[FFmpeg-devel] [PATCH 1/4] libavcodec: Implementation of AC3 fixed point decoder.
Babic, Nedeljko
nbabic at mips.com
Wed Oct 3 15:17:03 CEST 2012
Hi Vitor,
>> AC3 fixed point decoder is based on AC3 floating point
>> decoder that is already part of FFmpeg.
>>
>> It does not use FFmpegs FFT. It uses FFT developed for
>> optimization of floating point AC3 decoder and because
>> of that currently some of the files that implement this
>> FFT are located in libavcodec/mips folder.
>
>> diff --git a/libavcodec/Makefile b/libavcodec/Makefile
>> index 9b86f7c..cf88e48 100644
>> --- a/libavcodec/Makefile
>> +++ b/libavcodec/Makefile
>> @@ -80,7 +80,8 @@ OBJS-$(CONFIG_AAC_ENCODER) += aacenc.o aaccoder.o \
>> mpeg4audio.o kbdwin.o \
>> audio_frame_queue.o
>> OBJS-$(CONFIG_AASC_DECODER) += aasc.o msrledec.o
>> -OBJS-$(CONFIG_AC3_DECODER) += ac3dec.o ac3dec_data.o ac3.o kbdwin.o
>> +OBJS-$(CONFIG_AC3_DECODER) += ac3dec_float.o ac3dec_data.o ac3.o kbdwin.o
>> +OBJS-$(CONFIG_AC3_FIXED_DECODER) += fft_ac3_init_tab.o fft_ac3_fixed.o ac3dec_fixed.o ac3dec_data.o ac3.o kbdwin.o
>> OBJS-$(CONFIG_AC3_ENCODER) += ac3enc_float.o ac3enc.o ac3tab.o \
>> ac3.o kbdwin.o
>> OBJS-$(CONFIG_AC3_FIXED_ENCODER) += ac3enc_fixed.o ac3enc.o ac3tab.o ac3.o
>> diff --git a/libavcodec/ac3.h b/libavcodec/ac3.h
>> index b9f34b9..0a1ff23 100644
>> --- a/libavcodec/ac3.h
>> +++ b/libavcodec/ac3.h
>> @@ -27,6 +27,10 @@
>> #ifndef AVCODEC_AC3_H
>> #define AVCODEC_AC3_H
>>
>> +#ifndef CONFIG_AC3_FIXED
>> +# define CONFIG_AC3_FIXED 0
>> +#endif
>
>I don't think this is needed.
But how can I distinguish between fixed and floating point macros without this?
Should I use just #ifdef CONFIG_AC3_FIXED?
>
>> #define AC3_MAX_CODED_FRAME_SIZE 3840 /* in bytes */
>> #define AC3_MAX_CHANNELS 7 /**< maximum number of channels, including coupling channel */
>> #define CPL_CH 0 /**< coupling channel index */
>> @@ -51,6 +55,41 @@
>> #define EXP_D25 2
>> #define EXP_D45 3
>>
>> +#if CONFIG_AC3_FIXED
>> +
>> +#define CONFIG_FFT_FLOAT 0
>> +
>> +/* pre-defined gain values */
>> +#define LEVEL_PLUS_3DB 5793
>> +#define LEVEL_PLUS_1POINT5DB 4871
>> +#define LEVEL_MINUS_1POINT5DB 3444
>> +#define LEVEL_MINUS_3DB 2896
>> +#define LEVEL_MINUS_4POINT5DB 2435
>> +#define LEVEL_MINUS_6DB 2048
>> +#define LEVEL_MINUS_9DB 1448
>> +#define LEVEL_ZERO 0
>> +#define LEVEL_ONE 4096
>> +
>> +#define MUL_BIAS1 65536
>> +#define MUL_BIAS2 2147418112
>
>If this is just the floating-point values converted to fixed, it is
>better to use a macro to define it only once. See FIXR() macro in
>mpegaudiodec.c.
>
You are correct and I will create macro for this.
>> +#define AC3_RENAME(x) x ## _fixed
>
>> +#define AC3_CENTER(x) center_levels[x]
>> +#define AC3_SURROUND(x) surround_levels[x]
>
>Those two looks unused.
>
They are leftovers from previous version of code. Sorry.
I will remove them.
>> +#define AC3_LEVEL(x) ((x)*23170 + 0x4000) >> 15
>> +#define AC3_NORM(x,norm) ((x)<<12)/norm
>
>> +#define AC3_DYNAMIC_RANGE(x) (x)
>
>This looks very different from the float version, how can it work?
>
The way in which this is used in fixed point version of code is different
from the way it is used in floating point version. You can see in
function scale_coefs how dynamic range is used.
>> +#define AC3_SPX_BLEND(x) (x)
>
>> +#define TYPE_PREFIX(x) fixed_ ## x
>
>Why not use sulfix everywhere, for simplicity and consistency?
I didn't want to change the name of the function float_to_int16_interleave
that was created before this decoder (and its name is more descriptive this way).
>
>> +#define AC3_DYNAMIC_RANGE1 0
>
>Hmm, that is strange.
Similar as above. The way that dynamic range is used in this decoder is
different from the way it is used in floating point decoder.
>
>> +#define INTFLOAT int
>> +#define SHORTFLOAT int16_t
>> +
>> +#define ROUND12(x) ((x)+2048)>>12
>> +
>> +#else
>> +
>> /* pre-defined gain values */
>> #define LEVEL_PLUS_3DB 1.4142135623730950
>> #define LEVEL_PLUS_1POINT5DB 1.1892071150027209
>> @@ -62,6 +101,26 @@
>> #define LEVEL_ZERO 0.0000000000000000
>> #define LEVEL_ONE 1.0000000000000000
>>
>> +#define MUL_BIAS1 1.0f
>> +#define MUL_BIAS2 32767.0f
>> +
>> +#define AC3_RENAME(x) x
>> +#define AC3_CENTER(x) (x)
>> +#define AC3_SURROUND(x) (x)
>> +#define AC3_LEVEL(x) (x)*LEVEL_MINUS_3DB
>> +#define AC3_NORM(x,norm) (x)*(1.0f/norm)
>> +#define AC3_DYNAMIC_RANGE(x) ((dynamic_range_tab[x] - 1.0) * s->drc_scale) + 1.0
>> +#define AC3_SPX_BLEND(x) (x)* (1.0f/32)
>> +#define TYPE_PREFIX(x) float_ ## x
>> +
>> +#define AC3_DYNAMIC_RANGE1 1.0f
>> +#define INTFLOAT float
>> +#define SHORTFLOAT float
>> +
>> +#define ROUND12(x) (x)
>> +
>> +#endif /* CONFIG_AC3_FIXED */
>> +
>> /** Delta bit allocation strategy */
>> typedef enum {
>> DBA_REUSE = 0,
>> diff --git a/libavcodec/ac3dec.c b/libavcodec/ac3dec.c
>> index c608de8..5d82e23 100644
>> --- a/libavcodec/ac3dec.c
>> +++ b/libavcodec/ac3dec.c
>> @@ -64,7 +64,7 @@ static const uint8_t quantization_tab[16] = {
>> static float dynamic_range_tab[256];
>>
>> /** Adjustments in dB gain */
>> -static const float gain_levels[9] = {
>> +static const INTFLOAT AC3_RENAME(gain_levels)[9] = {
>> LEVEL_PLUS_3DB,
>> LEVEL_PLUS_1POINT5DB,
>> LEVEL_ONE,
>> @@ -157,27 +157,32 @@ static av_cold void ac3_tables_init(void)
>> /**
>> * AVCodec initialization
>> */
>> -static av_cold int ac3_decode_init(AVCodecContext *avctx)
>> +static av_cold int AC3_RENAME(ac3_decode_init)(AVCodecContext *avctx)
>> {
>> AC3DecodeContext *s = avctx->priv_data;
>> s->avctx = avctx;
>>
>> ff_ac3_common_init();
>> ac3_tables_init();
>> - ff_mdct_init(&s->imdct_256, 8, 1, 1.0);
>> - ff_mdct_init(&s->imdct_512, 9, 1, 1.0);
>> - ff_kbd_window_init(s->window, 5.0, 256);
>> + AC3_RENAME(ff_mdct_init)(&s->imdct_256, 8, 1, 1.0);
>
>> + AC3_RENAME(ff_mdct_init)(&s->imdct_512, 9, 1, 1.0);\
>
>Why the backslash?
>
My mistake. I will remove it.
>> + AC3_RENAME(ff_kbd_window_init)(s->window, 5.0, 256);
>> ff_dsputil_init(&s->dsp, avctx);
>> ff_ac3dsp_init(&s->ac3dsp, avctx->flags & CODEC_FLAG_BITEXACT);
>> ff_fmt_convert_init(&s->fmt_conv, avctx);
>> av_lfg_init(&s->dith_state, 0);
>>
>> +#if CONFIG_AC3_FIXED
>> + ff_ac3_fft_init_fixed(&s->imdct_256);
>> + ff_ac3_fft_init_fixed(&s->imdct_512);
>> +#endif
>> +
>> /* set scale value for float to int16 conversion */
>> if (avctx->request_sample_fmt == AV_SAMPLE_FMT_FLT) {
>> - s->mul_bias = 1.0f;
>> + s->mul_bias = MUL_BIAS1;
>> avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
>> } else {
>> - s->mul_bias = 32767.0f;
>> + s->mul_bias = MUL_BIAS2;
>> avctx->sample_fmt = AV_SAMPLE_FMT_S16;
>> }
>>
>> @@ -300,23 +305,23 @@ static int parse_frame_header(AC3DecodeContext *s)
>> * Set stereo downmixing coefficients based on frame header info.
>> * reference: Section 7.8.2 Downmixing Into Two Channels
>> */
>> -static void set_downmix_coeffs(AC3DecodeContext *s)
>> +static void AC3_RENAME(set_downmix_coeffs)(AC3DecodeContext *s)
>> {
>> int i;
>> - float cmix = gain_levels[s-> center_mix_level];
>> - float smix = gain_levels[s->surround_mix_level];
>> - float norm0, norm1;
>> + INTFLOAT cmix = AC3_RENAME(gain_levels)[s-> center_mix_level];
>> + INTFLOAT smix = AC3_RENAME(gain_levels)[s->surround_mix_level];
>> + INTFLOAT norm0, norm1;
>>
>> for (i = 0; i < s->fbw_channels; i++) {
>> - s->downmix_coeffs[i][0] = gain_levels[ac3_default_coeffs[s->channel_mode][i][0]];
>> - s->downmix_coeffs[i][1] = gain_levels[ac3_default_coeffs[s->channel_mode][i][1]];
>> + s->downmix_coeffs[i][0] = AC3_RENAME(gain_levels)[ac3_default_coeffs[s->channel_mode][i][0]];
>> + s->downmix_coeffs[i][1] = AC3_RENAME(gain_levels)[ac3_default_coeffs[s->channel_mode][i][1]];
>> }
>> if (s->channel_mode > 1 && s->channel_mode & 1) {
>> s->downmix_coeffs[1][0] = s->downmix_coeffs[1][1] = cmix;
>> }
>> if (s->channel_mode == AC3_CHMODE_2F1R || s->channel_mode == AC3_CHMODE_3F1R) {
>> int nf = s->channel_mode - 2;
>> - s->downmix_coeffs[nf][0] = s->downmix_coeffs[nf][1] = smix * LEVEL_MINUS_3DB;
>> + s->downmix_coeffs[nf][0] = s->downmix_coeffs[nf][1] = AC3_LEVEL(smix);
>> }
>> if (s->channel_mode == AC3_CHMODE_2F2R || s->channel_mode == AC3_CHMODE_3F2R) {
>> int nf = s->channel_mode - 4;
>> @@ -324,22 +329,20 @@ static void set_downmix_coeffs(AC3DecodeContext *s)
>> }
>>
>> /* renormalize */
>> - norm0 = norm1 = 0.0;
>> + norm0 = norm1 = (INTFLOAT)0.0;
>
>norm0 = norm1 = FIXR(0.0);
>
>so this construct can also work for other constants.
>
Will be used.
>> for (i = 0; i < s->fbw_channels; i++) {
>> norm0 += s->downmix_coeffs[i][0];
>> norm1 += s->downmix_coeffs[i][1];
>> }
>
>> - norm0 = 1.0f / norm0;
>> - norm1 = 1.0f / norm1;
>> +
>> for (i = 0; i < s->fbw_channels; i++) {
>> - s->downmix_coeffs[i][0] *= norm0;
>> - s->downmix_coeffs[i][1] *= norm1;
>> + s->downmix_coeffs[i][0] = AC3_NORM(s->downmix_coeffs[i][0],norm0);
>> + s->downmix_coeffs[i][1] = AC3_NORM(s->downmix_coeffs[i][1],norm1);
>> }
>
>Division is much slower than multiplication. You can do the same trick
>in fixed point:
>
>norm0 = (1 << bits) / norm0;
>s->downmix_coeffs[i][0] = MUL(s->downmix_coeffs[i][0],norm0);
>
>where
>
>#define MUL(a,b) (((int64_t) (a)) * (b)) >> (s))
>
I am not sure that accuracy of fixed point code would be ok if division is moved out
from the loop
>> if (s->output_mode == AC3_CHMODE_MONO) {
>> for (i = 0; i < s->fbw_channels; i++)
>> - s->downmix_coeffs[i][0] = (s->downmix_coeffs[i][0] +
>> - s->downmix_coeffs[i][1]) * LEVEL_MINUS_3DB;
>> + s->downmix_coeffs[i][0] = AC3_LEVEL(s->downmix_coeffs[i][0] + s->downmix_coeffs[i][1]);
>> }
>> }
>
>Hmm, is this correct?
Not exactly, although it passes fate tests. I will have to add back multiplication with LEVEL_MINUS_3DB.
>
>> @@ -602,51 +605,25 @@ static inline void do_imdct(AC3DecodeContext *s, int channels)
>> for (ch = 1; ch <= channels; ch++) {
>> if (s->block_switch[ch]) {
>> int i;
>> - float *x = s->tmp_output + 128;
>> + FFTSample *x = s->tmp_output+128;
>> for (i = 0; i < 128; i++)
>> x[i] = s->transform_coeffs[ch][2 * i];
>> - s->imdct_256.imdct_half(&s->imdct_256, s->tmp_output, x);
>> - s->dsp.vector_fmul_window(s->output[ch - 1], s->delay[ch - 1],
>> + s->imdct_256.AC3_RENAME(imdct_half)(&s->imdct_256, s->tmp_output, x);
>> + s->dsp.AC3_RENAME(vector_fmul_window)(s->output[ch - 1], s->delay[ch - 1],
>> s->tmp_output, s->window, 128);
>> for (i = 0; i < 128; i++)
>> x[i] = s->transform_coeffs[ch][2 * i + 1];
>> - s->imdct_256.imdct_half(&s->imdct_256, s->delay[ch - 1], x);
>> + s->imdct_256.AC3_RENAME(imdct_half)(&s->imdct_256, s->delay[ch - 1], x);
>> } else {
>> - s->imdct_512.imdct_half(&s->imdct_512, s->tmp_output, s->transform_coeffs[ch]);
>> - s->dsp.vector_fmul_window(s->output[ch - 1], s->delay[ch - 1],
>> + s->imdct_512.AC3_RENAME(imdct_half)(&s->imdct_512, s->tmp_output, s->transform_coeffs[ch]);
>> + s->dsp.AC3_RENAME(vector_fmul_window)(s->output[ch - 1], s->delay[ch - 1],
>> s->tmp_output, s->window, 128);
>> - memcpy(s->delay[ch - 1], s->tmp_output + 128, 128 * sizeof(float));
>> + memcpy(s->delay[ch - 1], s->tmp_output + 128, 128 * sizeof(FFTSample));
>> }
>> }
>> }
>>
>> -/**
>> - * Upmix delay samples from stereo to original channel layout.
>> - */
>> -static void ac3_upmix_delay(AC3DecodeContext *s)
>> -{
>> - int channel_data_size = sizeof(s->delay[0]);
>> - switch (s->channel_mode) {
>> - case AC3_CHMODE_DUALMONO:
>> - case AC3_CHMODE_STEREO:
>> - /* upmix mono to stereo */
>> - memcpy(s->delay[1], s->delay[0], channel_data_size);
>> - break;
>> - case AC3_CHMODE_2F2R:
>> - memset(s->delay[3], 0, channel_data_size);
>> - case AC3_CHMODE_2F1R:
>> - memset(s->delay[2], 0, channel_data_size);
>> - break;
>> - case AC3_CHMODE_3F2R:
>> - memset(s->delay[4], 0, channel_data_size);
>> - case AC3_CHMODE_3F1R:
>> - memset(s->delay[3], 0, channel_data_size);
>> - case AC3_CHMODE_3F:
>> - memcpy(s->delay[2], s->delay[1], channel_data_size);
>> - memset(s->delay[1], 0, channel_data_size);
>> - break;
>> - }
>> -}
>> +
>
>??
The sequence of imdct and downmix is changed in fixed point code for better accuracy,
so this function is not needed.
>
>> /**
>> * Decode band structure for coupling, spectral extension, or enhanced coupling.
>> @@ -748,10 +725,9 @@ static int decode_audio_block(AC3DecodeContext *s, int blk)
>> i = !s->channel_mode;
>> do {
>> if (get_bits1(gbc)) {
>> - s->dynamic_range[i] = ((dynamic_range_tab[get_bits(gbc, 8)] - 1.0) *
>> - s->drc_scale) + 1.0;
>> + s->dynamic_range[i] = AC3_DYNAMIC_RANGE(get_bits(gbc, 8));
>> } else if (blk == 0) {
>> - s->dynamic_range[i] = 1.0f;
>> + s->dynamic_range[i] = AC3_DYNAMIC_RANGE1;
>> }
>> } while (i--);
>>
>> @@ -777,6 +753,10 @@ static int decode_audio_block(AC3DecodeContext *s, int blk)
>> if (start_subband > 7)
>> start_subband += start_subband - 7;
>> end_subband = get_bits(gbc, 3) + 5;
>> +#if CONFIG_AC3_FIXED
>> + s->spx_dst_end_freq = end_freq_inv_tab[end_subband];
>> + end_subband += 5;
>> +#endif
>> if (end_subband > 7)
>> end_subband += end_subband - 7;
>> dst_start_freq = dst_start_freq * 12 + 25;
>> @@ -797,7 +777,9 @@ static int decode_audio_block(AC3DecodeContext *s, int blk)
>>
>> s->spx_dst_start_freq = dst_start_freq;
>> s->spx_src_start_freq = src_start_freq;
>> +#if !CONFIG_AC3_FIXED
>> s->spx_dst_end_freq = dst_end_freq;
>> +#endif
>>
>> decode_band_structure(gbc, blk, s->eac3, 0,
>> start_subband, end_subband,
>> @@ -817,18 +799,45 @@ static int decode_audio_block(AC3DecodeContext *s, int blk)
>> for (ch = 1; ch <= fbw_channels; ch++) {
>> if (s->channel_uses_spx[ch]) {
>> if (s->first_spx_coords[ch] || get_bits1(gbc)) {
>> - float spx_blend;
>> + INTFLOAT spx_blend;
>> int bin, master_spx_coord;
>>
>> s->first_spx_coords[ch] = 0;
>> - spx_blend = get_bits(gbc, 5) * (1.0f/32);
>> + spx_blend = AC3_SPX_BLEND(get_bits(gbc, 5));
>> master_spx_coord = get_bits(gbc, 2) * 3;
>>
>> bin = s->spx_src_start_freq;
>> for (bnd = 0; bnd < s->num_spx_bands; bnd++) {
>> int bandsize;
>> int spx_coord_exp, spx_coord_mant;
>> - float nratio, sblend, nblend, spx_coord;
>> + INTFLOAT nratio, sblend, nblend;
>> +#if CONFIG_AC3_FIXED
>> + int64_t accu;
>> + /* calculate blending factors */
>> + bandsize = s->spx_band_sizes[bnd];
>> + accu = (long long)((bin << 23) + (bandsize << 22)) * s->spx_dst_end_freq;
>> + nratio = (int)(accu >> 32);
>> + nratio -= spx_blend << 18;
>> +
>> + if (nratio < 0)
>> + {
>> + nblend = 0;
>> + sblend = 0x800000;
>> + }
>> + else if (nratio > 0x7fffff)
>> + {
>> + nblend = 0x800000;
>> + sblend = 0;
>> + }
>> + else
>> + {
>> + nblend = ac3_fixed_sqrt(nratio);
>> + accu = (long long)nblend * 1859775393;
>> + nblend = (int)((accu + (1<<29)) >> 30);
>> + sblend = ac3_fixed_sqrt(0x800000 - nratio);
>> + }
>> +#else
>
>Indentation:
>
>if (...) {
>} else if (...) {
>}
I will fix indentation.
-Nedeljko
More information about the ffmpeg-devel
mailing list