[Build] QNN EP build : implicit conversion loses integer precision

[john-dance]

Describe the issue

When building with use_qnn (clang, targeting Android), several files in the QNN Execution Provider sources use size_t and unit64_t interchangeably. Clang disagrees with that.

base_op_builder.cc - this one is a simple fix, just change line 212 to use auto

simple_op_builder.cc
qnn_backend_manager.cc

• In these files, unt64_t is returned from some functions, but passed as size_t to other methods

simple_op_builder.cc:175:68: error: implicit conversion loses integer precision: 'int64_t' (aka 'long long') to 'size_t' (aka 'unsigned int') [-Werror,-Wshorten-64-to-32]
    ParQuantizeLinearStd(&tensor_data.alpha, unpacked_data.data(), num_of_elements, scale, zero_point, thread_pool);

qnn_backend_manager.cc:361:87: error: implicit conversion loses integer precision: 'uint64_t' (aka 'unsigned long long') to 'size_t' (aka 'unsigned int') [-Werror,-Wshorten-64-to-32]
  std::unique_ptr<unsigned char[]> context_buffer = std::make_unique<unsigned char[]>(required_buffer_size);

qnn_backend_manager.cc:413:66: error: implicit conversion loses integer precision: 'uint64_t' (aka 'unsigned long long') to 'std::streamsize' (aka 'int') [-Werror,-Wshorten-64-to-32]
  of_stream.write(reinterpret_cast<char*>(context_buffer.get()), written_buffer_size);

qnn_backend_manager.cc:440:79: error: implicit conversion loses integer precision: 'uint64_t' (aka 'unsigned long long') to 'size_t' (aka 'unsigned int') [-Werror,-Wshorten-64-to-32]
  std::unique_ptr<unsigned char[]> buffer = std::make_unique<unsigned char[]>(buffer_size);

qnn_backend_manager.cc:444:84: error: implicit conversion loses integer precision: 'uint64_t' (aka 'unsigned long long') to 'std::streamsize' (aka 'int') [-Werror,-Wshorten-64-to-32]
  const auto& read_result = cache_file.read(reinterpret_cast<char*>(buffer.get()), buffer_size);


### Urgency

_No response_

### Target platform

Android

### Build script

 python3 tools/ci_build/build.py \
        --cmake_extra_defines "CMAKE_VERBOSE_MAKEFILE=ON" \
        --cmake_extra_defines "CMAKE_C_FLAGS=$extra_cflags" \
        --cmake_extra_defines "CMAKE_CXX_FLAGS=$extra_cflags" \
        --build_dir "$build_dir" \
        --config $build_type \
        --build_shared_lib \
        --parallel \
        --android \
        --android_abi $build_target \
        --android_api $VERSION \
        --android_sdk_path "$(get_android_home)" \
        --android_ndk_path "[...]" \
        --android_cpp_shared \
        --use_cache \
        --use_xnnpack \
        --use_nnapi \
        --use_qnn \
        --qnn_home="[...]" \
        --skip_tests \
        --path_to_protoc_exe "$PROTOC" \

### Error / output

(See above)

### Visual Studio Version

_No response_

### GCC / Compiler Version

_No response_

[adrianlizarraga]

Thanks for reporting @john-dance. It looks like this an architecture with a 32-bit size_t. Can you please provide more details about the build target?

I also went ahead and created a draft PR that properly casts uint64/int64 to size_t in the locations you listed. Can you please try to compile with the PR's branch?

• Draft PR: [QNN EP] Handle implicit conversion loses when size_t is 32-bit #17808
• branch: adrianl/qnn-issue-17778

[john-dance]

@adrianlizarraga - Thanks for reaching out.

The build target is armabi-v7a, so it is a 32-bit target.

Your draft handles most of the issues, but a few remain. As an aside, I previously didn't have an issue with conv.cc so either these were changes post 1.16.0 or changes in your diff.

src/onnxruntime/core/providers/xnnpack/nn/conv.cc:71:65: warning: implicit conversion loses integer precision: 'const int64_t' (aka 'const long long') to 'size_t' (aka 'unsigned int') [-Wshorten-64-to-32]
status = xnn_setup_convolution2d_nhwc_f32(op0_.get(), N, H, W, X.Data(), Y->MutableData(),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
src/onnxruntime/core/providers/xnnpack/nn/conv.cc:71:62: warning: implicit conversion loses integer precision: 'const int64_t' (aka 'const long long') to 'size_t' (aka 'unsigned int') [-Wshorten-64-to-32]
status = xnn_setup_convolution2d_nhwc_f32(op0_.get(), N, H, W, X.Data(), Y->MutableData(),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
src/onnxruntime/core/providers/xnnpack/nn/conv.cc:71:59: warning: implicit conversion loses integer precision: 'const int64_t' (aka 'const long long') to 'size_t' (aka 'unsigned int') [-Wshorten-64-to-32]
status = xnn_setup_convolution2d_nhwc_f32(op0_.get(), N, H, W, X.Data(), Y->MutableData(),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
src/onnxruntime/core/providers/xnnpack/nn/conv.cc:74:65: warning: implicit conversion loses integer precision: 'const int64_t' (aka 'const long long') to 'size_t' (aka 'unsigned int') [-Wshorten-64-to-32]
status = xnn_setup_convolution2d_nhwc_qs8(op0_.get(), N, H, W, X.Data<int8_t>(), Y->MutableData<int8_t>(),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
src/onnxruntime/core/providers/xnnpack/nn/conv.cc:74:62: warning: implicit conversion loses integer precision: 'const int64_t' (aka 'const long long') to 'size_t' (aka 'unsigned int') [-Wshorten-64-to-32]
status = xnn_setup_convolution2d_nhwc_qs8(op0_.get(), N, H, W, X.Data<int8_t>(), Y->MutableData<int8_t>(),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
src/onnxruntime/core/providers/xnnpack/nn/conv.cc:74:59: warning: implicit conversion loses integer precision: 'const int64_t' (aka 'const long long') to 'size_t' (aka 'unsigned int') [-Wshorten-64-to-32]
status = xnn_setup_convolution2d_nhwc_qs8(op0_.get(), N, H, W, X.Data<int8_t>(), Y->MutableData<int8_t>(),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
src/onnxruntime/core/providers/xnnpack/nn/conv.cc:77:65: warning: implicit conversion loses integer precision: 'const int64_t' (aka 'const long long') to 'size_t' (aka 'unsigned int') [-Wshorten-64-to-32]
status = xnn_setup_convolution2d_nhwc_qu8(op0_.get(), N, H, W, X.Data<uint8_t>(), Y->MutableData<uint8_t>(),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
src/onnxruntime/core/providers/xnnpack/nn/conv.cc:77:62: warning: implicit conversion loses integer precision: 'const int64_t' (aka 'const long long') to 'size_t' (aka 'unsigned int') [-Wshorten-64-to-32]
status = xnn_setup_convolution2d_nhwc_qu8(op0_.get(), N, H, W, X.Data<uint8_t>(), Y->MutableData<uint8_t>(),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
src/onnxruntime/core/providers/xnnpack/nn/conv.cc:77:59: warning: implicit conversion loses integer precision: 'const int64_t' (aka 'const long long') to 'size_t' (aka 'unsigned int') [-Wshorten-64-to-32]
status = xnn_setup_convolution2d_nhwc_qu8(op0_.get(), N, H, W, X.Data<uint8_t>(), Y->MutableData<uint8_t>(),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
src/onnxruntime/core/providers/xnnpack/nn/conv.cc:80:65: warning: implicit conversion loses integer precision: 'const int64_t' (aka 'const long long') to 'size_t' (aka 'unsigned int') [-Wshorten-64-to-32]
status = xnn_setup_convolution2d_nhwc_qc8(op0_.get(), N, H, W, X.Data<int8_t>(), Y->MutableData<int8_t>(),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
src/onnxruntime/core/providers/xnnpack/nn/conv.cc:80:62: warning: implicit conversion loses integer precision: 'const int64_t' (aka 'const long long') to 'size_t' (aka 'unsigned int') [-Wshorten-64-to-32]
status = xnn_setup_convolution2d_nhwc_qc8(op0_.get(), N, H, W, X.Data<int8_t>(), Y->MutableData<int8_t>(),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
src/onnxruntime/core/providers/xnnpack/nn/conv.cc:80:59: warning: implicit conversion loses integer precision: 'const int64_t' (aka 'const long long') to 'size_t' (aka 'unsigned int') [-Wshorten-64-to-32]
status = xnn_setup_convolution2d_nhwc_qc8(op0_.get(), N, H, W, X.Data<int8_t>(), Y->MutableData<int8_t>(),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
12 warnings generated.
src/onnxruntime/core/providers/qnn/builder/opbuilder/base_op_builder.cc:213:28: error: implicit conversion loses integer precision: 'int64_t' (aka 'long long') to 'std::vector::value_type' (aka 'unsigned int') [-Werror,-Wshorten-64-to-32]
permutations.push_back(p);
~~~~~~~~~ ^
src/onnxruntime/core/providers/qnn/builder/opbuilder/base_op_builder.cc:214:55: error: implicit conversion loses integer precision: 'int64_t' (aka 'long long') to 'size_t' (aka 'unsigned int') [-Werror,-Wshorten-64-to-32]
new_tensor_shape_dims.push_back(tensor_shape_dims[p]);
~~~~~~~~~~~~~~~~~ ^