| Message ID | GV1SPRMB0063AFB9D369C21EDE6DB4778F532@GV1SPRMB0063.EURP250.PROD.OUTLOOK.COM |
|---|---|
| State | Accepted |
| Commit | 42f6dfc42e81725f466001b64f0536622a6682ed |
| Headers | show |
| Series | [FFmpeg-devel,1/4] avcodec/jpeg2000dec, j2kenc: Constify where appropriate | expand |
| Context | Check | Description |
|---|---|---|
| yinshiyou/make_loongarch64 | success | Make finished |
| yinshiyou/make_fate_loongarch64 | success | Make fate finished |
| andriy/make_x86 | success | Make finished |
| andriy/make_fate_x86 | success | Make fate finished |
lör 2024-02-17 klockan 21:05 +0100 skrev Andreas Rheinhardt: > The call to ff_exp2fi() here always uses arguments in the normal > range, so that the branches in ff_exp2fi() are unnecessary. > This is so because JPEG2000 itself only supports up to > 128 bits per component per pixel (we only support far less); > furthermore, expn is always 0..31 for the decoder and also > sane for the encoder, so that the difference between these > two values is always in the normal range of -126..128. Any measurable improvement in decode speed? /Tomas
Tomas Härdin: > lör 2024-02-17 klockan 21:05 +0100 skrev Andreas Rheinhardt: >> The call to ff_exp2fi() here always uses arguments in the normal >> range, so that the branches in ff_exp2fi() are unnecessary. >> This is so because JPEG2000 itself only supports up to >> 128 bits per component per pixel (we only support far less); >> furthermore, expn is always 0..31 for the decoder and also >> sane for the encoder, so that the difference between these >> two values is always in the normal range of -126..128. > > Any measurable improvement in decode speed? > I don't think it is measurable; my aim was to move ff_exp2fi() out of internal.h (where it does not belong). - Andreas
mån 2024-02-19 klockan 15:14 +0100 skrev Andreas Rheinhardt: > Tomas Härdin: > > lör 2024-02-17 klockan 21:05 +0100 skrev Andreas Rheinhardt: > > > The call to ff_exp2fi() here always uses arguments in the normal > > > range, so that the branches in ff_exp2fi() are unnecessary. > > > This is so because JPEG2000 itself only supports up to > > > 128 bits per component per pixel (we only support far less); > > > furthermore, expn is always 0..31 for the decoder and also > > > sane for the encoder, so that the difference between these > > > two values is always in the normal range of -126..128. > > > > Any measurable improvement in decode speed? > > > > I don't think it is measurable; my aim was to move ff_exp2fi() out of > internal.h (where it does not belong). Ah. Well, fewer branches is almost always good so, looks good to me. Maybe pal has an opinion? /Tomas
diff --git a/libavcodec/jpeg2000.c b/libavcodec/jpeg2000.c index 0aa984bc53..d6ffb02319 100644 --- a/libavcodec/jpeg2000.c +++ b/libavcodec/jpeg2000.c @@ -32,7 +32,6 @@ #include "libavutil/mem.h" #include "libavutil/thread.h" #include "avcodec.h" -#include "internal.h" #include "jpeg2000.h" #define SHL(a, n) ((n) >= 0 ? (a) << (n) : (a) >> -(n)) @@ -201,6 +200,17 @@ void ff_jpeg2000_set_significance(Jpeg2000T1Context *t1, int x, int y, // static const uint8_t lut_gain[2][4] = { { 0, 0, 0, 0 }, { 0, 1, 1, 2 } }; (unused) +/** + * 2^(x) for integer x in the range -126..128. + * @return correctly rounded float + */ +static av_always_inline float exp2fi(int x) +{ + av_assert2(-126 <= x && x <= 128); + /* Normal range */ + return av_int2float((x+127) << 23); +} + static void init_band_stepsize(AVCodecContext *avctx, Jpeg2000Band *band, Jpeg2000CodingStyle *codsty, @@ -230,7 +240,7 @@ static void init_band_stepsize(AVCodecContext *avctx, * R_b = R_I + log2 (gain_b ) * see ISO/IEC 15444-1:2002 E.1.1 eqn. E-3 and E-4 */ gain = cbps; - band->f_stepsize = ff_exp2fi(gain - qntsty->expn[gbandno]); + band->f_stepsize = exp2fi(gain - qntsty->expn[gbandno]); band->f_stepsize *= qntsty->mant[gbandno] / 2048.0 + 1.0; break; default: @@ -391,7 +401,7 @@ static int init_band(AVCodecContext *avctx, Jpeg2000CodingStyle *codsty, Jpeg2000QuantStyle *qntsty, int bandno, int gbandno, int reslevelno, - int cbps, int dx, int dy) + const int cbps, int dx, int dy) { Jpeg2000Band *band = reslevel->band + bandno; uint8_t log2_band_prec_width, log2_band_prec_height; @@ -466,7 +476,7 @@ static int init_band(AVCodecContext *avctx, int ff_jpeg2000_init_component(Jpeg2000Component *comp, Jpeg2000CodingStyle *codsty, Jpeg2000QuantStyle *qntsty, - int cbps, int dx, int dy, + const int cbps, int dx, int dy, AVCodecContext *avctx) { int reslevelno, bandno, gbandno = 0, ret, i, j;
The call to ff_exp2fi() here always uses arguments in the normal range, so that the branches in ff_exp2fi() are unnecessary. This is so because JPEG2000 itself only supports up to 128 bits per component per pixel (we only support far less); furthermore, expn is always 0..31 for the decoder and also sane for the encoder, so that the difference between these two values is always in the normal range of -126..128. Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com> --- libavcodec/jpeg2000.c | 18 ++++++++++++++---- 1 file changed, 14 insertions(+), 4 deletions(-)