Reduce variability between compile outputs and transfertuning requests

[ramonwirsch]

Description

• dot_viz sorts edges deterministically. Makes .dot files more deterministic

• Context can also produce a shared_ptr of the context
• RPCNodeTransform can optionally dump the cutout and response into sub-folders for debugging (default off, using output_dir metadata on sdfg)

• updated llvm frontend to use DoccTargets and construct RpcLoopScheduler fresh when needed
• updated llvm frontend to use the docc context for access to DoccTargets (no it vectorizes again)

• added "applied_to_element_id" field to RPCNodeTransform reporting, so that we can identify which level of a loop nest was actually tuned
  ~ replaced global PassReportCollector with thread-safe management that creates reports per module.

Related Issues

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Refactor / internal change (no functional change)
• Documentation update

Quality Checklist

Please ensure the following are completed before requesting review:

• Runtime-compatible (no externally visible existing method has its parameters changed, was deleted or has its exception behavior changed)
• Compile-time compatible (new arguments & properties added have default values, but users need to recompile against the new version to work)

IR:

• Introduces new types into the SDFG IR
• Changed / added properties on SDFG IR Elements (LibraryNodes, ControlFlowNodes DataFlowNodes etc.)
  • New / modified types have serialization / deserialization support
  • Deserialization is backward compatible (new serialization can deserialize old format and represent it in the new format if applicable)
  • Serialized format is backward compatible (previous deserializer can read new serialized format without incorrectness (just lacking additional guarantees that are optional or did not exist before)
  • Serializer has an option to emit the previous format

Unit Tests

• New unit tests have been added for the changes.
• Existing unit tests pass locally.
• Edge cases and error paths are covered.

Integration Tests

• New integration tests have been added (or existing ones updated).
• Integration tests pass locally.
• Interactions with other components/services are verified.

Documentation

• Public APIs, options, and behavior changes are documented.
• README / docs site / inline comments updated where applicable.
• Changelog / release notes updated (if applicable).

Test Coverage

• Test coverage has not decreased as a result of this change.
• Coverage report reviewed and acceptable.
• Critical / new code paths are covered.

How Has This Been Tested?

[daisytuner]

Daisytuner Report - python_npbench (zinnia)

@@                                            Benchmarks                                            @@
=======================================================================================================
  Benchmark              Speedup (Time) ±%           Energy      Regions   Types (M/F/R/O)   % Peak    
=======================================================================================================
# adi                    
#   numpy                119.00 ms                   11.82 J     —         —                 N/A       
#   docc-sequential      2.27x (52.52 ms)            8.43 J      2         1M 1F             1%        
#   docc-omp             2.06x (57.88 ms)            8.74 J      2         1M 1F             1%        
#   docc-cuda            2.46x (48.34 ms)            4.80 J      6         1M 1F             1%        
# atax                   
#   numpy                46.22 ms                    8.59 J      —         —                 N/A       
#   docc-sequential      2.63x (17.57 ms)            3.74 J      3         1M                1%        
#   docc-omp             2.03x (22.79 ms)            4.91 J      3         1M                0%        
#   docc-cuda            1.52x (30.37 ms)            3.06 J      6         2M                N/A       
# gemm                   
#   numpy                92.87 ms                    18.38 J     —         —                 N/A       
#   docc-sequential      3.67x (25.34 ms)            7.25 J      1         —                 N/A       
#   docc-omp             3.65x (25.42 ms)            7.24 J      1         —                 N/A       
#   docc-cuda            0.12x (752.25 ms)           74.28 J     1         —                 N/A       
# gesummv                
#   numpy                119.33 ms                   20.91 J     —         —                 N/A       
#   docc-sequential      0.68x (174.53 ms)           33.06 J     5         3M                3%        
#   docc-omp             1.73x (69.05 ms)            14.82 J     5         3M                0%        
#   docc-cuda            6.08x (19.61 ms)            2.00 J      10        4M                N/A       
# gemver                 
#   numpy                89.59 ms                    16.20 J     —         —                 N/A       
#   docc-sequential      10.50x (8.53 ms)            1.79 J      7         3M                1%        
#   docc-omp             5.63x (15.93 ms)            3.33 J      7         3M                0%        
#   docc-cuda            7.25x (12.35 ms)            1.29 J      12        6M                N/A       
# k2mm                   
#   numpy                96.57 ms                    19.41 J     —         —                 N/A       
#   docc-sequential      0.71x (135.25 ms)           24.61 J     3         1M                3%        
#   docc-omp             0.71x (135.74 ms)           26.75 J     3         1M                0%        
#   docc-cuda            0.12x (820.35 ms)           81.01 J     7         1M                N/A       
# k3mm                   
#   numpy                74.80 ms                    17.48 J     —         —                 N/A       
#   docc-sequential      0.24x (307.17 ms)           50.54 J     4         1M                4%        
#   docc-omp             0.23x (332.07 ms)           57.47 J     4         1M                0%        
#   docc-cuda            0.06x (1.30 s)              128.42 J    8         1M                N/A       
# mvt                    
#   numpy                50.61 ms                    9.63 J      —         —                 N/A       
#   docc-sequential      2.76x (18.33 ms)            3.83 J      2         —                 N/A       
#   docc-omp             2.75x (18.39 ms)            3.85 J      2         —                 N/A       
#   docc-cuda            1.66x (30.48 ms)            3.07 J      1         1M                N/A       
# symm                   
#   numpy                63.80 ms                    6.34 J      —         —                 N/A       
#   docc-sequential      4.31x (14.81 ms)            3.17 J      3         2M 1F             3%        
#   docc-omp             3.55x (18.00 ms)            4.00 J      3         2M 1F             2%        
# syr2k                  
#   numpy                74.61 ms                    7.44 J      —         —                 N/A       
#   docc-sequential      2.02x (36.91 ms)            7.46 J      1         1F                0%        
#   docc-omp             2.07x (36.05 ms)            7.17 J      1         1F                0%        
#   docc-cuda            2.52x (29.60 ms)            2.97 J      1         1F                1%        
# syrk                   
#   numpy                62.94 ms                    6.25 J      —         —                 N/A       
#   docc-sequential      2.40x (26.17 ms)            5.31 J      1         1F                1%        
#   docc-omp             2.36x (26.62 ms)            5.33 J      1         1F                1%        
#   docc-cuda            2.78x (22.65 ms)            2.29 J      1         1F                1%        
# trmm                   
#   numpy                70.98 ms                    7.13 J      —         —                 N/A       
#   docc-sequential      4.93x (14.39 ms)            3.02 J      3         2M 1F             14%       
#   docc-omp             3.89x (18.23 ms)            3.90 J      3         2M 1F             7%        

[daisytuner]

Daisytuner Report - mlir_torch_models (chamomile)

@@                                            Benchmarks                                            @@
=======================================================================================================
  Benchmark              Speedup (Time) ±%           Energy      Regions   Types (M/F/R/O)   % Peak    
=======================================================================================================
# resnet18               
#   torch                90.91 ms                    23.42 J     —         —                 N/A       
#   torch-cuda           18.29 ms                    3.68 J      —         —                 N/A       
#   docc-omp             0.14x (649.53 ms)           207.84 J    45        44M               24%       
#   docc-cuda            0.36x (51.26 ms)            9.74 J      101       72M               N/A       
# segformer              
#   torch-cuda           38.43 ms                    7.56 J      —         —                 N/A       
#   docc-cuda            0.28x (139.06 ms)           26.63 J     435       337M              N/A       

[Moehre2]

Please fix and/or adapt the verifiers of llvm/integration/polybench_sequential_test.py

[NoraHagmeyer]

Needs a rebase

[Atrisan]

Verifier update for polybench sequential is missing

[Atrisan]

Why not delete the file if its basically empty now