[Bug Report] 带split op的模型报错

[XiaotaoChen]

背景

1. 我们在尝试将minicpm-v-2-6部署到bm1684x上，在尝试用自己的工具分别导出: vit, embedding, lm_head, block, block_cache模块；
2. 分别使用model_transform.py 和 model_deploy.py导出到mlir和bmodel；
3. 模型转换问题主要发生在bf16数据格式的vit模型，问题原因为：bfloat16格式的onnx模型貌似onnx相关优化会有问题，导致vit部分split相关算子没有被优化掉；
4. 在model_transform.py脚本转换vit模型时，首先在python层遇到split op的转换问题，原因在于我们导出的vit模型为bfloat16数据的onnx，split op输入是一个constant，而tpu-mlir中OnnxConverter.py：convert_split_op只考虑输入是activation tensor的情况，代码修改如下所示：

# op = self.getOperand(onnx_node.inputs[0])
# split input maybe a constant tensor
op = self.getOp(onnx_node.inputs[0])

5. 修复该问题后，报tpuc-opt相关的优化pass错误，如下面所示；
6. 可以复现该问题的最小模型见：vision_tower_bf16-cut.onnx

具体错误log

2024/12/23 17:05:04 - INFO : TPU-MLIR v1.12.beta.0-29-g5dacf2a47-20241223
2024/12/23 17:05:04 - INFO : 
         _____________________________________________________ 
        | preprocess:                                           |
        |   (x - mean) * scale                                  |
        '-------------------------------------------------------'
  config Preprocess args : 
        resize_dims           : same to net input dims
        keep_aspect_ratio     : False
        keep_ratio_mode       : letterbox
        pad_value             : 0
        pad_type              : center
        --------------------------
        mean                  : [0.0, 0.0, 0.0]
        scale                 : [1.0, 1.0, 1.0]
        --------------------------
        pixel_format          : bgr
        channel_format        : nchw

2024/12/23 17:05:04 - INFO : Input_shape assigned
2024/12/23 17:05:04 - WARNING : ConstantFolding failed.
2024/12/23 17:05:04 - INFO : ConstantFolding finished
2024/12/23 17:05:04 - INFO : skip_fuse_bn:False
2024/12/23 17:05:04 - INFO : Onnxsim opt finished
2024/12/23 17:05:04 - WARNING : ConstantFolding failed.
2024/12/23 17:05:04 - INFO : ConstantFolding finished
name:/visual/Constant_1_output_0
2024/12/23 17:08:20 - INFO : Save mlir file: /workspace/tpu-mlir/tmp_origin.mlir
[Running]: tpuc-opt /workspace/tpu-mlir/tmp_origin.mlir --shape-infer --canonicalize --extra-optimize -o /workspace/tpu-mlir/tmp.mlir
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.      Program arguments: tpuc-opt /workspace/tpu-mlir/tmp_origin.mlir --init --shape-infer --canonicalize --extra-optimize --deinit --mlir-print-debuginfo -o /workspace/tpu-mlir/tmp.mlir
 #0 0x000062af85a6ee87 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (/workspace/tpu-mlir/install/bin/tpuc-opt+0x869e87)
 #1 0x000062af85a6cbae llvm::sys::RunSignalHandlers() (/workspace/tpu-mlir/install/bin/tpuc-opt+0x867bae)
 #2 0x000062af85a6f80a SignalHandler(int) Signals.cpp:0:0
 #3 0x00007177ab158520 (/lib/x86_64-linux-gnu/libc.so.6+0x42520)
 #4 0x000062af86eefe98 tpu_mlir::top::RangeOp::shape_inference() (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1ceae98)
 #5 0x000062af86e00e23 tpu_mlir::detail::ShapeInterfaceInterfaceTraits::Model<tpu_mlir::top::RangeOp>::shape_inference(tpu_mlir::detail::ShapeInterfaceInterfaceTraits::Concept const*, mlir::Operation*) (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1bfbe23)
 #6 0x000062af86f78910 tpu_mlir::top::ShapeInferPass::runOnOperation()::'lambda'(tpu_mlir::ShapeInterface)::operator()(tpu_mlir::ShapeInterface) const (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1d73910)
 #7 0x000062af85bc0c5e void mlir::detail::walk<mlir::ForwardIterator>(mlir::Operation*, llvm::function_ref<void (mlir::Operation*)>, mlir::WalkOrder) (/workspace/tpu-mlir/install/bin/tpuc-opt+0x9bbc5e)
 #8 0x000062af86f76964 tpu_mlir::top::ShapeInferPass::runOnOperation() (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1d71964)
 #9 0x000062af8703c904 mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1e37904)
#10 0x000062af8703cf31 mlir::detail::OpToOpPassAdaptor::runPipeline(mlir::OpPassManager&, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int, mlir::PassInstrumentor*, mlir::PassInstrumentation::PipelineParentInfo const*) (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1e37f31)
#11 0x000062af8703f3d8 mlir::PassManager::run(mlir::Operation*) (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1e3a3d8)
#12 0x000062af85a6053b performActions(llvm::raw_ostream&, std::shared_ptr<llvm::SourceMgr> const&, mlir::MLIRContext*, mlir::MlirOptMainConfig const&) MlirOptMain.cpp:0:0
#13 0x000062af85a5f904 mlir::LogicalResult llvm::function_ref<mlir::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&)::$_2>(long, std::unique_ptr<llvm::MemoryBuffer, std::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&) MlirOptMain.cpp:0:0
#14 0x000062af872532c8 mlir::splitAndProcessBuffer(std::unique_ptr<llvm::MemoryBuffer, std::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (std::unique_ptr<llvm::MemoryBuffer, std::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>, llvm::raw_ostream&, bool, bool) (/workspace/tpu-mlir/install/bin/tpuc-opt+0x204e2c8)
#15 0x000062af85a59c0a mlir::MlirOptMain(llvm::raw_ostream&, std::unique_ptr<llvm::MemoryBuffer, std::default_delete<llvm::MemoryBuffer>>, mlir::DialectRegistry&, mlir::MlirOptMainConfig const&) (/workspace/tpu-mlir/install/bin/tpuc-opt+0x854c0a)
#16 0x000062af85a5a0d4 mlir::MlirOptMain(int, char**, llvm::StringRef, mlir::DialectRegistry&) (/workspace/tpu-mlir/install/bin/tpuc-opt+0x8550d4)
#17 0x000062af85a58b1a main (/workspace/tpu-mlir/install/bin/tpuc-opt+0x853b1a)
#18 0x00007177ab13fd90 (/lib/x86_64-linux-gnu/libc.so.6+0x29d90)
#19 0x00007177ab13fe40 __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x29e40)
#20 0x000062af85a57f25 _start (/workspace/tpu-mlir/install/bin/tpuc-opt+0x852f25)
Floating point exception (core dumped)

[XiaotaoChen]

貌似是lib/Dialect/Top/Interfaces/Range.cpp 中对start, delta, limit参数，将float转为int64时，存在数值差异导致。比如66行limit的值，0.999998987 被转成了0。