diff options
37 files changed, 960 insertions, 962 deletions
diff --git a/README.draft b/README.draft index 22159663..db4e35e3 100644 --- a/README.draft +++ b/README.draft @@ -9,12 +9,12 @@ in the Makefile. An up-to-date implementation conforming to this standard is available in a Git repository at git://git.xiph.org/opus.git or on a website at: http://opus-codec.org/ -However, although that implementation is expected to remain conformant -with the standard, it is the code in this RFC that shall remain normative. +However, although that implementation is expected to remain conformant +with the standard, it is the code in this RFC that shall remain normative. To build from the git repository instead of using this RFC, follow these steps: -1) Clone the repository (latest implementation of this standard at the time +1) Clone the repository (latest implementation of this standard at the time of publication) % git clone git://git.opus-codec.org/opus.git diff --git a/celt/dump_modes/Makefile b/celt/dump_modes/Makefile index c2de3181..74d527e1 100644 --- a/celt/dump_modes/Makefile +++ b/celt/dump_modes/Makefile @@ -4,7 +4,7 @@ INCLUDES=-I. -I../ -I../.. -I../../include all: dump_modes dump_modes: - $(CC) $(CFLAGS) $(INCLUDES) -DCUSTOM_MODES_ONLY -DCUSTOM_MODES dump_modes.c ../modes.c ../cwrs.c ../rate.c ../entenc.c ../entdec.c ../mathops.c ../mdct.c ../kiss_fft.c -o dump_modes -lm + $(CC) $(CFLAGS) $(INCLUDES) -DCUSTOM_MODES_ONLY -DCUSTOM_MODES dump_modes.c ../modes.c ../cwrs.c ../rate.c ../entenc.c ../entdec.c ../mathops.c ../mdct.c ../kiss_fft.c -o dump_modes -lm clean: rm -f dump_modes diff --git a/celt/dump_modes/dump_modes.c b/celt/dump_modes/dump_modes.c index 4759ae1e..ae6a8c15 100644 --- a/celt/dump_modes/dump_modes.c +++ b/celt/dump_modes/dump_modes.c @@ -171,18 +171,18 @@ void dump_modes(FILE *file, CELTMode **modes, int nb_modes) fprintf(file, "#define FFT_STATE%d_%d_%d\n", mode->Fs, mdctSize, k); fprintf (file, "static const kiss_fft_state fft_state%d_%d_%d = {\n", mode->Fs, mdctSize, k); - fprintf (file, "%d,\t/* nfft */\n", mode->mdct.kfft[k]->nfft); - fprintf (file, WORD16 ",\t/* scale */\n", mode->mdct.kfft[k]->scale); + fprintf (file, "%d, /* nfft */\n", mode->mdct.kfft[k]->nfft); + fprintf (file, WORD16 ", /* scale */\n", mode->mdct.kfft[k]->scale); #ifdef FIXED_POINT - fprintf (file, "%d,\t/* scale_shift */\n", mode->mdct.kfft[k]->scale_shift); + fprintf (file, "%d, /* scale_shift */\n", mode->mdct.kfft[k]->scale_shift); #endif - fprintf (file, "%d,\t/* shift */\n", mode->mdct.kfft[k]->shift); + fprintf (file, "%d, /* shift */\n", mode->mdct.kfft[k]->shift); fprintf (file, "{"); for (j=0;j<2*MAXFACTORS;j++) fprintf (file, "%d, ", mode->mdct.kfft[k]->factors[j]); - fprintf (file, "},\t/* factors */\n"); - fprintf (file, "fft_bitrev%d,\t/* bitrev */\n", mode->mdct.kfft[k]->nfft); - fprintf (file, "fft_twiddles%d_%d,\t/* bitrev */\n", mode->Fs, mdctSize); + fprintf (file, "}, /* factors */\n"); + fprintf (file, "fft_bitrev%d, /* bitrev */\n", mode->mdct.kfft[k]->nfft); + fprintf (file, "fft_twiddles%d_%d, /* bitrev */\n", mode->Fs, mdctSize); fprintf (file, "};\n"); fprintf(file, "#endif\n"); @@ -208,37 +208,37 @@ void dump_modes(FILE *file, CELTMode **modes, int nb_modes) /* Print the actual mode data */ fprintf(file, "static const CELTMode mode%d_%d_%d = {\n", mode->Fs, mdctSize, mode->overlap); - fprintf(file, INT32 ",\t/* Fs */\n", mode->Fs); - fprintf(file, "%d,\t/* overlap */\n", mode->overlap); - fprintf(file, "%d,\t/* nbEBands */\n", mode->nbEBands); - fprintf(file, "%d,\t/* effEBands */\n", mode->effEBands); + fprintf(file, INT32 ", /* Fs */\n", mode->Fs); + fprintf(file, "%d, /* overlap */\n", mode->overlap); + fprintf(file, "%d, /* nbEBands */\n", mode->nbEBands); + fprintf(file, "%d, /* effEBands */\n", mode->effEBands); fprintf(file, "{"); for (j=0;j<4;j++) fprintf(file, WORD16 ", ", mode->preemph[j]); - fprintf(file, "},\t/* preemph */\n"); + fprintf(file, "}, /* preemph */\n"); if (standard) - fprintf(file, "eband5ms,\t/* eBands */\n"); + fprintf(file, "eband5ms, /* eBands */\n"); else - fprintf(file, "eBands%d_%d,\t/* eBands */\n", mode->Fs, mdctSize); + fprintf(file, "eBands%d_%d, /* eBands */\n", mode->Fs, mdctSize); - fprintf(file, "%d,\t/* maxLM */\n", mode->maxLM); - fprintf(file, "%d,\t/* nbShortMdcts */\n", mode->nbShortMdcts); - fprintf(file, "%d,\t/* shortMdctSize */\n", mode->shortMdctSize); + fprintf(file, "%d, /* maxLM */\n", mode->maxLM); + fprintf(file, "%d, /* nbShortMdcts */\n", mode->nbShortMdcts); + fprintf(file, "%d, /* shortMdctSize */\n", mode->shortMdctSize); - fprintf(file, "%d,\t/* nbAllocVectors */\n", mode->nbAllocVectors); + fprintf(file, "%d, /* nbAllocVectors */\n", mode->nbAllocVectors); if (standard) - fprintf(file, "band_allocation,\t/* allocVectors */\n"); + fprintf(file, "band_allocation, /* allocVectors */\n"); else - fprintf(file, "allocVectors%d_%d,\t/* allocVectors */\n", mode->Fs, mdctSize); + fprintf(file, "allocVectors%d_%d, /* allocVectors */\n", mode->Fs, mdctSize); - fprintf(file, "logN%d,\t/* logN */\n", framerate); - fprintf(file, "window%d,\t/* window */\n", mode->overlap); + fprintf(file, "logN%d, /* logN */\n", framerate); + fprintf(file, "window%d, /* window */\n", mode->overlap); fprintf(file, "{%d, %d, {", mode->mdct.n, mode->mdct.maxshift); for (k=0;k<=mode->mdct.maxshift;k++) fprintf(file, "&fft_state%d_%d_%d, ", mode->Fs, mdctSize, k); - fprintf (file, "}, mdct_twiddles%d},\t/* mdct */\n", mdctSize); + fprintf (file, "}, mdct_twiddles%d}, /* mdct */\n", mdctSize); - fprintf(file, "{%d, cache_index%d, cache_bits%d, cache_caps%d},\t/* cache */\n", + fprintf(file, "{%d, cache_index%d, cache_bits%d, cache_caps%d}, /* cache */\n", mode->cache.size, mode->Fs/mdctSize, mode->Fs/mdctSize, mode->Fs/mdctSize); fprintf(file, "};\n"); } diff --git a/celt/mips/fixed_generic_mipsr1.h b/celt/mips/fixed_generic_mipsr1.h index cf885856..4a05efbf 100644 --- a/celt/mips/fixed_generic_mipsr1.h +++ b/celt/mips/fixed_generic_mipsr1.h @@ -35,38 +35,38 @@ #undef MULT16_32_Q15_ADD static inline int MULT16_32_Q15_ADD(int a, int b, int c, int d) { - int m; - asm volatile("MULT $ac1, %0, %1" : : "r" ((int)a), "r" ((int)b)); - asm volatile("madd $ac1, %0, %1" : : "r" ((int)c), "r" ((int)d)); - asm volatile("EXTR.W %0,$ac1, %1" : "=r" (m): "i" (15)); - return m; + int m; + asm volatile("MULT $ac1, %0, %1" : : "r" ((int)a), "r" ((int)b)); + asm volatile("madd $ac1, %0, %1" : : "r" ((int)c), "r" ((int)d)); + asm volatile("EXTR.W %0,$ac1, %1" : "=r" (m): "i" (15)); + return m; } #undef MULT16_32_Q15_SUB static inline int MULT16_32_Q15_SUB(int a, int b, int c, int d) { - int m; - asm volatile("MULT $ac1, %0, %1" : : "r" ((int)a), "r" ((int)b)); - asm volatile("msub $ac1, %0, %1" : : "r" ((int)c), "r" ((int)d)); - asm volatile("EXTR.W %0,$ac1, %1" : "=r" (m): "i" (15)); - return m; + int m; + asm volatile("MULT $ac1, %0, %1" : : "r" ((int)a), "r" ((int)b)); + asm volatile("msub $ac1, %0, %1" : : "r" ((int)c), "r" ((int)d)); + asm volatile("EXTR.W %0,$ac1, %1" : "=r" (m): "i" (15)); + return m; } #undef MULT16_16_Q15_ADD static inline int MULT16_16_Q15_ADD(int a, int b, int c, int d) { - int m; - asm volatile("MULT $ac1, %0, %1" : : "r" ((int)a), "r" ((int)b)); - asm volatile("madd $ac1, %0, %1" : : "r" ((int)c), "r" ((int)d)); - asm volatile("EXTR.W %0,$ac1, %1" : "=r" (m): "i" (15)); - return m; + int m; + asm volatile("MULT $ac1, %0, %1" : : "r" ((int)a), "r" ((int)b)); + asm volatile("madd $ac1, %0, %1" : : "r" ((int)c), "r" ((int)d)); + asm volatile("EXTR.W %0,$ac1, %1" : "=r" (m): "i" (15)); + return m; } #undef MULT16_16_Q15_SUB static inline int MULT16_16_Q15_SUB(int a, int b, int c, int d) { - int m; - asm volatile("MULT $ac1, %0, %1" : : "r" ((int)a), "r" ((int)b)); - asm volatile("msub $ac1, %0, %1" : : "r" ((int)c), "r" ((int)d)); - asm volatile("EXTR.W %0,$ac1, %1" : "=r" (m): "i" (15)); - return m; + int m; + asm volatile("MULT $ac1, %0, %1" : : "r" ((int)a), "r" ((int)b)); + asm volatile("msub $ac1, %0, %1" : : "r" ((int)c), "r" ((int)d)); + asm volatile("EXTR.W %0,$ac1, %1" : "=r" (m): "i" (15)); + return m; } diff --git a/celt/mips/kiss_fft_mipsr1.h b/celt/mips/kiss_fft_mipsr1.h index 73d39634..400ca4de 100644 --- a/celt/mips/kiss_fft_mipsr1.h +++ b/celt/mips/kiss_fft_mipsr1.h @@ -37,20 +37,20 @@ #undef S_MUL_ADD static inline int S_MUL_ADD(int a, int b, int c, int d) { - int m; - asm volatile("MULT $ac1, %0, %1" : : "r" ((int)a), "r" ((int)b)); - asm volatile("madd $ac1, %0, %1" : : "r" ((int)c), "r" ((int)d)); - asm volatile("EXTR.W %0,$ac1, %1" : "=r" (m): "i" (15)); - return m; + int m; + asm volatile("MULT $ac1, %0, %1" : : "r" ((int)a), "r" ((int)b)); + asm volatile("madd $ac1, %0, %1" : : "r" ((int)c), "r" ((int)d)); + asm volatile("EXTR.W %0,$ac1, %1" : "=r" (m): "i" (15)); + return m; } #undef S_MUL_SUB static inline int S_MUL_SUB(int a, int b, int c, int d) { - int m; - asm volatile("MULT $ac1, %0, %1" : : "r" ((int)a), "r" ((int)b)); - asm volatile("msub $ac1, %0, %1" : : "r" ((int)c), "r" ((int)d)); - asm volatile("EXTR.W %0,$ac1, %1" : "=r" (m): "i" (15)); - return m; + int m; + asm volatile("MULT $ac1, %0, %1" : : "r" ((int)a), "r" ((int)b)); + asm volatile("msub $ac1, %0, %1" : : "r" ((int)c), "r" ((int)d)); + asm volatile("EXTR.W %0,$ac1, %1" : "=r" (m): "i" (15)); + return m; } #undef C_MUL diff --git a/celt/mips/pitch_mipsr1.h b/celt/mips/pitch_mipsr1.h index b85338af..9c5dc34d 100644 --- a/celt/mips/pitch_mipsr1.h +++ b/celt/mips/pitch_mipsr1.h @@ -46,11 +46,11 @@ static inline void dual_inner_prod(const opus_val16 *x, const opus_val16 *y01, c /* Compute the norm of X+Y and X-Y as |X|^2 + |Y|^2 +/- sum(xy) */ for (j=0;j<N;j++) { - asm volatile("MADD $ac1, %0, %1" : : "r" ((int)x[j]), "r" ((int)y01[j])); - asm volatile("MADD $ac2, %0, %1" : : "r" ((int)x[j]), "r" ((int)y02[j])); - ++j; - asm volatile("MADD $ac1, %0, %1" : : "r" ((int)x[j]), "r" ((int)y01[j])); - asm volatile("MADD $ac2, %0, %1" : : "r" ((int)x[j]), "r" ((int)y02[j])); + asm volatile("MADD $ac1, %0, %1" : : "r" ((int)x[j]), "r" ((int)y01[j])); + asm volatile("MADD $ac2, %0, %1" : : "r" ((int)x[j]), "r" ((int)y02[j])); + ++j; + asm volatile("MADD $ac1, %0, %1" : : "r" ((int)x[j]), "r" ((int)y01[j])); + asm volatile("MADD $ac2, %0, %1" : : "r" ((int)x[j]), "r" ((int)y02[j])); } asm volatile ("mflo %0, $ac1": "=r"(xy01)); asm volatile ("mflo %0, $ac2": "=r"(xy02)); diff --git a/celt/static_modes_fixed.h b/celt/static_modes_fixed.h index c822929e..60f0f066 100644 --- a/celt/static_modes_fixed.h +++ b/celt/static_modes_fixed.h @@ -426,10 +426,10 @@ static const opus_int16 fft_bitrev60[60] = { #define FFT_STATE48000_960_0 static const kiss_fft_state fft_state48000_960_0 = { 480, /* nfft */ -17476, /* scale */ +17476, /* scale */ 8, /* scale_shift */ -1, /* shift */ -{5, 96, 3, 32, 4, 8, 2, 4, 4, 1, 0, 0, 0, 0, 0, 0, }, /* factors */ +{5, 96, 3, 32, 4, 8, 2, 4, 4, 1, 0, 0, 0, 0, 0, 0, }, /* factors */ fft_bitrev480, /* bitrev */ fft_twiddles48000_960, /* bitrev */ }; @@ -439,10 +439,10 @@ fft_twiddles48000_960, /* bitrev */ #define FFT_STATE48000_960_1 static const kiss_fft_state fft_state48000_960_1 = { 240, /* nfft */ -17476, /* scale */ +17476, /* scale */ 7, /* scale_shift */ 1, /* shift */ -{5, 48, 3, 16, 4, 4, 4, 1, 0, 0, 0, 0, 0, 0, 0, 0, }, /* factors */ +{5, 48, 3, 16, 4, 4, 4, 1, 0, 0, 0, 0, 0, 0, 0, 0, }, /* factors */ fft_bitrev240, /* bitrev */ fft_twiddles48000_960, /* bitrev */ }; @@ -452,10 +452,10 @@ fft_twiddles48000_960, /* bitrev */ #define FFT_STATE48000_960_2 static const kiss_fft_state fft_state48000_960_2 = { 120, /* nfft */ -17476, /* scale */ +17476, /* scale */ 6, /* scale_shift */ 2, /* shift */ -{5, 24, 3, 8, 2, 4, 4, 1, 0, 0, 0, 0, 0, 0, 0, 0, }, /* factors */ +{5, 24, 3, 8, 2, 4, 4, 1, 0, 0, 0, 0, 0, 0, 0, 0, }, /* factors */ fft_bitrev120, /* bitrev */ fft_twiddles48000_960, /* bitrev */ }; @@ -465,10 +465,10 @@ fft_twiddles48000_960, /* bitrev */ #define FFT_STATE48000_960_3 static const kiss_fft_state fft_state48000_960_3 = { 60, /* nfft */ -17476, /* scale */ +17476, /* scale */ 5, /* scale_shift */ 3, /* shift */ -{5, 12, 3, 4, 4, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, /* factors */ +{5, 12, 3, 4, 4, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, /* factors */ fft_bitrev60, /* bitrev */ fft_twiddles48000_960, /* bitrev */ }; diff --git a/celt/static_modes_float.h b/celt/static_modes_float.h index c4a5938d..2fadb625 100644 --- a/celt/static_modes_float.h +++ b/celt/static_modes_float.h @@ -440,7 +440,7 @@ static const kiss_fft_state fft_state48000_960_1 = { 240, /* nfft */ 0.004166667f, /* scale */ 1, /* shift */ -{5, 48, 3, 16, 4, 4, 4, 1, 0, 0, 0, 0, 0, 0, 0, 0, }, /* factors */ +{5, 48, 3, 16, 4, 4, 4, 1, 0, 0, 0, 0, 0, 0, 0, 0, }, /* factors */ fft_bitrev240, /* bitrev */ fft_twiddles48000_960, /* bitrev */ }; @@ -452,7 +452,7 @@ static const kiss_fft_state fft_state48000_960_2 = { 120, /* nfft */ 0.008333333f, /* scale */ 2, /* shift */ -{5, 24, 3, 8, 2, 4, 4, 1, 0, 0, 0, 0, 0, 0, 0, 0, }, /* factors */ +{5, 24, 3, 8, 2, 4, 4, 1, 0, 0, 0, 0, 0, 0, 0, 0, }, /* factors */ fft_bitrev120, /* bitrev */ fft_twiddles48000_960, /* bitrev */ }; @@ -464,7 +464,7 @@ static const kiss_fft_state fft_state48000_960_3 = { 60, /* nfft */ 0.016666667f, /* scale */ 3, /* shift */ -{5, 12, 3, 4, 4, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, /* factors */ +{5, 12, 3, 4, 4, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, /* factors */ fft_bitrev60, /* bitrev */ fft_twiddles48000_960, /* bitrev */ }; diff --git a/celt/tests/test_unit_entropy.c b/celt/tests/test_unit_entropy.c index bd839866..ff926586 100644 --- a/celt/tests/test_unit_entropy.c +++ b/celt/tests/test_unit_entropy.c @@ -66,10 +66,10 @@ int main(int _argc,char **_argv){ const char *env_seed; ret=0; entropy=0; - if (_argc > 2) { - fprintf(stderr, "Usage: %s [<seed>]\n", _argv[0]); - return 1; - } + if (_argc > 2) { + fprintf(stderr, "Usage: %s [<seed>]\n", _argv[0]); + return 1; + } env_seed = getenv("SEED"); if (_argc > 1) seed = atoi(_argv[1]); diff --git a/celt/tests/test_unit_mathops.c b/celt/tests/test_unit_mathops.c index 5060f7f2..5296f324 100644 --- a/celt/tests/test_unit_mathops.c +++ b/celt/tests/test_unit_mathops.c @@ -224,7 +224,7 @@ void testexp2(void) float error2 = fabs(exp(0.6931471805599453094*x/1024.0)-celt_exp2(x)/65536.0); if (error1>0.0002&&error2>0.00004) { - fprintf (stderr, "celt_exp2 failed: x = "WORD", error1 = %f, error2 = %f\n", x,error1,error2); + fprintf (stderr, "celt_exp2 failed: x = "WORD", error1 = %f, error2 = %f\n", x,error1,error2); ret = 1; } } diff --git a/doc/Doxyfile.in b/doc/Doxyfile.in index 8a63d5e8..0c016f32 100644 --- a/doc/Doxyfile.in +++ b/doc/Doxyfile.in @@ -611,7 +611,7 @@ WARN_LOGFILE = # with spaces. INPUT = @top_srcdir@/include/opus.h \ - @top_srcdir@/include/opus_types.h \ + @top_srcdir@/include/opus_types.h \ @top_srcdir@/include/opus_defines.h \ @top_srcdir@/include/opus_multistream.h \ @top_srcdir@/include/opus_custom.h diff --git a/doc/customdoxygen.css b/doc/customdoxygen.css index 4bce44db..70047787 100644 --- a/doc/customdoxygen.css +++ b/doc/customdoxygen.css @@ -1,62 +1,62 @@ /* The standard CSS for doxygen */ body, table, div, p, dl { - font-family: Lucida Grande, Verdana, Geneva, Arial, sans-serif; - font-size: 13px; - line-height: 1.3; + font-family: Lucida Grande, Verdana, Geneva, Arial, sans-serif; + font-size: 13px; + line-height: 1.3; } /* @group Heading Levels */ h1 { - font-size: 150%; + font-size: 150%; } .title { - font-size: 150%; - font-weight: bold; - margin: 10px 2px; + font-size: 150%; + font-weight: bold; + margin: 10px 2px; } h2 { - font-size: 120%; + font-size: 120%; } h3 { - font-size: 100%; + font-size: 100%; } dt { - font-weight: bold; + font-weight: bold; } div.multicol { - -moz-column-gap: 1em; - -webkit-column-gap: 1em; - -moz-column-count: 3; - -webkit-column-count: 3; + -moz-column-gap: 1em; + -webkit-column-gap: 1em; + -moz-column-count: 3; + -webkit-column-count: 3; } p.startli, p.startdd, p.starttd { - margin-top: 2px; + margin-top: 2px; } p.endli { - margin-bottom: 0px; + margin-bottom: 0px; } p.enddd { - margin-bottom: 4px; + margin-bottom: 4px; } p.endtd { - margin-bottom: 2px; + margin-bottom: 2px; } /* @end */ caption { - font-weight: bold; + font-weight: bold; } span.legend { @@ -70,45 +70,45 @@ h3.version { } div.qindex, div.navtab{ - background-color: #F1F1F1; - border: 1px solid #BDBDBD; - text-align: center; + background-color: #F1F1F1; + border: 1px solid #BDBDBD; + text-align: center; } div.qindex, div.navpath { - width: 100%; - line-height: 140%; + width: 100%; + line-height: 140%; } div.navtab { - margin-right: 15px; + margin-right: 15px; } /* @group Link Styling */ a { - color: #646464; - font-weight: normal; - text-decoration: none; + color: #646464; + font-weight: normal; + text-decoration: none; } .contents a:visited { - color: #747474; + color: #747474; } a:hover { - text-decoration: underline; + text-decoration: underline; } a.qindex { - font-weight: bold; + font-weight: bold; } a.qindexHL { - font-weight: bold; - background-color: #B8B8B8; - color: #ffffff; - border: 1px double #A8A8A8; + font-weight: bold; + background-color: #B8B8B8; + color: #ffffff; + border: 1px double #A8A8A8; } .contents a.qindexHL:visited { @@ -116,181 +116,181 @@ a.qindexHL { } a.el { - font-weight: bold; + font-weight: bold; } a.elRef { } a.code, a.code:visited { - color: #4665A2; + color: #4665A2; } a.codeRef, a.codeRef:visited { - color: #4665A2; + color: #4665A2; } /* @end */ dl.el { - margin-left: -1cm; + margin-left: -1cm; } .fragment { - font-family: monospace, fixed; - font-size: 105%; + font-family: monospace, fixed; + font-size: 105%; } pre.fragment { - border: 1px solid #D5D5D5; - background-color: #FCFCFC; - padding: 4px 6px; - margin: 4px 8px 4px 2px; - overflow: auto; - word-wrap: break-word; - font-size: 9pt; - line-height: 125%; + border: 1px solid #D5D5D5; + background-color: #FCFCFC; + padding: 4px 6px; + margin: 4px 8px 4px 2px; + overflow: auto; + word-wrap: break-word; + font-size: 9pt; + line-height: 125%; } div.ah { - background-color: black; - font-weight: bold; - color: #ffffff; - margin-bottom: 3px; - margin-top: 3px; - padding: 0.2em; - border: solid thin #333; - border-radius: 0.5em; - -webkit-border-radius: .5em; - -moz-border-radius: .5em; - box-shadow: 2px 2px 3px #999; - -webkit-box-shadow: 2px 2px 3px #999; - -moz-box-shadow: rgba(0, 0, 0, 0.15) 2px 2px 2px; - background-image: -webkit-gradient(linear, left top, left bottom, from(#eee), to(#000),color-stop(0.3, #444)); - background-image: -moz-linear-gradient(center top, #eee 0%, #444 40%, #000); + background-color: black; + font-weight: bold; + color: #ffffff; + margin-bottom: 3px; + margin-top: 3px; + padding: 0.2em; + border: solid thin #333; + border-radius: 0.5em; + -webkit-border-radius: .5em; + -moz-border-radius: .5em; + box-shadow: 2px 2px 3px #999; + -webkit-box-shadow: 2px 2px 3px #999; + -moz-box-shadow: rgba(0, 0, 0, 0.15) 2px 2px 2px; + background-image: -webkit-gradient(linear, left top, left bottom, from(#eee), to(#000),color-stop(0.3, #444)); + background-image: -moz-linear-gradient(center top, #eee 0%, #444 40%, #000); } div.groupHeader { - margin-left: 16px; - margin-top: 12px; - font-weight: bold; + margin-left: 16px; + margin-top: 12px; + font-weight: bold; } div.groupText { - margin-left: 16px; - font-style: italic; + margin-left: 16px; + font-style: italic; } body { - background-color: white; - color: black; + background-color: white; + color: black; margin: 0; } div.contents { - margin-top: 10px; - margin-left: 8px; - margin-right: 8px; + margin-top: 10px; + margin-left: 8px; + margin-right: 8px; } td.indexkey { - background-color: #F1F1F1; - font-weight: bold; - border: 1px solid #D5D5D5; - margin: 2px 0px 2px 0; - padding: 2px 10px; + background-color: #F1F1F1; + font-weight: bold; + border: 1px solid #D5D5D5; + margin: 2px 0px 2px 0; + padding: 2px 10px; white-space: nowrap; vertical-align: top; } td.indexvalue { - background-color: #F1F1F1; - border: 1px solid #D5D5D5; - padding: 2px 10px; - margin: 2px 0px; + background-color: #F1F1F1; + border: 1px solid #D5D5D5; + padding: 2px 10px; + margin: 2px 0px; } tr.memlist { - background-color: #F2F2F2; + background-color: #F2F2F2; } p.formulaDsp { - text-align: center; + text-align: center; } img.formulaDsp { - + } img.formulaInl { - vertical-align: middle; + vertical-align: middle; } div.center { - text-align: center; + text-align: center; margin-top: 0px; margin-bottom: 0px; padding: 0px; } div.center img { - border: 0px; + border: 0px; } address.footer { - text-align: right; - padding-right: 12px; + text-align: right; + padding-right: 12px; } img.footer { - border: 0px; - vertical-align: middle; + border: 0px; + vertical-align: middle; } /* @group Code Colorization */ span.keyword { - color: #008000 + color: #008000 } span.keywordtype { - color: #604020 + color: #604020 } span.keywordflow { - color: #e08000 + color: #e08000 } span.comment { - color: #800000 + color: #800000 } span.preprocessor { - color: #806020 + color: #806020 } span.stringliteral { - color: #002080 + color: #002080 } span.charliteral { - color: #008080 + color: #008080 } -span.vhdldigit { - color: #ff00ff +span.vhdldigit { + color: #ff00ff } -span.vhdlchar { - color: #000000 +span.vhdlchar { + color: #000000 } -span.vhdlkeyword { - color: #700070 +span.vhdlkeyword { + color: #700070 } -span.vhdllogic { - color: #ff0000 +span.vhdllogic { + color: #ff0000 } blockquote { @@ -304,71 +304,71 @@ blockquote { /* .search { - color: #003399; - font-weight: bold; + color: #003399; + font-weight: bold; } form.search { - margin-bottom: 0px; - margin-top: 0px; + margin-bottom: 0px; + margin-top: 0px; } input.search { - font-size: 75%; - color: #000080; - font-weight: normal; - background-color: #e8eef2; + font-size: 75%; + color: #000080; + font-weight: normal; + background-color: #e8eef2; } */ td.tiny { - font-size: 75%; + font-size: 75%; } .dirtab { - padding: 4px; - border-collapse: collapse; - border: 1px solid #BDBDBD; + padding: 4px; + border-collapse: collapse; + border: 1px solid #BDBDBD; } th.dirtab { - background: #F1F1F1; - font-weight: bold; + background: #F1F1F1; + font-weight: bold; } hr { - height: 0px; - border: none; - border-top: 1px solid #7A7A7A; + height: 0px; + border: none; + border-top: 1px solid #7A7A7A; } hr.footer { - height: 1px; + height: 1px; } /* @group Member Descriptions */ table.memberdecls { - border-spacing: 0px; - padding: 0px; + border-spacing: 0px; + padding: 0px; } .mdescLeft, .mdescRight, .memItemLeft, .memItemRight, .memTemplItemLeft, .memTemplItemRight, .memTemplParams { - background-color: #FAFAFA; - border: none; - margin: 4px; - padding: 1px 0 0 8px; + background-color: #FAFAFA; + border: none; + margin: 4px; + padding: 1px 0 0 8px; } .mdescLeft, .mdescRight { - padding: 0px 8px 4px 8px; - color: #555; + padding: 0px 8px 4px 8px; + color: #555; } .memItemLeft, .memItemRight, .memTemplParams { - border-top: 1px solid #D5D5D5; + border-top: 1px solid #D5D5D5; } .memItemLeft, .memTemplItemLeft { @@ -376,11 +376,11 @@ table.memberdecls { } .memItemRight { - width: 100%; + width: 100%; } .memTemplParams { - color: #747474; + color: #747474; white-space: nowrap; } @@ -391,29 +391,29 @@ table.memberdecls { /* Styles for detailed member documentation */ .memtemplate { - font-size: 80%; - color: #747474; - font-weight: normal; - margin-left: 9px; + font-size: 80%; + color: #747474; + font-weight: normal; + margin-left: 9px; } .memnav { - background-color: #F1F1F1; - border: 1px solid #BDBDBD; - text-align: center; - margin: 2px; - margin-right: 15px; - padding: 2px; + background-color: #F1F1F1; + border: 1px solid #BDBDBD; + text-align: center; + margin: 2px; + margin-right: 15px; + padding: 2px; } .mempage { - width: 100%; + width: 100%; } .memitem { - padding: 0; - margin-bottom: 10px; - margin-right: 5px; + padding: 0; + margin-bottom: 10px; + margin-right: 5px; } .memname { @@ -449,9 +449,9 @@ table.memberdecls { } .memdoc, dl.reflist dd { - border-bottom: 1px solid #C0C0C0; - border-left: 1px solid #C0C0C0; - border-right: 1px solid #C0C0C0; + border-bottom: 1px solid #C0C0C0; + border-left: 1px solid #C0C0C0; + border-right: 1px solid #C0C0C0; padding: 2px 5px; background-color: #FCFCFC; border-top-width: 0; @@ -481,35 +481,35 @@ dl.reflist dd { } .paramkey { - text-align: right; + text-align: right; } .paramtype { - white-space: nowrap; + white-space: nowrap; } .paramname { - color: #602020; - white-space: nowrap; + color: #602020; + white-space: nowrap; } .paramname em { - font-style: normal; + font-style: normal; } .params, .retval, .exception, .tparams { border-spacing: 6px 2px; -} +} .params .paramname, .retval .paramname { font-weight: bold; vertical-align: top; } - + .params .paramtype { font-style: italic; vertical-align: top; -} - +} + .params .paramdir { font-family: "courier new",courier,monospace; vertical-align: top; @@ -525,22 +525,22 @@ dl.reflist dd { /* for the tree view */ .ftvtree { - font-family: sans-serif; - margin: 0px; + font-family: sans-serif; + margin: 0px; } /* these are for tree view when used as main index */ .directory { - font-size: 9pt; - font-weight: bold; - margin: 5px; + font-size: 9pt; + font-weight: bold; + margin: 5px; } .directory h3 { - margin: 0px; - margin-top: 1em; - font-size: 11pt; + margin: 0px; + margin-top: 1em; + font-size: 11pt; } /* @@ -552,62 +552,62 @@ proper pixel height of your image. /* .directory h3.swap { - height: 61px; - background-repeat: no-repeat; - background-image: url("yourimage.gif"); + height: 61px; + background-repeat: no-repeat; + background-image: url("yourimage.gif"); } .directory h3.swap span { - display: none; + display: none; } */ .directory > h3 { - margin-top: 0; + margin-top: 0; } .directory p { - margin: 0px; - white-space: nowrap; + margin: 0px; + white-space: nowrap; } .directory div { - display: none; - margin: 0px; + display: none; + margin: 0px; } .directory img { - vertical-align: -30%; + vertical-align: -30%; } /* these are for tree view when not used as main index */ .directory-alt { - font-size: 100%; - font-weight: bold; + font-size: 100%; + font-weight: bold; } .directory-alt h3 { - margin: 0px; - margin-top: 1em; - font-size: 11pt; + margin: 0px; + margin-top: 1em; + font-size: 11pt; } .directory-alt > h3 { - margin-top: 0; + margin-top: 0; } .directory-alt p { - margin: 0px; - white-space: nowrap; + margin: 0px; + white-space: nowrap; } .directory-alt div { - display: none; - margin: 0px; + display: none; + margin: 0px; } .directory-alt img { - vertical-align: -30%; + vertical-align: -30%; } /* @end */ @@ -617,27 +617,27 @@ div.dynheader { } address { - font-style: normal; - color: #464646; + font-style: normal; + color: #464646; } table.doxtable { - border-collapse:collapse; + border-collapse:collapse; margin-top: 4px; margin-bottom: 4px; } table.doxtable td, table.doxtable th { - border: 1px solid #4A4A4A; - padding: 3px 7px 2px; + border: 1px solid #4A4A4A; + padding: 3px 7px 2px; } table.doxtable th { - background-color: #5B5B5B; - color: #FFFFFF; - font-size: 110%; - padding-bottom: 4px; - padding-top: 5px; + background-color: #5B5B5B; + color: #FFFFFF; + font-size: 110%; + padding-bottom: 4px; + padding-top: 5px; } table.fieldtable { @@ -693,52 +693,52 @@ table.fieldtable { .tabsearch { - top: 0px; - left: 10px; - height: 36px; - background-image: url('tab_b.png'); - z-index: 101; - overflow: hidden; - font-size: 13px; + top: 0px; + left: 10px; + height: 36px; + background-image: url('tab_b.png'); + z-index: 101; + overflow: hidden; + font-size: 13px; } .navpath ul { - font-size: 11px; - background-image:url('tab_b.png'); - background-repeat:repeat-x; - height:30px; - line-height:30px; - color:#ABABAB; - border:solid 1px #D3D3D3; - overflow:hidden; - margin:0px; - padding:0px; + font-size: 11px; + background-image:url('tab_b.png'); + background-repeat:repeat-x; + height:30px; + line-height:30px; + color:#ABABAB; + border:solid 1px #D3D3D3; + overflow:hidden; + margin:0px; + padding:0px; } .navpath li { - list-style-type:none; - float:left; - padding-left:10px; - padding-right:15px; - background-image:url('bc_s.png'); - background-repeat:no-repeat; - background-position:right; - color:#595959; + list-style-type:none; + float:left; + padding-left:10px; + padding-right:15px; + background-image:url('bc_s.png'); + background-repeat:no-repeat; + background-position:right; + color:#595959; } .navpath li.navelem a { - height:32px; - display:block; - text-decoration: none; - outline: none; + height:32px; + display:block; + text-decoration: none; + outline: none; } .navpath li.navelem a:hover { - color:#929292; + color:#929292; } .navpath li.footer @@ -757,44 +757,44 @@ table.fieldtable { div.summary { - float: right; - font-size: 8pt; - padding-right: 5px; - width: 50%; - text-align: right; -} + float: right; + font-size: 8pt; + padding-right: 5px; + width: 50%; + text-align: right; +} div.summary a { - white-space: nowrap; + white-space: nowrap; } div.ingroups { - margin-left: 5px; - font-size: 8pt; - padding-left: 5px; - width: 50%; - text-align: left; + margin-left: 5px; + font-size: 8pt; + padding-left: 5px; + width: 50%; + text-align: left; } div.ingroups a { - white-space: nowrap; + white-space: nowrap; } div.header { background-image:url('nav_h.png'); background-repeat:repeat-x; - background-color: #FAFAFA; - margin: 0px; - border-bottom: 1px solid #D5D5D5; + background-color: #FAFAFA; + margin: 0px; + border-bottom: 1px solid #D5D5D5; } div.headertitle { - padding: 5px 5px 5px 7px; + padding: 5px 5px 5px 7px; } dl @@ -845,49 +845,49 @@ dl.bug } dl.section dd { - margin-bottom: 6px; + margin-bottom: 6px; } #projectlogo { - text-align: center; - vertical-align: bottom; - border-collapse: separate; + text-align: center; + vertical-align: bottom; + border-collapse: separate; } - + #projectlogo img -{ - border: 0px none; +{ + border: 0px none; } - + #projectname { - font: 300% Tahoma, Arial,sans-serif; - margin: 0px; - padding: 2px 0px; + font: 300% Tahoma, Arial,sans-serif; + margin: 0px; + padding: 2px 0px; } - + #projectbrief { - font: 120% Tahoma, Arial,sans-serif; - margin: 0px; - padding: 0px; + font: 120% Tahoma, Arial,sans-serif; + margin: 0px; + padding: 0px; } #projectnumber { - font: 100% Tahoma, Arial,sans-serif; - margin: 0px; - padding: 0px; + font: 100% Tahoma, Arial,sans-serif; + margin: 0px; + padding: 0px; } #titlearea { - padding: 0px; - margin: 0px; - width: 100%; - border-bottom: 1px solid #848484; + padding: 0px; + margin: 0px; + width: 100%; + border-bottom: 1px solid #848484; } .image @@ -907,12 +907,12 @@ dl.section dd { .caption { - font-weight: bold; + font-weight: bold; } div.zoom { - border: 1px solid #AFAFAF; + border: 1px solid #AFAFAF; } dl.citelist { @@ -953,7 +953,7 @@ div.toc li { div.toc h3 { font: bold 12px/1.2 Arial,FreeSans,sans-serif; - color: #747474; + color: #747474; border-bottom: 0 none; margin: 0; } @@ -962,7 +962,7 @@ div.toc ul { list-style: none outside none; border: medium none; padding: 0px; -} +} div.toc li.level1 { margin-left: 0px; @@ -1009,4 +1009,3 @@ div.toc li.level4 { word-wrap: break-word; /* IE 5.5+ */ } } - diff --git a/doc/draft-ietf-codec-opus-update.xml b/doc/draft-ietf-codec-opus-update.xml index 54f15364..74147221 100644 --- a/doc/draft-ietf-codec-opus-update.xml +++ b/doc/draft-ietf-codec-opus-update.xml @@ -58,7 +58,7 @@ <middle> <section title="Introduction"> <t>This document addresses minor issues that were discovered in the reference - implementation of the Opus codec that serves as the specification in + implementation of the Opus codec that serves as the specification in <xref target="RFC6716">RFC 6716</xref>. Only issues affecting the decoder are listed here. An up-to-date implementation of the Opus encoder can be found at http://opus-codec.org/. The updated specification remains fully compatible with @@ -84,7 +84,7 @@ for( n = 0; n < DECODER_NUM_CHANNELS; n++ ) { ret = silk_init_decoder( &channel_state[ n ] ); } -+ silk_memset(&((silk_decoder *)decState)->sStereo, 0, ++ silk_memset(&((silk_decoder *)decState)->sStereo, 0, + sizeof(((silk_decoder *)decState)->sStereo)); + /* Not strictly needed, but it's cleaner that way */ + ((silk_decoder *)decState)->prev_decode_only_middle = 0; @@ -218,16 +218,16 @@ RESAMPLER_ORDER_FIR_12 * sizeof( opus_int16 ) ); in the patch above are split using a backslash character. The backslashes at the end of a line and the white space at the beginning of the following line are not part of the patch. A properly formatted patch - including the three changes above is available at + including the three changes above is available at <eref target="http://jmvalin.ca/misc_stuff/opus_update.patch"/>. </t> </section> - + <section title="Downmix to Mono"> <t>The last issue is not strictly a bug, but it is an issue that has been reported when downmixing an Opus decoded stream to mono, whether this is done inside the decoder or as a post-processing step on the stereo decoder output. Opus intensity stereo allows - optionally coding the two channels 180-degrees out of phase on a per-band basis. + optionally coding the two channels 180-degrees out of phase on a per-band basis. This provides better stereo quality than forcing the two channels to be in phase, but when the output is downmixed to mono, the energy in the affected bands is cancelled sometimes resulting in audible artefacts. diff --git a/doc/draft-ietf-payload-rtp-opus.xml b/doc/draft-ietf-payload-rtp-opus.xml index 3c40324f..4a207487 100644 --- a/doc/draft-ietf-payload-rtp-opus.xml +++ b/doc/draft-ietf-payload-rtp-opus.xml @@ -419,7 +419,7 @@ </figure> <t> - <xref target='opus-packetization'/> shows supported frame sizes in + <xref target='opus-packetization'/> shows supported frame sizes in milliseconds of encoded speech or audio data for the speech and audio modes (Mode) and sampling rates (fs) of Opus and shows how the timestamp is incremented for packetization (ts incr). If the Opus encoder @@ -427,7 +427,7 @@ increment is the sum of the increments for the individual frames. </t> - <texttable anchor='opus-packetization' title="Supported Opus frame + <texttable anchor='opus-packetization' title="Supported Opus frame sizes and timestamp increments"> <ttcol align='center'>Mode</ttcol> <ttcol align='center'>fs</ttcol> @@ -517,7 +517,7 @@ usage and encoding complexity, so an encoder SHOULD NOT encode frequencies above the audio bandwidth specified by maxplaybackrate. This parameter can take any value between 8000 and 48000, although - commonly the value will match one of the Opus bandwidths + commonly the value will match one of the Opus bandwidths (<xref target="bandwidth_definitions"/>). By default, the receiver is assumed to have no limitations, i.e. 48000. <vspace blankLines='1'/> @@ -531,7 +531,7 @@ This parameter is useful to avoid wasting receiver resources by operating the audio processing pipeline (e.g. echo cancellation) at a higher rate than necessary. This parameter can take any value between 8000 and 48000, although - commonly the value will match one of the Opus bandwidths + commonly the value will match one of the Opus bandwidths (<xref target="bandwidth_definitions"/>). By default, the sender is assumed to have no limitations, i.e. 48000. <vspace blankLines='1'/> @@ -708,9 +708,9 @@ mapped to "a=ptime" and "a=maxptime" attributes, respectively, in the SDP.</t> - <t>The OPTIONAL media type parameters "maxaveragebitrate", - "maxplaybackrate", "minptime", "stereo", "cbr", "useinbandfec", and - "usedtx", when present, MUST be included in the "a=fmtp" attribute + <t>The OPTIONAL media type parameters "maxaveragebitrate", + "maxplaybackrate", "minptime", "stereo", "cbr", "useinbandfec", and + "usedtx", when present, MUST be included in the "a=fmtp" attribute in the SDP, expressed as a media type string in the form of a semicolon-separated list of parameter=value pairs (e.g., maxaveragebitrate=20000). They MUST NOT be specified in an @@ -838,7 +838,7 @@ of the other side MUST NOT send with an average bitrate higher than "maxaveragebitrate" as it might overload the network and/or receiver. The "maxaveragebitrate" parameter typically will not - compromise interoperability; however, some values might cause + compromise interoperability; however, some values might cause application performance to suffer, and ought to be set with care.</t> diff --git a/doc/opus_in_isobmff.css b/doc/opus_in_isobmff.css index bffe8f45..3fbb1de5 100644 --- a/doc/opus_in_isobmff.css +++ b/doc/opus_in_isobmff.css @@ -11,13 +11,13 @@ /* Boxes */ .pre { - white-space: pre; /* CSS 2.0 */ - white-space: pre-wrap; /* CSS 2.1 */ - white-space: -pre-wrap; /* Opera 4-6 */ - white-space: -o-pre-wrap; /* Opera 7 */ - white-space: -moz-pre-wrap; /* Mozilla */ - white-space: -hp-pre-wrap; /* HP Printers */ - word-wrap : break-word; /* IE 5+ */ + white-space: pre; /* CSS 2.0 */ + white-space: pre-wrap; /* CSS 2.1 */ + white-space: -pre-wrap; /* Opera 4-6 */ + white-space: -o-pre-wrap; /* Opera 7 */ + white-space: -moz-pre-wrap; /* Mozilla */ + white-space: -hp-pre-wrap; /* HP Printers */ + word-wrap : break-word; /* IE 5+ */ } .title_box diff --git a/doc/opus_logo.svg b/doc/opus_logo.svg index 97112afd..db2879ef 100644 --- a/doc/opus_logo.svg +++ b/doc/opus_logo.svg @@ -59,11 +59,11 @@ y1="95.107399" x2="194.53169" y2="9.9475983e-14"> - <stop + <stop offset="0.0056" style="stop-color:#8E8E8E" id="stop7" /> - <stop + <stop offset="1" style="stop-color:#B5B5B5" id="stop9" /> @@ -76,11 +76,11 @@ y1="116.208" x2="229.61819" y2="164.46291"> - <stop + <stop offset="0.0056" style="stop-color:#494748" id="stop14" /> - <stop + <stop offset="1" style="stop-color:#000000" id="stop16" /> @@ -93,11 +93,11 @@ y1="115.4395" x2="43.9897" y2="165.2314"> - <stop + <stop offset="0.0056" style="stop-color:#494748" id="stop21" /> - <stop + <stop offset="1" style="stop-color:#000000" id="stop23" /> @@ -110,11 +110,11 @@ y1="115.7188" x2="311.2847" y2="165.2822"> - <stop + <stop offset="0.0056" style="stop-color:#494748" id="stop28" /> - <stop + <stop offset="1" style="stop-color:#000000" id="stop30" /> @@ -127,11 +127,11 @@ y1="115.5791" x2="129.1987" y2="204.4863"> - <stop + <stop offset="0.0056" style="stop-color:#494748" id="stop35" /> - <stop + <stop offset="1" style="stop-color:#000000" id="stop37" /> diff --git a/doc/release.txt b/doc/release.txt index 0758c75b..3eb8db6b 100644 --- a/doc/release.txt +++ b/doc/release.txt @@ -10,7 +10,7 @@ - Verify 'make distcheck' produces a tarball with the desired name. - Push tag to public repo. -- Upload source package 'opus-${version}.tar.gz' +- Upload source package 'opus-${version}.tar.gz' - Add to https://svn.xiph.org/releases/opus/ - Update checksum files - svn commit @@ -128,16 +128,16 @@ void silk_CNG( ALLOC( CNG_sig_Q10, length + MAX_LPC_ORDER, opus_int32 ); /* Generate CNG excitation */ - gain_Q16 = silk_SMULWW( psDec->sPLC.randScale_Q14, psDec->sPLC.prevGain_Q16[1] ); - if( gain_Q16 >= (1 << 21) || psCNG->CNG_smth_Gain_Q16 > (1 << 23) ) { - gain_Q16 = silk_SMULTT( gain_Q16, gain_Q16 ); - gain_Q16 = silk_SUB_LSHIFT32(silk_SMULTT( psCNG->CNG_smth_Gain_Q16, psCNG->CNG_smth_Gain_Q16 ), gain_Q16, 5 ); - gain_Q16 = silk_LSHIFT32( silk_SQRT_APPROX( gain_Q16 ), 16 ); - } else { - gain_Q16 = silk_SMULWW( gain_Q16, gain_Q16 ); - gain_Q16 = silk_SUB_LSHIFT32(silk_SMULWW( psCNG->CNG_smth_Gain_Q16, psCNG->CNG_smth_Gain_Q16 ), gain_Q16, 5 ); - gain_Q16 = silk_LSHIFT32( silk_SQRT_APPROX( gain_Q16 ), 8 ); - } + gain_Q16 = silk_SMULWW( psDec->sPLC.randScale_Q14, psDec->sPLC.prevGain_Q16[1] ); + if( gain_Q16 >= (1 << 21) || psCNG->CNG_smth_Gain_Q16 > (1 << 23) ) { + gain_Q16 = silk_SMULTT( gain_Q16, gain_Q16 ); + gain_Q16 = silk_SUB_LSHIFT32(silk_SMULTT( psCNG->CNG_smth_Gain_Q16, psCNG->CNG_smth_Gain_Q16 ), gain_Q16, 5 ); + gain_Q16 = silk_LSHIFT32( silk_SQRT_APPROX( gain_Q16 ), 16 ); + } else { + gain_Q16 = silk_SMULWW( gain_Q16, gain_Q16 ); + gain_Q16 = silk_SUB_LSHIFT32(silk_SMULWW( psCNG->CNG_smth_Gain_Q16, psCNG->CNG_smth_Gain_Q16 ), gain_Q16, 5 ); + gain_Q16 = silk_LSHIFT32( silk_SQRT_APPROX( gain_Q16 ), 8 ); + } silk_CNG_exc( CNG_sig_Q10 + MAX_LPC_ORDER, psCNG->CNG_exc_buf_Q14, gain_Q16, length, &psCNG->rand_seed ); /* Convert CNG NLSF to filter representation */ diff --git a/silk/VQ_WMat_EC.c b/silk/VQ_WMat_EC.c index 9e4ede4b..7983f1db 100644 --- a/silk/VQ_WMat_EC.c +++ b/silk/VQ_WMat_EC.c @@ -55,7 +55,7 @@ void silk_VQ_WMat_EC_c( *rate_dist_Q14 = silk_int32_MAX; cb_row_Q7 = cb_Q7; for( k = 0; k < L; k++ ) { - gain_tmp_Q7 = cb_gain_Q7[k]; + gain_tmp_Q7 = cb_gain_Q7[k]; diff_Q14[ 0 ] = in_Q14[ 0 ] - silk_LSHIFT( cb_row_Q7[ 0 ], 7 ); diff_Q14[ 1 ] = in_Q14[ 1 ] - silk_LSHIFT( cb_row_Q7[ 1 ], 7 ); @@ -66,8 +66,8 @@ void silk_VQ_WMat_EC_c( /* Weighted rate */ sum1_Q14 = silk_SMULBB( mu_Q9, cl_Q5[ k ] ); - /* Penalty for too large gain */ - sum1_Q14 = silk_ADD_LSHIFT32( sum1_Q14, silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 10 ); + /* Penalty for too large gain */ + sum1_Q14 = silk_ADD_LSHIFT32( sum1_Q14, silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 10 ); silk_assert( sum1_Q14 >= 0 ); @@ -111,7 +111,7 @@ void silk_VQ_WMat_EC_c( if( sum1_Q14 < *rate_dist_Q14 ) { *rate_dist_Q14 = sum1_Q14; *ind = (opus_int8)k; - *gain_Q7 = gain_tmp_Q7; + *gain_Q7 = gain_tmp_Q7; } /* Go to next cbk vector */ diff --git a/silk/fixed/burg_modified_FIX.c b/silk/fixed/burg_modified_FIX.c index d0a07f94..4878553b 100644 --- a/silk/fixed/burg_modified_FIX.c +++ b/silk/fixed/burg_modified_FIX.c @@ -73,13 +73,13 @@ void silk_burg_modified_c( rshifts = 32 + 1 + N_BITS_HEAD_ROOM - lz; if (rshifts > MAX_RSHIFTS) rshifts = MAX_RSHIFTS; if (rshifts < MIN_RSHIFTS) rshifts = MIN_RSHIFTS; - + if (rshifts > 0) { - C0 = (opus_int32)silk_RSHIFT64(C0_64, rshifts ); + C0 = (opus_int32)silk_RSHIFT64(C0_64, rshifts ); } else { C0 = silk_LSHIFT32((opus_int32)C0_64, -rshifts ); } - + CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); if( rshifts > 0 ) { diff --git a/silk/fixed/encode_frame_FIX.c b/silk/fixed/encode_frame_FIX.c index 1561f238..5ef44b03 100644 --- a/silk/fixed/encode_frame_FIX.c +++ b/silk/fixed/encode_frame_FIX.c @@ -289,7 +289,7 @@ opus_int silk_encode_frame_FIX( for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { sEncCtrl.Gains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], gainMult_Q8 ), 8 ); } - + /* Quantize gains */ psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev; silk_gains_quant( psEnc->sCmn.indices.GainsIndices, sEncCtrl.Gains_Q16, diff --git a/silk/fixed/find_pred_coefs_FIX.c b/silk/fixed/find_pred_coefs_FIX.c index 9706eb8a..d308e9cf 100644 --- a/silk/fixed/find_pred_coefs_FIX.c +++ b/silk/fixed/find_pred_coefs_FIX.c @@ -119,16 +119,16 @@ void silk_find_pred_coefs_FIX( silk_memset( psEncCtrl->LTPCoef_Q14, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( opus_int16 ) ); psEncCtrl->LTPredCodGain_Q7 = 0; - psEnc->sCmn.sum_log_gain_Q7 = 0; + psEnc->sCmn.sum_log_gain_Q7 = 0; } /* Limit on total predictive coding gain */ if( psEnc->sCmn.first_frame_after_reset ) { minInvGain_Q30 = SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN_AFTER_RESET, 30 ); - } else { + } else { minInvGain_Q30 = silk_log2lin( silk_SMLAWB( 16 << 7, (opus_int32)psEncCtrl->LTPredCodGain_Q7, SILK_FIX_CONST( 1.0 / 3, 16 ) ) ); /* Q16 */ - minInvGain_Q30 = silk_DIV32_varQ( minInvGain_Q30, - silk_SMULWW( SILK_FIX_CONST( MAX_PREDICTION_POWER_GAIN, 0 ), + minInvGain_Q30 = silk_DIV32_varQ( minInvGain_Q30, + silk_SMULWW( SILK_FIX_CONST( MAX_PREDICTION_POWER_GAIN, 0 ), silk_SMLAWB( SILK_FIX_CONST( 0.25, 18 ), SILK_FIX_CONST( 0.75, 18 ), psEncCtrl->coding_quality_Q14 ) ), 14 ); } diff --git a/silk/float/find_LPC_FLP.c b/silk/float/find_LPC_FLP.c index 61c1ad95..fcfe1c36 100644 --- a/silk/float/find_LPC_FLP.c +++ b/silk/float/find_LPC_FLP.c @@ -99,6 +99,6 @@ void silk_find_LPC_FLP( silk_A2NLSF_FLP( NLSF_Q15, a, psEncC->predictLPCOrder ); } - silk_assert( psEncC->indices.NLSFInterpCoef_Q2 == 4 || + silk_assert( psEncC->indices.NLSFInterpCoef_Q2 == 4 || ( psEncC->useInterpolatedNLSFs && !psEncC->first_frame_after_reset && psEncC->nb_subfr == MAX_NB_SUBFR ) ); } diff --git a/silk/float/find_pred_coefs_FLP.c b/silk/float/find_pred_coefs_FLP.c index 1d2dd72c..1af4fe5f 100644 --- a/silk/float/find_pred_coefs_FLP.c +++ b/silk/float/find_pred_coefs_FLP.c @@ -91,13 +91,13 @@ void silk_find_pred_coefs_FLP( } silk_memset( psEncCtrl->LTPCoef, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( silk_float ) ); psEncCtrl->LTPredCodGain = 0.0f; - psEnc->sCmn.sum_log_gain_Q7 = 0; + psEnc->sCmn.sum_log_gain_Q7 = 0; } /* Limit on total predictive coding gain */ if( psEnc->sCmn.first_frame_after_reset ) { minInvGain = 1.0f / MAX_PREDICTION_POWER_GAIN_AFTER_RESET; - } else { + } else { minInvGain = (silk_float)pow( 2, psEncCtrl->LTPredCodGain / 3 ) / MAX_PREDICTION_POWER_GAIN; minInvGain /= 0.25f + 0.75f * psEncCtrl->coding_quality; } diff --git a/silk/float/pitch_analysis_core_FLP.c b/silk/float/pitch_analysis_core_FLP.c index e58f041b..d0e637a2 100644 --- a/silk/float/pitch_analysis_core_FLP.c +++ b/silk/float/pitch_analysis_core_FLP.c @@ -182,8 +182,8 @@ opus_int silk_pitch_analysis_core_FLP( /* O Voicing estimate: 0 voiced, /* Calculate first vector products before loop */ cross_corr = xcorr[ max_lag_4kHz - min_lag_4kHz ]; - normalizer = silk_energy_FLP( target_ptr, sf_length_8kHz ) + - silk_energy_FLP( basis_ptr, sf_length_8kHz ) + + normalizer = silk_energy_FLP( target_ptr, sf_length_8kHz ) + + silk_energy_FLP( basis_ptr, sf_length_8kHz ) + sf_length_8kHz * 4000.0f; C[ 0 ][ min_lag_4kHz ] += (silk_float)( 2 * cross_corr / normalizer ); diff --git a/silk/log2lin.c b/silk/log2lin.c index a692e009..b7c48e47 100644 --- a/silk/log2lin.c +++ b/silk/log2lin.c @@ -33,7 +33,7 @@ POSSIBILITY OF SUCH DAMAGE. /* Approximation of 2^() (very close inverse of silk_lin2log()) */ /* Convert input to a linear scale */ -opus_int32 silk_log2lin( +opus_int32 silk_log2lin( const opus_int32 inLog_Q7 /* I input on log scale */ ) { @@ -42,8 +42,8 @@ opus_int32 silk_log2lin( if( inLog_Q7 < 0 ) { return 0; } else if ( inLog_Q7 >= 3967 ) { - return silk_int32_MAX; - } + return silk_int32_MAX; + } out = silk_LSHIFT( 1, silk_RSHIFT( inLog_Q7, 7 ) ); frac_Q7 = inLog_Q7 & 0x7F; diff --git a/silk/main.h b/silk/main.h index 444542d0..24d7e66c 100644 --- a/silk/main.h +++ b/silk/main.h @@ -208,7 +208,7 @@ void silk_quant_LTP_gains( opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* I/O (un)quantized LTP gains */ opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook Index */ opus_int8 *periodicity_index, /* O Periodicity Index */ - opus_int32 *sum_gain_dB_Q7, /* I/O Cumulative max prediction gain */ + opus_int32 *sum_gain_dB_Q7, /* I/O Cumulative max prediction gain */ const opus_int32 W_Q18[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Error Weights in Q18 */ opus_int mu_Q9, /* I Mu value (R/D tradeoff) */ opus_int lowComplexity, /* I Flag for low complexity */ diff --git a/silk/mips/NSQ_del_dec_mipsr1.h b/silk/mips/NSQ_del_dec_mipsr1.h index 0a3b4992..f6afd923 100644 --- a/silk/mips/NSQ_del_dec_mipsr1.h +++ b/silk/mips/NSQ_del_dec_mipsr1.h @@ -127,7 +127,7 @@ static inline void silk_noise_shape_quantizer_del_dec( temp64 = __builtin_mips_madd( temp64, pred_lag_ptr[ -2 ], b_Q14_2 ); temp64 = __builtin_mips_madd( temp64, pred_lag_ptr[ -3 ], b_Q14_3 ); temp64 = __builtin_mips_madd( temp64, pred_lag_ptr[ -4 ], b_Q14_4 ); - temp64 += 32768; + temp64 += 32768; LTP_pred_Q14 = __builtin_mips_extr_w(temp64, 16); LTP_pred_Q14 = silk_LSHIFT( LTP_pred_Q14, 1 ); /* Q13 -> Q14 */ pred_lag_ptr++; diff --git a/silk/quant_LTP_gains.c b/silk/quant_LTP_gains.c index f10faee3..513a8c44 100644 --- a/silk/quant_LTP_gains.c +++ b/silk/quant_LTP_gains.c @@ -36,7 +36,7 @@ void silk_quant_LTP_gains( opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* I/O (un)quantized LTP gains */ opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook Index */ opus_int8 *periodicity_index, /* O Periodicity Index */ - opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */ + opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */ const opus_int32 W_Q18[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Error Weights in Q18 */ opus_int mu_Q9, /* I Mu value (R/D tradeoff) */ opus_int lowComplexity, /* I Flag for low complexity */ @@ -52,7 +52,7 @@ void silk_quant_LTP_gains( const opus_int16 *b_Q14_ptr; const opus_int32 *W_Q18_ptr; opus_int32 rate_dist_Q14_subfr, rate_dist_Q14, min_rate_dist_Q14; - opus_int32 sum_log_gain_tmp_Q7, best_sum_log_gain_Q7, max_gain_Q7, gain_Q7; + opus_int32 sum_log_gain_tmp_Q7, best_sum_log_gain_Q7, max_gain_Q7, gain_Q7; /***************************************************/ /* iterate over different codebooks with different */ @@ -75,22 +75,22 @@ void silk_quant_LTP_gains( b_Q14_ptr = B_Q14; rate_dist_Q14 = 0; - sum_log_gain_tmp_Q7 = *sum_log_gain_Q7; + sum_log_gain_tmp_Q7 = *sum_log_gain_Q7; for( j = 0; j < nb_subfr; j++ ) { - max_gain_Q7 = silk_log2lin( ( SILK_FIX_CONST( MAX_SUM_LOG_GAIN_DB / 6.0, 7 ) - sum_log_gain_tmp_Q7 ) - + SILK_FIX_CONST( 7, 7 ) ) - gain_safety; + max_gain_Q7 = silk_log2lin( ( SILK_FIX_CONST( MAX_SUM_LOG_GAIN_DB / 6.0, 7 ) - sum_log_gain_tmp_Q7 ) + + SILK_FIX_CONST( 7, 7 ) ) - gain_safety; silk_VQ_WMat_EC( &temp_idx[ j ], /* O index of best codebook vector */ &rate_dist_Q14_subfr, /* O best weighted quantization error + mu * rate */ - &gain_Q7, /* O sum of absolute LTP coefficients */ + &gain_Q7, /* O sum of absolute LTP coefficients */ b_Q14_ptr, /* I input vector to be quantized */ W_Q18_ptr, /* I weighting matrix */ cbk_ptr_Q7, /* I codebook */ cbk_gain_ptr_Q7, /* I codebook effective gains */ cl_ptr_Q5, /* I code length for each codebook vector */ mu_Q9, /* I tradeoff between weighted error and rate */ - max_gain_Q7, /* I maximum sum of absolute LTP coefficients */ + max_gain_Q7, /* I maximum sum of absolute LTP coefficients */ cbk_size, /* I number of vectors in codebook */ arch /* I Run-time architecture */ ); @@ -110,7 +110,7 @@ void silk_quant_LTP_gains( min_rate_dist_Q14 = rate_dist_Q14; *periodicity_index = (opus_int8)k; silk_memcpy( cbk_index, temp_idx, nb_subfr * sizeof( opus_int8 ) ); - best_sum_log_gain_Q7 = sum_log_gain_tmp_Q7; + best_sum_log_gain_Q7 = sum_log_gain_tmp_Q7; } /* Break early in low-complexity mode if rate distortion is below threshold */ @@ -125,6 +125,5 @@ void silk_quant_LTP_gains( B_Q14[ j * LTP_ORDER + k ] = silk_LSHIFT( cbk_ptr_Q7[ cbk_index[ j ] * LTP_ORDER + k ], 7 ); } } - *sum_log_gain_Q7 = best_sum_log_gain_Q7; + *sum_log_gain_Q7 = best_sum_log_gain_Q7; } - diff --git a/silk/resampler_rom.c b/silk/resampler_rom.c index 2d502706..5e6b0447 100644 --- a/silk/resampler_rom.c +++ b/silk/resampler_rom.c @@ -41,36 +41,36 @@ POSSIBILITY OF SUCH DAMAGE. /* Tables with IIR and FIR coefficients for fractional downsamplers (123 Words) */ silk_DWORD_ALIGN const opus_int16 silk_Resampler_3_4_COEFS[ 2 + 3 * RESAMPLER_DOWN_ORDER_FIR0 / 2 ] = { - -20694, -13867, - -49, 64, 17, -157, 353, -496, 163, 11047, 22205, - -39, 6, 91, -170, 186, 23, -896, 6336, 19928, - -19, -36, 102, -89, -24, 328, -951, 2568, 15909, + -20694, -13867, + -49, 64, 17, -157, 353, -496, 163, 11047, 22205, + -39, 6, 91, -170, 186, 23, -896, 6336, 19928, + -19, -36, 102, -89, -24, 328, -951, 2568, 15909, }; silk_DWORD_ALIGN const opus_int16 silk_Resampler_2_3_COEFS[ 2 + 2 * RESAMPLER_DOWN_ORDER_FIR0 / 2 ] = { - -14457, -14019, - 64, 128, -122, 36, 310, -768, 584, 9267, 17733, - 12, 128, 18, -142, 288, -117, -865, 4123, 14459, + -14457, -14019, + 64, 128, -122, 36, 310, -768, 584, 9267, 17733, + 12, 128, 18, -142, 288, -117, -865, 4123, 14459, }; silk_DWORD_ALIGN const opus_int16 silk_Resampler_1_2_COEFS[ 2 + RESAMPLER_DOWN_ORDER_FIR1 / 2 ] = { - 616, -14323, - -10, 39, 58, -46, -84, 120, 184, -315, -541, 1284, 5380, 9024, + 616, -14323, + -10, 39, 58, -46, -84, 120, 184, -315, -541, 1284, 5380, 9024, }; silk_DWORD_ALIGN const opus_int16 silk_Resampler_1_3_COEFS[ 2 + RESAMPLER_DOWN_ORDER_FIR2 / 2 ] = { - 16102, -15162, - -13, 0, 20, 26, 5, -31, -43, -4, 65, 90, 7, -157, -248, -44, 593, 1583, 2612, 3271, + 16102, -15162, + -13, 0, 20, 26, 5, -31, -43, -4, 65, 90, 7, -157, -248, -44, 593, 1583, 2612, 3271, }; silk_DWORD_ALIGN const opus_int16 silk_Resampler_1_4_COEFS[ 2 + RESAMPLER_DOWN_ORDER_FIR2 / 2 ] = { - 22500, -15099, - 3, -14, -20, -15, 2, 25, 37, 25, -16, -71, -107, -79, 50, 292, 623, 982, 1288, 1464, + 22500, -15099, + 3, -14, -20, -15, 2, 25, 37, 25, -16, -71, -107, -79, 50, 292, 623, 982, 1288, 1464, }; silk_DWORD_ALIGN const opus_int16 silk_Resampler_1_6_COEFS[ 2 + RESAMPLER_DOWN_ORDER_FIR2 / 2 ] = { - 27540, -15257, - 17, 12, 8, 1, -10, -22, -30, -32, -22, 3, 44, 100, 168, 243, 317, 381, 429, 455, + 27540, -15257, + 17, 12, 8, 1, -10, -22, -30, -32, -22, 3, 44, 100, 168, 243, 317, 381, 429, 455, }; silk_DWORD_ALIGN const opus_int16 silk_Resampler_2_3_COEFS_LQ[ 2 + 2 * 2 ] = { @@ -81,16 +81,16 @@ silk_DWORD_ALIGN const opus_int16 silk_Resampler_2_3_COEFS_LQ[ 2 + 2 * 2 ] = { /* Table with interplation fractions of 1/24, 3/24, 5/24, ... , 23/24 : 23/24 (46 Words) */ silk_DWORD_ALIGN const opus_int16 silk_resampler_frac_FIR_12[ 12 ][ RESAMPLER_ORDER_FIR_12 / 2 ] = { - { 189, -600, 617, 30567 }, - { 117, -159, -1070, 29704 }, - { 52, 221, -2392, 28276 }, - { -4, 529, -3350, 26341 }, - { -48, 758, -3956, 23973 }, - { -80, 905, -4235, 21254 }, - { -99, 972, -4222, 18278 }, - { -107, 967, -3957, 15143 }, - { -103, 896, -3487, 11950 }, - { -91, 773, -2865, 8798 }, - { -71, 611, -2143, 5784 }, - { -46, 425, -1375, 2996 }, + { 189, -600, 617, 30567 }, + { 117, -159, -1070, 29704 }, + { 52, 221, -2392, 28276 }, + { -4, 529, -3350, 26341 }, + { -48, 758, -3956, 23973 }, + { -80, 905, -4235, 21254 }, + { -99, 972, -4222, 18278 }, + { -107, 967, -3957, 15143 }, + { -103, 896, -3487, 11950 }, + { -91, 773, -2865, 8798 }, + { -71, 611, -2143, 5784 }, + { -46, 425, -1375, 2996 }, }; diff --git a/silk/structs.h b/silk/structs.h index 1826b36a..827829dc 100644 --- a/silk/structs.h +++ b/silk/structs.h @@ -171,7 +171,7 @@ typedef struct { opus_int32 pitchEstimationThreshold_Q16; /* Threshold for pitch estimator */ opus_int LTPQuantLowComplexity; /* Flag for low complexity LTP quantization */ opus_int mu_LTP_Q9; /* Rate-distortion tradeoff in LTP quantization */ - opus_int32 sum_log_gain_Q7; /* Cumulative max prediction gain */ + opus_int32 sum_log_gain_Q7; /* Cumulative max prediction gain */ opus_int NLSF_MSVQ_Survivors; /* Number of survivors in NLSF MSVQ */ opus_int first_frame_after_reset; /* Flag for deactivating NLSF interpolation, pitch prediction */ opus_int controlled_since_last_payload; /* Flag for ensuring codec_control only runs once per packet */ diff --git a/silk/tuning_parameters.h b/silk/tuning_parameters.h index e1057bba..5b8f4042 100644 --- a/silk/tuning_parameters.h +++ b/silk/tuning_parameters.h @@ -64,7 +64,7 @@ extern "C" #define MU_LTP_QUANT_WB 0.02f /* Max cumulative LTP gain */ -#define MAX_SUM_LOG_GAIN_DB 250.0f +#define MAX_SUM_LOG_GAIN_DB 250.0f /***********************/ /* High pass filtering */ @@ -41,36 +41,36 @@ #if 0 static OPUS_INLINE opus_val16 tansig_approx(opus_val32 _x) /* Q19 */ { - int i; - opus_val16 xx; /* Q11 */ - /*double x, y;*/ - opus_val16 dy, yy; /* Q14 */ - /*x = 1.9073e-06*_x;*/ - if (_x>=QCONST32(8,19)) - return QCONST32(1.,14); - if (_x<=-QCONST32(8,19)) - return -QCONST32(1.,14); - xx = EXTRACT16(SHR32(_x, 8)); - /*i = lrint(25*x);*/ - i = SHR32(ADD32(1024,MULT16_16(25, xx)),11); - /*x -= .04*i;*/ - xx -= EXTRACT16(SHR32(MULT16_16(20972,i),8)); - /*x = xx*(1./2048);*/ - /*y = tansig_table[250+i];*/ - yy = tansig_table[250+i]; - /*y = yy*(1./16384);*/ - dy = 16384-MULT16_16_Q14(yy,yy); - yy = yy + MULT16_16_Q14(MULT16_16_Q11(xx,dy),(16384 - MULT16_16_Q11(yy,xx))); - return yy; + int i; + opus_val16 xx; /* Q11 */ + /*double x, y;*/ + opus_val16 dy, yy; /* Q14 */ + /*x = 1.9073e-06*_x;*/ + if (_x>=QCONST32(8,19)) + return QCONST32(1.,14); + if (_x<=-QCONST32(8,19)) + return -QCONST32(1.,14); + xx = EXTRACT16(SHR32(_x, 8)); + /*i = lrint(25*x);*/ + i = SHR32(ADD32(1024,MULT16_16(25, xx)),11); + /*x -= .04*i;*/ + xx -= EXTRACT16(SHR32(MULT16_16(20972,i),8)); + /*x = xx*(1./2048);*/ + /*y = tansig_table[250+i];*/ + yy = tansig_table[250+i]; + /*y = yy*(1./16384);*/ + dy = 16384-MULT16_16_Q14(yy,yy); + yy = yy + MULT16_16_Q14(MULT16_16_Q11(xx,dy),(16384 - MULT16_16_Q11(yy,xx))); + return yy; } #else /*extern const float tansig_table[501];*/ static OPUS_INLINE float tansig_approx(float x) { - int i; - float y, dy; - float sign=1; - /* Tests are reversed to catch NaNs */ + int i; + float y, dy; + float sign=1; + /* Tests are reversed to catch NaNs */ if (!(x<8)) return 1; if (!(x>-8)) @@ -80,43 +80,43 @@ static OPUS_INLINE float tansig_approx(float x) if (celt_isnan(x)) return 0; #endif - if (x<0) - { - x=-x; - sign=-1; - } - i = (int)floor(.5f+25*x); - x -= .04f*i; - y = tansig_table[i]; - dy = 1-y*y; - y = y + x*dy*(1 - y*x); - return sign*y; + if (x<0) + { + x=-x; + sign=-1; + } + i = (int)floor(.5f+25*x); + x -= .04f*i; + y = tansig_table[i]; + dy = 1-y*y; + y = y + x*dy*(1 - y*x); + return sign*y; } #endif #if 0 void mlp_process(const MLP *m, const opus_val16 *in, opus_val16 *out) { - int j; - opus_val16 hidden[MAX_NEURONS]; - const opus_val16 *W = m->weights; - /* Copy to tmp_in */ - for (j=0;j<m->topo[1];j++) - { - int k; - opus_val32 sum = SHL32(EXTEND32(*W++),8); - for (k=0;k<m->topo[0];k++) - sum = MAC16_16(sum, in[k],*W++); - hidden[j] = tansig_approx(sum); - } - for (j=0;j<m->topo[2];j++) - { - int k; - opus_val32 sum = SHL32(EXTEND32(*W++),14); - for (k=0;k<m->topo[1];k++) - sum = MAC16_16(sum, hidden[k], *W++); - out[j] = tansig_approx(EXTRACT16(PSHR32(sum,17))); - } + int j; + opus_val16 hidden[MAX_NEURONS]; + const opus_val16 *W = m->weights; + /* Copy to tmp_in */ + for (j=0;j<m->topo[1];j++) + { + int k; + opus_val32 sum = SHL32(EXTEND32(*W++),8); + for (k=0;k<m->topo[0];k++) + sum = MAC16_16(sum, in[k],*W++); + hidden[j] = tansig_approx(sum); + } + for (j=0;j<m->topo[2];j++) + { + int k; + opus_val32 sum = SHL32(EXTEND32(*W++),14); + for (k=0;k<m->topo[1];k++) + sum = MAC16_16(sum, hidden[k], *W++); + out[j] = tansig_approx(EXTRACT16(PSHR32(sum,17))); + } } #else void mlp_process(const MLP *m, const float *in, float *out) @@ -31,9 +31,9 @@ #include "arch.h" typedef struct { - int layers; - const int *topo; - const float *weights; + int layers; + const int *topo; + const float *weights; } MLP; void mlp_process(const MLP *m, const float *in, float *out); diff --git a/src/mlp_train.c b/src/mlp_train.c index 769f503d..e02fea93 100644 --- a/src/mlp_train.c +++ b/src/mlp_train.c @@ -39,70 +39,70 @@ int stopped = 0; void handler(int sig) { - stopped = 1; - signal(sig, handler); + stopped = 1; + signal(sig, handler); } MLPTrain * mlp_init(int *topo, int nbLayers, float *inputs, float *outputs, int nbSamples) { - int i, j, k; - MLPTrain *net; - int inDim, outDim; - net = malloc(sizeof(*net)); - net->topo = malloc(nbLayers*sizeof(net->topo[0])); - for (i=0;i<nbLayers;i++) - net->topo[i] = topo[i]; - inDim = topo[0]; - outDim = topo[nbLayers-1]; - net->in_rate = malloc((inDim+1)*sizeof(net->in_rate[0])); - net->weights = malloc((nbLayers-1)*sizeof(net->weights)); - net->best_weights = malloc((nbLayers-1)*sizeof(net->weights)); - for (i=0;i<nbLayers-1;i++) - { - net->weights[i] = malloc((topo[i]+1)*topo[i+1]*sizeof(net->weights[0][0])); - net->best_weights[i] = malloc((topo[i]+1)*topo[i+1]*sizeof(net->weights[0][0])); - } - double inMean[inDim]; - for (j=0;j<inDim;j++) - { - double std=0; - inMean[j] = 0; - for (i=0;i<nbSamples;i++) - { - inMean[j] += inputs[i*inDim+j]; - std += inputs[i*inDim+j]*inputs[i*inDim+j]; - } - inMean[j] /= nbSamples; - std /= nbSamples; - net->in_rate[1+j] = .5/(.0001+std); - std = std-inMean[j]*inMean[j]; - if (std<.001) - std = .001; - std = 1/sqrt(inDim*std); - for (k=0;k<topo[1];k++) - net->weights[0][k*(topo[0]+1)+j+1] = randn(std); - } - net->in_rate[0] = 1; - for (j=0;j<topo[1];j++) - { - double sum = 0; - for (k=0;k<inDim;k++) - sum += inMean[k]*net->weights[0][j*(topo[0]+1)+k+1]; - net->weights[0][j*(topo[0]+1)] = -sum; - } - for (j=0;j<outDim;j++) - { - double mean = 0; - double std; - for (i=0;i<nbSamples;i++) - mean += outputs[i*outDim+j]; - mean /= nbSamples; - std = 1/sqrt(topo[nbLayers-2]); - net->weights[nbLayers-2][j*(topo[nbLayers-2]+1)] = mean; - for (k=0;k<topo[nbLayers-2];k++) - net->weights[nbLayers-2][j*(topo[nbLayers-2]+1)+k+1] = randn(std); - } - return net; + int i, j, k; + MLPTrain *net; + int inDim, outDim; + net = malloc(sizeof(*net)); + net->topo = malloc(nbLayers*sizeof(net->topo[0])); + for (i=0;i<nbLayers;i++) + net->topo[i] = topo[i]; + inDim = topo[0]; + outDim = topo[nbLayers-1]; + net->in_rate = malloc((inDim+1)*sizeof(net->in_rate[0])); + net->weights = malloc((nbLayers-1)*sizeof(net->weights)); + net->best_weights = malloc((nbLayers-1)*sizeof(net->weights)); + for (i=0;i<nbLayers-1;i++) + { + net->weights[i] = malloc((topo[i]+1)*topo[i+1]*sizeof(net->weights[0][0])); + net->best_weights[i] = malloc((topo[i]+1)*topo[i+1]*sizeof(net->weights[0][0])); + } + double inMean[inDim]; + for (j=0;j<inDim;j++) + { + double std=0; + inMean[j] = 0; + for (i=0;i<nbSamples;i++) + { + inMean[j] += inputs[i*inDim+j]; + std += inputs[i*inDim+j]*inputs[i*inDim+j]; + } + inMean[j] /= nbSamples; + std /= nbSamples; + net->in_rate[1+j] = .5/(.0001+std); + std = std-inMean[j]*inMean[j]; + if (std<.001) + std = .001; + std = 1/sqrt(inDim*std); + for (k=0;k<topo[1];k++) + net->weights[0][k*(topo[0]+1)+j+1] = randn(std); + } + net->in_rate[0] = 1; + for (j=0;j<topo[1];j++) + { + double sum = 0; + for (k=0;k<inDim;k++) + sum += inMean[k]*net->weights[0][j*(topo[0]+1)+k+1]; + net->weights[0][j*(topo[0]+1)] = -sum; + } + for (j=0;j<outDim;j++) + { + double mean = 0; + double std; + for (i=0;i<nbSamples;i++) + mean += outputs[i*outDim+j]; + mean /= nbSamples; + std = 1/sqrt(topo[nbLayers-2]); + net->weights[nbLayers-2][j*(topo[nbLayers-2]+1)] = mean; + for (k=0;k<topo[nbLayers-2];k++) + net->weights[nbLayers-2][j*(topo[nbLayers-2]+1)+k+1] = randn(std); + } + return net; } #define MAX_NEURONS 100 @@ -110,75 +110,75 @@ MLPTrain * mlp_init(int *topo, int nbLayers, float *inputs, float *outputs, int double compute_gradient(MLPTrain *net, float *inputs, float *outputs, int nbSamples, double *W0_grad, double *W1_grad, double *error_rate) { - int i,j; - int s; - int inDim, outDim, hiddenDim; - int *topo; - double *W0, *W1; - double rms=0; - int W0_size, W1_size; - double hidden[MAX_NEURONS]; - double netOut[MAX_NEURONS]; - double error[MAX_NEURONS]; + int i,j; + int s; + int inDim, outDim, hiddenDim; + int *topo; + double *W0, *W1; + double rms=0; + int W0_size, W1_size; + double hidden[MAX_NEURONS]; + double netOut[MAX_NEURONS]; + double error[MAX_NEURONS]; - topo = net->topo; - inDim = net->topo[0]; - hiddenDim = net->topo[1]; - outDim = net->topo[2]; - W0_size = (topo[0]+1)*topo[1]; - W1_size = (topo[1]+1)*topo[2]; - W0 = net->weights[0]; - W1 = net->weights[1]; - memset(W0_grad, 0, W0_size*sizeof(double)); - memset(W1_grad, 0, W1_size*sizeof(double)); - for (i=0;i<outDim;i++) - netOut[i] = outputs[i]; - for (i=0;i<outDim;i++) - error_rate[i] = 0; - for (s=0;s<nbSamples;s++) - { - float *in, *out; - in = inputs+s*inDim; - out = outputs + s*outDim; - for (i=0;i<hiddenDim;i++) - { - double sum = W0[i*(inDim+1)]; - for (j=0;j<inDim;j++) - sum += W0[i*(inDim+1)+j+1]*in[j]; - hidden[i] = tansig_approx(sum); - } - for (i=0;i<outDim;i++) - { - double sum = W1[i*(hiddenDim+1)]; - for (j=0;j<hiddenDim;j++) - sum += W1[i*(hiddenDim+1)+j+1]*hidden[j]; - netOut[i] = tansig_approx(sum); - error[i] = out[i] - netOut[i]; - rms += error[i]*error[i]; - error_rate[i] += fabs(error[i])>1; - /*error[i] = error[i]/(1+fabs(error[i]));*/ - } - /* Back-propagate error */ - for (i=0;i<outDim;i++) - { - float grad = 1-netOut[i]*netOut[i]; - W1_grad[i*(hiddenDim+1)] += error[i]*grad; - for (j=0;j<hiddenDim;j++) - W1_grad[i*(hiddenDim+1)+j+1] += grad*error[i]*hidden[j]; - } - for (i=0;i<hiddenDim;i++) - { - double grad; - grad = 0; - for (j=0;j<outDim;j++) - grad += error[j]*W1[j*(hiddenDim+1)+i+1]; - grad *= 1-hidden[i]*hidden[i]; - W0_grad[i*(inDim+1)] += grad; - for (j=0;j<inDim;j++) - W0_grad[i*(inDim+1)+j+1] += grad*in[j]; - } - } - return rms; + topo = net->topo; + inDim = net->topo[0]; + hiddenDim = net->topo[1]; + outDim = net->topo[2]; + W0_size = (topo[0]+1)*topo[1]; + W1_size = (topo[1]+1)*topo[2]; + W0 = net->weights[0]; + W1 = net->weights[1]; + memset(W0_grad, 0, W0_size*sizeof(double)); + memset(W1_grad, 0, W1_size*sizeof(double)); + for (i=0;i<outDim;i++) + netOut[i] = outputs[i]; + for (i=0;i<outDim;i++) + error_rate[i] = 0; + for (s=0;s<nbSamples;s++) + { + float *in, *out; + in = inputs+s*inDim; + out = outputs + s*outDim; + for (i=0;i<hiddenDim;i++) + { + double sum = W0[i*(inDim+1)]; + for (j=0;j<inDim;j++) + sum += W0[i*(inDim+1)+j+1]*in[j]; + hidden[i] = tansig_approx(sum); + } + for (i=0;i<outDim;i++) + { + double sum = W1[i*(hiddenDim+1)]; + for (j=0;j<hiddenDim;j++) + sum += W1[i*(hiddenDim+1)+j+1]*hidden[j]; + netOut[i] = tansig_approx(sum); + error[i] = out[i] - netOut[i]; + rms += error[i]*error[i]; + error_rate[i] += fabs(error[i])>1; + /*error[i] = error[i]/(1+fabs(error[i]));*/ + } + /* Back-propagate error */ + for (i=0;i<outDim;i++) + { + float grad = 1-netOut[i]*netOut[i]; + W1_grad[i*(hiddenDim+1)] += error[i]*grad; + for (j=0;j<hiddenDim;j++) + W1_grad[i*(hiddenDim+1)+j+1] += grad*error[i]*hidden[j]; + } + for (i=0;i<hiddenDim;i++) + { + double grad; + grad = 0; + for (j=0;j<outDim;j++) + grad += error[j]*W1[j*(hiddenDim+1)+i+1]; + grad *= 1-hidden[i]*hidden[i]; + W0_grad[i*(inDim+1)] += grad; + for (j=0;j<inDim;j++) + W0_grad[i*(inDim+1)+j+1] += grad*in[j]; + } + } + return rms; } #define NB_THREADS 8 @@ -187,315 +187,315 @@ sem_t sem_begin[NB_THREADS]; sem_t sem_end[NB_THREADS]; struct GradientArg { - int id; - int done; - MLPTrain *net; - float *inputs; - float *outputs; - int nbSamples; - double *W0_grad; - double *W1_grad; - double rms; - double error_rate[MAX_OUT]; + int id; + int done; + MLPTrain *net; + float *inputs; + float *outputs; + int nbSamples; + double *W0_grad; + double *W1_grad; + double rms; + double error_rate[MAX_OUT]; }; void *gradient_thread_process(void *_arg) { - int W0_size, W1_size; - struct GradientArg *arg = _arg; - int *topo = arg->net->topo; - W0_size = (topo[0]+1)*topo[1]; - W1_size = (topo[1]+1)*topo[2]; - double W0_grad[W0_size]; - double W1_grad[W1_size]; - arg->W0_grad = W0_grad; - arg->W1_grad = W1_grad; - while (1) - { - sem_wait(&sem_begin[arg->id]); - if (arg->done) - break; - arg->rms = compute_gradient(arg->net, arg->inputs, arg->outputs, arg->nbSamples, arg->W0_grad, arg->W1_grad, arg->error_rate); - sem_post(&sem_end[arg->id]); - } - fprintf(stderr, "done\n"); - return NULL; + int W0_size, W1_size; + struct GradientArg *arg = _arg; + int *topo = arg->net->topo; + W0_size = (topo[0]+1)*topo[1]; + W1_size = (topo[1]+1)*topo[2]; + double W0_grad[W0_size]; + double W1_grad[W1_size]; + arg->W0_grad = W0_grad; + arg->W1_grad = W1_grad; + while (1) + { + sem_wait(&sem_begin[arg->id]); + if (arg->done) + break; + arg->rms = compute_gradient(arg->net, arg->inputs, arg->outputs, arg->nbSamples, arg->W0_grad, arg->W1_grad, arg->error_rate); + sem_post(&sem_end[arg->id]); + } + fprintf(stderr, "done\n"); + return NULL; } float mlp_train_backprop(MLPTrain *net, float *inputs, float *outputs, int nbSamples, int nbEpoch, float rate) { - int i, j; - int e; - float best_rms = 1e10; - int inDim, outDim, hiddenDim; - int *topo; - double *W0, *W1, *best_W0, *best_W1; - double *W0_old, *W1_old; - double *W0_old2, *W1_old2; - double *W0_grad, *W1_grad; - double *W0_oldgrad, *W1_oldgrad; - double *W0_rate, *W1_rate; - double *best_W0_rate, *best_W1_rate; - int W0_size, W1_size; - topo = net->topo; - W0_size = (topo[0]+1)*topo[1]; - W1_size = (topo[1]+1)*topo[2]; - struct GradientArg args[NB_THREADS]; - pthread_t thread[NB_THREADS]; - int samplePerPart = nbSamples/NB_THREADS; - int count_worse=0; - int count_retries=0; + int i, j; + int e; + float best_rms = 1e10; + int inDim, outDim, hiddenDim; + int *topo; + double *W0, *W1, *best_W0, *best_W1; + double *W0_old, *W1_old; + double *W0_old2, *W1_old2; + double *W0_grad, *W1_grad; + double *W0_oldgrad, *W1_oldgrad; + double *W0_rate, *W1_rate; + double *best_W0_rate, *best_W1_rate; + int W0_size, W1_size; + topo = net->topo; + W0_size = (topo[0]+1)*topo[1]; + W1_size = (topo[1]+1)*topo[2]; + struct GradientArg args[NB_THREADS]; + pthread_t thread[NB_THREADS]; + int samplePerPart = nbSamples/NB_THREADS; + int count_worse=0; + int count_retries=0; - topo = net->topo; - inDim = net->topo[0]; - hiddenDim = net->topo[1]; - outDim = net->topo[2]; - W0 = net->weights[0]; - W1 = net->weights[1]; - best_W0 = net->best_weights[0]; - best_W1 = net->best_weights[1]; - W0_old = malloc(W0_size*sizeof(double)); - W1_old = malloc(W1_size*sizeof(double)); - W0_old2 = malloc(W0_size*sizeof(double)); - W1_old2 = malloc(W1_size*sizeof(double)); - W0_grad = malloc(W0_size*sizeof(double)); - W1_grad = malloc(W1_size*sizeof(double)); - W0_oldgrad = malloc(W0_size*sizeof(double)); - W1_oldgrad = malloc(W1_size*sizeof(double)); - W0_rate = malloc(W0_size*sizeof(double)); - W1_rate = malloc(W1_size*sizeof(double)); - best_W0_rate = malloc(W0_size*sizeof(double)); - best_W1_rate = malloc(W1_size*sizeof(double)); - memcpy(W0_old, W0, W0_size*sizeof(double)); - memcpy(W0_old2, W0, W0_size*sizeof(double)); - memset(W0_grad, 0, W0_size*sizeof(double)); - memset(W0_oldgrad, 0, W0_size*sizeof(double)); - memcpy(W1_old, W1, W1_size*sizeof(double)); - memcpy(W1_old2, W1, W1_size*sizeof(double)); - memset(W1_grad, 0, W1_size*sizeof(double)); - memset(W1_oldgrad, 0, W1_size*sizeof(double)); - - rate /= nbSamples; - for (i=0;i<hiddenDim;i++) - for (j=0;j<inDim+1;j++) - W0_rate[i*(inDim+1)+j] = rate*net->in_rate[j]; - for (i=0;i<W1_size;i++) - W1_rate[i] = rate; - - for (i=0;i<NB_THREADS;i++) - { - args[i].net = net; - args[i].inputs = inputs+i*samplePerPart*inDim; - args[i].outputs = outputs+i*samplePerPart*outDim; - args[i].nbSamples = samplePerPart; - args[i].id = i; - args[i].done = 0; - sem_init(&sem_begin[i], 0, 0); - sem_init(&sem_end[i], 0, 0); - pthread_create(&thread[i], NULL, gradient_thread_process, &args[i]); - } - for (e=0;e<nbEpoch;e++) - { - double rms=0; - double error_rate[2] = {0,0}; - for (i=0;i<NB_THREADS;i++) - { - sem_post(&sem_begin[i]); - } - memset(W0_grad, 0, W0_size*sizeof(double)); - memset(W1_grad, 0, W1_size*sizeof(double)); - for (i=0;i<NB_THREADS;i++) - { - sem_wait(&sem_end[i]); - rms += args[i].rms; - error_rate[0] += args[i].error_rate[0]; + topo = net->topo; + inDim = net->topo[0]; + hiddenDim = net->topo[1]; + outDim = net->topo[2]; + W0 = net->weights[0]; + W1 = net->weights[1]; + best_W0 = net->best_weights[0]; + best_W1 = net->best_weights[1]; + W0_old = malloc(W0_size*sizeof(double)); + W1_old = malloc(W1_size*sizeof(double)); + W0_old2 = malloc(W0_size*sizeof(double)); + W1_old2 = malloc(W1_size*sizeof(double)); + W0_grad = malloc(W0_size*sizeof(double)); + W1_grad = malloc(W1_size*sizeof(double)); + W0_oldgrad = malloc(W0_size*sizeof(double)); + W1_oldgrad = malloc(W1_size*sizeof(double)); + W0_rate = malloc(W0_size*sizeof(double)); + W1_rate = malloc(W1_size*sizeof(double)); + best_W0_rate = malloc(W0_size*sizeof(double)); + best_W1_rate = malloc(W1_size*sizeof(double)); + memcpy(W0_old, W0, W0_size*sizeof(double)); + memcpy(W0_old2, W0, W0_size*sizeof(double)); + memset(W0_grad, 0, W0_size*sizeof(double)); + memset(W0_oldgrad, 0, W0_size*sizeof(double)); + memcpy(W1_old, W1, W1_size*sizeof(double)); + memcpy(W1_old2, W1, W1_size*sizeof(double)); + memset(W1_grad, 0, W1_size*sizeof(double)); + memset(W1_oldgrad, 0, W1_size*sizeof(double)); + + rate /= nbSamples; + for (i=0;i<hiddenDim;i++) + for (j=0;j<inDim+1;j++) + W0_rate[i*(inDim+1)+j] = rate*net->in_rate[j]; + for (i=0;i<W1_size;i++) + W1_rate[i] = rate; + + for (i=0;i<NB_THREADS;i++) + { + args[i].net = net; + args[i].inputs = inputs+i*samplePerPart*inDim; + args[i].outputs = outputs+i*samplePerPart*outDim; + args[i].nbSamples = samplePerPart; + args[i].id = i; + args[i].done = 0; + sem_init(&sem_begin[i], 0, 0); + sem_init(&sem_end[i], 0, 0); + pthread_create(&thread[i], NULL, gradient_thread_process, &args[i]); + } + for (e=0;e<nbEpoch;e++) + { + double rms=0; + double error_rate[2] = {0,0}; + for (i=0;i<NB_THREADS;i++) + { + sem_post(&sem_begin[i]); + } + memset(W0_grad, 0, W0_size*sizeof(double)); + memset(W1_grad, 0, W1_size*sizeof(double)); + for (i=0;i<NB_THREADS;i++) + { + sem_wait(&sem_end[i]); + rms += args[i].rms; + error_rate[0] += args[i].error_rate[0]; error_rate[1] += args[i].error_rate[1]; - for (j=0;j<W0_size;j++) - W0_grad[j] += args[i].W0_grad[j]; - for (j=0;j<W1_size;j++) - W1_grad[j] += args[i].W1_grad[j]; - } + for (j=0;j<W0_size;j++) + W0_grad[j] += args[i].W0_grad[j]; + for (j=0;j<W1_size;j++) + W1_grad[j] += args[i].W1_grad[j]; + } - float mean_rate = 0, min_rate = 1e10; - rms = (rms/(outDim*nbSamples)); - error_rate[0] = (error_rate[0]/(nbSamples)); + float mean_rate = 0, min_rate = 1e10; + rms = (rms/(outDim*nbSamples)); + error_rate[0] = (error_rate[0]/(nbSamples)); error_rate[1] = (error_rate[1]/(nbSamples)); - fprintf (stderr, "%f %f (%f %f) ", error_rate[0], error_rate[1], rms, best_rms); - if (rms < best_rms) - { - best_rms = rms; - for (i=0;i<W0_size;i++) - { - best_W0[i] = W0[i]; - best_W0_rate[i] = W0_rate[i]; - } - for (i=0;i<W1_size;i++) - { - best_W1[i] = W1[i]; - best_W1_rate[i] = W1_rate[i]; - } - count_worse=0; - count_retries=0; - } else { - count_worse++; - if (count_worse>30) - { - count_retries++; - count_worse=0; - for (i=0;i<W0_size;i++) - { - W0[i] = best_W0[i]; - best_W0_rate[i] *= .7; - if (best_W0_rate[i]<1e-15) best_W0_rate[i]=1e-15; - W0_rate[i] = best_W0_rate[i]; - W0_grad[i] = 0; - } - for (i=0;i<W1_size;i++) - { - W1[i] = best_W1[i]; - best_W1_rate[i] *= .8; - if (best_W1_rate[i]<1e-15) best_W1_rate[i]=1e-15; - W1_rate[i] = best_W1_rate[i]; - W1_grad[i] = 0; - } - } - } - if (count_retries>10) - break; - for (i=0;i<W0_size;i++) - { - if (W0_oldgrad[i]*W0_grad[i] > 0) - W0_rate[i] *= 1.01; - else if (W0_oldgrad[i]*W0_grad[i] < 0) - W0_rate[i] *= .9; - mean_rate += W0_rate[i]; - if (W0_rate[i] < min_rate) - min_rate = W0_rate[i]; - if (W0_rate[i] < 1e-15) - W0_rate[i] = 1e-15; - /*if (W0_rate[i] > .01) - W0_rate[i] = .01;*/ - W0_oldgrad[i] = W0_grad[i]; - W0_old2[i] = W0_old[i]; - W0_old[i] = W0[i]; - W0[i] += W0_grad[i]*W0_rate[i]; - } - for (i=0;i<W1_size;i++) - { - if (W1_oldgrad[i]*W1_grad[i] > 0) - W1_rate[i] *= 1.01; - else if (W1_oldgrad[i]*W1_grad[i] < 0) - W1_rate[i] *= .9; - mean_rate += W1_rate[i]; - if (W1_rate[i] < min_rate) - min_rate = W1_rate[i]; - if (W1_rate[i] < 1e-15) - W1_rate[i] = 1e-15; - W1_oldgrad[i] = W1_grad[i]; - W1_old2[i] = W1_old[i]; - W1_old[i] = W1[i]; - W1[i] += W1_grad[i]*W1_rate[i]; - } - mean_rate /= (topo[0]+1)*topo[1] + (topo[1]+1)*topo[2]; - fprintf (stderr, "%g %d", mean_rate, e); - if (count_retries) - fprintf(stderr, " %d", count_retries); - fprintf(stderr, "\n"); - if (stopped) - break; - } - for (i=0;i<NB_THREADS;i++) - { - args[i].done = 1; - sem_post(&sem_begin[i]); - pthread_join(thread[i], NULL); - fprintf (stderr, "joined %d\n", i); - } - free(W0_old); - free(W1_old); - free(W0_grad); - free(W1_grad); - free(W0_rate); - free(W1_rate); - return best_rms; + fprintf (stderr, "%f %f (%f %f) ", error_rate[0], error_rate[1], rms, best_rms); + if (rms < best_rms) + { + best_rms = rms; + for (i=0;i<W0_size;i++) + { + best_W0[i] = W0[i]; + best_W0_rate[i] = W0_rate[i]; + } + for (i=0;i<W1_size;i++) + { + best_W1[i] = W1[i]; + best_W1_rate[i] = W1_rate[i]; + } + count_worse=0; + count_retries=0; + } else { + count_worse++; + if (count_worse>30) + { + count_retries++; + count_worse=0; + for (i=0;i<W0_size;i++) + { + W0[i] = best_W0[i]; + best_W0_rate[i] *= .7; + if (best_W0_rate[i]<1e-15) best_W0_rate[i]=1e-15; + W0_rate[i] = best_W0_rate[i]; + W0_grad[i] = 0; + } + for (i=0;i<W1_size;i++) + { + W1[i] = best_W1[i]; + best_W1_rate[i] *= .8; + if (best_W1_rate[i]<1e-15) best_W1_rate[i]=1e-15; + W1_rate[i] = best_W1_rate[i]; + W1_grad[i] = 0; + } + } + } + if (count_retries>10) + break; + for (i=0;i<W0_size;i++) + { + if (W0_oldgrad[i]*W0_grad[i] > 0) + W0_rate[i] *= 1.01; + else if (W0_oldgrad[i]*W0_grad[i] < 0) + W0_rate[i] *= .9; + mean_rate += W0_rate[i]; + if (W0_rate[i] < min_rate) + min_rate = W0_rate[i]; + if (W0_rate[i] < 1e-15) + W0_rate[i] = 1e-15; + /*if (W0_rate[i] > .01) + W0_rate[i] = .01;*/ + W0_oldgrad[i] = W0_grad[i]; + W0_old2[i] = W0_old[i]; + W0_old[i] = W0[i]; + W0[i] += W0_grad[i]*W0_rate[i]; + } + for (i=0;i<W1_size;i++) + { + if (W1_oldgrad[i]*W1_grad[i] > 0) + W1_rate[i] *= 1.01; + else if (W1_oldgrad[i]*W1_grad[i] < 0) + W1_rate[i] *= .9; + mean_rate += W1_rate[i]; + if (W1_rate[i] < min_rate) + min_rate = W1_rate[i]; + if (W1_rate[i] < 1e-15) + W1_rate[i] = 1e-15; + W1_oldgrad[i] = W1_grad[i]; + W1_old2[i] = W1_old[i]; + W1_old[i] = W1[i]; + W1[i] += W1_grad[i]*W1_rate[i]; + } + mean_rate /= (topo[0]+1)*topo[1] + (topo[1]+1)*topo[2]; + fprintf (stderr, "%g %d", mean_rate, e); + if (count_retries) + fprintf(stderr, " %d", count_retries); + fprintf(stderr, "\n"); + if (stopped) + break; + } + for (i=0;i<NB_THREADS;i++) + { + args[i].done = 1; + sem_post(&sem_begin[i]); + pthread_join(thread[i], NULL); + fprintf (stderr, "joined %d\n", i); + } + free(W0_old); + free(W1_old); + free(W0_grad); + free(W1_grad); + free(W0_rate); + free(W1_rate); + return best_rms; } int main(int argc, char **argv) { - int i, j; - int nbInputs; - int nbOutputs; - int nbHidden; - int nbSamples; - int nbEpoch; - int nbRealInputs; - unsigned int seed; - int ret; - float rms; - float *inputs; - float *outputs; - if (argc!=6) - { - fprintf (stderr, "usage: mlp_train <inputs> <hidden> <outputs> <nb samples> <nb epoch>\n"); - return 1; - } - nbInputs = atoi(argv[1]); - nbHidden = atoi(argv[2]); - nbOutputs = atoi(argv[3]); - nbSamples = atoi(argv[4]); - nbEpoch = atoi(argv[5]); - nbRealInputs = nbInputs; - inputs = malloc(nbInputs*nbSamples*sizeof(*inputs)); - outputs = malloc(nbOutputs*nbSamples*sizeof(*outputs)); - - seed = time(NULL); + int i, j; + int nbInputs; + int nbOutputs; + int nbHidden; + int nbSamples; + int nbEpoch; + int nbRealInputs; + unsigned int seed; + int ret; + float rms; + float *inputs; + float *outputs; + if (argc!=6) + { + fprintf (stderr, "usage: mlp_train <inputs> <hidden> <outputs> <nb samples> <nb epoch>\n"); + return 1; + } + nbInputs = atoi(argv[1]); + nbHidden = atoi(argv[2]); + nbOutputs = atoi(argv[3]); + nbSamples = atoi(argv[4]); + nbEpoch = atoi(argv[5]); + nbRealInputs = nbInputs; + inputs = malloc(nbInputs*nbSamples*sizeof(*inputs)); + outputs = malloc(nbOutputs*nbSamples*sizeof(*outputs)); + + seed = time(NULL); /*seed = 1361480659;*/ - fprintf (stderr, "Seed is %u\n", seed); - srand(seed); - build_tansig_table(); - signal(SIGTERM, handler); - signal(SIGINT, handler); - signal(SIGHUP, handler); - for (i=0;i<nbSamples;i++) - { - for (j=0;j<nbRealInputs;j++) - ret = scanf(" %f", &inputs[i*nbInputs+j]); - for (j=0;j<nbOutputs;j++) - ret = scanf(" %f", &outputs[i*nbOutputs+j]); - if (feof(stdin)) - { - nbSamples = i; - break; - } - } - int topo[3] = {nbInputs, nbHidden, nbOutputs}; - MLPTrain *net; + fprintf (stderr, "Seed is %u\n", seed); + srand(seed); + build_tansig_table(); + signal(SIGTERM, handler); + signal(SIGINT, handler); + signal(SIGHUP, handler); + for (i=0;i<nbSamples;i++) + { + for (j=0;j<nbRealInputs;j++) + ret = scanf(" %f", &inputs[i*nbInputs+j]); + for (j=0;j<nbOutputs;j++) + ret = scanf(" %f", &outputs[i*nbOutputs+j]); + if (feof(stdin)) + { + nbSamples = i; + break; + } + } + int topo[3] = {nbInputs, nbHidden, nbOutputs}; + MLPTrain *net; - fprintf (stderr, "Got %d samples\n", nbSamples); - net = mlp_init(topo, 3, inputs, outputs, nbSamples); - rms = mlp_train_backprop(net, inputs, outputs, nbSamples, nbEpoch, 1); - printf ("#include \"mlp.h\"\n\n"); - printf ("/* RMS error was %f, seed was %u */\n\n", rms, seed); - printf ("static const float weights[%d] = {\n", (topo[0]+1)*topo[1] + (topo[1]+1)*topo[2]); - printf ("\n/* hidden layer */\n"); - for (i=0;i<(topo[0]+1)*topo[1];i++) - { - printf ("%gf, ", net->weights[0][i]); - if (i%5==4) - printf("\n"); - } - printf ("\n/* output layer */\n"); - for (i=0;i<(topo[1]+1)*topo[2];i++) - { - printf ("%g, ", net->weights[1][i]); - if (i%5==4) - printf("\n"); - } - printf ("};\n\n"); - printf ("static const int topo[3] = {%d, %d, %d};\n\n", topo[0], topo[1], topo[2]); - printf ("const MLP net = {\n"); - printf ("\t3,\n"); - printf ("\ttopo,\n"); - printf ("\tweights\n};\n"); - return 0; + fprintf (stderr, "Got %d samples\n", nbSamples); + net = mlp_init(topo, 3, inputs, outputs, nbSamples); + rms = mlp_train_backprop(net, inputs, outputs, nbSamples, nbEpoch, 1); + printf ("#include \"mlp.h\"\n\n"); + printf ("/* RMS error was %f, seed was %u */\n\n", rms, seed); + printf ("static const float weights[%d] = {\n", (topo[0]+1)*topo[1] + (topo[1]+1)*topo[2]); + printf ("\n/* hidden layer */\n"); + for (i=0;i<(topo[0]+1)*topo[1];i++) + { + printf ("%gf, ", net->weights[0][i]); + if (i%5==4) + printf("\n"); + } + printf ("\n/* output layer */\n"); + for (i=0;i<(topo[1]+1)*topo[2];i++) + { + printf ("%g, ", net->weights[1][i]); + if (i%5==4) + printf("\n"); + } + printf ("};\n\n"); + printf ("static const int topo[3] = {%d, %d, %d};\n\n", topo[0], topo[1], topo[2]); + printf ("const MLP net = {\n"); + printf ("\t3,\n"); + printf ("\ttopo,\n"); + printf ("\tweights\n};\n"); + return 0; } diff --git a/src/mlp_train.h b/src/mlp_train.h index 1857f644..2786b40d 100644 --- a/src/mlp_train.h +++ b/src/mlp_train.h @@ -32,31 +32,31 @@ #include <stdlib.h> double tansig_table[501]; -static inline double tansig_double(double x) +static inline double tansig_double(double x) { - return 2./(1.+exp(-2.*x)) - 1.; + return 2./(1.+exp(-2.*x)) - 1.; } static inline void build_tansig_table() { - int i; - for (i=0;i<501;i++) - tansig_table[i] = tansig_double(.04*(i-250)); + int i; + for (i=0;i<501;i++) + tansig_table[i] = tansig_double(.04*(i-250)); } static inline double tansig_approx(double x) { - int i; - double y, dy; - if (x>=10) - return 1; - if (x<=-10) - return -1; - i = lrint(25*x); - x -= .04*i; - y = tansig_table[250+i]; - dy = 1-y*y; - y = y + x*dy*(1 - y*x); - return y; + int i; + double y, dy; + if (x>=10) + return 1; + if (x<=-10) + return -1; + i = lrint(25*x); + x -= .04*i; + y = tansig_table[250+i]; + dy = 1-y*y; + y = y + x*dy*(1 - y*x); + return y; } inline float randn(float sd) @@ -75,11 +75,11 @@ inline float randn(float sd) typedef struct { - int layers; - int *topo; - double **weights; - double **best_weights; - double *in_rate; + int layers; + int *topo; + double **weights; + double **best_weights; + double *in_rate; } MLPTrain; diff --git a/src/opus_encoder.c b/src/opus_encoder.c index d836843e..cdcee083 100644 --- a/src/opus_encoder.c +++ b/src/opus_encoder.c @@ -232,7 +232,7 @@ int opus_encoder_init(OpusEncoder* st, opus_int32 Fs, int channels, int applicat st->lsb_depth = 24; st->variable_duration = OPUS_FRAMESIZE_ARG; - /* Delay compensation of 4 ms (2.5 ms for SILK's extra look-ahead + /* Delay compensation of 4 ms (2.5 ms for SILK's extra look-ahead + 1.5 ms for SILK resamplers and stereo prediction) */ st->delay_compensation = st->Fs/250; @@ -2125,7 +2125,7 @@ int opus_encoder_ctl(OpusEncoder *st, int request, ...) case OPUS_SET_MAX_BANDWIDTH_REQUEST: { opus_int32 value = va_arg(ap, opus_int32); - if (value < OPUS_BANDWIDTH_NARROWBAND || value > OPUS_BANDWIDTH_FULLBAND) + if (value < OPUS_BANDWIDTH_NARROWBAND || value > OPUS_BANDWIDTH_FULLBAND) { goto bad_arg; } |
