[Roadmap] AutoTensorIR

[junrushao]

CPU Rules

• [x] Multi-level-tiling
• [x] Multi-level-tiling-with-fusion
• [x] Factorize-reduction
• [ ] Simplify-compute-with-const-tensor
• [x] Add-cache-write
• [x] Always-inline
• [x] Parallelize outer, vectorize inner

GPU Rules

• [x] Multi-level-tiling
• [x] Multi-level-tiling-with-fusion
• [x] Cross-thread-reduction
• [x] Simplify compute with const tensor
• [ ] Special-compute-location
• [x] Add-cache-write
• [x] Add-cache-read
• [x] Always-inline

Init rules

• [x] Init-parallel
• [x] Init-vectorize
• [x] Init-fill-tile-size
• [x] Change-compute-location
• [x] Init-thread-bind
• [x] Init-unroll

Mutation rules:

• [x] Mutate tile size
• [x] Mutate compute location
• [x] Mutate parallel
• [x] Mutate max unroll factor

Cost model, features, search policy

• [x] Random
• [x] XGBModel
• [x] Feature extraction
• [x] Evolutionary search
• [x] Random search
• [x] Serialization

Misc

• [x] Sampler re-design, reproducibility
• [x] Re-design mutation
• [x] Tiling size factorization
• [x] Evolutionary search API re-design
• [x] Merge builder and runner in AutoTune API
• [x] Analysis module in TIR
• [x] Search with stages
• [x] Evolutionary search mixup, don’t have to sample random
• [x] RPC format
• [x] Organize workloads into workload.py
• [x] Instruction attrs
• [x] Buffer
• [x] AnnotateLoopType => MarkLoop, AnnotateBlockType => MarkBlock, loop_type => mark
• [x] Measure Serialization / Deserialization
• [x] Use TE instead of TIR workloads
• [x] Fix sample_fusible_loops
• [x] FLOPs counter
• [x] Refactor SampleFusibleLoops and bug fix
• [x] PrintAsPythonAPI
• [x] Remove useless TContextInfo in search rules
• [x] Exception handling
• [ ] Rename blocks
• [ ] RenameVars
• [ ] RemoveUnitLoop
• [x] Use meta schedule primitives in postproc

[junrushao]

It's done