Last active
May 29, 2024 17:10
-
-
Save renxida/cc2da8a9f23ee70bfd1b77b87fbdd6ed to your computer and use it in GitHub Desktop.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| # Description: This script is used to test the model deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir | |
| # run original model and print ir after failure | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir -o deit-small-distilled-patch16-224.default.stripped.vmfb --mlir-print-debuginfo --mlir-print-ir-after-failure |& gh gist create - -d "native_layer_norm ir dump after failure" | |
| # run again with --debug and grep for `(tensor<198xf32>) -> tensor<?x198xf32>` and pass names | |
| # grep patterns: | |
| # `(tensor<198xf32>) -> tensor<?x198xf32>` | |
| # `IR Dump After` | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir -o deit-small-distilled-patch16-224.default.stripped.vmfb --mlir-print-debuginfo --mlir-print-ir-after-all --debug |& grep -E "(tensor<198xf32>) -> tensor<?x198xf32>|IR Dump After" |& gh gist create - -d "native_layer_norm ir dump after failure with debug" | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir -o deit-small-distilled-patch16-224.default.stripped.vmfb --mlir-print-debuginfo --mlir-print-ir-before-all --mlir-print-skip-regions | |
| # find out which pass created the problem | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir -o deit-small-distilled-patch16-224.default.stripped.vmfb --mlir-print-debuginfo --mlir-print-ir-after-all &> printirafterall.mlir | |
| cat printirafterall.mlir | grep -E '\(tensor<198xf32>\)|IR Dump' | |
| # output: | |
| # // -----// IR Dump After RaiseSpecialOps (iree-global-opt-raise-special-ops) //----- // | |
| # // -----// IR Dump After DecomposeConcat (iree-global-opt-decompose-concat) //----- // | |
| # // -----// IR Dump After GeneralizeLinalgNamedOps (iree-global-opt-generalize-linalg-named-ops) //----- // | |
| # // -----// IR Dump After FoldUnitExtentDims Failed (iree-flow-fold-unit-extent-dims) //----- // | |
| # %73 = "tensor.expand_shape"(%72) <{reassociation = [[0, 1]]}> : (tensor<198xf32>) -> tensor<?x198xf32> | |
| # %86 = "tensor.expand_shape"(%85) <{reassociation = [[0, 1]]}> : (tensor<198xf32>) -> tensor<?x198xf32> | |
| # stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir:364:36: note: see current operation: %73 = "tensor.expand_shape"(%72) <{reassociation = [[0, 1]]}> : (tensor<198xf32>) -> tensor<?x198xf32> | |
| # -- the problem seems to be with FoldUnitExtentDims pass | |
| # print ir before and after FoldUnitExtentDims | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir -o deit-small-distilled-patch16-224.default.stripped.vmfb --mlir-print-debuginfo --mlir-print-ir-before=iree-flow-fold-unit-extent-dims --mlir-print-ir-after=iree-flow-fold-unit-extent-dims |& gh gist create - -d "native_layer_norm ir before and after FoldUnitExtentDims" | |
| # same, but with minimal.mlir | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/minimal.mlir -o minimal.vmfb --mlir-print-debuginfo --mlir-print-ir-before=iree-flow-fold-unit-extent-dims --mlir-print-ir-after=iree-flow-fold-unit-extent-dims |& gh gist create - -d "minimal ir before and after FoldUnitExtentDims" | |
| # print ir before and after FoldUnitExtentDims, but with stack trace | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir -o deit-small-distilled-patch16-224.default.stripped.vmfb --mlir-print-debuginfo --mlir-print-ir-before=iree-flow-fold-unit-extent-dims --mlir-print-ir-after=iree-flow-fold-unit-extent-dims --mlir-print-stacktrace-on-diagnostic |& gh gist create - -d "native_layer_norm ir before and after FoldUnitExtentDims" | |
| ### meeting with stanley | |
| # --pass-pipeline="builtin.module(iree-flow-fold-unit-extent-dims)" | |
| # ┌─[0]─[azureuser@xida-cpu-0]─[~/SHARK-TestSuite/e2eshark/test-onnx/pytorch/models/deit-small-distilled-patch16-224] | |
| # └──╼ $/home/azureuser/iree-build/tools/iree-opt ./even_bigger.mlir --pass-pipeline="builtin.module(iree-flow-fold-unit-extent-dims)" | |
| # /home/azureuser/iree-build/tools/iree-opt: /home/azureuser/miniconda/lib/libtinfo.so.6: no version information available (required by /home/azureuser/iree-build/lib/libIREECompiler.so) | |
| # <unknown>:0: error: unable to schedule pass 'FoldUnitExtentDims' on a PassManager intended to run on 'builtin.module'! | |
| ### Trying to find how the 1->? happened | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir -o deit-small-distilled-patch16-224.default.stripped.vmfb --mlir-print-debuginfo --mlir-print-ir-after-all &> printirafterall.mlir | |
| cat printirafterall.mlir | grep -E 'ins\(%[0-9]+ : tensor<1x198x384xf32>\) outs\(%[0-9]+ : tensor<\?x198x1xf32>\)|IR Dump' | |
| ## stop at linalg | |
| /home/azureuser/iree-build/tools/iree-compile --compile-to=input stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir -o deit-small-distilled-patch16-224.default.stripped.vmfb --mlir-print-debuginfo --mlir-print-ir-after-all &> printirafterall-input.mlir | |
| cat printirafterall-input.mlir | grep -E 'ins\(%[0-9]+ : tensor<1x198x384xf32>\) outs\(%[0-9]+ : tensor<\?x198x1xf32>\)|IR Dump' | |
| # same but on stripped/minimal.mlir | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/minimal.mlir -o minimal.vmfb --mlir-print-debuginfo --mlir-print-ir-after-all &> printirafterall-minimal.mlir | |
| cat printirafterall-minimal.mlir | grep -E 'ins\(%[0-9]+ : tensor<1x198x384xf32>\) outs\(%[0-9]+ : tensor<\?x198x1xf32>\)|IR Dump' | |
| # print ir before and after convert-elementwise-to-linalg | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir -o deit-small-distilled-patch16-224.default.stripped.vmfb --mlir-print-debuginfo --mlir-print-ir-before=convert-elementwise-to-linalg &> convert-elementwise-to-linalg-before.mlir | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir -o deit-small-distilled-patch16-224.default.stripped.vmfb --mlir-print-debuginfo --mlir-print-ir-after=convert-elementwise-to-linalg &> convert-elementwise-to-linalg-after.mlir | |
| gh gist create convert-elementwise-to-linalg-before.mlir convert-elementwise-to-linalg-after.mlir -d "convert-elementwise-to-linalg before and after" | |
| # trying to get a better trace | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir -o deit-small-distilled-patch16-224.default.stripped.vmfb --mlir-print-debuginfo --mlir-print-ir-after-failure --mlir-print-ir-module-scope --mlir-disable-threading | |
| --mlir-print-debug-info | |
| # save ir after torch-decompose-complex-ops | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu stripped/deit-small-distilled-patch16-224.default.pytorch.torch.stripped.mlir -o deit-small-distilled-patch16-224.default.stripped.vmfb --mlir-print-debuginfo --mlir-print-ir-after=torch-decompose-complex-ops &> torch-decompose-complex-ops.log | |
| # decomposed.mlir contains the mlir | |
| # run with --mlir-print-debug-info | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu decomposed.mlir -o decomposed.vmfb --mlir-print-debuginfo --mlir-print-debuginfo &> torch-decompose-complex-ops-debuginfo.log | |
| # now try atenrepro.mlir | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu atenrepro.mlir -o atenrepro.vmfb --mlir-print-debuginfo --mlir-print-debuginfo &> atenrepro-debuginfo.log | |
| # try to canonicalize the decomposed.mlir | |
| torch-mlir-opt --canonicalize decomposed.mlir decomposed-canonicalized.mlir | |
| # run with --mlir-print-debug-info | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu decomposed-canonicalized.mlir -o decomposed-canonicalized.vmfb --mlir-print-debuginfo --mlir-print-debuginfo &> torch-decompose-complex-ops-debuginfo.log | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu decomposed-canonicalized.mlir -o /tmp/tmp.vmfb --mlir-print-ir-after=memref-expand &> ./after-memref-expand.mlir | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu after-memref-expand.mlir -o /tmp/tmp.vmfb --mlir-print-ir-after=canonicalize --mlir-print-debuginfo &> rest-of-the-way.mlir | |
| # grep | |
| cat rest-of-the-way.mlir | grep -E 'ins\(%[0-9]+ : tensor<1x198x384xf32>\) outs\(%[0-9]+ : tensor<\?x198x1xf32>\)|IR Dump' | |
| # Is the linalg.generic coming directly from native_layer_norm or is some intermediate op spawning it? | |
| /home/azureuser/iree-build/tools/iree-compile --iree-input-demote-i64-to-i32 --iree-hal-target-backends=llvm-cpu decomposed.mlir -o /tmp/tmp.vmfb --mlir-print-ir-after-all --mlir-print-debuginfo &> decomposed-all.mlir | |
| # grep for the linalg.generic with the ? -> 1 shape | |
| cat decomposed-all.mlir | grep -E 'ins\(%[0-9]+ : tensor<1x198x384xf32>\) outs\(%[0-9]+ : tensor<\?x198x1xf32>\)|IR Dump' | |
| # Now debugging onnx.if | |
| Test onnx/models/KeypointRCNN_vaiq_int8 failed [torch-mlir] | |
| onnx if | |
| Test onnx/models/retinanet_resnet50_fpn_vaiq_int8 failed [torch-mlir] | |
| onnx if | |
| cd /home/azureuser/SHARK-TestSuite/e2eshark/ | |
| PYTHONPATH="/home/azureuser/torch-mlir/build/tools/torch-mlir/python_packages/torch_mlir" HF_TOKEN=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 python run.py --cachedir="~/.cache/" --torchtolinalg --torchmlirbuild "~/torch-mlir/build/" --ireebuild ~/iree-build --mode=onnx --framework=onnx --tests onnx/models/KeypointRCNN_vaiq_int8 --verbose | |
| # strip /home/azureuser/SHARK-TestSuite/e2eshark/test-run/onnx/models/KeypointRCNN_vaiq_int8/KeypointRCNN_vaiq_int8.default.torch-onnx.mlir | |
| torch-mlir-opt --mlir-elide-resource-strings-if-larger=16 /home/azureuser/SHARK-TestSuite/e2eshark/test-run/onnx/models/KeypointRCNN_vaiq_int8/KeypointRCNN_vaiq_int8.default.torch-onnx.mlir -o /home/azureuser/SHARK-TestSuite/e2eshark/test-run/onnx/models/KeypointRCNN_vaiq_int8/KeypointRCNN_vaiq_int8.default.torch-onnx.stripped.mlir | |
| ### now debugging onnx.if issue | |
| cd /home/azureuser/SHARK-TestSuite/e2eshark/test-run/onnx/models/KeypointRCNN_vaiq_int8 | |
| # try to lower the stripped model to torch | |
| torch-mlir-opt --convert-torch-onnx-to-torch /home/azureuser/SHARK-TestSuite/e2eshark/test-run/onnx/models/KeypointRCNN_vaiq_int8/KeypointRCNN_vaiq_int8.default.torch-onnx.mlir -o /home/azureuser/SHARK-TestSuite/e2eshark/test-run/onnx/models/KeypointRCNN_vaiq_int8/KeypointRCNN_vaiq_int8.default.torch-onnx.torch.mlir --mlir-print-ir-after-failure --mlir-elide-resource-strings-if-larger=16 &> /tmp/after_failure.log | |
| torch-mlir-opt --convert-torch-onnx-to-torch /home/azureuser/SHARK-TestSuite/e2eshark/test-run/onnx/models/KeypointRCNN_vaiq_int8/KeypointRCNN_vaiq_int8.default.torch-onnx.mlir -o /home/azureuser/SHARK-TestSuite/e2eshark/test-run/onnx/models/KeypointRCNN_vaiq_int8/KeypointRCNN_vaiq_int8.default.torch-onnx.torch.mlir --mlir-print-ir-after-failure --debug &> /tmp/debug.log | |
| cat <<-EOF > ./rcnn_repro.mlir | |
| func.func @minimal_example(%arg0: !torch.vtensor<[1],i1>, %arg1: !torch.vtensor<[?,4],f32>, %arg2: !torch.vtensor<[1],si64>, %arg3: !torch.vtensor<[?],f32>) -> !torch.vtensor<[?],si64> attributes {torch.onnx_meta.ir_version = 8 : si64, torch.onnx_meta.opset_version = 17 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "1.13.1"} { | |
| %0 = torch.operator "onnx.If"(%arg0) : (!torch.vtensor<[1],i1>) -> !torch.vtensor<[?],si64> { | |
| %1 = torch.operator "onnx.ReduceMax"(%arg1) {torch.onnx.keepdims = 0 : si64} : (!torch.vtensor<[?,4],f32>) -> !torch.vtensor<[],f32> | |
| %2 = torch.operator "onnx.Cast"(%arg2) {torch.onnx.to = 1 : si64} : (!torch.vtensor<[1],si64>) -> !torch.vtensor<[1],f32> | |
| %3 = torch.operator "onnx.Mul"(%2, %1) : (!torch.vtensor<[1],f32>, !torch.vtensor<[],f32>) -> !torch.vtensor<[1],f32> | |
| %4 = torch.operator "onnx.Unsqueeze"(%3, %arg2) : (!torch.vtensor<[1],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1,1],f32> | |
| %5 = torch.operator "onnx.Add"(%arg1, %4) : (!torch.vtensor<[?,4],f32>, !torch.vtensor<[1,1],f32>) -> !torch.vtensor<[?,4],f32> | |
| %6 = torch.operator "onnx.Unsqueeze"(%5, %arg2) : (!torch.vtensor<[?,4],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1,?,4],f32> | |
| %7 = torch.operator "onnx.Unsqueeze"(%arg3, %arg2) : (!torch.vtensor<[?],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1,?],f32> | |
| %8 = torch.operator "onnx.Unsqueeze"(%7, %arg2) : (!torch.vtensor<[1,?],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1,1,?],f32> | |
| %9 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<1> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> | |
| %10 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<0.5> : tensor<1xf32>} : () -> !torch.vtensor<[1],f32> | |
| %11 = torch.operator "onnx.NonMaxSuppression"(%6, %8, %9, %10) : (!torch.vtensor<[1,?,4],f32>, !torch.vtensor<[1,1,?],f32>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],f32>) -> !torch.vtensor<[?,3],si64> | |
| %12 = torch.operator "onnx.Gather"(%11, %arg2) {torch.onnx.axis = 1 : si64} : (!torch.vtensor<[?,3],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[?,1],si64> | |
| %13 = torch.operator "onnx.Squeeze"(%12, %arg2) : (!torch.vtensor<[?,1],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[?],si64> | |
| torch.operator_terminator %13 : !torch.vtensor<[?],si64> | |
| }, { | |
| %1 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<> : tensor<0xsi64>} : () -> !torch.vtensor<[0],si64> | |
| torch.operator_terminator %1 : !torch.vtensor<[0],si64> | |
| } | |
| return %0 : !torch.vtensor<[?],si64> | |
| } | |
| EOF | |
| torch-mlir-opt --convert-torch-onnx-to-torch repro.mlir -o repro.torch.mlir --debug |& tee repro.log | |
| # //===-------------------------------------------===// | |
| # Legalizing operation : 'torch.operator'(0x55914d9c55f0) { | |
| # %33 = "torch.operator"(%32, %arg2) <{name = "onnx.Unsqueeze"}> : (!torch.vtensor<[?,4],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1,?,4],f32> | |
| # * Fold { | |
| # } -> FAILURE : unable to fold | |
| # * Pattern : 'torch.operator -> ()' { | |
| # Trying to match "" | |
| # ** Failure : only support constant int axes values | |
| # : conversion failed to apply: "onnx.Unsqueeze", sinceVersion=13 | |
| # ** Failure : no matching versioned converter | |
| # "" result 0 | |
| # } -> FAILURE : pattern failed to match | |
| # } -> FAILURE : no matched legalization pattern | |
| # //===-------------------------------------------===// | |
| # repro.mlir:8:14: error: failed to legalize operation 'torch.operator' that was explicitly marked illegal | |
| # %6 = torch.operator "onnx.Unsqueeze"(%5, %arg2) : (!torch.vtensor<[?,4],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1,?,4],f32> | |
| # ^ | |
| # repro.mlir:8:14: note: see current operation: %33 = "torch.operator"(%32, %arg2) <{name = "onnx.Unsqueeze"}> : (!torch.vtensor<[?,4],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1,?,4],f32> | |
| # ImplicitTypeIDRegistry::lookupOrInsert(mlir::detail::PreservedAnalyses::AllAnalysesType) | |
| cat <<-EOF > ./rcnn_repro.mlir | |
| func.func @minimal_example(%arg0: !torch.vtensor<[1],i1>, %arg1: !torch.vtensor<[?,4],f32>, %arg3: !torch.vtensor<[?],f32>) -> !torch.vtensor<[?],si64> attributes {torch.onnx_meta.ir_version = 8 : si64, torch.onnx_meta.opset_version = 17 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "1.13.1"} { | |
| %0 = torch.operator "onnx.If"(%arg0) : (!torch.vtensor<[1],i1>) -> !torch.vtensor<[?],si64> { | |
| %c0 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<0> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> | |
| %1 = torch.operator "onnx.ReduceMax"(%arg1) {torch.onnx.keepdims = 0 : si64} : (!torch.vtensor<[?,4],f32>) -> !torch.vtensor<[],f32> | |
| %2 = torch.operator "onnx.Cast"(%c0) {torch.onnx.to = 1 : si64} : (!torch.vtensor<[1],si64>) -> !torch.vtensor<[1],f32> | |
| %3 = torch.operator "onnx.Mul"(%2, %1) : (!torch.vtensor<[1],f32>, !torch.vtensor<[],f32>) -> !torch.vtensor<[1],f32> | |
| %4 = torch.operator "onnx.Unsqueeze"(%3, %c0) : (!torch.vtensor<[1],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1,1],f32> | |
| %5 = torch.operator "onnx.Add"(%arg1, %4) : (!torch.vtensor<[?,4],f32>, !torch.vtensor<[1,1],f32>) -> !torch.vtensor<[?,4],f32> | |
| %6 = torch.operator "onnx.Unsqueeze"(%5, %c0) : (!torch.vtensor<[?,4],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1,?,4],f32> | |
| %7 = torch.operator "onnx.Unsqueeze"(%arg3, %c0) : (!torch.vtensor<[?],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1,?],f32> | |
| %8 = torch.operator "onnx.Unsqueeze"(%7, %c0) : (!torch.vtensor<[1,?],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1,1,?],f32> | |
| %9 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<1> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> | |
| %10 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<0.5> : tensor<1xf32>} : () -> !torch.vtensor<[1],f32> | |
| %11 = torch.operator "onnx.NonMaxSuppression"(%6, %8, %9, %10) : (!torch.vtensor<[1,?,4],f32>, !torch.vtensor<[1,1,?],f32>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],f32>) -> !torch.vtensor<[?,3],si64> | |
| %12 = torch.operator "onnx.Gather"(%11, %c0) {torch.onnx.axis = 1 : si64} : (!torch.vtensor<[?,3],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[?,1],si64> | |
| %13 = torch.operator "onnx.Squeeze"(%12, %c0) : (!torch.vtensor<[?,1],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[?],si64> | |
| torch.operator_terminator %13 : !torch.vtensor<[?],si64> | |
| }, { | |
| %1 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<> : tensor<0xsi64>} : () -> !torch.vtensor<[0],si64> | |
| torch.operator_terminator %1 : !torch.vtensor<[0],si64> | |
| } | |
| return %0 : !torch.vtensor<[?],si64> | |
| } | |
| EOF | |
| torch-mlir-opt --convert-torch-onnx-to-torch rcnn_repro.mlir -o rcnn_repro.torch.mlir --debug |& tee rcnn_repro.log | |
| # simple if test | |
| cat <<-EOF > ./rcnn_repro.mlir | |
| func.func @minimal_example(%arg0: !torch.vtensor<[1],i1>, %arg1: !torch.vtensor<[?,4],f32>, %arg3: !torch.vtensor<[?],f32>) -> !torch.vtensor<[?],si64> attributes {torch.onnx_meta.ir_version = 8 : si64, torch.onnx_meta.opset_version = 17 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "1.13.1"} { | |
| %0 = torch.operator "onnx.If"(%arg0) : (!torch.vtensor<[1],i1>) -> !torch.vtensor<[?],si64> { | |
| %c0 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<0> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> | |
| torch.operator_terminator %13 : !torch.vtensor<[?],si64> | |
| }, { | |
| %1 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<> : tensor<0xsi64>} : () -> !torch.vtensor<[0],si64> | |
| torch.operator_terminator %1 : !torch.vtensor<[0],si64> | |
| } | |
| return %0 : !torch.vtensor<[?],si64> | |
| } | |
| EOF | |
| torch-mlir-opt --convert-torch-onnx-to-torch rcnn_repro.mlir -o rcnn_repro.torch.mlir --debug |& tee rcnn_repro.log | |
| torch-mlir-opt --convert-torch-onnx-to-torch /home/azureuser/SHARK-TestSuite/e2eshark/test-run/onnx/models/KeypointRCNN_vaiq_int8/KeypointRCNN_vaiq_int8.default.torch-onnx.mlir -o /home/azureuser/SHARK-TestSuite/e2eshark/test-run/onnx/models/KeypointRCNN_vaiq_int8/KeypointRCNN_vaiq_int8.default.torch-onnx.torch.mlir --mlir-print-ir-after-failure --mlir-elide-resource-strings-if-larger=16 --debug &> /tmp/after_failure.log | |
| torch-mlir-opt --convert-torch-onnx-to-torch /tmp/afterfailure.mlir -o /tmp/afterfailure.torch.mlir --mlir-print-ir-after-failure | |
| torch-mlir-opt --convert-torch-onnx-to-torch ./repro2.mlir -o /tmp/afterfailure.torch.mlir --mlir-print-ir-after-failure |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment