two questions about the indirection in dwconv about xnnpack HOT 7 CLOSED

I'd need to look closer into 1.

Regarding 2. input_x and input_y are computed in unsigned arithmetic, where negative numbers wrap and become very big positive numbers. Thus, the comparison input_x < input_width is correct: if input_x is negative when viewed as a signed number, it is a very large positive number when viewed as an unsigned number, and input_x < input_width (unsigned comparison) is false.

from xnnpack.

@Maratyszcza Thanks for your reply. I didn't notice the 'size_t' type. So using unsigned type here can reduce one branch which should have been added, right?

from xnnpack.

Right, using unsigned type removes half of comparisons (and branches)

from xnnpack.

Thansk, I get this brilliant trick. What about the first problem?

from xnnpack.

The first problem is a bug. It should be kernel_size + (output_width-1) * step_width * kernel_height as you suggested.

from xnnpack.

@Maratyszcza got it. Thanks!

from xnnpack.

Fixed in 03ff294

from xnnpack.