similar to: Possible bug in "pitch_downsample"

I wrote a proof of concept regarding the cpu capabilities runtime detection and choice of optimized function. I follow design which had been discussed on IRC. Also, i notice a little drawback: we must propagate the arch index through functions which don't have codec state as argument. However, if it's look good, i will continue to implement it. Best regards, -- Aur?lien Zanelli

> On Nov 16, 2015, at 4:42 PM, Jonathan Lennox <jonathan at vidyo.com> wrote: > > I haven?t yet tried replacing SIG2WORD16 (or silk_ADD_SAT32/silk_SUB_SAT32) with Neon intrinsics. That?s an obvious next step. This doesn?t show any appreciable speed difference in my tests, but the code is obviously better by inspection (all three of these map directly to a single Aarch64

Should call celt_inner_prod(). --- celt/bands.c | 7 ++++--- celt/bands.h | 2 +- celt/celt_encoder.c | 6 +++--- celt/pitch.c | 2 +- src/opus_multistream_encoder.c | 2 +- 5 files changed, 10 insertions(+), 9 deletions(-) diff --git a/celt/bands.c b/celt/bands.c index bbe8a4c..1ab24aa 100644 --- a/celt/bands.c +++ b/celt/bands.c

> On Nov 19, 2015, at 5:47 PM, John Ridges <jridges at masque.com> wrote: > > Any speedup from the intrinsics may just be swamped by the rest of the encode/decode process. But I think you really want SIG2WORD16 to be (vqmovns_s32(PSHR32((x), SIG_SHIFT))) Yes, you?re right. I forgot to run the vectors under qemu with my previous version (oh, the embarrassment!) Fixed forthcoming

- Use MAC16_16 macros instead of (sum += a*b) and unroll a loop by 2. It increase performance when using optimized macros (ex: ARMv5E). A possible side effect of loop unroll is that i don't check for odd length here. - Add NEON version of FIR filter and autocorr -- Aur?lien Zanelli Parrot SA 174, quai de Jemmapes 75010 Paris France -------------- next part -------------- diff --git

> I was wrong, it is the encoder that is not working, and it stopped working > in build 11103. The log message for this build is "another 640 bytes > removed from the encoder state (using the input data instead of copying it > to st->frame/st->inBuf)". Only nb_celp.c/h are changed in this build. What > I am seeing out of the decoder is an extremely low signal

Hi Jean-Marc, Hi Jim, I have also seen some problems with the 1.1.8 release on the C55x. So far I have boiled down the issues to the following: 1) We need our own "fixed_xx.h" header file. I don't know why, and haven't had time to investigate, but there is a definite improvement when I use the attached fixed_c55x.h file which has turned all the maths into inline functions.

From: Thorvald Natvig <slicer at users.sourceforge.net> Some optimizing compilers miscompile the current SSE optimizations when full optimizations are enabled. By using output value pointer instead of a return value, we can bypass this misbehaviour. --- libspeex/resample.c | 8 ++++---- libspeex/resample_sse.h | 24 ++++++++---------------- 2 files changed, 12 insertions(+), 20

Please ignore my previous mail and patch, there is a new version :). Patch changes are: - Use MAC16_16 macros instead of (sum += a*b) and unroll a loop by 2. It increase performance when using optimized macros (ex: ARMv5E). A possible side effect of loop unroll is that i don't check for odd length here. - Add NEON version of FIR filter and autocorr - Add a section in autoconf in order to

Hi! It seems that using a different compiler (I assume the development is done with MSC) uncovers various issues with the floating point. Here is the second one I'm seeing - DOMAIN error from sqrt(). In the open_loop_nbest_pitch(), this line: g = DIV32(corr[i-start], 10+SHR32(MULT16_16(spx_sqrt(e0),spx_sqrt(energy[i-start]) ),6)); The values are: e0 is 7.2 i is 142, start is 17

Speex tells me that the decoder is always 5 ms, but it says that the encoder is 5 ms for NB, 8.9375 ms for WB, and 10.90625 ms for UWB. Is there an extra frame of delay in the encoder that isn't otherwise accounted for? John Ridges Jean-Marc Valin wrote: > Quoting John Ridges <jridges at masque.com>: > >> I also need to know the precise delays from Speex but I used

Thanks, I look forward to seeing what you find out. BTW, I was wondering if you tried replacing the SIG2WORD16 macro using the vqmovns_s32 intrinsic? I'm sure it would be faster than the C code, but in the grand scheme of things it might not make much difference. On 11/13/2015 12:15 PM, Jonathan Lennox wrote: >> On Nov 13, 2015, at 1:51 PM, John Ridges <jridges at masque.com>

By my reckoning the confluent hypergoemetric functions should have the following values: M(-.25;1;-.5) = 1.11433 M(-.25;1;-1) = 1.21088 M(-.25;1;-1.5) = 1.29385 M(-.25;1;-2) = 1.36627 M(-.25;1;-2.5) = 1.43038 M(-.25;1;-3) = 1.48784 M(-.25;1;-3.5) = 1.53988 M(-.25;1;-4) = 1.58747 M(-.25;1;-4.5) = 1.63134 M(-.25;1;-5) = 1.67206 M(-.25;1;-5.5) = 1.71009 M(-.25;1;-6) = 1.74579 M(-.25;1;-6.5) =

I got the approximation from a Google book: http://books.google.com/books?id=2CAqsF-RebgC&pg=PA385 Page 392, formula (10.33) Using this formula, you're right, hypergeom_gain() would *not* converge to 1 for large x, but would instead be gamma(1.25)/sqrt(sqrt(x)) which would approach zero. Now if the formula for the hypergeometric gain were instead gamma(1.5) * M(-.5;1;-x) / sqrt(x)

Hi all, This is Linfeng Zhang from Google. I'll work on ARM NEON optimization in the next few months. I'm submitting 2 patches in the following couple of emails, which have the new created celt_fir_neon(). I revised celt_fir_c() to not pass in argument "mem" in Patch 1. If there are concerns to this change, please let me know. Many thanks to your comments. Linfeng Zhang

Hello Jean-Marc, > Got any luck with the Blackfin stuff? One thing I forgot mentioning that > would probably help a lot in narrowing down the bug is to simply disable > assembly functions one by one and see which one breaks Speex. sorry that I didn't reply for some days. Unfortunately I had some other serious problems with my software which are not related to Speex. I hope to solve

Hi Jonathan, I'm sorry to bring this up again, and I don't want to beat a dead horse, but I was very surprised by your benchmarks so I took a little closer look. I think what's happening is that it's a little unfair to compare the ARM64 inline assembly to the C code, because looking at the C macros in "fixed_generic.h" for MULT16_32_Q16 and MULT16_32_Q15 you find

hi, I am using speex1.2beta2 on a TI 54x on narrow band I have been trying to get speex to work for a while now, and it's been a real teeter-totter ride. For a long time I noticed that I will get a project to work and then without changing any code and programming it to an eprom/flash the project will not work. It turns out it was a value called innov_save. I found this bugger by zero

JM, I also need to know the precise delays from Speex but I used the SPEEX_GET_LOOKAHEAD control requests to determine them (plus the "speex_resampler_get_output_latency" function from the resampler). The returned values from the Speex lookahead request don't seem to match with the values you gave Alexander. Am I doing this wrong? Thanks, John Ridges speex-dev-request at

Hi Jim, Thanks a lot for investigating. It definitely makes sense now. I'll fix the problem now. Is there any other place where you see that same (or similar) problem? Jean-Marc Jim Crichton a ?crit : > Jean-Marc, > > I dug into this further, and found that the warning occurred when PSHR32 > had a shift greater than 15. > > in fixed_generic.h, PSHR32 is defined as: >