osamu620 / jpegenc Goto Github PK

Hello again @osamu620, it is great to see your nice SIMD code :)

We'd be happy to link to your project from the Highway readme.

Some thoughts on the code, in the hope that they are useful:

In quantization.cc and block_coding.cpp maybe we could HWY_DYNAMIC_DISPATCH only once because it is an indirect function call, which is relatively expensive. We could move the for(;;) loop inside quantize_fwd?

In dct.cpp, ZipLower(s32, ..) is the same as BitCast(s32, InterleaveLower(s16, ..)) and it could shorten the code a bit.

In color.cpp, it looks like coeffs is 8 int16 constants, and then copied into each 128-bit block.
We can shrink the array to 8 elements if we use hn::LoadDup128 to load it. And Broadcast<> could be used instead of BroadcastLane<>: it is a bit cheaper on AVX2 because it is allowed to assume that all 128-bit blocks are equal, which is true here.

I'm also curious whether you see anything we can improve or better document in the Highway library?

Currently, there is a restriction to the size of input images; width and height shall be multiple of 16.