Add SIMD version of grayscale #2214
Conversation
yunline
added
the
Performance
| mm_zero = _mm_setzero_si128(); | ||
| mm_alpha_mask = _mm_cvtsi32_si128(amask); | ||
| mm_rgb_mask = _mm_cvtsi32_si128(rgbmask); | ||
| mm_two_five_fives = _mm_set_epi64x(0x00FF00FF00FF00FF, 0x00FF00FF00FF00FF); |
| while (pixel_batch_counter--) { | ||
| mm_src = _mm_cvtsi32_si128(*srcp); | ||
| /*mm_src = 0x000000000000000000000000AARRGGBB*/ | ||
| mm_alpha = _mm_subs_epu8(mm_src, mm_rgb_mask); |
There was a problem hiding this comment.
I've never been super clear about what saturation means in the intrinsics, so this is a very interesting example.
https://en.wikipedia.org/wiki/Saturation_arithmetic tells me it is arithmetic clamped to a range.
So rgb could be anything, but you subtract 255 with saturation--> meaning it will always end as zero. Very clever stuff. I wonder if there's masking in other SIMD stuff that could be converted to this for a small efficiency gain.
|
|
||
| mm_dst = _mm_adds_epu8( | ||
| _mm_adds_epu8( | ||
| _mm_shufflelo_epi16(mm_dst, _MM_SHUFFLE(0, 0, 0, 0)), |
There was a problem hiding this comment.
Maybe we should provide our own version of _MM_SHUFFLE so we're not dependent on this implementation detail of sse2neon.
It took me a bit to parse whether this was something by us, by the intrinsics header, or by sse2neon.
|
Closing this as splitting it up into multiple PRs did not seem to work very well. |
Splitting #2042 into two parts. This will just be the grayscale implementation excluding the setup #2042 should be resolved first.