Raisetolinalg - #412

Draft

arpitj1 wants to merge 305 commits into

llvm:raisetolinalgfrom

arpitj1:raisetolinalg

This pull request is big! We're only showing the most recent 250 commits

Commits on Jun 12, 2025

Working pattern matching and replacement for linalg generics
arpitj1
committed
7f9d00f
Partial changes for different files for kernel and input
arpitj1
committed
d765bb9

Commits on Jun 13, 2025

Crash fix
arpitj1
committed
15ef84e

Commits on Jun 26, 2025

Improved lib
arpitj1
committed
44fed6c
Removing redundant file
arpitj1
committed
4a95c7f
Renamed kernel lib
arpitj1
committed
f1e5f02
Added min_abs_index test
arpitj1
committed
e941c5e

Commits on Jun 27, 2025

Fixed a bunch of bugs in raiseToLinalg while raising polybench
arpitj1
committed
a99fad9

Commits on Jun 28, 2025

Fixed raise to linalg and canonicalizer to generate subview
arpitj1
committed
4e782d5

Commits on Jul 31, 2025

Fixed submap simplification, improved raisedToLinalg to work with non constant bounds, added debufferization to work with allocs
arpitj1
committed
bd15b6d

Commits on Aug 1, 2025

Added parallel fission pass
arpitj1
committed
cb34836
Added pattern for parallel to seq for loops
arpitj1
committed
53c5d14
Added raise-to-linalg-pipeline
arpitj1
committed
60b81d2
Added linalgGenericEliminateSubmaps and commented out submapToSubviewOp
arpitj1
committed
7b2f5d9
Canonicalization fix
arpitj1
committed
71e441f
bug fix for non nullptr in submap creation
arpitj1
committed
e421a86
Fix in linalg debufferizer - failure return and only insert memref.copy if current!=totensor
arpitj1
committed
56724a5
improved matcher to create a dependency graph and use it for matching
arpitj1
committed
c3c2700

Commits on Aug 3, 2025

Runtime failure but match happening correctly to kernel dialect
arpitj1
committed
ca12291
Working match for linalg kernel match for gemm
arpitj1
committed
7c204f2
Added debug prints
arpitj1
committed
37dd847

Commits on Aug 4, 2025

Able to raise gemv
arpitj1
committed
7e3f0d0

Commits on Oct 15, 2025

Commits on Oct 17, 2025

Update RemoveIterArgs to work with chain ops before store for affine.for
arpitj1
committed
ed30a14
Added int op support
arpitj1
committed
a816708
remote iter args improved and test added
arpitj1
committed
0edd38e

Commits on Jan 10, 2026

Implemented improvement in linalg debufferize to work through inversesubmap, and a more shophisticated approach taken for tracking current tensors via a tree of regions
arpitj1
committed
3b8c43b

Commits on May 13, 2026

Commits on May 14, 2026

Commits on May 15, 2026

Commits on May 23, 2026

Commits on May 24, 2026

Commits on May 25, 2026

Commits on May 26, 2026

PVA backend: lower kernel.launch to libpva_operator for int8/int16 conv2d + 4 image filters on Jetson Orin

arpitj1
and
claude
committed
a1961ce

Commits on May 29, 2026

Make pipeline paths portable
arpitj1
committed
6363ac5
Add fused im2col GEMM raising path
arpitj1
committed
837dc46

Commits on May 30, 2026

Add tensor LLM lowering paths
arpitj1
committed
c7162d1

Commits on May 31, 2026

Fold guarded stores for tensor softmax
arpitj1
committed
d534c23
Add larger LLM forward benchmark
arpitj1
committed
927275c

Commits on Jun 1, 2026

Add Llama runtime matching and Jetson benchmarks
arpitj1
committed
5de644e
Add Llama extended and stencil benchmark tracking
arpitj1
committed
952731c
Add 5x5 stencil matching coverage
arpitj1
committed
fa802e3

Commits on Jun 2, 2026

Record Jetson 5x5 stencil timings
arpitj1
committed
b2feced
Add generalized ntap stencil matching
arpitj1
committed
6009047
Use tensor ntap stencil matching
arpitj1
committed
f867eb9
Fix cgeist memref and record lowering crashes
arpitj1
committed
1db4fc5
Fix cgeist scalar conditions and pointer subtraction
arpitj1
committed
3d616d0
Add Whisper kernel fixtures and RMSNorm matcher family
arpitj1
committed
7143325

Commits on Jun 5, 2026

Fix proxy app raising failures
arpitj1
committed
984c9bf
Add proxy app kernel extraction probes
arpitj1
committed
e671cd0
Fix proxy kernel linalg raising gaps
arpitj1
committed
dd0abae
Verify proxy kernel correctness progress
arpitj1
committed
e64e381
Add easy proxy pipeline fixtures
arpitj1
committed
9ffc4a0

Commits on Jun 7, 2026

Add backend-capable kernel candidate matching
arpitj1
committed
3c61b86
Add custom CUDA 7-point stencil backend
arpitj1
committed
66b05d1
Add optional CPU BLAS runtime backend
arpitj1
committed
ccaf958

Commits on Jul 21, 2026

Raise nested reductions and lower tensor products to cuTensorNet
arpitj1
committed
ec6e187

Commits on Jul 24, 2026

Expand ATen raising corpus and split IR explorer
arpitj1
committed
69cfbac
Add MFEM raising results to IR explorer
arpitj1
committed
ec6d123
Expand ATen raising and lower MFEM contractions
arpitj1
committed
44eb20a

Commits on Jul 25, 2026

Generalize MFEM tensor contraction matching
arpitj1
committed
049547d

Commits on Jul 31, 2026

Add raised MFEM application benchmarks
arpitj1
committed
d466485
Link MFEM application extracted C in explorer
arpitj1
committed
45b092f

Commits on Aug 8, 2026

Complete ATen raising and CUDA library matching phases
arpitj1
committed
e48140d

Commits on Aug 10, 2026

Generalize device-resident library lowering and publish kernel results
arpitj1
committed
3b102a4

Commits on Aug 11, 2026

Expand ATen CUDA library matching and publish silicon results
arpitj1
committed
a9c7d62

Commits on Aug 14, 2026

Generalize CUDA library lowering and refresh ATen results
arpitj1
committed
c925e6f
Compose tensor contractions into cuTensorNet networks
arpitj1
committed
59f48ff

Commits on Sep 3, 2026

Support device-resident residual GPU pipelines
arpitj1
committed
c953d2d
Record ATen device-resident validation data
arpitj1
committed
86288d5
Regenerate MFEM residual pipeline validation
arpitj1
committed
a26445e
Remove stale RMSNorm lowering expectations
arpitj1
committed
5910546
Test reordered GEMM matching end to end
arpitj1
committed
5313662
Use Egglog for semantic kernel matching
arpitj1
committed
b553110

Commits on Sep 4, 2026

Add complete FP32 GEMM lowering support
arpitj1
committed
9861598
Add executable loop-aware benchmark idiom lowering
arpitj1
committed
9956228

Commits on Sep 5, 2026

Complete cross-compiled Ginsbach idiom validation
arpitj1
committed
2716beb
Use external CUDA libraries for Ginsbach idioms
arpitj1
committed
b19db31
Add Ginsbach audit to IR viewer
arpitj1
committed
4f1b8fb
Add modified-kernel audit page and related-work survey
arpitj1
committed
283bf93
Raise histograms and dense factorizations
arpitj1
committed
fd7d43c
Normalize dynamic loop nests for Linalg raising
arpitj1
committed
c4dda7f
Validate histogram and solver library raising
arpitj1
committed
f313920

Commits on Sep 6, 2026

Restore ATen dot and zero-init SGEMM matches
arpitj1
committed
21c38a0
Evaluate PolyBench Section 4.2 pipelines
arpitj1
committed
89aad9d
Match fixed cuDNN ATen convolution routes
arpitj1
committed
0da3fd6
Raise and validate Ginsbach external library idioms
arpitj1
committed
e77f674
Match fixed cuDNN average pool3d routes
arpitj1
committed
3842c16
Refresh ATen matcher snapshots for viewer
arpitj1
committed
d7adfd0
Match ATen Kronecker products with cuTENSOR
arpitj1
committed
2ac2fb9
Consolidate PolyBench four-runtime evaluation
arpitj1
committed
f85ad6e
Match remaining ATen CUDA library candidates
arpitj1
committed
38cb074
Match next ATen library mapping batch
arpitj1
committed
581e961
Batch tiny BLAS patterns and refresh Ginsbach results
arpitj1
committed
6b2cf6c
Fix and validate PolyBench raised GPU paths
arpitj1
committed
70a4857
Complete next ATen library mapping batch
arpitj1
committed
a62ca08
Match Parboil JDS SpMV through cuSPARSE
arpitj1
committed
b9a73bd
Publish validated ATen raised GPU timings
arpitj1
committed
37b37b0
Complete available PolyBenchGPU baselines
arpitj1
committed
1de7275
Record final PolyBench viewer audit
arpitj1
committed
d012774
Fix expanded ATen library compositions
arpitj1
committed
b28469b
Recover expanded ATen convolution mappings
arpitj1
committed
f114e74
Recover expanded ATen 3D convolution mappings
arpitj1
committed
d2c8208
Recover expanded ATen QKV composition
arpitj1
committed
9220a08
Recover tensorized ATen CSR SpMV mapping
arpitj1
committed
95574e4
Map tensorized ATen CSR SpMM to cuSPARSE
arpitj1
committed
f459568
Map ATen COO sparse-dense products to cuSPARSE
arpitj1
committed
64ca07c
Map ATen block-sparse matvec to cuSPARSE
arpitj1
committed
8202cc3
Map ATen sampled CSR addmm to cuSPARSE
arpitj1
committed
444feb4
Map ATen sparse index conversions to cuSPARSE
arpitj1
committed
0aa83c4
Map ATen quantized column offsets to CUB
arpitj1
committed
0a53555
Map ATen adjacent difference to CUB
arpitj1
committed
9b2339d
Map ATen embedding counts to CUB histogram
arpitj1
committed
8bbfe05
Map aligned ATen bilinear upsampling to cuDNN
arpitj1
committed
4fbc967
Map ATen block layout conversion to cuTENSOR
arpitj1
committed
799ea0f
Map dynamic ATen all-any reduction to CUB
arpitj1
committed
c1b3180
Map ATen row-wise arg reductions to CUB
arpitj1
committed
542e88f
Map ATen nansum composition to CUB
arpitj1
committed
e85bc90
Map ATen sparse norm composition to cuBLAS
arpitj1
committed
b63d3d4
Map ATen joint scaling composition to cuBLAS
arpitj1
committed
c928c18
Correct stale ATen library candidates
arpitj1
committed
ab39955
Correct embedding bag backward library routes
arpitj1
committed
3939e7a
Correct rowwise prune library route
arpitj1
committed
0733ede
Correct weighted bincount library route
arpitj1
committed
28f3afc
Correct indexed max-pool library route
arpitj1
committed
d38a4e3
Correct cumprod backward library route
arpitj1
committed
7b1279e
Refine whole-algorithm library routes
arpitj1
committed
ed6c23b
Refine embedding bag max library route
arpitj1
committed
22d1312
Correct mixed FP32 INT8 matmul route
arpitj1
committed
dbc4ac4
Correct packed INT4 matmul route
arpitj1
committed
a0d1fc0
Lower NPB MG stencils through cuDNN
arpitj1
committed
44960a7
Refresh ATen audit after structured rematch
arpitj1
committed
87ed442
Refresh ATen library match viewer
arpitj1
committed
0a716fa
Update Llama runtime comparison
arpitj1
committed
9051c35

Commits on Sep 7, 2026

Fix Llama submap library lowering on CUDA
arpitj1
committed
8e11321
Capture mixed GPU pipelines in one CUDA graph
arpitj1
committed
af531d2
Update viewer with Llama CUDA graph results
arpitj1
committed
5a5bb5e
Record ATen resident GPU benchmark campaign
arpitj1
committed
ae59a05
Record Llama resident session performance
arpitj1
committed
2912d37
Unify ATen resident benchmark reporting
arpitj1
committed
509b8b7
Remeasure PolyBench raised GPU pipelines
arpitj1
committed
1b25437
Publish direct Llama GPU timing
arpitj1
committed
d398dc6
Coalesce repeated GPU pipeline scopes
arpitj1
committed
c376ba8
Remeasure coalesced PolyBench GPU kernels
arpitj1
committed
70340a7
Document portable C to CPU GPU and PVA flow
arpitj1
committed
65435b8
Propagate zero initialization into BLAS parameters
arpitj1
committed
4c9915f
Fix and publish MFEM full-pipeline validation
arpitj1
committed
ec5d223
Generalize constant propagation for library lowering
arpitj1
committed
3838e51
Update ATen and MFEM evaluation reporting
arpitj1
committed
d9cf726
Label MFEM CPU hardware in results viewer
arpitj1
committed
5cd2887

Commits on Sep 8, 2026

Add PolyBench symmetric and triangular library matches
arpitj1
committed
6514a00
Update Ginsbach reduction audit results
arpitj1
committed
7658421
Compose projected tensor contraction networks
arpitj1
committed
f8186a3
Adjudicate ATen native benchmark comparisons
arpitj1
committed
c58b6f1
Measure persistent MFEM workspace lowering
arpitj1
committed
e4464a6
Fix PolyBench raising and library evaluation gaps
arpitj1
committed
cfb0e97
test: audit MFEM cuTensorNet matcher safety
arpitj1
committed
794dac6
Complete ATen native comparison cohort
arpitj1
committed
90dac2b
Publish correctness-gated Llama evaluation
arpitj1
committed
5464d88
Add MFEM Section 4.2 publication analysis
arpitj1
committed
cd909cd
Add PolyBench Section 4.2 paper analysis
arpitj1
committed
7b6437b
Publish fresh Orin GEMM results
arpitj1
committed
2602974
Validate MFEM derived application pipelines
arpitj1
committed
acfa022
Add Ginsbach Orin exact-region benchmark results
arpitj1
committed
183ff79
Publish FP32 and FP64 GEMM results
arpitj1
committed
3ba6692

Commits on Sep 9, 2026

Publish expanded ATen Section 4.2 results
arpitj1
committed
81dd59d
Rework paper comparison visualizations
arpitj1
committed
4f16405
Fix MFEM 3D integration and publish matcher audit
arpitj1
committed
e581348
Publish logarithmic PolyBench comparison figures
arpitj1
committed
b353847
Restore linear PolyBench GPU figure
arpitj1
committed
9ac6925
Expand ATen vendor-library benchmark coverage
arpitj1
committed
1a70b05
Fix PolyBench selected builds and publish Orin reruns
arpitj1
committed
4bd99e3
Publish MFEM library mapping table
arpitj1
committed
1256b1d
Document Llama raising and match coverage
arpitj1
committed
430fa91
Add equality saturation paper study
arpitj1
committed
c40c9b7
Add complete FP32 PolyBench library raising
arpitj1
committed
f37f06f

Commits on Sep 10, 2026

Publish PolyBench whole-build cost study
arpitj1
committed
2cb29da
Add controlled PolyBench equality saturation study
arpitj1
committed
ca4700b
Add ATen paper analysis figures and coverage tables
arpitj1
committed
39227ea
Add ATen paper analysis figures and coverage tables
arpitj1
committed
ba2d410
Add compile coverage for 22 MFEM PA families
arpitj1
committed
7abfd5e
Add compile coverage for 22 MFEM PA families
arpitj1
committed
cd07d64
Add controlled PolyBench equality saturation study
arpitj1
committed
27840b7
Show both equality saturation studies
arpitj1
committed
fe1c3cf
Add half-native guide to ATen GPU plot
arpitj1
committed
d9eb5aa
Add half-native guide to ATen GPU plot
arpitj1
committed
26fb1d3
Refine ATen non-pointwise performance corpus
arpitj1
committed
590d8be
Refine ATen non-pointwise performance corpus
arpitj1
committed
7352053
Add native ATen CUDA library audit provenance
arpitj1
committed
5415b6a
Add native ATen CUDA library audit provenance
arpitj1
committed
98b3972
Fix strided and transposed GEMM lowering
arpitj1
committed
7d69dcc
Merge raisetolinalg updates
arpitj1
committed
a57a32f
Add structural PVA image matching and silicon validation
arpitj1
committed
ae447b3
Complete typed PVA numerical contract validation
arpitj1
committed
bc39400

Commits on Sep 11, 2026

Merge typed PVA semantic matching
arpitj1
committed
5642a64