Advance llvm-project and stablehlo. (#2619)

llvm-project: bbd2b08b95fe76bea138c1b03c1cd42ed3ee04df
stablehlo: ab709fe48de88c67717abfbd7ef17425eb95ddaf

These commits were chosen in order to account for an MLIR API break from
llvm/llvm-project@3dbac2c
which required a patch to stablehlo. We integrate a bit beyond that
commit to deal with some revert/reapply cycles in the intervening range
which were discovered in another downstream.

Further, it requires adaptation to the stablehlo API breaks introduced
from openxla/stablehlo#1872 which are along for
the ride.

Since some stablehlo builders were changed to directly take int64_t
array refs, also traced that up some call stacks to eliminate some
signed/unsigned mismatches that result.

Also adds a few TOSA tests to the passing set that seem to work now.

include/torch-mlir/Conversion/TorchToStablehlo/StablehloLegalizeUtils.h

@@ -51,7 +51,7 @@ Value promoteType(PatternRewriter &rewriter, Location loc, Value input,
 Value promoteAndBroadcast(ConversionPatternRewriter &rewriter, Value input,
                           TensorType outType);
 
-SmallVector<size_t> toPositiveDims(ArrayRef<int64_t> dims, int64_t rank);
+SmallVector<int64_t> toPositiveDims(ArrayRef<int64_t> dims, int64_t rank);
 
 // Get the dimension sizes of the input tensor, given the dimension axes
 FailureOr<SmallVector<Value, 4>> getDimSizesOfTensor(PatternRewriter &rewriter,

lib/Conversion/TorchToStablehlo/Basic.cpp

@@ -615,12 +615,8 @@ class ConvertAtenTransposeIntOp
     SmallVector<int64_t> permValues(inputRank);
     std::iota(std::begin(permValues), std::end(permValues), 0);
     std::swap(permValues[dim0], permValues[dim1]);
-    DenseIntElementsAttr permutation = DenseIntElementsAttr::get(
-        RankedTensorType::get({static_cast<long int>(permValues.size())},
-                              rewriter.getI64Type()),
-        permValues);
     rewriter.replaceOpWithNewOp<stablehlo::TransposeOp>(op, outType, self,
-                                                        permutation);
+                                                        permValues);
     return success();
   }
 };
@@ -793,12 +789,8 @@ LogicalResult ConvertAtenOp<AtenPermuteOp>::matchAndRewrite(
       return op.emitError("not all dims are valid");
   }
 
-  DenseIntElementsAttr permutation = DenseIntElementsAttr::get(
-      RankedTensorType::get({static_cast<long int>(permValues.size())},
-                            rewriter.getI64Type()),
-      permValues);
   rewriter.replaceOpWithNewOp<stablehlo::TransposeOp>(op, outType, self,
-                                                      permutation);
+                                                      permValues);
   return success();
 }
 
@@ -1755,8 +1747,7 @@ LogicalResult ConvertAtenOp<AtenFlipOp>::matchAndRewrite(
     }
   }
 
-  rewriter.replaceOpWithNewOp<stablehlo::ReverseOp>(
-      op, outType, self, rewriter.getI64TensorAttr(dims));
+  rewriter.replaceOpWithNewOp<stablehlo::ReverseOp>(op, outType, self, dims);
   return success();
 }
 

lib/Conversion/TorchToStablehlo/Linear.cpp

@@ -62,13 +62,9 @@ Value getPermutedTensor(PatternRewriter &rewriter, Operation *op, Value input,
     newShape.push_back(inpShape[d]);
   }
 
-  auto attrTy = RankedTensorType::get({static_cast<int64_t>(transDims.size())},
-                                      rewriter.getIntegerType(64));
-  auto permuteAttr = DenseIntElementsAttr::get(attrTy, transDims);
-
   auto outTy = RankedTensorType::get(newShape, inputTy.getElementType());
   auto result = rewriter.create<stablehlo::TransposeOp>(op->getLoc(), outTy,
-                                                        input, permuteAttr);
+                                                        input, transDims);
   return result.getResult();
 }
 
@@ -500,8 +496,8 @@ class ConvertAtenConvolutionOp : public ConvertAtenOp<AtenConvolutionOp> {
     for (int64_t i = 0; i <= rank; i++)
       transposeDims[i] = i;
     std::swap(transposeDims[rank - 1], transposeDims[rank - 2]);
-    weight = rewriter.create<stablehlo::TransposeOp>(
-        op->getLoc(), weight, rewriter.getI64TensorAttr(transposeDims));
+    weight = rewriter.create<stablehlo::TransposeOp>(op->getLoc(), weight,
+                                                     transposeDims);
 
     // 3. [H, W, ..., G, OC, IC//G] => [H, W, ..., G*OC, IC//G]
     weightShapeInt.erase(weightShapeInt.end() - 2);
@@ -546,12 +542,10 @@ class ConvertAtenConvolutionOp : public ConvertAtenOp<AtenConvolutionOp> {
     }
     auto transposeTy =
         RankedTensorType::get(transposeShape, weightTy.getElementType());
-    DenseIntElementsAttr permAttr = DenseIntElementsAttr::get(
-        RankedTensorType::get({nDims}, rewriter.getI64Type()), perm);
     auto transposeOp = rewriter.create<stablehlo::TransposeOp>(
-        op->getLoc(), transposeTy, weight, permAttr);
+        op->getLoc(), transposeTy, weight, perm);
     auto reverseOp = rewriter.create<stablehlo::ReverseOp>(
-        op->getLoc(), transposeOp, rewriter.getI64TensorAttr({0, 1}));
+        op->getLoc(), transposeOp, ArrayRef<int64_t>{0, 1});
 
     // Prepare for transposed convolution
     SmallVector<int64_t> stablehloStrideVec(nSpatialDims, 1);

lib/Conversion/TorchToStablehlo/StablehloLegalizeUtils.cpp

@@ -250,12 +250,12 @@ Value promoteAndBroadcast(ConversionPatternRewriter &rewriter, Value input,
   return bcast_op.getResult();
 }
 
-SmallVector<size_t> toPositiveDims(ArrayRef<int64_t> dims, int64_t rank) {
-  SmallVector<size_t> posDims;
+SmallVector<int64_t> toPositiveDims(ArrayRef<int64_t> dims, int64_t rank) {
+  SmallVector<int64_t> posDims;
   posDims.reserve(rank);
   std::transform(
       dims.begin(), dims.end(), std::back_inserter(posDims),
-      [rank](int64_t d) -> size_t { return toPositiveDim(d, rank); });
+      [rank](int64_t d) -> int64_t { return toPositiveDim(d, rank); });
   return posDims;
 }
 
@@ -316,10 +316,10 @@ FailureOr<Value> unsqueezeTensor(PatternRewriter &rewriter, Operation *op,
         op, "failed to get dimension sizes of the input");
 
   auto dimSizes = *dimSizesInfo;
-  auto rank = dimSizes.size();
-  size_t newRank = rank + inputUnsqzDims.size();
+  int64_t rank = dimSizes.size();
+  int64_t newRank = rank + inputUnsqzDims.size();
   auto unsqzDims = toPositiveDims(inputUnsqzDims, newRank);
-  for (size_t k = 0, sz = unsqzDims.size(); k < sz; ++k)
+  for (int64_t k = 0, sz = unsqzDims.size(); k < sz; ++k)
     if (k > 1 && unsqzDims[k] <= unsqzDims[k - 1])
       return rewriter.notifyMatchFailure(
           op, "unsqueeze dimensions must be specified in order");
@@ -335,8 +335,8 @@ FailureOr<Value> unsqueezeTensor(PatternRewriter &rewriter, Operation *op,
   std::vector<int64_t> newShape;
   newDimSizes.reserve(newRank);
   newShape.reserve(newRank);
-  for (size_t k = 0, i = 0, j = 0; k < newRank; ++k) {
-    if (j < unsqzDims.size() && unsqzDims[j] == k) {
+  for (int64_t k = 0, i = 0, j = 0; k < newRank; ++k) {
+    if (j < static_cast<int64_t>(unsqzDims.size()) && unsqzDims[j] == k) {
       newDimSizes.push_back(one);
       newShape.push_back(1);
       j++;

projects/pt1/e2e_testing/xfail_sets.py

@@ -13,6 +13,8 @@
 from torch_mlir_e2e_test.test_suite import COMMON_TORCH_MLIR_LOWERING_XFAILS
 from torch_mlir._version import torch_version_for_comparison, version
 
+print(f"TORCH_VERSION_FOR_COMPARISON =", torch_version_for_comparison())
+
 LINALG_XFAIL_SET = COMMON_TORCH_MLIR_LOWERING_XFAILS | {
     # Lowering Torch Backend IR -> Linalg-on-Tensors Backend IR failed
     # 'linalg.depthwise_conv_2d_nchw_chw' op inferred input/output operand #1 has shape's dimension #0 to be 4, but found 8
@@ -21,6 +23,14 @@
     "IscloseStaticModuleTrue_basic"
 }
 
+if torch_version_for_comparison() >= version.parse("2.2.0.dev20231204"):
+    LINALG_XFAIL_SET |= {
+        "Conv2dWithPaddingDilationStrideStaticModule_grouped",
+        "Conv2dWithPaddingDilationStrideStaticModule_grouped_multiplier",
+        "ConvolutionModule2DGroups_basic",
+    }
+
+
 TORCHDYNAMO_XFAIL_SET = {
     #### General TorchDynamo/PyTorch errors
 
@@ -306,10 +316,11 @@
     "ArangeStartOutViewModule_basic",
 }
 
-if torch_version_for_comparison() < version.parse("2.1.0.dev"):
-    TORCHDYNAMO_XFAIL_SET -= {
-        "ScaledDotProductAttentionSameModule_basic",
-        "ScaledDotProductAttentionDifferentModule_basic",
+if torch_version_for_comparison() >= version.parse("2.2.0.dev20231204"):
+    TORCHDYNAMO_XFAIL_SET |= {
+        "Conv2dWithPaddingDilationStrideStaticModule_grouped",
+        "Conv2dWithPaddingDilationStrideStaticModule_grouped_multiplier",
+        "ConvolutionModule2DGroups_basic",
     }
 
 TORCHDYNAMO_CRASHING_SET = {
@@ -1305,6 +1316,10 @@
     "MeanModule_basic",
     "ArangeStartOutModule_basic",
     "ArangeStartOutViewModule_basic",
+    "Conv2dBiasNoPaddingModule_basic",
+    "Conv2dNoPaddingModule_basic",
+    "Conv2dWithPaddingDilationStrideModule_basic",
+    "Conv2dWithPaddingModule_basic",
 }
 
 MAKE_FX_TOSA_PASS_SET = (TOSA_PASS_SET | {
@@ -1335,20 +1350,12 @@
     # failed to legalize operation 'torch.aten.to.dtype' that was explicitly marked illegal
     "AtenEyeModuleInt2D_basic",
     "AtenEyeMModuleInt2D_basic",
-}
 
-if torch_version_for_comparison() < version.parse("2.1.0.dev"):
-    MAKE_FX_TOSA_PASS_SET -= {
-        # 'tensor.expand_shape' op expected rank expansion, but found source rank 1 >= result rank 1
-        "ReshapeCollapseModule_basic",
-
-        # failed to lower torch.aten.empty.memory_format
-        "BatchNorm1DModule_basic",
-        "BatchNorm1DWith2DInputModule_basic",
-        "BatchNorm2DModule_basic",
-        "BatchNorm3DModule_basic",
-        "BatchNorm1DStaticShapeModule_basic",
-    }
+    "Conv2dBiasNoPaddingModule_basic",
+    "Conv2dNoPaddingModule_basic",
+    "Conv2dWithPaddingDilationStrideModule_basic",
+    "Conv2dWithPaddingModule_basic",
+}
 
 LTC_CRASHING_SET = {
     # TODO: update test to move all inputs to the lazy device. Otherwise test fails with: