@vpisarev MaxPool int8 tests pass now. Tests with maxpooling as the last layer are commented as they compute max indices which is not supported in int8 version.

Although I still don't get why maxpooling as the last layer should compute max index as well by default.

@vpisarev MaxPool int8 tests pass now. Tests with maxpooling as the last layer are commented as they compute max indices which is not supported in int8 version.

Although I still don't get why maxpooling as the last layer should compute max index as well by default.

opencv/modules/dnn/src/layers/pooling_layer.cpp

Line 1314 in4c2dff8

int numOutputs = requiredOutputs ? requiredOutputs : (type == MAX ?2 :1);

I actually suggest to support computing indices (in FP32, as before) together with computing the max value in INT8. Yes, it will be slower, but it will let us to provide 100% compatibility.

support computing indices (in FP32, as before) together with computing the max value in INT8

This is not possible right now as a single variable determines the datatypes of all the outputs for a layer. So both outputs[0] and outputs[1] can either be CV_32F or CV_8S. One as CV_32F and another as CV_8S is not possible currently.

Some build issues :

1. AVX-512 path in int8layers/layers_common.simd.hpp causes build warnings and convolution tests failure (segmentation fault). The same tests are passed locally and in some other builders so I suspect there is an issue with that specific path. Also, I cannot reproduce this locally as my CPU only supports up to AVX2.
2. DNN Tests failure in OpenCL builders. From what I remember, I haven't modified any OpenCL related code, so don't know whats causing the failures.

Ofcourse changing "dtype" to std::vector would solve it, but that will introduce a lot of complexity in allocating blobs and a lot of work for a feature which is rarely used. Maybe we can keep it as low priority and look at it later.

ok, sounds good to me. Let's keep it as a low-priority item

Some build issues :

1. AVX-512 path in int8layers/layers_common.simd.hpp causes build warnings and convolution tests failure (segmentation fault). The same tests are passed locally and in some other builders so I suspect there is an issue with that specific path. Also, I cannot reproduce this locally as my CPU only supports up to AVX2.

I suggest to comment off AVX-512 branches for now (in your newly added code, not everywhere)

2. DNN Tests failure in OpenCL builders. From what I remember, I haven't modified any OpenCL related code, so don't know whats causing the failures.

well, you need to figure that out. If you stuck at that, we can look at it together

@alalek@vpisarev How do I ensure only dnn module tests are run in CI Linux OpenCL builder. I edited my original comment but it looks like tests for all modules is being checked.

@vpisarev As discussed, int8 layers which had mostly duplicated code (concat, flatten, padding) have been removed and the original fp32 layers are modified to support 8-bit inputs as well.

In some parallel_for(), I have used templates to support multiple datatypes, please check the latest commit to see if any changes have to be made.

@SamFC10, could you please fix the merge conflicts once again? And then squash commits? We will try to merge your pull request quickly.

And then squash commits

Looks like I messed up something.
Command used :

git reset --soft HEAD~38git commit -m ""git push -f origin int8

You can rollback changes:

git checkout -B int8 79eca09675fecd2b58c237233a0aaaa7197ace6f

Managed to restore my commits and squashed them. Thanks for the help@alalek@vpisarev

👍