Closed zihaomu closed 2 years ago
@zihaomu, this is not a bug. Please, check the condition on the next line:
if (j + 8 > out_width) {
...
this is so-called "Halide trick" of handling loop tails
Hi @vpisarev, I get your point now. A very useful trick so that the end of the loop can still use SIMD acceleration. That means there some tails will be computed twice.
And when pbptr
is true, it may cause bugs if you still use another trick you told me before (pb_data
was saved as Conv's output_data
).
If pb_data
is different from output_data
, it works fine.
Fix possible out-of-bounds bug