Shintarou Okada

Hi @skysley. Sorry for my late reply. hash1, hash2 and hash5 share a same issue. `hash256_hex_string` calculates a hash so `picosha2::hash256_hex_string(hashVec, hexHash);` means calculating a hash of hashVec. So the...

`hash256_hex_string()` is functionally equal to `hash256()` + `bytes_to_hex_string()`. So calling `hash256_hex_string()` following `hash256()` means calculating hash twice. You can modify like: ``` picosha2::hash256(inputFile, hashVec.begin(), hashVec.end()); //picosha2::hash256_hex_string(hashVec, hexHash); picosha2::bytes_to_hex_string(hashVec.begin(), hashVec.end(), hexHash);...

@WilliamVenner Sorry for my late reply. The line in your "Read from file" code; ```picosha2::hash256(idt, s.begin(), s.end());``` means "hashing the hashed data". So the printed hex string is different.

To determine which nodes ares useless, menoh has to know which outputs are needed. Construction of variable_profile_table reqires all nodes are known to menoh. So we can't add ignoring feature...

Or we can simply add reducing feature in `menoh_build_variable_profile_table`.

It is correct. Sorry, > Construction of variable_profile_table reqires all nodes are known to menoh. is incorrect. So we can add ignoring-if-possible feature into `menoh_build_variable_profle_table()`.

Users who doesn't use mkldnn but tensorrt needn't link libmkldnn. So we need to change `@MKLDNN_LIBRARY@` part depending on `ENABLE_MKLDNN`.

Hi @yurivict. As @take-cheeze said, current menoh doesn't support mkldnn>0.17 . I recommend you to use mkldnn==0.16 because there is known issue about small buffer allocation for 0.17 . Or,...

Thank you for your report @WilliamTambellini and sorry for the trouble. Checking at cmake time is suitable. We would like to take your PR.

Thank you! I found the wrong code. I will fix it. We should add such a test for other operators too.