多输入模型转换时添加--optimizePrefer 2选项 并基于此使用MNNPythonOfflineQuant离线量化，在arm cpu后端推理输出异常，与x86 cpu推理不对齐

[LXWLZU]

问题描述：

多输入模型转换时添加--optimizePrefer 2选项，Convert MatMul Convolution use shared const B inputs,
提升了推理速度，但模型增大，我期望通过量化减小模型大小以降低内存占用
转换命令如下：
./MNNConvert -f ONNX --modelFile small.onnx --MNNModel small_opt2.mnn --bizCode biz --optimizePrefer 2 --forTraining
日志：
The device support i8sdot:0, support fp16:0, support i8mm: 0
Start to Convert Other Model Format To MNN Model..., target version: 2.9
[14:33:55] /home/lixw/MNN/tools/converter/source/onnx/onnxConverter.cpp:46: ONNX Model ir version: 4
[14:33:55] /home/lixw/MNN/tools/converter/source/onnx/onnxConverter.cpp:47: ONNX Model opset version: 9
Start to Optimize the MNN Net...
[14:33:55] /home/lixw/MNN/tools/converter/source/optimizer/PostConverter.cpp:225: convert model for training, reserve BatchNorm and Dropout
Convert MatMul Convolution use shared const B inputs, may increase the model size
[14:33:55] /home/lixw/MNN/tools/converter/source/optimizer/PostConverter.cpp:225: convert model for training, reserve BatchNorm and Dropout
inputTensors : [ cahce_c0, cahce_spec, feat_spec, cahce_erb, feat_erb, h0, h2, h1, spec, ]
outputTensors: [ cahce_c0o, cahce_erbo, cahce_speco, df_coefs, ho0, ho1, ho2, speco, ]
Converted Success!
基于该步骤转化后的模型进行多输入模型的量化，量化后的模型在x86 cpu 和 arm cpu 后端推理结果不对齐，而且有的输出值为nan，
arm cpu后端推理似乎在处理使用共享常量B的卷积以及反量化过程中存在问题？（x86cpu的应该是正常的）

单独测试，
（1）仅转换时添加--optimizePrefer 2选项以转换MatMul Convolution使用共享常量B，不量化，x86 cpu与arm cpu可基本对齐
（2）不添加--optimizePrefer 2选项，直接量化，x86 cpu与arm cpu也能基本对齐
因此，只有当二者同时执行，即--optimizePrefer 2 加 量化，会出现arm cpu 推理错误的情况？这是否是bug呢？

我提供了简化版模型以及复现该现象的记录
google https://drive.google.com/file/d/1Z1Fy9ClfhhZcrRJXsrPbRkwmK22wtnAE/view?usp=sharing
baidu链接：https://pan.baidu.com/s/1X4h32GqvkY1NgkiZuPxZ5g?pwd=umxf
提取码：umxf

平台:

x86 CPU 及 ARMv8 CPU

Github版本:

2.9.0

编译方式:

ARM CPU MNN 2.9.0编译
set(CMAKE_SYSTEM_NAME Linux)
set(CMAKE_SYSTEM_PROCESSOR aarch64)
set(CMAKE_TRY_COMPILE_TARGET_TYPE "STATIC_LIBRARY")
set(CMAKE_BUILD_TYPE Release CACHE STRING "build release" FORCE)
set(CMAKE_INSTALL_PREFIX package CACHE STRING "install path" FORCE)
set(MNN_ARM82 OFF CACHE STRING "build arm82" FORCE)
set(MNN_FORBID_MULTI_THREAD OFF CACHE STRING "build single thread" FORCE)
set(MNN_USE_THREAD_POOL ON CACHE STRING "build thread pool" FORCE)
set(MNN_SUPPORT_BF16 OFF CACHE STRING "build bf16" FORCE)
set(MNN_BUILD_SHARED_LIBS OFF CACHE STRING "build static" FORCE)
set(MNN_SUPPORT_TFLITE_QUAN OFF CACHE STRING "" FORCE)
set(MNN_SEP_BUILD OFF CACHE STRING "build sep" FORCE)
set(MNN_USE_SSE OFF CACHE STRING "use sse" FORCE)
set(MNN_USE_LOGCAT OFF CACHE STRING "sue log cat" FORCE)

x86 CPU MNN 2.9.0编译
cmake .. -DMNN_BUILD_SHARED_LIBS=OFF -DMNN_BUILD_TOOLS=OFF && make -j4

编译日志:

ARM CPU MNN 2.9.0编译日志，编译成功
-- The C compiler identification is GNU 7.5.0
-- The CXX compiler identification is GNU 7.5.0
-- The ASM compiler identification is GNU
-- Found assembler: /usr/bin/cc
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Use Threadpool, forbid openmp
-- >>>>>>>>>>>>>
-- MNN BUILD INFO:
-- System: Linux
-- Processor: aarch64
-- Version: 2.9.0
-- Metal: OFF
-- OpenCL: OFF
-- OpenGL: OFF
-- Vulkan: OFF
-- ARM82: OFF
-- oneDNN: OFF
-- TensorRT: OFF
-- CoreML: OFF
-- NNAPI: OFF
-- CUDA: OFF
-- OpenMP: OFF
-- BF16: OFF
-- ThreadPool: ON
-- Hidden: TRUE
-- Build Path: /sdcard/lixiangwei/MNN/build_static
-- CUDA PROFILE: OFF
-- Enabling AArch64 Assemblies
-- Looking for pthread.h
-- Looking for pthread.h - found
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Success
-- Found Threads: TRUE
-- Configuring done
-- Generating done
-- Build files have been written to: /sdcard/lixiangwei/MNN/build_static

x86 CPU MNN 2.9.0编译日志，编译成功
-- Use Threadpool, forbid openmp
-- >>>>>>>>>>>>>
-- MNN BUILD INFO:
-- System: Linux
-- Processor: x86_64
-- Version: 2.9.0
-- Metal: OFF
-- OpenCL: OFF
-- OpenGL: OFF
-- Vulkan: OFF
-- ARM82: OFF
-- oneDNN: OFF
-- TensorRT: OFF
-- CoreML: OFF
-- NNAPI: OFF
-- CUDA: OFF
-- OpenMP: OFF
-- BF16: OFF
-- ThreadPool: ON
-- Hidden: TRUE
-- Build Path: /home/lixw/MNN/build_static
-- CUDA PROFILE: OFF
-- WIN_USE_ASM:
-- x86_64: Open SSE
-- MNN_AVX512:OFF
-- Configuring done
-- Generating done
-- Build files have been written to: /home/lixw/MNN/build_static

[LXWLZU]

补充一下，通过打印中间operator的输入输出，发现在x86 cpu上 onnx::Split_317__matmul_converted 这个op的输出也是nan，只不过模型最终输出里没有nan，这说明 （Convert MatMul Convolution use shared const B inputs）+ 量化 这种方式存在问题？或者是相关算子在x86 cpu后端推理时也存在问题？

[github-actions]

Marking as stale. No activity in 60 days.