[CPU][ONNX] Onnx test failures after pulling in torch-mlir changes

[Max191]

There were 3 new test failures after pulling in a torch-mlir patch: llvm/torch-mlir@55ff110

The following tests failed:

// RUN: iree-compile /tmp/test.mlir --iree-hal-target-backends=llvm-cpu --iree-llvmcpu-target-cpu=generic --iree-input-demote-f64-to-f32=false --mlir-disable-threading --mlir-print-ir-after-all -o /tmp/out.vmfb &> /tmp/dump.mlir

module {
  func.func @test_tfidfvectorizer_tf_batch_onlybigrams_skip0(%arg0: !torch.vtensor<[2,6],si32>) -> !torch.vtensor<[2,7],f32> attributes {torch.onnx_meta.ir_version = 4 : si64, torch.onnx_meta.opset_version = 17 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
    %none = torch.constant.none
    %0 = torch.operator "onnx.TfIdfVectorizer"(%arg0) {torch.onnx.max_gram_length = 2 : si64, torch.onnx.max_skip_count = 0 : si64, torch.onnx.min_gram_length = 2 : si64, torch.onnx.mode = "TF", torch.onnx.ngram_counts = [0 : si64, 4 : si64], torch.onnx.ngram_indexes = [0 : si64, 1 : si64, 2 : si64, 3 : si64, 4 : si64, 5 : si64, 6 : si64], torch.onnx.pool_int64s = [2 : si64, 3 : si64, 5 : si64, 4 : si64, 5 : si64, 6 : si64, 7 : si64, 8 : si64, 6 : si64, 7 : si64]} : (!torch.vtensor<[2,6],si32>) -> !torch.vtensor<[2,7],f32> 
    return %0 : !torch.vtensor<[2,7],f32>
  }
}

// -----

// RUN: iree-compile /tmp/test.mlir --iree-hal-target-backends=llvm-cpu --iree-llvmcpu-target-cpu=generic --iree-input-demote-f64-to-f32=false --mlir-disable-threading --mlir-print-ir-after-all -o /tmp/out.vmfb &> /tmp/dump.mlir

module {
  func.func @test_tfidfvectorizer_tf_batch_onlybigrams_skip5(%arg0: !torch.vtensor<[2,6],si32>) -> !torch.vtensor<[2,7],f32> attributes {torch.onnx_meta.ir_version = 4 : si64, torch.onnx_meta.opset_version = 17 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
    %none = torch.constant.none
    %0 = torch.operator "onnx.TfIdfVectorizer"(%arg0) {torch.onnx.max_gram_length = 2 : si64, torch.onnx.max_skip_count = 5 : si64, torch.onnx.min_gram_length = 2 : si64, torch.onnx.mode = "TF", torch.onnx.ngram_counts = [0 : si64, 4 : si64], torch.onnx.ngram_indexes = [0 : si64, 1 : si64, 2 : si64, 3 : si64, 4 : si64, 5 : si64, 6 : si64], torch.onnx.pool_int64s = [2 : si64, 3 : si64, 5 : si64, 4 : si64, 5 : si64, 6 : si64, 7 : si64, 8 : si64, 6 : si64, 7 : si64]} : (!torch.vtensor<[2,6],si32>) -> !torch.vtensor<[2,7],f32> 
    return %0 : !torch.vtensor<[2,7],f32>
  }
}

// -----

// RUN: iree-compile /tmp/test.mlir --iree-hal-target-backends=llvm-cpu --iree-llvmcpu-target-cpu=generic --iree-input-demote-f64-to-f32=false --mlir-disable-threading --mlir-print-ir-after-all -o /tmp/out.vmfb &> /tmp/dump.mlir

module {
  func.func @test_tfidfvectorizer_tf_batch_uniandbigrams_skip5(%arg0: !torch.vtensor<[2,6],si32>) -> !torch.vtensor<[2,7],f32> attributes {torch.onnx_meta.ir_version = 4 : si64, torch.onnx_meta.opset_version = 17 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
    %none = torch.constant.none
    %0 = torch.operator "onnx.TfIdfVectorizer"(%arg0) {torch.onnx.max_gram_length = 2 : si64, torch.onnx.max_skip_count = 5 : si64, torch.onnx.min_gram_length = 1 : si64, torch.onnx.mode = "TF", torch.onnx.ngram_counts = [0 : si64, 4 : si64], torch.onnx.ngram_indexes = [0 : si64, 1 : si64, 2 : si64, 3 : si64, 4 : si64, 5 : si64, 6 : si64], torch.onnx.pool_int64s = [2 : si64, 3 : si64, 5 : si64, 4 : si64, 5 : si64, 6 : si64, 7 : si64, 8 : si64, 6 : si64, 7 : si64]} : (!torch.vtensor<[2,6],si32>) -> !torch.vtensor<[2,7],f32> 
    return %0 : !torch.vtensor<[2,7],f32>
  }
}

The patch is reverted in IREE for now, so to reproduce the failures, use this branch that has the patch reapplied: https://github.com/Max191/iree/tree/onnx-cpu-unrolling-fail

[zjgarvey]

For context, I actually added this patch as a means to resolve an issue when unrolling loops for other tests see #18867 (comment).

In my opinion, it seems bad to indiscriminately unroll loops in the IR, so something actually needs to be addressed in the test examples. Why do the batched examples of TFIDF fail, but the unbatched ones pass?

It seems like IREE fails to handle arith.sitofp : i64 -> f64 when the input is the result of an scf.for loop, but not when those loops are unrolled?

For operations like this, I don't expect the result to be performant, but I just don't know what the constraints are from the IREE side. Do we straight up disallow any scalar scf loops?

[zjgarvey]

Maybe this comment from Ben is related: #18268 (comment)