| Message ID | 20170705103318.21826-1-mfcc64@gmail.com |
|---|---|
| State | Superseded |
| Headers | show |
On 5 July 2017 at 11:33, Muhammad Faiz <mfcc64@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 <mfcc64@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. > return 0; > } > diff --git a/libavcodec/rdft.h b/libavcodec/rdft.h > index 37c40e7..0b3b675 100644 > --- a/libavcodec/rdft.h > +++ b/libavcodec/rdft.h > @@ -25,29 +25,6 @@ > #include "config.h" > #include "fft.h" > > -#if CONFIG_HARDCODED_TABLES > -# define SINTABLE_CONST const > -#else > -# define SINTABLE_CONST > -#endif > - > -#define SINTABLE(size) \ > - SINTABLE_CONST DECLARE_ALIGNED(16, FFTSample, ff_sin_##size)[size/2] > - > -extern SINTABLE(16); > -extern SINTABLE(32); > -extern SINTABLE(64); > -extern SINTABLE(128); > -extern SINTABLE(256); > -extern SINTABLE(512); > -extern SINTABLE(1024); > -extern SINTABLE(2048); > -extern SINTABLE(4096); > -extern SINTABLE(8192); > -extern SINTABLE(16384); > -extern SINTABLE(32768); > -extern SINTABLE(65536); > - > struct RDFTContext { > int nbits; > int inverse; > @@ -55,7 +32,7 @@ struct RDFTContext { > > /* pre/post rotation tables */ > const FFTSample *tcos; > - SINTABLE_CONST FFTSample *tsin; > + const FFTSample *tsin; > FFTContext fft; > void (*rdft_calc)(struct RDFTContext *s, FFTSample *z); > }; > -- > 2.9.3 > > _______________________________________________ > ffmpeg-devel mailing list > ffmpeg-devel@ffmpeg.org > http://ffmpeg.org/mailman/listinfo/ffmpeg-devel >
On Wed, Jul 5, 2017 at 10:22 PM, Rostislav Pehlivanov <atomnuker@gmail.com> wrote: > On 5 July 2017 at 11:33, Muhammad Faiz <mfcc64@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 <mfcc64@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.
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); return 0; } diff --git a/libavcodec/rdft.h b/libavcodec/rdft.h index 37c40e7..0b3b675 100644 --- a/libavcodec/rdft.h +++ b/libavcodec/rdft.h @@ -25,29 +25,6 @@ #include "config.h" #include "fft.h" -#if CONFIG_HARDCODED_TABLES -# define SINTABLE_CONST const -#else -# define SINTABLE_CONST -#endif - -#define SINTABLE(size) \ - SINTABLE_CONST DECLARE_ALIGNED(16, FFTSample, ff_sin_##size)[size/2] - -extern SINTABLE(16); -extern SINTABLE(32); -extern SINTABLE(64); -extern SINTABLE(128); -extern SINTABLE(256); -extern SINTABLE(512); -extern SINTABLE(1024); -extern SINTABLE(2048); -extern SINTABLE(4096); -extern SINTABLE(8192); -extern SINTABLE(16384); -extern SINTABLE(32768); -extern SINTABLE(65536); - struct RDFTContext { int nbits; int inverse; @@ -55,7 +32,7 @@ struct RDFTContext { /* pre/post rotation tables */ const FFTSample *tcos; - SINTABLE_CONST FFTSample *tsin; + const FFTSample *tsin; FFTContext fft; void (*rdft_calc)(struct RDFTContext *s, FFTSample *z); };
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 <mfcc64@gmail.com> --- libavcodec/Makefile | 3 +-- libavcodec/rdft.c | 70 +++++++++++++++++++---------------------------------- libavcodec/rdft.h | 25 +------------------ 3 files changed, 27 insertions(+), 71 deletions(-)