int8 model compilation run into unexpected failure

I'm using the develop branch, latest commit: aa7ff911d4e6b2a3539ff29b4a05f1200b4250e0. (comparatively speaking it is doing fine if I use a preloaded onnx file)

The command I used is:

./bin/driver run --int8 --model resnet50 --batch 256

Compilation failure:

main.cpp:17:17: error: use of undeclared identifier 'inf'
auto z6 = float(inf);
                 ^
In file included from main.cpp:3:
./migraphx/kernels/pointwise.hpp:42:66: error: no matching function for call to object of type '(lambda at main.cpp:49:112)'
                      [&](auto i) { out[i] = implicit_conversion(f(xs[i]...)); });
                                                                 ^
...

For details please refer to the attached failure log: migraphx_int8_resnet50_compilation_failure.txt

As a reminder, the below errors are expected:

MIOpen(HIP): Error [EvaluateInvokers] /root/MIOpen/src/gemm_v2.cpp:649: rocBlas error encountered
MIOpen Error: /root/MIOpen/src/ocl/convolutionocl.cpp:518: Forward Convolution cannot be executed due to incorrect params
MIOpen(HIP): Warning [SetupFloats] Unsupported data types configuration: INT8x4xINT8x4xINT32
MIOpen(HIP): Warning [SetupFloats] Unsupported data types configuration: INT8x4xINT8x4xINT32

It happened when there's no applicable solution found for int8 format and therefore the quant_convolution fall back to int8x4. You can safely ignore this chunk of the log.

Start by reproducing ./bin/driver run --int8 --model resnet50 --batch 256 at commit aa7ff911d4e6b2a3539ff29b4a05f1200b4250e0 and review the attached failure log. Read main.cpp and migraphx/kernels/pointwise.hpp around the reported errors, while separating the expected MIOpen messages from the compilation failure. Done means the int8 ResNet-50 compilation completes without the reported errors.