[Protopipe] Workaround for INT64 input type #31972
Conversation
There was a problem hiding this comment.
I can't see where you actually convert from int64 to int32. Neither can I see for the output where you convert from int32 to int64. Please do a test with a int64 model with int64 precision where you check that you get the same output for the same input as benchmark app or singleimagetest.
| // workaround for I64 input type precision | ||
| if (mat.elemSize() == sizeof(int32_t) && file_byte_size == mat_byte_size * 2) { | ||
| std::size_t elem_size = mat.elemSize(); | ||
| for (std::size_t i = 0; i < mat_byte_size; i += elem_size) { |
There was a problem hiding this comment.
I can't see where you actually convert from int64 to int32. Neither can I see for the output where you convert from int32 to int64. Please do a test with a int64 model with int64 precision where you check that you get the same output for the same input as benchmark app or singleimagetest.
There was a problem hiding this comment.
The conversion is made here, when reading from the file, it reads lower 4 bytes from each 8 byte value
ifs.read(mat.ptr<char>() + i, elem_size); ifs.seekg(elem_size, std::ios::cur);
There was a problem hiding this comment.
C++ has functions to make the conversion from int64 to int32 explicit. Also trucking the upper half is not a common way for a conversion. Still I don't see the reverse conversion when saving the output.
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
Details:
Tickets: