[Tcp] Merge main into mlir-tcp

.github/workflows/RollPyTorch.yml

@@ -24,9 +24,21 @@ jobs:
     - name: Get torch-mlir
       uses: actions/checkout@v3
       with:
-        submodules: 'true'
+        submodules: 'false'
         token: ${{ secrets.WORKFLOW_INVOCATION_TOKEN }}
 
+    - name: Get LLVM and StableHlo submodules
+      run: |
+        set -eo pipefail
+        cd ${GITHUB_WORKSPACE}
+
+        # Fetching the submodules concurrently may cause problems, so we fetch
+        # them one after another.
+        rm -f .git/modules/externals/llvm-project/index.lock
+        rm -f .git/modules/externals/stablehlo/index.lock
+        git submodule update --init --recursive externals/llvm-project
+        git submodule update --init --recursive externals/stablehlo
+
     - name: Setup ccache
       uses: ./.github/actions/setup-build
       with:
@@ -71,15 +83,14 @@ jobs:
         echo "PTVISION_RELEASE=${VISION_RELEASE}" >> ${GITHUB_ENV}
         echo "PT_HASH_CHANGED=${PT_HASH_CHANGED}" >> ${GITHUB_ENV}
 
-    - name: Build and test (in-tree), also update ODS and abstract interpretation library
+    - name: Build and test (out-of-tree), also update ODS and abstract interpretation library
       if: env.PT_HASH_CHANGED != '0'
       run: |
         cd ${GITHUB_WORKSPACE}
-        TM_PACKAGES="in-tree" TM_USE_PYTORCH_BINARY="OFF" \
+        TM_PACKAGES="out-of-tree" TM_USE_PYTORCH_BINARY="OFF" \
         TORCH_MLIR_SRC_PYTORCH_BRANCH="${{ env.PT_HASH }}" \
         TORCH_MLIR_SRC_PYTORCH_RELEASE="${{ env.PT_RELEASE }}" \
         TM_UPDATE_ODS_AND_ABSTRACT_INTERP_LIB="ON" \
-        TM_PYTHON_VERSIONS="cp311-cp311" \
         ./build_tools/python_deploy/build_linux_packages.sh
 
     - name: Post issue comment on build failure

.github/workflows/buildRelease.yml

@@ -115,7 +115,7 @@ jobs:
         cd $GITHUB_WORKSPACE
         TM_PACKAGE_VERSION=${{ github.event.inputs.python_package_version }}
         printf "TORCH_MLIR_PYTHON_PACKAGE_VERSION=%s\n" $TM_PACKAGE_VERSION > ./torch_mlir_package_version
-        TM_PYTHON_VERSIONS=${{ matrix.py_version }} TM_PACKAGES=${{ matrix.package }} TM_TORCH_VERSION="stable" ./build_tools/python_deploy/build_linux_packages.sh
+        TM_PYTHON_VERSIONS=${{ matrix.py_version }} TM_PACKAGES=${{ matrix.package }} TORCH_MLIR_ENABLE_LTC='0' ./build_tools/python_deploy/build_linux_packages.sh
 
     # If we were given a release_id, then upload the package we just built
     # to the github releases page.

README.md

@@ -38,8 +38,17 @@ We have few paths to lower down to the Torch MLIR Dialect.
 - `#torch-mlir` channel on the LLVM [Discord](https://discord.gg/xS7Z362) - this is the most active communication channel
 - Github issues [here](https://github.com/llvm/torch-mlir/issues)
 - [`torch-mlir` section](https://llvm.discourse.group/c/projects-that-want-to-become-official-llvm-projects/torch-mlir/41) of LLVM Discourse
-- Weekly meetings on Mondays 9AM PST. See [here](https://discourse.llvm.org/t/community-meeting-developer-hour-refactoring-recurring-meetings/62575) for more information.
-- Weekly op office hours on Thursdays 8:30-9:30AM PST. See [here](https://discourse.llvm.org/t/announcing-torch-mlir-office-hours/63973/2) for more information.
+
+### Meetings
+
+Community Meeting / Developer Hour:
+- 1st and 3rd Monday of the month at 9 am PST
+- 2nd and 4th Monday of the month at 5 pm PST
+
+Office Hours:
+- Every Thursday at 8:30 am PST
+
+Meeting links can be found [here](https://discourse.llvm.org/t/new-community-meeting-developer-hour-schedule/73868).
 
 ## Install torch-mlir snapshot
 
@@ -61,7 +70,7 @@ python -m pip install --upgrade pip
 Then, we can install torch-mlir with the corresponding torch and torchvision nightlies.
 ```
 pip install --pre torch-mlir torchvision \
-  -f https://llvm.github.io/torch-mlir/package-index/
+  -f https://llvm.github.io/torch-mlir/package-index/ \
   --extra-index-url https://download.pytorch.org/whl/nightly/cpu
 ```
 

build_tools/autogen_ltc_backend.py

@@ -467,7 +467,8 @@ def gen_fallback_code(*args, **kwargs):
             node_base="torch::lazy::TorchMlirNode",
             node_base_hdr=str(self.backend_path.joinpath("mlir_node.h")),
             tensor_class=self.tensor_class,
-            tensor_class_hdr="torch/csrc/lazy/core/tensor.h",
+            tensor_class_hdr="torch_mlir/csrc/base_lazy_backend/tensor.h",
+            create_aten_from_ltc_tensor="CreateFunctionalizedAtenFromLtcTensor",
             shape_inference_hdr=str(self.generated_path.joinpath("shape_inference.h")),
             lazy_ir_generator=GenMlirLazyIr,
         )

build_tools/autogen_ltc_backend.yaml

@@ -3,12 +3,6 @@ blacklist:
 # It also doesn't have confusing `unsafe` argument.
 - _index_put_impl
 
-# Ops with list of tensors output
-- split.Tensor
-- split_with_sizes
-- unbind.int
-- chunk
-
 # Additional ops which autogen is supported for but don't compile yet
 - _convolution
 - detach
@@ -18,42 +12,28 @@ blacklist:
 
 # Disabled for consistency with TS backend
 - lift_fresh_copy
-- new_empty
 - rsub
-- slice.Tensor  # Disabled in favour of slice_copy.Tensor
-- zeros
-- ones
-- arange
-- arange.start
-- arange.start_step
-- fill.Scalar
-- scalar_tensor
 
 # Disabled in favour of functionalized alternatives
 - _reshape_alias
-- expand
 - permute
 - select.int
-- squeeze
 - squeeze.dim
-- t
 - transpose.int
+- expand
+- squeeze
 - unsqueeze
 - view
+- slice.Tensor
+- split.Tensor
+- split_with_sizes
+- unbind.int
 
-whitelist:
-# Enabled for consistency with TS backend
-- arange.start_out
 
 # List of supported ops that we don't want to do the full codegen for
 supported:
-# - bernoulli
-# - bernoulli_
 - _to_copy
 - clone
-- empty.memory_format
-- empty_strided
-- fill_.Scalar
 - _unsafe_view
 - unbind_copy.int
 - split_copy.Tensor
@@ -80,18 +60,17 @@ supported:
 - _trilinear
 - linalg_pinv.atol_rtol_tensor
 - logsumexp.out
+- t
 
 # List of ops that will take in symints for the size instead of ints
 symint:
-- empty.memory_format
 - new_empty_strided
 - expand_copy
 - narrow_copy
 - slice_backward
 - slice_copy.Tensor
 - split_copy.Tensor
 - slice_scatter
-- view
 - view_copy
 - as_strided_copy
 - as_strided_scatter

build_tools/python_deploy/build_linux_packages.sh

@@ -178,6 +178,12 @@ function run_in_docker() {
         out-of-tree)
           setup_venv "$python_version" "$TM_TORCH_VERSION"
           build_out_of_tree "$TM_USE_PYTORCH_BINARY" "$python_version" "$TM_TORCH_VERSION"
+          if [ "${TM_UPDATE_ODS_AND_ABSTRACT_INTERP_LIB}" == "ON" ]; then
+            pushd /main_checkout/torch-mlir
+            TORCH_MLIR_BUILD_DIR=/main_checkout/torch-mlir/build_oot ./build_tools/update_torch_ods.sh
+            TORCH_MLIR_BUILD_DIR=/main_checkout/torch-mlir/build_oot ./build_tools/update_abstract_interp_lib.sh
+            popd
+          fi
           if [ "${TM_SKIP_TESTS}" == "OFF" ]; then
             test_out_of_tree
           fi

e2e_testing/xfail_sets.py

@@ -17,6 +17,9 @@
     # 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
     "Conv2dWithPaddingDilationStrideStaticModule_depthwise_multiplier",
+    "UnflattenStaticModule_basic",
+    "IscloseStaticModule_basic",
+    "IscloseStaticModuleTrue_basic",
 }
 
 TORCHDYNAMO_XFAIL_SET = {
@@ -288,6 +291,12 @@
 
     # AssertionError: Unregistered operation: torch.aten._embedding_bag_forward_only
     "AtenEmbeddingBagStaticModule_basic",
+
+    # Lowering not present for this case
+    "ElementwiseToDtypeI64ToUI8Module_basic",
+
+    # torch._dynamo.exc.TorchRuntimeError: Failed running call_function <built-in method add of type object at 0x7f4f8b05a720>(*(FakeTensor(..., size=(3, 4), dtype=torch.int8), 3, 2), **{}): Tensor with dtype torch.int64 is not the expected dtype of torch.int8!
+    "ElementwiseAddScalarInt8Module_basic",
 }
 
 if torch_version_for_comparison() < version.parse("2.1.0.dev"):
@@ -827,7 +836,6 @@
     "ReshapeAliasCollapseModule_basic",
     "ReshapeAliasExpandModule_basic",
     "ReshapeExpandModule_basic",
-    "RollModule_basic",
     "TestMultipleTensorReturn_basic",
     "AdaptiveAvgPool1dUnitOutputSizeStaticModule_basic",
     "AdaptiveAvgPool2dUnitOutputSizeStaticModule_basic",
@@ -1022,6 +1030,8 @@
     "TypePromotionSameCategoryDifferentWidthModule_basic",
     "TypePromotionSameCategoryZeroRankWider_basic",
     "TypePromotionZeroRankHigherCategoryModule_basic",
+    "ElementwiseAddScalarInt8Module_basic",
+    "ElementwiseSubTensorInt8Module_basic",
     "ElementwiseMulScalarModule_basic",
     "ElementwiseMulScalarModule_float",
     "ElementwiseMulScalarModule_int",
@@ -1039,13 +1049,17 @@
     "BatchNorm1DStaticShapeModule_basic",
     "BroadcastListConstructWithMinusOneModule_basic",
     "ElementwiseAtenDivIntScalarModule_basic",
+    "ElementwiseToDtypeI64ToI8Module_basic",
+    "ToDtypeLayoutCPUModule_basic",
     "TupleModule_basic",
     "TypeAsDifferentModule_basic",
 }
 
 # Write the TOSA set as a "passing" set as it is very early in development
 # and very few tests work yet.
 TOSA_PASS_SET = {
+    "IscloseStaticModule_basic",
+    "IscloseStaticModuleTrue_basic",
     "TileBigDimsSizeModule_basic",
     "TileSmallDimsSizeModule_basic",
     "IndexPutImpl2DNoneIndexStaticModule_basic",
@@ -1175,6 +1189,7 @@
     "BatchNorm3DModule_basic",
     "BatchNorm1DStaticShapeModule_basic",
     "FlattenStaticModule_basic",
+    "UnflattenStaticModule_basic",
     "FlattenRank0Module_basic",
     "ElementwiseFlattenBroadcastModule_basic",
     "SquareModule_basic",
@@ -1383,6 +1398,8 @@
     "SoftmaxIntNegDimModule_basic",
     "_LogSoftmaxModule_basic",
     "_SoftmaxModule_basic",
+    "ElementwiseAddScalarInt8Module_basic",
+    "ElementwiseSubTensorInt8Module_basic",
 }
 
 MAKE_FX_TOSA_PASS_SET = (TOSA_PASS_SET | {
@@ -1441,10 +1458,6 @@
     "_ConvolutionDeprecated2DBenchmarkModule_basic",
     "_ConvolutionDeprecated2DCudnnModule_basic",
     "_ConvolutionDeprecated2DDeterministicModule_basic",
-    "AdaptiveAvgPool1dNonUnitOutputSizeDynamicModule_basic",
-    "AdaptiveAvgPool1dNonUnitOutputSizeStaticModule_basic",
-    "AdaptiveAvgPool2dNonUnitOutputSizeDynamicModule_basic",
-    "AdaptiveAvgPool2dNonUnitOutputSizeStaticModule_basic",
     "AddIntModule_basic",
     "AtenIntBoolOpModule_basic",
     "BernoulliTensorModule_basic",
@@ -1480,7 +1493,6 @@
     "NeFloatIntModule_basic",
     "NeIntModule_basic",
     "QuantizedMLP_basic",
-    "RollModule_basic",
     "ScalarImplicitFloatModule_basic",
     "ScalarImplicitIntModule_basic",
     "SliceEndSleStartModule_basic",
@@ -1512,7 +1524,6 @@
     "ConvolutionBackwardModule2DPadded_basic",
     "VarMeanCorrectionModule_basic",
     "VarMeanCorrectionNoneModule_basic",
-    "PrimsConvertElementTypeModule_basic",
     "ElementwisePreluModule_basic",
     "VarMeanBiasedModule_basic",
     "VarMeanUnbiasedModule_basic",
@@ -1547,4 +1558,7 @@
     "UniformStaticShapeModule_basic",
     "AtenEmbeddingBagStaticModule_basic",
     "EmptyStridedModule_basic",
+    "ElementwiseBitwiseAndScalarInt64Module_basic",
+    "ElementwiseBitwiseAndScalarInt32Module_basic",
+    "ElementwiseBitwiseAndScalarInt8Module_basic",
 }

externals/llvm-external-projects/torch-mlir-dialects/lib/Conversion/TcpToLinalg/Elementwise.cpp

@@ -72,11 +72,11 @@ createLinalgPayloadForElementwiseOp(Operation *op,
       auto minFloat = clampOp.getMinFloat();
       auto maxFloat = clampOp.getMaxFloat();
       if (minFloat)
-        result = b.create<arith::MaxFOp>(
+        result = b.create<arith::MaximumFOp>(
             loc, result,
             b.create<arith::ConstantFloatOp>(loc, *minFloat, b.getF32Type()));
       if (maxFloat)
-        result = b.create<arith::MinFOp>(
+        result = b.create<arith::MinimumFOp>(
             loc, result,
             b.create<arith::ConstantFloatOp>(loc, *maxFloat, b.getF32Type()));
     } else if (elemType.isa<mlir::IntegerType>()) {

externals/llvm-external-projects/torch-mlir-dialects/lib/Dialect/TMTensor/IR/TMTensorOps.cpp

@@ -233,7 +233,7 @@ LogicalResult AttentionOp::generateScalarImplementation(OpBuilder &b,
                  loc, init,
                  [&](OpBuilder &b, Location loc, Value elem, Value acc) {
                    Value x = b.create<memref::LoadOp>(loc, weight, localIVs);
-                   Value max = b.create<arith::MaxFOp>(loc, x, acc);
+                   Value max = b.create<arith::MaximumFOp>(loc, x, acc);
                    b.create<scf::ReduceReturnOp>(loc, max);
                  });
            })

externals/llvm-external-projects/torch-mlir-dialects/test/Conversion/TcpToLinalg/unary.mlir

@@ -43,9 +43,9 @@ func.func @tanh(%arg0 : tensor<?x?xf32>) -> tensor<?x?xf32> {
 // CHECK-SAME:                        outs(%[[EMPTY_TENSOR]] : tensor<?x?xf32>) {
 // CHECK:         ^bb0(%[[BBARG0:.*]]: f32, %{{.*}}: f32):
 // CHECK:           %[[CST0:.*]] = arith.constant 1.000000e-01 : f32
-// CHECK:           %[[MAX:.*]] = arith.maxf %[[BBARG0]], %[[CST0]] : f32
+// CHECK:           %[[MAX:.*]] = arith.maximumf %[[BBARG0]], %[[CST0]] : f32
 // CHECK:           %[[CST1:.*]] = arith.constant 1.024000e+03 : f32
-// CHECK:           %[[MIN:.*]] = arith.minf %[[MAX]], %[[CST1]] : f32
+// CHECK:           %[[MIN:.*]] = arith.minimumf %[[MAX]], %[[CST1]] : f32
 // CHECK:           linalg.yield %[[MIN]] : f32
 // CHECK:         } -> tensor<?x?xf32>
 // CHECK:         return %[[GENERIC]] : tensor<?x?xf32>

include/torch-mlir/Conversion/Utils/Utils.h

@@ -87,7 +87,8 @@ mlir::RankedTensorType GetTypeFromTensorShape(llvm::ArrayRef<int64_t> shape,
 // from a tensor or a scalar in the pytorch dialect. Both the scalar and dtype
 // should be converted builtin types.
 Value convertScalarToDtype(OpBuilder &b, Location loc, Value scalar, Type dtype,
-                           std::optional<Type> srcOriginalDtype = std::nullopt);
+                           std::optional<Type> srcOriginalDtype = std::nullopt,
+                           std::optional<Type> dstOriginalDtype = std::nullopt);
 
 Value toPositiveValidDim(ConversionPatternRewriter &rewriter, Location loc,
                          Value torchOptionalInt, Value builtinInt,