On Sat, Dec 19, 2015 at 4:21 AM, Andreas Cadhalpun
<andreas.cadhal...@googlemail.com> wrote:
> On 18.12.2015 16:52, Ganesh Ajjanagadde wrote:
>> On Tue, Dec 15, 2015 at 9:53 PM, Ganesh Ajjanagadde
>> <gajjanaga...@gmail.com> wrote:
>>> This uses ff_exp2fi to get a speedup (~ 6x).
>>>
>>> sample benchmark (Haswell, GNU/Linux):
>>> old:
>>> 19102 decicycles in sbr_dequant, 1023 runs, 1 skips
>>> 19002 decicycles in sbr_dequant, 2045 runs, 3 skips
>>> 17638 decicycles in sbr_dequant, 4093 runs, 3 skips
>>> 15825 decicycles in sbr_dequant, 8189 runs, 3 skips
>>> 16404 decicycles in sbr_dequant, 16379 runs, 5 skips
>>>
>>> new:
>>> 3063 decicycles in sbr_dequant, 1024 runs, 0 skips
>>> 3049 decicycles in sbr_dequant, 2048 runs, 0 skips
>>> 2968 decicycles in sbr_dequant, 4096 runs, 0 skips
>>> 2818 decicycles in sbr_dequant, 8191 runs, 1 skips
>>> 2853 decicycles in sbr_dequant, 16383 runs, 1 skips
>>>
>>> Signed-off-by: Ganesh Ajjanagadde <gajjanaga...@gmail.com>
>>> ---
>>> libavcodec/aacsbr.c | 34 ++++++++++++++++++++++------------
>>> 1 file changed, 22 insertions(+), 12 deletions(-)
>>>
>>> diff --git a/libavcodec/aacsbr.c b/libavcodec/aacsbr.c
>>> index d1e3a91..15956e3 100644
>>> --- a/libavcodec/aacsbr.c
>>> +++ b/libavcodec/aacsbr.c
>>> @@ -33,6 +33,7 @@
>>> #include "aacsbrdata.h"
>>> #include "aacsbr_tablegen.h"
>>> #include "fft.h"
>>> +#include "internal.h"
>>> #include "aacps.h"
>>> #include "sbrdsp.h"
>>> #include "libavutil/internal.h"
>>> @@ -73,15 +74,22 @@ static void sbr_dequant(SpectralBandReplication *sbr,
>>> int id_aac)
>>> {
>>> int k, e;
>>> int ch;
>>> -
>>> + static const double exp2_tab[2] = {1, M_SQRT2};
>>> if (id_aac == TYPE_CPE && sbr->bs_coupling) {
>>> - float alpha = sbr->data[0].bs_amp_res ? 1.0f : 0.5f;
>>> - float pan_offset = sbr->data[0].bs_amp_res ? 12.0f : 24.0f;
>>> + int pan_offset = sbr->data[0].bs_amp_res ? 12 : 24;
>>> for (e = 1; e <= sbr->data[0].bs_num_env; e++) {
>>> for (k = 0; k < sbr->n[sbr->data[0].bs_freq_res[e]]; k++) {
>>> - float temp1 = exp2f(sbr->data[0].env_facs_q[e][k] * alpha
>>> + 7.0f);
>>> - float temp2 = exp2f((pan_offset -
>>> sbr->data[1].env_facs_q[e][k]) * alpha);
>>> - float fac;
>>> + float temp1, temp2, fac;
>>> + if (sbr->data[0].bs_amp_res) {
>>> + temp1 = ff_exp2fi(sbr->data[0].env_facs_q[e][k] + 7);
>>> + temp2 = ff_exp2fi(pan_offset -
>>> sbr->data[1].env_facs_q[e][k]);
>>> + }
>>> + else {
>>> + temp1 = ff_exp2fi((sbr->data[0].env_facs_q[e][k]>>1) +
>>> 7) *
>>> + exp2_tab[sbr->data[0].env_facs_q[e][k] & 1];
>>> + temp2 = ff_exp2fi((pan_offset -
>>> sbr->data[1].env_facs_q[e][k])>>1) *
>>> + exp2_tab[(pan_offset -
>>> sbr->data[1].env_facs_q[e][k]) & 1];
>>> + }
>>> if (temp1 > 1E20) {
>>> av_log(NULL, AV_LOG_ERROR, "envelope scalefactor
>>> overflow in dequant\n");
>>> temp1 = 1;
>>> @@ -93,8 +101,8 @@ static void sbr_dequant(SpectralBandReplication *sbr,
>>> int id_aac)
>>> }
>>> for (e = 1; e <= sbr->data[0].bs_num_noise; e++) {
>>> for (k = 0; k < sbr->n_q; k++) {
>>> - float temp1 = exp2f(NOISE_FLOOR_OFFSET -
>>> sbr->data[0].noise_facs_q[e][k] + 1);
>>> - float temp2 = exp2f(12 - sbr->data[1].noise_facs_q[e][k]);
>>> + float temp1 = ff_exp2fi(NOISE_FLOOR_OFFSET -
>>> sbr->data[0].noise_facs_q[e][k] + 1);
>>> + float temp2 = ff_exp2fi(12 -
>>> sbr->data[1].noise_facs_q[e][k]);
>>> float fac;
>>> av_assert0(temp1 <= 1E20);
>>> fac = temp1 / (1.0f + temp2);
>>> @@ -104,11 +112,13 @@ static void sbr_dequant(SpectralBandReplication *sbr,
>>> int id_aac)
>>> }
>>> } else { // SCE or one non-coupled CPE
>>> for (ch = 0; ch < (id_aac == TYPE_CPE) + 1; ch++) {
>>> - float alpha = sbr->data[ch].bs_amp_res ? 1.0f : 0.5f;
>>> for (e = 1; e <= sbr->data[ch].bs_num_env; e++)
>>> for (k = 0; k < sbr->n[sbr->data[ch].bs_freq_res[e]]; k++){
>>> - sbr->data[ch].env_facs[e][k] =
>>> - exp2f(alpha * sbr->data[ch].env_facs_q[e][k] +
>>> 6.0f);
>>> + if (sbr->data[ch].bs_amp_res)
>>> + sbr->data[ch].env_facs[e][k] =
>>> ff_exp2fi(sbr->data[ch].env_facs_q[e][k] + 6);
>>> + else
>>> + sbr->data[ch].env_facs[e][k] =
>>> ff_exp2fi((sbr->data[ch].env_facs_q[e][k]>>1) + 6)
>>> + *
>>> exp2_tab[sbr->data[ch].env_facs_q[e][k] & 1];
>>> if (sbr->data[ch].env_facs[e][k] > 1E20) {
>>> av_log(NULL, AV_LOG_ERROR, "envelope scalefactor
>>> overflow in dequant\n");
>>> sbr->data[ch].env_facs[e][k] = 1;
>>> @@ -118,7 +128,7 @@ static void sbr_dequant(SpectralBandReplication *sbr,
>>> int id_aac)
>>> for (e = 1; e <= sbr->data[ch].bs_num_noise; e++)
>>> for (k = 0; k < sbr->n_q; k++)
>>> sbr->data[ch].noise_facs[e][k] =
>>> - exp2f(NOISE_FLOOR_OFFSET -
>>> sbr->data[ch].noise_facs_q[e][k]);
>>> + ff_exp2fi(NOISE_FLOOR_OFFSET -
>>> sbr->data[ch].noise_facs_q[e][k]);
>>> }
>>> }
>>> }
>>> --
>>> 2.6.4
>>>
>>
>> ping for aac maintainers: Rostislav? thanks.
>
> I'm not an aac maintainer, but the change looks correct and the speedup nice,
> so this
> patch should be OK.
>
> One caveat is that NOISE_FLOOR_OFFSET is defined as FIXR(6.0f) and is thus a
> float here.
> Since this patch makes that unnecessary, it's probably a good idea to change
> that
> to simply 6.
Good point, don't want the overhead/impreciseness of possible float to
int conversion, though I think a good compiler can remove it since 6.0
is exactly representable.
Pushed, thanks.
>
> Best regards,
> Andreas
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
