<div dir="ltr">Hi Jean-Marc,<div><br></div><div>Attached is the silk_warped_autocorrelation_FIX_neon() which implements your idea.</div><div><span style="color:rgb(53,53,53);font-family:sans-serif;white-space:pre-wrap"><br></span></div><div><span style="color:rgb(53,53,53);font-family:sans-serif;white-space:pre-wrap">Speed improvement vs the previous optimization:</span></div><div><span style="color:rgb(53,53,53);font-family:sans-serif;white-space:pre-wrap"><br></span></div><div><p style="white-space:pre-wrap;color:rgb(53,53,53);font-family:sans-serif">Complexity 0-4: Doesn't call this function.
Complexity 5: 2.1% (order = 16)
Complexity 6: 1.0% (order = 20)
Complexity 8: 0.1% (order = 24)
Complexity 10: 0.1% (order = 24)</p><p style="white-space:pre-wrap;color:rgb(53,53,53);font-family:sans-serif">Code size of <span style="font-family:arial,sans-serif;color:rgb(34,34,34)">silk_warped_autocorrelation_FIX_neon()</span> changes from 2,644 bytes to 3,228 bytes.</p><p style="white-space:pre-wrap;color:rgb(53,53,53);font-family:sans-serif">The reason of larger code size is that the new optimization specializes order 16, 20 and 24. If only keeping order 24 specialization, the code still works and the code size is smaller, but the encoder speed will drop 4.0% for Complexity 5 and 2.0% for Complexity 6. Anyway, the new code is easier to understand and maintain.</p><p style="white-space:pre-wrap;color:rgb(53,53,53);font-family:sans-serif">Thanks,<br></p><p style="white-space:pre-wrap;color:rgb(53,53,53);font-family:sans-serif">Linfeng</p></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Tue, Feb 7, 2017 at 8:58 AM, Linfeng Zhang <span dir="ltr"><<a href="mailto:linfengz@google.com" target="_blank">linfengz@google.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr">Hi Jean-Marc,<div><br></div><div>Thanks for your suggestions. Will get back to you once we have some updates.</div><span class="HOEnZb"><font color="#888888"><div><br></div><div>Linfeng</div></font></span></div><div class="HOEnZb"><div class="h5"><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Feb 6, 2017 at 5:47 PM, Jean-Marc Valin <span dir="ltr"><<a href="mailto:jmvalin@jmvalin.ca" target="_blank">jmvalin@jmvalin.ca</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Hi Linfeng,<br>
<span><br>
On 06/02/17 07:18 PM, Linfeng Zhang wrote:<br>
> This is a great idea. But the order (psEncC->shapingLPCOrder) can be<br>
> configured to 12, 14, 16, 20 and 24 according to complexity parameter.<br>
><br>
> It's hard to get a universal function to handle all these orders<br>
> efficiently. Any suggestions?<br>
<br>
</span>I can think of two ways of handling larger orders. The obvious one is<br>
simply to add an inner loop of the form:<br>
for (i=0;i<order;i+=VECTOR_SIZE)<br>
I think what may be more efficient is to simply have a small "order-N"<br>
(N=4 or 8) kernel that not only computes the correlation of order N, but<br>
then spits out the signal after the N-stage all-pass is applied. The<br>
kernel would look like:<br>
<br>
void autocorr_kernel4(int *corr, int *orig, int *input, int *output, int<br>
len) {<br>
/* Implement vectorized order-4 filter (could also be order 8)<br>
as described in previous email and outputs the filtered signal.<br>
*/<br>
}<br>
<br>
and then the full function would run the kernel multiple times and look<br>
like:<br>
<br>
void full_autocorr(int *corr, int *orig, int len, int order) {<br>
int i;<br>
int tmp[MAX_SIZE];<br>
int *in = orig;<br>
for (i=0;i<order;i+=4) {<br>
autocorr_kernel4(corr+i, orig, in, tmp, len);<br>
/* Make subsequent calls use the filtered signal as input. */<br>
in = tmp;<br>
}<br>
}<br>
<br>
I think the should not only reduce/eliminate the prologue/epilogue<br>
problem, but it should also be more efficient since almost all vectors<br>
processed would use the full size.<br>
<br>
Maybe a third option (not sure it's a good idea, but still mentioning<br>
it) would be to have a function that hardcodes order=24 and discards the<br>
larger values that aren't needed. Since the smallest order seems to be<br>
16, it wouldn't be much of a waste and the code might end up running<br>
faster for the higher orders.<br>
<br>
Cheers,<br>
<br>
Jean-Marc<br>
<span><br>
<br>
> Thanks,<br>
> Linfeng<br>
><br>
> On Mon, Feb 6, 2017 at 12:40 PM, Jean-Marc Valin <<a href="mailto:jmvalin@jmvalin.ca" target="_blank">jmvalin@jmvalin.ca</a><br>
</span><div><div class="m_-6695158637204629633h5">> <mailto:<a href="mailto:jmvalin@jmvalin.ca" target="_blank">jmvalin@jmvalin.ca</a>>> wrote:<br>
><br>
> Hi Linfeng,<br>
><br>
> On 06/02/17 02:51 PM, Linfeng Zhang wrote:<br>
> > However, the critical thing is that all the states in each stage when<br>
> > processing input[i] are reused by the next input[i+1]. That is<br>
> > input[i+1] must wait input[i] for 1 stage, and input[i+2] must wait<br>
> > input[i+1] for 1 stage, etc.<br>
><br>
> That is indeed the tricky part... and the one I think you could do<br>
> slightly differently. If you approach the problem in terms of computing<br>
> chunks of the inputs N samples at a time, then indeed the approach you<br>
> are describing is the only solution. What I was proposing though is to<br>
> instead chop the "order" in chunks of N. Using your notation, you would<br>
> be doing:<br>
><br>
> PROC( in0(s0))<br>
> PROC( in0(s1) in1(s0))<br>
> PROC( in0(s2) in1(s1) in2(s0))<br>
> PROC( in0(s3) in1(s2) in2(s1) in3(s0))<br>
> PROC( in0(s4) in1(s3) in2(s2) in3(s1) in4(s0))<br>
> PROC( in0(s5) in1(s4) in2(s3) in3(s2) in4(s1) in5(s0))<br>
> PROC( in0(s6) in1(s5) in2(s4) in3(s3) in4(s2) in5(s1) in6(s0))<br>
> PROC(in0(s7) in1(s6) in2(s5) in3(s4) in4(s3) in5(s2) in6(s1) in7(s0))<br>
> PROC(in1(s7) in2(s6) in3(s5) in4(s4) in5(s3) in6(s2) in7(s1) in8(s0))<br>
> PROC(in2(s7) in3(s6) in4(s5) in5(s4) in6(s3) in7(s2) in8(s1) in9(s0))<br>
> PROC(in3(s7) in4(s6) in5(s5) in6(s4) in7(s3) in8(s2) in9(s1)in10(s0))<br>
> PROC(in4(s7) in5(s6) in6(s5) in7(s4) in8(s3) in9(s2)in10(s1)in11(s0))<br>
> ...and so on until the end of the input vector<br>
><br>
> The difference is that it's now the input vector that "slides" and the<br>
> "state" values sy that remain in the same place. There's still a<br>
> prologue, but you can easily get rid of it by (implicitly) zero-padding<br>
> the in vector during the initialization phase (start with a zero vector<br>
> and real one value at a time). Getting rid of the epilogue is a little<br>
> trickier, but I think it can be done.<br>
><br>
> Cheers,<br>
><br>
> Jean-Marc<br>
><br>
> > Then it becomes this<br>
> ><br>
> > FOR in=0 to N WITH in+=8<br>
> > PROC(in0(s0)) /* prolog 0 */<br>
> > PROC(in0(s1) in1(s0)) /* prolog 1 */<br>
> > PROC(in0(s2) in1(s1) in2(s0)) /* prolog 2 */<br>
> > PROC(in0(s3) in1(s2) in2(s1) in3(s0)) /* prolog 3 */<br>
> > PROC(in0(s4) in1(s3) in2(s2) in3(s1) in4(s0)) /* prolog 4 */<br>
> > PROC(in0(s5) in1(s4) in2(s3) in3(s2) in4(s1) in5(s0)) /* prolog 5 */<br>
> > PROC(in0(s6) in1(s5) in2(s4) in3(s3) in4(s2) in5(s1) in6(s0)) /*<br>
> > prolog 6 */<br>
> > PROC(in0(s7) in1(s6) in2(s5) in3(s4) in4(s3) in5(s2) in6(s1)<br>
> in7(s0))<br>
> > /* fully process 8 inputs */<br>
> > PROC(in0(s8) in1(s7) in2(s6) in3(s5) in4(s4) in5(s3) in6(s2)<br>
> in7(s1))<br>
> > /* continue */<br>
> > PROC(in0(s9) in1(s8) in2(s7) in3(s6) in4(s5) in5(s4) in6(s3)<br>
> in7(s2))<br>
> > /* continue */<br>
> > PROC(in0(s10) in1(s9) in2(s8) in3(s7) in4(s6) in5(s5) in6(s4)<br>
> in7(s3))<br>
> > /* continue */<br>
> > PROC(in1(s10) in2(s9) in3(s8) in4(s7) in5(s6) in6(s5) in7(s4)) /*<br>
> > epilog 0 */<br>
> > PROC(in2(s10) in3(s9) in4(s8) in5(s7) in6(s6) in7(s5)) /* epilog<br>
> 1 */<br>
> > PROC(in3(s10) in4(s9) in5(s8) in6(s7) in7(s6)) /* epilog 2 */<br>
> > PROC(in4(s10) in5(s9) in6(s8) in7(s7)) /* epilog 3 */<br>
> > PROC(in5(s10) in6(s9) in7(s8)) /* epilog 4 */<br>
> > PROC(in6(s10) in7(s9)) /* epilog 5 */<br>
> > PROC(in7(s10)) /* epilog 6 */<br>
> > END FOR<br>
> ><br>
> > And<br>
> > PROC(in0(s7) in1(s6) in2(s5) in3(s4) in4(s3) in5(s2) in6(s1)<br>
> in7(s0))<br>
> > /* fully process 8 inputs */<br>
> > PROC(in0(s8) in1(s7) in2(s6) in3(s5) in4(s4) in5(s3) in6(s2)<br>
> in7(s1))<br>
> > /* continue */<br>
> > PROC(in0(s9) in1(s8) in2(s7) in3(s6) in4(s5) in5(s4) in6(s3)<br>
> in7(s2))<br>
> > /* continue */<br>
> > is actually the expansion of the kernel loop<br>
> > FOR i=0 TO order-6 WITH i++<br>
> > PROC(in0(si+7) in1(si+6) in2(si+5) in3(si+4) in4(si+3) in5(si+2)<br>
> > in6(si+1) in7(si+0))<br>
> > END FOR<br>
> ><br>
> > The worst thing is that corr_QC[] is so sensitive that any extra<br>
> > processing will make them wrong and propagate to the next loop (next 8<br>
> > inputs). state_QS[] is a little better but still very sensitive. For<br>
> > instance, if adding PROC(in0(s11') in1(s10) in2(s9) in3(s8) in4(s7)<br>
> > in5(s6) in6(s5) in7(s4)) to the kernel loop (by looping one more time)<br>
> > and remove epilog 0, then all final results will be wrong.<br>
> ><br>
> > That's why the prolog and epilog cannot be saved to the best of my<br>
> > knowledge.<br>
> ><br>
> > The assembly size of silk_warped_autocorrelation_FI<wbr>X_neon() is about<br>
> > 2,744 bytes. Compared with the C code size (about 452 bytes), it's 2.3<br>
> > KB larger. Considering silk_warped_autocorrelation_FI<wbr>X_c() is the<br>
> second<br>
> > place CPU heavy function in fixed-point, and our testing shows up<br>
> to 7%<br>
> > CPU run time saving of the total encoder with this optimization (at<br>
> > Complexity 8), maybe we can take the I-cache burden even if finally we<br>
> > still cannot remove the big chunk of prolog and epilog.<br>
> ><br>
> > Thanks,<br>
> > Linfeng Zhang<br>
> ><br>
> > On Sat, Feb 4, 2017 at 4:17 PM, Jean-Marc Valin<br>
> <<a href="mailto:jmvalin@jmvalin.ca" target="_blank">jmvalin@jmvalin.ca</a> <mailto:<a href="mailto:jmvalin@jmvalin.ca" target="_blank">jmvalin@jmvalin.ca</a>><br>
</div></div><div><div class="m_-6695158637204629633h5">> > <mailto:<a href="mailto:jmvalin@jmvalin.ca" target="_blank">jmvalin@jmvalin.ca</a> <mailto:<a href="mailto:jmvalin@jmvalin.ca" target="_blank">jmvalin@jmvalin.ca</a>>>> wrote:<br>
> ><br>
> > Hi Felicia,<br>
> ><br>
> > I've had time to work through the math in the original<br>
> function and I'm<br>
> > pretty sure it's possible to vectorize this without the huge<br>
> > prologue/epilogue.<br>
> ><br>
> > For the simple case where order = vector size = N (but it<br>
> should easily<br>
> > generalize to larger order), what I came up with is:<br>
> ><br>
> > initialize X, Y, M, C to vector of zeros<br>
> ><br>
> > for i=0 to N+order<br>
> > T = [x(i), Y(0:N-2)]<br>
> > Y = M + coeff * (Y - T)<br>
> > M = T<br>
> > X = [x(i), X(0:N-1)]<br>
> > C = C + X*Y<br>
> ><br>
> > I think something similar to this (assuming I didn't mess up any<br>
> > details) should give you the correlations in vector C. Did I miss<br>
> > anything?<br>
> ><br>
> > Cheers,<br>
> ><br>
> > Jean-Marc<br>
> ><br>
> ><br>
> > On 31/01/17 12:30 PM, Felicia Lim wrote:<br>
> > > Hi,<br>
> > ><br>
> > > Attached is a patch with arm neon optimizations for<br>
> > > silk_warped_autocorrelation_FI<wbr>X(). Please review.<br>
> > ><br>
> > > Thanks,<br>
> > > Felicia<br>
> > ><br>
> > ><br>
> > > ______________________________<wbr>_________________<br>
> > > opus mailing list<br>
</div></div>> > > <a href="mailto:opus@xiph.org" target="_blank">opus@xiph.org</a> <mailto:<a href="mailto:opus@xiph.org" target="_blank">opus@xiph.org</a>> <mailto:<a href="mailto:opus@xiph.org" target="_blank">opus@xiph.org</a><br>
<span>> <mailto:<a href="mailto:opus@xiph.org" target="_blank">opus@xiph.org</a>>><br>
> > > <a href="http://lists.xiph.org/mailman/listinfo/opus" rel="noreferrer" target="_blank">http://lists.xiph.org/mailman/<wbr>listinfo/opus</a><br>
> <<a href="http://lists.xiph.org/mailman/listinfo/opus" rel="noreferrer" target="_blank">http://lists.xiph.org/mailma<wbr>n/listinfo/opus</a>><br>
> > <<a href="http://lists.xiph.org/mailman/listinfo/opus" rel="noreferrer" target="_blank">http://lists.xiph.org/mailma<wbr>n/listinfo/opus</a><br>
> <<a href="http://lists.xiph.org/mailman/listinfo/opus" rel="noreferrer" target="_blank">http://lists.xiph.org/mailma<wbr>n/listinfo/opus</a>>><br>
> > ><br>
> > _____________________________<wbr>__________________<br>
> > opus mailing list<br>
</span>> > <a href="mailto:opus@xiph.org" target="_blank">opus@xiph.org</a> <mailto:<a href="mailto:opus@xiph.org" target="_blank">opus@xiph.org</a>> <mailto:<a href="mailto:opus@xiph.org" target="_blank">opus@xiph.org</a><br>
<span>> <mailto:<a href="mailto:opus@xiph.org" target="_blank">opus@xiph.org</a>>><br>
> > <a href="http://lists.xiph.org/mailman/listinfo/opus" rel="noreferrer" target="_blank">http://lists.xiph.org/<wbr>mailman/listinfo/opus</a><br>
> <<a href="http://lists.xiph.org/mailman/listinfo/opus" rel="noreferrer" target="_blank">http://lists.xiph.org/mailma<wbr>n/listinfo/opus</a>><br>
</span>> > <<a href="http://lists.xiph.org/mailman/listinfo/opus" rel="noreferrer" target="_blank">http://lists.xiph.org/mailma<wbr>n/listinfo/opus</a><br>
> <<a href="http://lists.xiph.org/mailman/listinfo/opus" rel="noreferrer" target="_blank">http://lists.xiph.org/mailma<wbr>n/listinfo/opus</a>>><br>
> ><br>
> ><br>
><br>
><br>
</blockquote></div><br></div>
</div></div></blockquote></div><br></div>