<div>Thanks for your reply, I also want to ask more about this issue, I worked on transplant ME into GPU, the cost of GPU memory can be ignored and super-sampling can also moved to GPU. so, It seems the penalty of pre-calculate is trivial.</div>
<div> </div>
<div>However, when call h->mc.get_ref each time, the different branch in get_ref function will lead to decrease the degree of parallelism, under this circumstance, Is it benefit if pre-calculate of Qpel?</div>
<div> </div>
<div> </div>
<div>2009/7/15 Jason Garrett-Glaser <span dir="ltr"><<a href="mailto:darkshikari@gmail.com">darkshikari@gmail.com</a>></span><br></div>
<div class="gmail_quote">
<blockquote class="gmail_quote" style="PADDING-LEFT: 1ex; MARGIN: 0px 0px 0px 0.8ex; BORDER-LEFT: #ccc 1px solid">
<div>
<div></div>
<div class="h5">On Tue, Jul 14, 2009 at 11:49 PM, kenter<<a href="mailto:kenter83@gmail.com">kenter83@gmail.com</a>> wrote:<br>> Hi, All<br>> I currently study about the ME of x264, I noticed when perform 1/4 and<br>
> 1/2 pel search, we need to call function h->mc.get_ref first, which used to<br>> getthe aveage value of two pixel when needed.<br>> I'm puzzeled that why not pre-calculate of 1/4 and 1/2 pel in the very<br>
> beginning of ME for the whole frame, so that when each time of spel ME, we<br>> can directly get from buffer which can save the computation power of cpu?<br><br></div></div>Halfpel is already precalculated; get_ref handles the interpolation<br>
from hpel samples.<br><br>Qpel is pointless to precalculate as the memory cost is gargantuan (4x<br>that of hpel) and provides no real benefit (qpel interpolation is<br>trivial, while the extra cache cost is not).<br><br>
Dark Shikari<br>_______________________________________________<br>x264-devel mailing list<br><a href="mailto:x264-devel@videolan.org">x264-devel@videolan.org</a><br><a href="http://mailman.videolan.org/listinfo/x264-devel" target="_blank">http://mailman.videolan.org/listinfo/x264-devel</a><br>
</blockquote></div><br>