Issues to use location information for performance investigation

[chenqiny]

I tried to get location information for resnet-50 model in onnx-model-zoo.

Command: onnx-mlir --preserveBitcode --preserveLocations --preserveMLIR /external/onnx-zoo/models/vision/classification/resnet/model/resnet50-caffe2-v1-7.onnx

Expect:

1. User can get mlir files for each pass
2. User can get location information in mlir files. The location attributes indicates source line number in the input mlir.
3. User can get location information in input mlir which indicates source line number/node id in onnx file (Optional)
4. User can get listing information includes llvm mlir source lines, instruction bytes, disassembly code and instruction addresses generated from that line. With these information, I can map instructions in perf sampling report with llvm mlir file, I can also map hot instructions with onnx file or mlir files from any pass. If possible I also recommend to output the passes information in a structurized file like json or txt.

Actual result There is only input and llvm mlir files.

And in the mlir files, the location information are looks like following:

Input mlir: %16 = "onnx.Relu"(%15) {onnx_node_name = ""} : (tensor<*xf32>) -> tensor<*xf32> loc(#loc) llvm mlir

  llvm.func @omTensorListGetOmtArray(!llvm.ptr<i8>) -> !llvm.ptr<ptr<i8>> loc(#loc0)
  llvm.func @omTensorSetDataType(!llvm.ptr<i8>, i32) loc(#loc0)
  llvm.func @omTensorGetDataType(!llvm.ptr<i8>) -> i32 loc(#loc0)
  llvm.func @omTensorGetStrides(!llvm.ptr<i8>) -> !llvm.ptr<i64> loc(#loc0)
  llvm.func @omTensorGetShape(!llvm.ptr<i8>) -> !llvm.ptr<i64> loc(#loc0)
  llvm.func @omTensorSetDataPtr(!llvm.ptr<i8>, i32, !llvm.ptr<i8>, !llvm.ptr<i8>) loc(#loc0)
  llvm.func @omTensorGetDataPtr(!llvm.ptr<i8>) -> !llvm.ptr<i8> loc(#loc0)
  llvm.func @omTensorCreateEmptyDeprecated(i64) -> !llvm.ptr<i8> loc(#loc0)
  llvm.func @omTensorListCreateWithOwnership(!llvm.ptr<ptr<i8>>, i32, i32) -> !llvm.ptr<i8> loc(#loc0)
  llvm.func @free(!llvm.ptr<i8>) loc(#loc0)

Passes: 'builtin.func' Pipeline {anonymous}::DecomposeONNXToONNXPass {anonymous}::ShapeInferencePass Canonicalizer {anonymous}::ShapeInferencePass 'builtin.func' Pipeline {anonymous}::ConstPropONNXToONNXPass {anonymous}::ONNXOpTransformPass SymbolDCE 'builtin.func' Pipeline {anonymous}::ONNXPreKrnlVerifyPass {anonymous}::InstrumentONNXPass {anonymous}::FrontendToKrnlLoweringPass Canonicalizer 'builtin.func' Pipeline {anonymous}::DisconnectKrnlDimFromAllocPass BufferDeallocation {anonymous}::KrnlEnableMemoryPoolPass {anonymous}::KrnlBundleMemoryPoolsPass Canonicalizer 'builtin.func' Pipeline {anonymous}::KrnlOptimizeMemoryPoolsPass Canonicalizer 'builtin.func' Pipeline {anonymous}::ConvertKrnlToAffinePass ConvertVectorToSCF ConvertAffineToStandard SCFToStandard {anonymous}::ConvertKrnlToLLVMPass ReconcileUnrealizedCasts Canonicalizer

[AlexandreEichenberger]

@chentong319 I recall you did work to get better line numbers. Is your work able to cover these cases? Thanks for looking into this.

[chentong319]

I tried to get location information for resnet-50 model in onnx-model-zoo.

Command: onnx-mlir --preserveBitcode --preserveLocations --preserveMLIR /external/onnx-zoo/models/vision/classification/resnet/model/resnet50-caffe2-v1-7.onnx

Expect:

1. User can get mlir files for each pass

To get mlir files for each pass, you need to use mlir flag "--print-ir-before-all" or "--print-ir-after-all"

2. User can get location information in mlir files. The location attributes indicates source line number in the input mlir.

Yes. Previously, user can only get location information when input is a .mlir file. Now input from .onnx input is fixed.

3. User can get location information in input mlir which indicates source line number/node id in onnx file (Optional)

Location inform is line number in *.input.mlir file, no node id.

4. User can get listing information includes llvm mlir source lines, instruction bytes, disassembly code and instruction addresses generated from that line. With these information, I can map instructions in perf sampling report with llvm mlir file, I can also map hot instructions with onnx file or mlir files from any pass.

Now the gdb can step through onnx operations. I think perf should be fine if it use the same mechanism as gdb.

If possible I also recommend to output the passes information in a structurized file like json or txt.

onnx-mlir uses mlir pass management tool, but we did not look into the debugging support of mlir. For example, the flag --preserveLocation should be a MLIR flag, not a onnx-mlir flag, to control the behavior of --print-ir-before/after-all. Does onnx-mlir work with perf now? We can try that.