On Wed, Jul 5, 2017 at 10:22 PM, Rostislav Pehlivanov
<atomnu...@gmail.com> wrote:
> On 5 July 2017 at 11:33, Muhammad Faiz <mfc...@gmail.com> wrote:
>
>> It is redundant with costable. The first half of sintable is
>> identical with the second half of costable. The second half
>> of sintable is negative value of the first half of sintable.
>>
>> The computation is changed to handle sign of sin values,
>> so disable arm optimization.
>>
>> Signed-off-by: Muhammad Faiz <mfc...@gmail.com>
>> ---
>> libavcodec/Makefile | 3 +--
>> libavcodec/rdft.c | 70 +++++++++++++++++++-----------
>> -----------------------
>> libavcodec/rdft.h | 25 +------------------
>> 3 files changed, 27 insertions(+), 71 deletions(-)
>>
>> diff --git a/libavcodec/Makefile b/libavcodec/Makefile
>> index b440a00..59029a8 100644
>> --- a/libavcodec/Makefile
>> +++ b/libavcodec/Makefile
>> @@ -122,8 +122,7 @@ OBJS-$(CONFIG_QSV) += qsv.o
>> OBJS-$(CONFIG_QSVDEC) += qsvdec.o
>> OBJS-$(CONFIG_QSVENC) += qsvenc.o
>> OBJS-$(CONFIG_RANGECODER) += rangecoder.o
>> -RDFT-OBJS-$(CONFIG_HARDCODED_TABLES) += sin_tables.o
>> -OBJS-$(CONFIG_RDFT) += rdft.o $(RDFT-OBJS-yes)
>> +OBJS-$(CONFIG_RDFT) += rdft.o
>> OBJS-$(CONFIG_RV34DSP) += rv34dsp.o
>> OBJS-$(CONFIG_SHARED) += log2_tab.o reverse.o
>> OBJS-$(CONFIG_SINEWIN) += sinewin.o sinewin_fixed.o
>> diff --git a/libavcodec/rdft.c b/libavcodec/rdft.c
>> index c318aa8..9590673 100644
>> --- a/libavcodec/rdft.c
>> +++ b/libavcodec/rdft.c
>> @@ -28,28 +28,6 @@
>> * (Inverse) Real Discrete Fourier Transforms.
>> */
>>
>> -/* sin(2*pi*x/n) for 0<=x<n/4, followed by n/2<=x<3n/4 */
>> -#if !CONFIG_HARDCODED_TABLES
>> -SINTABLE(16);
>> -SINTABLE(32);
>> -SINTABLE(64);
>> -SINTABLE(128);
>> -SINTABLE(256);
>> -SINTABLE(512);
>> -SINTABLE(1024);
>> -SINTABLE(2048);
>> -SINTABLE(4096);
>> -SINTABLE(8192);
>> -SINTABLE(16384);
>> -SINTABLE(32768);
>> -SINTABLE(65536);
>> -#endif
>> -static SINTABLE_CONST FFTSample * const ff_sin_tabs[] = {
>> - NULL, NULL, NULL, NULL,
>> - ff_sin_16, ff_sin_32, ff_sin_64, ff_sin_128, ff_sin_256, ff_sin_512,
>> ff_sin_1024,
>> - ff_sin_2048, ff_sin_4096, ff_sin_8192, ff_sin_16384, ff_sin_32768,
>> ff_sin_65536,
>> -};
>> -
>> /** Map one real FFT into two parallel real even and odd FFTs. Then
>> interleave
>> * the two real FFTs into one complex FFT. Unmangle the results.
>> * ref: http://www.engineeringproductivitytools.com/stuff/T0001/PT10.HTM
>> @@ -73,20 +51,29 @@ static void rdft_calc_c(RDFTContext *s, FFTSample
>> *data)
>> ev.re = data[0];
>> data[0] = ev.re+data[1];
>> data[1] = ev.re-data[1];
>> - for (i = 1; i < (n>>2); i++) {
>> - i1 = 2*i;
>> - i2 = n-i1;
>> - /* Separate even and odd FFTs */
>> - ev.re = k1*(data[i1 ]+data[i2 ]);
>> - od.im = -k2*(data[i1 ]-data[i2 ]);
>> - ev.im = k1*(data[i1+1]-data[i2+1]);
>> - od.re = k2*(data[i1+1]+data[i2+1]);
>> - /* Apply twiddle factors to the odd FFT and add to the even FFT */
>> - data[i1 ] = ev.re + od.re*tcos[i] - od.im*tsin[i];
>> - data[i1+1] = ev.im + od.im*tcos[i] + od.re*tsin[i];
>> - data[i2 ] = ev.re - od.re*tcos[i] + od.im*tsin[i];
>> - data[i2+1] = -ev.im + od.im*tcos[i] + od.re*tsin[i];
>> +
>> +#define RDFT_UNMANGLE(sign0, sign1)
>> \
>> + for (i = 1; i < (n>>2); i++) {
>> \
>> + i1 = 2*i;
>> \
>> + i2 = n-i1;
>> \
>> + /* Separate even and odd FFTs */
>> \
>> + ev.re = k1*(data[i1 ]+data[i2 ]);
>> \
>> + od.im = -k2*(data[i1 ]-data[i2 ]);
>> \
>> + ev.im = k1*(data[i1+1]-data[i2+1]);
>> \
>> + od.re = k2*(data[i1+1]+data[i2+1]);
>> \
>> + /* Apply twiddle factors to the odd FFT and add to the even FFT
>> */ \
>> + data[i1 ] = ev.re + od.re*tcos[i] sign0 od.im*tsin[i];
>> \
>> + data[i1+1] = ev.im + od.im*tcos[i] sign1 od.re*tsin[i];
>> \
>> + data[i2 ] = ev.re - od.re*tcos[i] sign1 od.im*tsin[i];
>> \
>> + data[i2+1] = -ev.im + od.im*tcos[i] sign1 od.re*tsin[i];
>> \
>> + }
>> +
>> + if (s->inverse != (s->sign_convention == 1)) {
>> + RDFT_UNMANGLE(-,+)
>> + } else {
>> + RDFT_UNMANGLE(+,-)
>> }
>> +
>> data[2*i+1]=s->sign_convention*data[2*i+1];
>> if (s->inverse) {
>> data[0] *= k1;
>> @@ -113,18 +100,11 @@ av_cold int ff_rdft_init(RDFTContext *s, int nbits,
>> enum RDFTransformType trans)
>>
>> ff_init_ff_cos_tabs(nbits);
>> s->tcos = ff_cos_tabs[nbits];
>> - s->tsin = ff_sin_tabs[nbits]+(trans == DFT_R2C || trans ==
>> DFT_C2R)*(n>>2);
>> -#if !CONFIG_HARDCODED_TABLES
>> - {
>> - int i;
>> - const double theta = (trans == DFT_R2C || trans == DFT_C2R ? -1 :
>> 1) * 2 * M_PI / n;
>> - for (i = 0; i < (n >> 2); i++)
>> - s->tsin[i] = sin(i * theta);
>> - }
>> -#endif
>> + s->tsin = ff_cos_tabs[nbits] + (n >> 2);
>> s->rdft_calc = rdft_calc_c;
>>
>> - if (ARCH_ARM) ff_rdft_init_arm(s);
>> + // FIXME change arm implementation
>> + //if (ARCH_ARM) ff_rdft_init_arm(s);
>>
>>
> You could just change the sintable back to what it used to be in the arm
> init function. Or you could try to figure out how to change the asm to work
> with it.
I tried to figure out how to change asm. Unfortunately I didn't test
it because I failed to setup cross compile environment. However I'll
post the patch.
Thank's.
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
