Upstreaming tracker

[mgehre-amd]

Tracks our progress of upstreaming fixes in this repository

Not related to TOSA:

• lib/Dialect/Torch/IR/TorchOps.cpp: Canonicalize rank changing broadcast_to into reshape + broadcast_to #40
• Support prims.sum / torch.cumsum #50
• Support Aten_IndexPutImplOp in 1D #60
• Torch legalization for im2col  #63 + e2e_testing/xfail_sets.py: Cleanup im2col from make_fx_tosa; this already fails in TOSA #129
• Legalize aten.repeat_interleave.Tensor for torch-mlir  #44 + TorchToTosa: Legalize aten.repeat_interleave operation. #96 + Derive output_size of repeat_interleave when inputs are broadcast(fill(x)) #109 + RecomposeComplexOps: Don't call erase after replaceOpWithNewOp #111
• Support aten::fake_quantize_per_tensor_affine #118
• Support aten._index_put_impl_.hacked_twin #128

Purely TorchToTOSA:

• bf16 dfixes -> TorchToTosa: Cast float constants to correct type to support bfloat16 llvm/torch-mlir#2239
• lib/Conversion/TorchToTosa/TosaLegalizeUtils.cpp: Just allow all cast… #28 (aten.to_dtype bf16/f64/i64, see function TorchToTosa)
• Support aten.sign #20 -> Decompose aten.sign and support resulting AtenGeScalarOp in TOSA llvm/torch-mlir#2205
• Don't crash when the input to aten.copy is unranked #19 -> Don't crash when the input to aten.copy is unranked llvm/torch-mlir#2307
• Tosa: Support aten.pow.scalar #17 (The non-TOSA parts already landed upstream)
• Legalize sqrt to tosa #18 -> TorchToTosa: Legalization for torch.aten.sqrt llvm/torch-mlir#2234
• Print name of the backend when tests fail to help debugging issues in CI #23 -> Print name of the backend when tests fail to help debugging issues in CI llvm/torch-mlir#2210
• Adds a folder for torch.aten.broadcast_to operation. #29 -> [MLIR][TOSA] Support unchanged dimensions in torch.aten.broadcast_to operation. llvm/torch-mlir#2204
• lib/Conversion/TorchToTosa/TorchToTosa.cpp: Fix legalization of compa… #36 -> TOSA: Fix legalization of comparions where the input type is bool llvm/torch-mlir#2304
• Support aten.split_tensor #22 + Some slice fixes #37 + Fix SliceOp::fold to check step (introduced by #37) #87 -> Fixes for split tensor and slice llvm/torch-mlir#2314
• Add support for aten.asin and aten.acos. #38
• [FXML-2172] Torch Div Conversion Fix #45 -> [MLIR][TOSA] Fix aten.div conversion for integer inputs llvm/torch-mlir#2200
• Tosa: Adds support for floating point types in torch.arange operation #47
• Support for tosa.custom_op operations. #58 (Lower aten.atan2 to tosa.custom)
• TorchToTosa: torch.aten.empty.memory_format #55
• TOSA: Support le.scalar & le.tensor #61
• lib/Conversion/TorchToTosa/TosaLegalizeCommon.cpp: Handle type promotion in reduction ops #64
• Adds support for aten.sin and aten.cost through tosa.custom_op. #62
• Torch: Fold RuntimeAssertOp #65 -> Torch: Fold RuntimeAssertOp when condition is true llvm/torch-mlir#2198
• Add make_fx_tosa variant to end2end tests #66 -> Add make_fx_tosa variant to end2end tests llvm/torch-mlir#2240
• Decomposition for torch.aten.index_select #67
• Fix version comparison against stable #68 -> Fix version comparison against stable llvm/torch-mlir#2209
• TorchToTosa: Support for 1D convolutions #72
• TorchToTosa: aten.embedding: Allow indices with any rank #77 -> TorchToTosa: aten.embedding: Allow indices with any rank llvm/torch-mlir#2327
• [TOSA] Support for group convolutions. #78
• [TOSA] Fix result type for Conv1D operations. #79
• [TOSA] Fix type for non-FP32 bias. #80
• [TOSA] Add support for aten.pow.Tensor_Tensor operation. #81
• [TOSA] Add support for bool in aten.empty.memory_format. #82
• TorchToTosa: Support AtenSliceTensorOp for any step #83 + TOSA: slice: Support start < 0, start < end and start + sizeOfDim < 0 #84
• python/torch_mlir/dynamo.py: Decompose index_select #85
• python/torch_mlir/dynamo.py: Decompose torch.aten.linalg_vector_norm. #86
• TOSA: Concat: promote input types #88
• TorchToTosa: support different variations of aten.clamp operation. #89
• TorchToTosa: attempt to fold tosa::add in AtenBroadcastToOp legalization. #94
• TorchToTosa: optimize aten.fill.Scalar legalization. #95
• TorchToTosa: legalize aten.erf operation #97
• TorchToTosa: Recompose index.Tensor #98
• [FXML-2294] Fix Constant Int Assembly Parser #113 -> [Torch] Fix torch.constant.int operation parsing llvm/torch-mlir#3476
• aten.eye #115
• Tosa: Support AtenRsubScalarOp #116
• Tosa: Don't fail on torch.prim.TupleConstruct; show the actual issue instead #124 -> Ignore constants in the legality error llvm/torch-mlir#2328
• torch.aten.max pool2d with indices #125
• test/CAPI/CMakeLists.txt: Depend on FileCheck #126 -> test/CAPI/CMakeLists.txt: Depend on FileCheck llvm/torch-mlir#2329