder's own FusedOpDecl::ti_op_name (see fusion.h) — empty means that op has no TI policy. TI is eligible for this chain iff BOTH are non-empty and transforms is empty; see tiSupportedChain() in nvrtc_warp_fusion.cu. This is the stage's own capability declaration, not a launcher-side name whitelist.