|
FZGPUModules 2.0
GPU-accelerated modular compression pipelines
|
Registry of fused implementations, keyed by the shape of a fusion group. More...
#include "backend/types.h"#include "stage/fusion.h"#include <cstddef>#include <string>#include <vector>Go to the source code of this file.
Classes | |
| struct | fz::FusedSideOutput |
| struct | fz::FusedSideInput |
| struct | fz::FusedRunContext |
| Everything a fused runner needs to compress one group in place. More... | |
| struct | fz::FusedImpl |
| A registered fused implementation and its matcher. More... | |
Namespaces | |
| namespace | fz |
Functions | |
| const FusedImpl * | fz::findFusedImpl (const std::vector< Stage * > &group, bool include_experimental=false) |
Registry of fused implementations, keyed by the shape of a fusion group.
The fusion planner (fusion_planner.h) finds maximal legality domains. The installer queries this registry for each contiguous subspan: "does a registered specialization strategy accept these stage declarations?". An eligible hit lets the DAG executor generate and run a fused kernel in place of the group's staged execute()s; misses and unselected overlaps remain staged. Normal Auto selection considers only auto-enabled strategies: registration is the current evidence gate, not a general predictive per-input profitability test. Compatible device operations can form new generated chains within a strategy without adding a per-chain registry entry.
See docs/codebase_notes.md CN-FUSE-PROOF / CN-FUSE-PLAN.