On Fri, Oct 24, 2014 at 2:06 PM, Ian Romanick <[email protected]> wrote: > On 10/24/2014 05:47 AM, Timothy Arceri wrote: >> + vec_count = count & ~0x3; >> + ui_indices_ptr = (__m128i*)ui_indices; >> + for (i = 0; i < vec_count / 4; i++) { >> + ui_indices4 = _mm_loadu_si128(&ui_indices_ptr[i]); > > How does this fare with unaligned data? My recollection is that > _mm_loadu_si128 could be quite a bit slower than _mm_load_si128. It > might be worth handling the first few values without SSE until the > pointer is aligned. > > Or my memory might be wrong.
Nope, that's a good suggestion. pixman does this a lot: http://cgit.freedesktop.org/pixman/tree/pixman/pixman-sse2.c#n582 _______________________________________________ mesa-dev mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/mesa-dev
