Hi, 2015-02-08 18:48 GMT+01:00 James Almer <jamr...@gmail.com>: >>> + %assign MMSIZE mmsize >> >> Why do that? Not a big deal: it's only for my education, if there's >> something I'm missing. > > For width 48, the COMPUTE macro is last run after an INIT_XMM cpuname, so > mmsize becomes > 16 and in the avx2 version the instructions would access the wrong data in > stack. > Doing %assign MMSIZE mmsize at the beginning of the function and using it > here makes sure > it's always 32 in avx2. > sse2 is unaffected by this, of course. > > And the reason I'm using INIT_XMM in the middle of the function for the avx2 > width 48 case > is because i couldn't find a nice and clean way to use the xm* reg aliases > with the COMPUTE > macros.
OK. Strange that it still compiled fine on Win32 here, but I haven't looked at the code generated, and I can't run the avx2 version. >>> +cglobal hevc_sao_band_filter_%1_8, 6, 6, 15, 8*mmsize*ARCH_X86_32, dst, >>> src, dststride, srcstride, offset, left >>> HEVC_SAO_BAND_FILTER_INIT 8 >> >> Why do you need room for 8 regs, and not 7? > > Remnant from before i realized i could keep m7 untouched. I'll change it. If that's the only change, then, unless someone complains, just push that version. Thanks, -- Christophe _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
