search

llvm / llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies.
http://llvm.org
Other
28.03k stars 11.58k forks source link

[mlir] [tosa] -tosa-to-tensor crashes in llvm-project/mlir/lib/Conversion/TosaToTensor/TosaToTensor.cpp:145: TensorType (anonymous namespace)::inferReshapeCollapsedType(TensorType, TensorType): Assertion `rhsShape[currRhsDim] == 1' failed. #107969

Open axeabc opened 1 week ago

axeabc commented 1 week ago

git version: 761bf333e378b52614c

system: Ubuntu 18.04.6 LTS

reproduce with: mlir-opt -tosa-to-tensor a.mlir

a.mlir:

module {
  func.func @test_reshape_3d(%arg0: tensor<1xf32>) -> tensor<?x?x?xf32> {
    %c0 = arith.constant 0 : index
    %0 = tosa.reshape %arg0 {new_shape = array<i64: 2, 3, 4>} :(tensor<1xf32>) -> tensor<?x?x?xf32>
    return %0 : tensor<?x?x?xf32>
  }
}

stack trace:

mlir-opt: /data/szy/MLIR/llvm-release/llvm-project/mlir/lib/Conversion/TosaToTensor/TosaToTensor.cpp:145: TensorType (anonymous namespace)::inferReshapeCollapsedType(TensorType, TensorType): Assertion `rhsShape[currRhsDim] == 1' failed.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.  Program arguments: /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt -tosa-to-tensor a.mlir
 #0 0x000055cf54970128 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x10d7128)
 #1 0x000055cf5496dc3e llvm::sys::RunSignalHandlers() (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x10d4c3e)
 #2 0x000055cf54970abd SignalHandler(int) Signals.cpp:0:0
 #3 0x00007f6af409e420 __restore_rt (/lib/x86_64-linux-gnu/libpthread.so.0+0x14420)
 #4 0x00007f6af36db00b raise /build/glibc-LcI20x/glibc-2.31/signal/../sysdeps/unix/sysv/linux/raise.c:51:1
 #5 0x00007f6af36ba859 abort /build/glibc-LcI20x/glibc-2.31/stdlib/abort.c:81:7
 #6 0x00007f6af36ba729 get_sysdep_segment_value /build/glibc-LcI20x/glibc-2.31/intl/loadmsgcat.c:509:8
 #7 0x00007f6af36ba729 _nl_load_domain /build/glibc-LcI20x/glibc-2.31/intl/loadmsgcat.c:970:34
 #8 0x00007f6af36cbfd6 (/lib/x86_64-linux-gnu/libc.so.6+0x33fd6)
 #9 0x000055cf57481427 (anonymous namespace)::ReshapeConverter::matchAndRewrite(mlir::tosa::ReshapeOp, mlir::tosa::ReshapeOpAdaptor, mlir::ConversionPatternRewriter&) const TosaToTensor.cpp:0:0
#10 0x000055cf57480350 mlir::OpConversionPattern<mlir::tosa::ReshapeOp>::matchAndRewrite(mlir::Operation*, llvm::ArrayRef<mlir::Value>, mlir::ConversionPatternRewriter&) const (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x3be7350)
#11 0x000055cf57ac0b11 mlir::ConversionPattern::matchAndRewrite(mlir::Operation*, mlir::PatternRewriter&) const (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x4227b11)
#12 0x000055cf5a44bc91 void llvm::function_ref<void ()>::callback_fn<mlir::PatternApplicator::matchAndRewrite(mlir::Operation*, mlir::PatternRewriter&, llvm::function_ref<bool (mlir::Pattern const&)>, llvm::function_ref<void (mlir::Pattern const&)>, llvm::function_ref<llvm::LogicalResult (mlir::Pattern const&)>)::$_0>(long) PatternApplicator.cpp:0:0
#13 0x000055cf5a44894b mlir::PatternApplicator::matchAndRewrite(mlir::Operation*, mlir::PatternRewriter&, llvm::function_ref<bool (mlir::Pattern const&)>, llvm::function_ref<void (mlir::Pattern const&)>, llvm::function_ref<llvm::LogicalResult (mlir::Pattern const&)>) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x6baf94b)
#14 0x000055cf57ac1b53 (anonymous namespace)::OperationLegalizer::legalize(mlir::Operation*, mlir::ConversionPatternRewriter&) DialectConversion.cpp:0:0
#15 0x000055cf57ac0bb7 mlir::OperationConverter::convert(mlir::ConversionPatternRewriter&, mlir::Operation*) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x4227bb7)
#16 0x000055cf57ac1d7f mlir::OperationConverter::convertOperations(llvm::ArrayRef<mlir::Operation*>) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x4228d7f)
#17 0x000055cf57ac996b mlir::applyPartialConversion(mlir::Operation*, mlir::ConversionTarget const&, mlir::FrozenRewritePatternSet const&, mlir::ConversionConfig) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x423096b)
#18 0x000055cf5747c6c8 (anonymous namespace)::TosaToTensor::runOnOperation() TosaToTensorPass.cpp:0:0
#19 0x000055cf57a622d6 mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x41c92d6)
#20 0x000055cf57a62c40 mlir::detail::OpToOpPassAdaptor::runPipeline(mlir::OpPassManager&, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int, mlir::PassInstrumentor*, mlir::PassInstrumentation::PipelineParentInfo const*) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x41c9c40)
#21 0x000055cf57a65282 mlir::PassManager::run(mlir::Operation*) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x41cc282)
#22 0x000055cf57a5dab1 performActions(llvm::raw_ostream&, std::shared_ptr<llvm::SourceMgr> const&, mlir::MLIRContext*, mlir::MlirOptMainConfig const&) MlirOptMain.cpp:0:0
#23 0x000055cf57a5d70b llvm::LogicalResult llvm::function_ref<llvm::LogicalResult (std::unique_ptr<llvm::MemoryBuffer, std::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>::callback_fn<mlir::MlirOptMain(llvm::raw_ostream&, std::unique_ptr<llvm::MemoryBuffer, std::default_delete<llvm::MemoryBuffer>>, mlir::DialectRegistry&, mlir::MlirOptMainConfig const&)::$_0>(long, std::unique_ptr<llvm::MemoryBuffer, std::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&) MlirOptMain.cpp:0:0
#24 0x000055cf57b0a3a5 mlir::splitAndProcessBuffer(std::unique_ptr<llvm::MemoryBuffer, std::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<llvm::LogicalResult (std::unique_ptr<llvm::MemoryBuffer, std::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>, llvm::raw_ostream&, llvm::StringRef, llvm::StringRef) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x42713a5)
#25 0x000055cf57a58b35 mlir::MlirOptMain(llvm::raw_ostream&, std::unique_ptr<llvm::MemoryBuffer, std::default_delete<llvm::MemoryBuffer>>, mlir::DialectRegistry&, mlir::MlirOptMainConfig const&) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x41bfb35)
#26 0x000055cf57a58ddf mlir::MlirOptMain(int, char**, llvm::StringRef, llvm::StringRef, mlir::DialectRegistry&) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x41bfddf)
#27 0x000055cf57a5910e mlir::MlirOptMain(int, char**, llvm::StringRef, mlir::DialectRegistry&) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x41c010e)
#28 0x000055cf54950d67 main (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x10b7d67)
#29 0x00007f6af36bc083 __libc_start_main /build/glibc-LcI20x/glibc-2.31/csu/../csu/libc-start.c:342:3
#30 0x000055cf549508ee _start (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x10b78ee)
CoTinker commented 1 week ago

I'm wondering if it's worth adding a verifier for tosa.reshape, such as:

if (inputType.hasStaticShape()) {
    int64_t inputElementsNum = inputType.getNumElements();
    if (outputType.hasStaticShape()) {
      int64_t outputElementsNum = outputType.getNumElements();
      if (inputElementsNum != outputElementsNum) {
        return emitOpError() << "cannot reshape " << inputElementsNum
                             << " elements into " << outputElementsNum;
      }
    }

    int64_t newShapeElementsNum = std::accumulate(
        getNewShape().begin(), getNewShape().end(), 1LL,
        [](int64_t acc, int64_t dim) { return (dim > 0) ? acc * dim : acc; });
    bool isStaticNewShape =
        llvm::all_of(getNewShape(), [](int64_t s) { return s > 0; });
    if ((isStaticNewShape && inputElementsNum != newShapeElementsNum) ||
        (!isStaticNewShape && newShapeElementsNum > inputElementsNum)) {
      return emitOpError() << "cannot reshape " << inputElementsNum
                           << " elements into " << newShapeElementsNum;
    }
  }

@joker-eph @eric-k256

joker-eph commented 1 week ago

It depends on the kind of type inference that TOSA is expected to do: the tricky part is that if you validate shapes in the verifier, then shape inference has to know about these validity rule before updating the result of an operation.