Compiler crashed when compiling a simple module with single conv layer

[Marrrb1e]

What happened?

Steps to reproduce your issue

iree crash.zip I put all the program into a zip, so you can reproduce. However, in the compilation script, be careful to change the corresponding path.

What component(s) does this issue relate to?

Compiler

Version information

I download the iree recently, maybe June 06. I am not sure about the exact version number.

Additional context

I was trying to compile the output of conv3*3 operators at a certain size with iree, and I wrote a network model with only one layer of convolution for this purpose, which I encountered errors after compiling.

[ScottTodd]

What is the output from iree-compile --version?

The "openxla" in the screenshot makes me suspect this is using the latest "stable" release from April, prior to https://github.com/iree-org/iree/commit/3f51a55daf4639473f36c42cbd9d775f2a76144e.

You can install a more recent nightly release by following https://iree.dev/reference/bindings/python/#__tabbed_2_2:

python -m pip install \
  --find-links https://iree.dev/pip-release-links.html \
  --upgrade \
  iree-compiler \
  iree-runtime

[Marrrb1e]

I checked my version and found that its version is 20240226.813, so I updated the iree version to 20240619.929 following your command. However, when I tried to compile the module, it still returned an error

[Marrrb1e]

The new error is shown below. The steps to reproduce are the same while the difference is the iree version is IREE (https://iree.dev): IREE compiler version 20240619.929 @ d792d2483a9d5d875a0f03b57c9d4f98e072c300 LLVM version 19.0.0git

[ScottTodd]

Thanks for checking with a more recent version. As a general point, please include logs as text instead of screenshots (or in addition to them). Text is searchable across issues and can be more easily referenced against source files. For best results you can use markdown code blocks: https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/creating-and-highlighting-code-blocks

The input program as .mlir would be useful. I'm not going to run compile_test.py myself for initial triage and test.log starts partway through compilation:

// -----// IR Dump Before LLVMCPUSelectLoweringStrategy (iree-llvmcpu-select-lowering-strategy) //----- //
hal.executable.variant public @embedded_elf_x86_64 target(<"llvm-cpu", "embedded-elf-x86_64", {cpu = "alderlake", cpu_features = "+prfchw,-cldemote,+avx,+aes,+sahf,+pclmul,-xop,+crc32,+xsaves,-avx512fp16,-usermsr,-sm4,-egpr,+sse4.1,-avx512ifma,+xsave,-avx512pf,+sse4.2,-tsxldtrk,-ptwrite,-widekl,-sm3,+invpcid,+64bit,+xsavec,-avx10.1-512,-avx512vpopcntdq,+cmov,-avx512vp2intersect,-avx512cd,+movbe,-avxvnniint8,-avx512er,-ccmp,-amx-int8,-kl,-avx10.1-256,-sha512,+avxvnni,-rtm,+adx,+avx2,-hreset,+movdiri,+serialize,+vpclmulqdq,-avx512vl,-uintr,-cf,+clflushopt,-raoint,-cmpccxadd,+bmi,-amx-tile,+sse,+gfni,-avxvnniint16,-amx-fp16,-ndd,+xsaveopt,+rdrnd,-avx512f,-amx-bf16,-avx512bf16,-avx512vnni,-push2pop2,+cx8,-avx512bw,+sse3,+pku,+fsgsbase,-clzero,-mwaitx,-lwp,+lzcnt,+sha,+movdir64b,-ppx,-wbnoinvd,-enqcmd,-prefetchwt1,-avxneconvert,-tbm,-pconfig,-amx-complex,+ssse3,+cx16,+bmi2,+fma,+popcnt,-avxifma,+f16c,-avx512bitalg,-rdpru,+clwb,+mmx,+sse2,+rdseed,-avx512vbmi2,-prefetchi,+rdpid,-fma4,-avx512vbmi,+shstk,+vaes,-waitpkg,-sgx,+fxsr,-avx512dq,-sse4a", data_layout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128", native_vector_size = 32 : i64, target_triple = "x86_64-unknown-unknown-eabi-elf"}>) {
  hal.executable.export public @jit_eval_dispatch_0_generic_64_f32 ordinal(0) layout(#hal.pipeline.layout<push_constants = 0, sets = [<0, bindings = [<0, storage_buffer, ReadOnly>, <1, storage_buffer>]>]>) attributes {hal.interface.bindings = [#hal.interface.binding<0, 0>, #hal.interface.binding<0, 1>]} {
  ^bb0(%arg0: !hal.device):
    %x, %y, %z = flow.dispatch.workgroup_count_from_slice 
    hal.return %x, %y, %z : index, index, index
  }
  builtin.module {
    func.func @jit_eval_dispatch_0_generic_64_f32() {
      %cst = arith.constant 9.99999974E-6 : f32
      %cst_0 = arith.constant 1.000000e+00 : f32
      %c0 = arith.constant 0 : index
      %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<64xf32>>
      %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<64xf32>>
      %2 = flow.dispatch.tensor.load %0, offsets = [0], sizes = [64], strides = [1] : !flow.dispatch.tensor<readonly:tensor<64xf32>> -> tensor<64xf32>
      %3 = tensor.empty() : tensor<64xf32>
      %4 = linalg.generic {indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> (d0)>], iterator_types = ["parallel"]} ins(%2 : tensor<64xf32>) outs(%3 : tensor<64xf32>) {
      ^bb0(%in: f32, %out: f32):
        %5 = arith.addf %in, %cst : f32
        %6 = math.sqrt %5 : f32
        %7 = arith.divf %cst_0, %6 : f32
        linalg.yield %7 : f32
      } -> tensor<64xf32>
      flow.dispatch.tensor.store %4, %1, offsets = [0], sizes = [64], strides = [1] : tensor<64xf32> -> !flow.dispatch.tensor<writeonly:tensor<64xf32>>
      return
    }
  }
}

The important part of the backtrace is:

 #3 0x00007ad6ad25cc5c mlir::DenseElementsAttr::getNumElements() const (/home/marble/iree-build/lib/libIREECompiler.so+0x1c5cc5c)
 #4 0x00007ad6b0882a98 mlir::iree_compiler::IREE::Stream::(anonymous namespace)::ConvertSplatConstantsIntoSplats::matchAndRewrite(mlir::iree_compiler::IREE::Stream::AsyncConstantOp, mlir::PatternRewriter&) const StreamOpFolders.cpp:0:0
 #5 0x00007ad6b0ae6655 mlir::PatternApplicator::matchAndRewrite(mlir::Operation*, mlir::PatternRewriter&, llvm::function_ref<bool (mlir::Pattern const&)>, llvm::function_ref<void (mlir::Pattern const&)>, llvm::function_ref<mlir::LogicalResult (mlir::Pattern const&)>) (/home/marble/iree-build/lib/libIREECompiler.so+0x54e6655)
 #6 0x00007ad6b0a9f26f (anonymous namespace)::GreedyPatternRewriteDriver::processWorklist() GreedyPatternRewriteDriver.cpp:0:0
 #7 0x00007ad6b0aa0c56 mlir::applyPatternsAndFoldGreedily(mlir::Region&, mlir::FrozenRewritePatternSet const&, mlir::GreedyRewriteConfig, bool*) (/home/marble/iree-build/lib/libIREECompiler.so+0x54a0c56)
 #8 0x00007ad6b0a54b7a (anonymous namespace)::Canonicalizer::runOnOperation() Canonicalizer.cpp:0:0

The program looks fairly normal to me only f32 data types (no exotic types like bf16, i4, etc.).

Possibly an issue with the Conv2d and BatchNorm2d op lowerings? We have a few models using convolution coming from PyTorch working without this issue though 🤔