[mlir] -tosa-optional-decompositions crashes

[Anonymous7i0em3yd]

git version: d097070

system: Ubuntu 18.04.6 LTS

reproduce with: mlir-opt -tosa-optional-decompositions a.mlir

a.mlir:

func.func @test_transpose_conv2D_pad_left(%arg0: tensor<1x32x 32x 16x f32>, %arg1: tensor<16x2x16x16xf32>, %arg2: tensor<16xf32>) -> tensor<1x32x 32x 16x f32> {
  %0 = "tosa.transpose_conv2d"(%arg0, %arg1, %arg2) {out_pad = array<i64: 0, 0, 8193, 0>, out_shape = array<i64: 1, 32, 8193, 16>, stride = array<i64: 1, 2>} :
              (tensor<1x32x 32x 16x f32>, tensor<16x2x16x16xf32>, tensor<16xf32>) -> tensor<1x32x 32x 16x f32>
  return %0 : tensor<1x32x 32x 16x f32>
}

stack trace:

<unknown>:0: error: invalid tensor dimension size
mlir-opt: /data/szy/MLIR/llvm-release/llvm-project/mlir/include/mlir/IR/StorageUniquerSupport.h:180: static ConcreteT mlir::detail::StorageUserBase<mlir::RankedTensorType, mlir::TensorType, mlir::detail::RankedTensorTypeStorage, mlir::detail::TypeUniquer, mlir::ShapedType::Trait, mlir::ValueSemantics>::get(MLIRContext *, Args &&...) [ConcreteT = mlir::RankedTensorType, BaseT = mlir::TensorType, StorageT = mlir::detail::RankedTensorTypeStorage, UniquerT = mlir::detail::TypeUniquer, Traits = <mlir::ShapedType::Trait, mlir::ValueSemantics>, Args = <llvm::ArrayRef<long> &, mlir::Type &, mlir::Attribute &>]: Assertion `succeeded( ConcreteT::verifyInvariants(getDefaultDiagnosticEmitFn(ctx), args...))' 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-optional-decompositions /data/szy/MLIR/seed/seed4/tmp.xkcd0pMA92.mlir
 #0 0x000055d5911b6f88 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x1185f88)
 #1 0x000055d5911b4a9e llvm::sys::RunSignalHandlers() (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x1183a9e)
 #2 0x000055d5911b791d SignalHandler(int) Signals.cpp:0:0
 #3 0x00007f93bf621420 __restore_rt (/lib/x86_64-linux-gnu/libpthread.so.0+0x14420)
 #4 0x00007f93bec5e00b raise /build/glibc-LcI20x/glibc-2.31/signal/../sysdeps/unix/sysv/linux/raise.c:51:1
 #5 0x00007f93bec3d859 abort /build/glibc-LcI20x/glibc-2.31/stdlib/abort.c:81:7
 #6 0x00007f93bec3d729 get_sysdep_segment_value /build/glibc-LcI20x/glibc-2.31/intl/loadmsgcat.c:509:8
 #7 0x00007f93bec3d729 _nl_load_domain /build/glibc-LcI20x/glibc-2.31/intl/loadmsgcat.c:970:34
 #8 0x00007f93bec4efd6 (/lib/x86_64-linux-gnu/libc.so.6+0x33fd6)
 #9 0x000055d59471e703 (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x46ed703)
#10 0x000055d59471e628 mlir::RankedTensorType::get(llvm::ArrayRef<long>, mlir::Type, mlir::Attribute) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x46ed628)
#11 0x000055d5938425d5 mlir::tosa::SliceOp mlir::tosa::CreateOpAndInferShape<mlir::tosa::SliceOp, mlir::Value&, mlir::detail::DenseArrayAttrImpl<long>&, mlir::detail::DenseArrayAttrImpl<long>&>(mlir::ImplicitLocOpBuilder&, mlir::Type, mlir::Value&, mlir::detail::DenseArrayAttrImpl<long>&, mlir::detail::DenseArrayAttrImpl<long>&) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x38115d5)
#12 0x000055d59383dbe6 (anonymous namespace)::TransposeConvStridedConverter::matchAndRewrite(mlir::tosa::TransposeConv2DOp, mlir::PatternRewriter&) const TosaDecomposeTransposeConv.cpp:0:0
#13 0x000055d597110a21 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&)>)::$_0>(long) PatternApplicator.cpp:0:0
#14 0x000055d59710d6eb mlir::PatternApplicator::matchAndRewrite(mlir::Operation*, mlir::PatternRewriter&, llvm::function_ref<bool (mlir::Pattern const&)>, llvm::function_ref<void (mlir::Pattern const&)>, llvm::function_ref<llvm::LogicalResult (mlir::Pattern const&)>) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x70dc6eb)
#15 0x000055d59466c11f (anonymous namespace)::GreedyPatternRewriteDriver::processWorklist() GreedyPatternRewriteDriver.cpp:0:0
#16 0x000055d594668aaf mlir::applyPatternsAndFoldGreedily(mlir::Region&, mlir::FrozenRewritePatternSet const&, mlir::GreedyRewriteConfig, bool*) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x4637aaf)
#17 0x000055d59382579b (anonymous namespace)::TosaOptionalDecompositions::runOnOperation() TosaOptionalDecompositions.cpp:0:0
#18 0x000055d5945f1936 mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x45c0936)
#19 0x000055d5945f2260 mlir::detail::OpToOpPassAdaptor::runPipeline(mlir::OpPassManager&, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int, mlir::PassInstrumentor*, mlir::PassInstrumentation::PipelineParentInfo const*) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x45c1260)
#20 0x000055d5945f760e auto void mlir::parallelForEach<__gnu_cxx::__normal_iterator<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo*, std::vector<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo, std::allocator<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo>>>, mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::$_0>(mlir::MLIRContext*, __gnu_cxx::__normal_iterator<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo*, std::vector<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo, std::allocator<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo>>>, __gnu_cxx::__normal_iterator<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo*, std::vector<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo, std::allocator<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo>>>, mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::$_0&&)::'lambda'(__gnu_cxx::__normal_iterator<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo*, std::vector<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo, std::allocator<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo>>>&&)::operator()<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo&>(__gnu_cxx::__normal_iterator<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo*, std::vector<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo, std::allocator<mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo>>>&&) const Pass.cpp:0:0
#21 0x000055d5945f38bb mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x45c28bb)
#22 0x000055d5945f1a8f mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x45c0a8f)
#23 0x000055d5945f2260 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+0x45c1260)
#24 0x000055d5945f4832 mlir::PassManager::run(mlir::Operation*) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x45c3832)
#25 0x000055d5945ed0da performActions(llvm::raw_ostream&, std::shared_ptr<llvm::SourceMgr> const&, mlir::MLIRContext*, mlir::MlirOptMainConfig const&) MlirOptMain.cpp:0:0
#26 0x000055d5945ecd2d 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
#27 0x000055d594698fa5 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+0x4667fa5)
#28 0x000055d5945e6d15 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+0x45b5d15)
#29 0x000055d5945e6fbf mlir::MlirOptMain(int, char**, llvm::StringRef, llvm::StringRef, mlir::DialectRegistry&) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x45b5fbf)
#30 0x000055d5945e72ee mlir::MlirOptMain(int, char**, llvm::StringRef, mlir::DialectRegistry&) (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x45b62ee)
#31 0x000055d591197a77 main (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x1166a77)
#32 0x00007f93bec3f083 __libc_start_main /build/glibc-LcI20x/glibc-2.31/csu/../csu/libc-start.c:342:3
#33 0x000055d5911975ee _start (/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt+0x11665ee)

[swote-git]

I have checked and reviewed your issue.

Here's Two Problem:

1. your mlir function is invalid.
   1-1. Missing Operands
   - in your code. %0 = "tosa.transpose_conv2d"(%arg0, %arg1, %arg2) is not satisfied mlir.tosa spec. there's 3(input, weight, bias) operands, but tosa.transpose_conv2d required 5(input, weight, bias, input_zp, weight_zp).
   - see: https://mlir.llvm.org/docs/Dialects/TOSA/#tosatranspose_conv2d-mlirtosatransposeconv2dop

   1-2. out_shape = array<i64: 1, 32, 8193, 16> out_shape is incorrect.
   - OH == (IH - 1) * stride_y + out_pad_top + out_pad_bottom + KH : should be 33, but 32.
   - OW == (IW - 1) * stride_x + out_pad_left + out_pad_right + KW : should be 8271, but 8193.

corrected mlir function should be:

func.func @test_transpose_conv2D_pad_left(%arg0: tensor<1x32x32x16xf32>, %arg1: tensor<16x2x16x16xf32>, %arg2: tensor<16xf32>) -> tensor<1x33x8271x16xf32> {
    %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
    %weight_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
    
    %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %input_zp, %weight_zp {
        acc_type = f32, 
        out_pad = array<i64: 0, 0, 8193, 0>, 
        out_shape = array<i64: 1, 33, 8271, 16>, 
        stride = array<i64: 1, 2>
    } : (tensor<1x32x32x16xf32>, tensor<16x2x16x16xf32>, tensor<16xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x33x8271x16xf32>
    
    return %0 : tensor<1x33x8271x16xf32>
}

2. Even if your mlir function is normal, you will still receive errors due to validation failure.

In, llvm/llvm-project/mlir/lib/Dialect/Tosa/Transforms/TosaValidation.cpp line 403~428. There's level check kernel sizes for kH and KW.

This explicitly checks that each padding value (p) satisfies the same constraints as kernel dimensions, ensuring pad <= MAX_KERNEL. This is to prevent serious waste of memory and computing resources due to large padding.

[lhutton1]

I tried with @swote-git's rewritten example on a newer checkout, but could not reproduce the issue:

$ mlir-opt test.mlir --tosa-optional-decompositions
module {
  func.func @test_transpose_conv2D_pad_left(%arg0: tensor<1x32x32x16xf32>, %arg1: tensor<16x2x16x16xf32>, %arg2: tensor<16xf32>) -> tensor<1x33x8271x16xf32> {
    %0 = tosa.const_shape  {values = dense<[1, 1, 1, 16]> : tensor<4xindex>} : () -> !tosa.shape<4>
    %1 = tosa.const_shape  {values = dense<[0, 0, 0, 0, 8193, 0, 0, 0]> : tensor<8xindex>} : () -> !tosa.shape<8>
    %2 = tosa.const_shape  {values = dense<[1, 33, 78, 16]> : tensor<4xindex>} : () -> !tosa.shape<4>
    %3 = tosa.const_shape  {values = dense<[1, 33, 39, 1, 2, 16]> : tensor<6xindex>} : () -> !tosa.shape<6>
    %4 = "tosa.const"() <{values = dense<0.000000e+00> : tensor<32xf32>}> : () -> tensor<32xf32>
    %5 = tosa.const_shape  {values = dense<[0, 0, 1, 1, 7, 7, 0, 0]> : tensor<8xindex>} : () -> !tosa.shape<8>
    %6 = tosa.const_shape  {values = dense<[32, 2, 8, 16]> : tensor<4xindex>} : () -> !tosa.shape<4>
    %7 = "tosa.const"() <{values = dense<0.000000e+00> : tensor<1xf32>}> : () -> tensor<1xf32>
    %8 = tosa.const_shape  {values = dense<[16, 2, 1, 8, 2, 16]> : tensor<6xindex>} : () -> !tosa.shape<6>
    %9 = tosa.reshape %arg1, %8 : (tensor<16x2x16x16xf32>, !tosa.shape<6>) -> tensor<16x2x1x8x2x16xf32>
    %10 = tosa.transpose %9 {perms = array<i32: 2, 4, 0, 1, 3, 5>} : (tensor<16x2x1x8x2x16xf32>) -> tensor<1x2x16x2x8x16xf32>
    %11 = tosa.reshape %10, %6 : (tensor<1x2x16x2x8x16xf32>, !tosa.shape<4>) -> tensor<32x2x8x16xf32>
    %12 = tosa.reverse %11 {axis = 1 : i32} : (tensor<32x2x8x16xf32>) -> tensor<32x2x8x16xf32>
    %13 = tosa.reverse %12 {axis = 2 : i32} : (tensor<32x2x8x16xf32>) -> tensor<32x2x8x16xf32>
    %14 = tosa.pad %arg0, %5, %7 : (tensor<1x32x32x16xf32>, !tosa.shape<8>, tensor<1xf32>) -> tensor<1x34x46x16xf32>
    %15 = tosa.conv2d %14, %13, %4, %7, %7 {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<1x34x46x16xf32>, tensor<32x2x8x16xf32>, tensor<32xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x33x39x32xf32>
    %16 = tosa.reshape %15, %3 : (tensor<1x33x39x32xf32>, !tosa.shape<6>) -> tensor<1x33x39x1x2x16xf32>
    %17 = tosa.transpose %16 {perms = array<i32: 0, 1, 3, 2, 4, 5>} : (tensor<1x33x39x1x2x16xf32>) -> tensor<1x33x1x39x2x16xf32>
    %18 = tosa.reshape %17, %2 : (tensor<1x33x1x39x2x16xf32>, !tosa.shape<4>) -> tensor<1x33x78x16xf32>
    %19 = tosa.pad %18, %1, %7 : (tensor<1x33x78x16xf32>, !tosa.shape<8>, tensor<1xf32>) -> tensor<1x33x8271x16xf32>
    %20 = tosa.reshape %arg2, %0 : (tensor<16xf32>, !tosa.shape<4>) -> tensor<1x1x1x16xf32>
    %21 = tosa.add %19, %20 : (tensor<1x33x8271x16xf32>, tensor<1x1x1x16xf32>) -> tensor<1x33x8271x16xf32>
    return %21 : tensor<1x33x8271x16xf32>
  }
}

shall we close it?

[psunn]

@Anonymous7i0em3yd Please reopen the issue if the problem persists.