virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2025-08-15 14:13:16 +02:00
Code Issues Releases Activity

vc2enc: replace quantization LUT with a smaller division LUT

Browse Source 
This commit replaces the huge and impractical LUT which converted coeffs
and a quantizer to bits to encode and instead uses a standard multiplication
and a shift to replace the division and then codes the values using the
regular golomb coding functions.

Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit is contained in:
Rostislav Pehlivanov
2018-02-27 23:12:33 +00:00
parent 950170bd3b
commit ea6973a573
1 changed files with 31 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files

118
libavcodec/vc2enc.c
View File

@@ -29,10 +29,6 @@
#include "vc2enc_dwt.h" #include "vc2enc_dwt.h"
#include "diractab.h" #include "diractab.h"
/* Total range is -COEF_LUT_TAB to +COEFF_LUT_TAB, but total tab size is half
* (COEF_LUT_TAB*DIRAC_MAX_QUANT_INDEX), as the sign is appended during encoding */
#define COEF_LUT_TAB 2048
/* The limited size resolution of each slice forces us to do this */ /* The limited size resolution of each slice forces us to do this */
#define SSIZE_ROUND(b) (FFALIGN((b), s->size_scaler) + 4 + s->prefix_bytes) #define SSIZE_ROUND(b) (FFALIGN((b), s->size_scaler) + 4 + s->prefix_bytes)
@@ -152,9 +148,8 @@ typedef struct VC2EncContext {
uint8_t quant[MAX_DWT_LEVELS][4]; uint8_t quant[MAX_DWT_LEVELS][4];
int custom_quant_matrix; int custom_quant_matrix;
/* Coefficient LUT */ /* Division LUT */
uint32_t *coef_lut_val; uint32_t qmagic_lut[116][2];
uint8_t *coef_lut_len;
int num_x; /* #slices horizontally */ int num_x; /* #slices horizontally */
int num_y; /* #slices vertically */ int num_y; /* #slices vertically */
@@ -229,37 +224,6 @@ static av_always_inline int count_vc2_ue_uint(uint32_t val)
return ff_log2(topbit)*2 + 1; return ff_log2(topbit)*2 + 1;
} }
static av_always_inline void get_vc2_ue_uint(int val, uint8_t *nbits,
uint32_t *eval)
{
int i;
int pbits = 0, bits = 0, topbit = 1, maxval = 1;
if (!val++) {
*nbits = 1;
*eval = 1;
return;
}
while (val > maxval) {
topbit <<= 1;
maxval <<= 1;
maxval |= 1;
}
bits = ff_log2(topbit);
for (i = 0; i < bits; i++) {
topbit >>= 1;
pbits <<= 2;
if (val & topbit)
pbits |= 0x1;
}
*nbits = bits*2 + 1;
*eval = (pbits << 1) | 1;
}
/* VC-2 10.4 - parse_info() */ /* VC-2 10.4 - parse_info() */
static void encode_parse_info(VC2EncContext *s, enum DiracParseCodes pcode) static void encode_parse_info(VC2EncContext *s, enum DiracParseCodes pcode)
{ {
@@ -557,7 +521,7 @@ static void encode_picture_start(VC2EncContext *s)
encode_wavelet_transform(s); encode_wavelet_transform(s);
} }
#define QUANT(c, qf) (((c) << 2)/(qf)) #define QUANT(c, mul, add, shift) (((mul) * (c) + (add)) >> (shift))
/* VC-2 13.5.5.2 - slice_band() */ /* VC-2 13.5.5.2 - slice_band() */
static void encode_subband(VC2EncContext *s, PutBitContext *pb, int sx, int sy, static void encode_subband(VC2EncContext *s, PutBitContext *pb, int sx, int sy,
@@ -570,24 +534,17 @@ static void encode_subband(VC2EncContext *s, PutBitContext *pb, int sx, int sy,
const int top = b->height * (sy+0) / s->num_y; const int top = b->height * (sy+0) / s->num_y;
const int bottom = b->height * (sy+1) / s->num_y; const int bottom = b->height * (sy+1) / s->num_y;
const int qfactor = ff_dirac_qscale_tab[quant];
const uint8_t *len_lut = &s->coef_lut_len[quant*COEF_LUT_TAB];
const uint32_t *val_lut = &s->coef_lut_val[quant*COEF_LUT_TAB];
dwtcoef *coeff = b->buf + top * b->stride; dwtcoef *coeff = b->buf + top * b->stride;
const uint64_t q_m = ((uint64_t)(s->qmagic_lut[quant][0])) << 2;
const uint64_t q_a = s->qmagic_lut[quant][1];
const int q_s = av_log2(ff_dirac_qscale_tab[quant]) + 32;
for (y = top; y < bottom; y++) { for (y = top; y < bottom; y++) {
for (x = left; x < right; x++) { for (x = left; x < right; x++) {
const int neg = coeff[x] < 0; uint32_t c_abs = QUANT(FFABS(coeff[x]), q_m, q_a, q_s);
uint32_t c_abs = FFABS(coeff[x]); put_vc2_ue_uint(pb, c_abs);
if (c_abs < COEF_LUT_TAB) { if (c_abs)
put_bits(pb, len_lut[c_abs], val_lut[c_abs] | neg); put_bits(pb, 1, coeff[x] < 0);
} else {
c_abs = QUANT(c_abs, qfactor);
put_vc2_ue_uint(pb, c_abs);
if (c_abs)
put_bits(pb, 1, neg);
}
} }
coeff += b->stride; coeff += b->stride;
} }
@@ -619,8 +576,9 @@ static int count_hq_slice(SliceArgs *slice, int quant_idx)
SubBand *b = &s->plane[p].band[level][orientation]; SubBand *b = &s->plane[p].band[level][orientation];
const int q_idx = quants[level][orientation]; const int q_idx = quants[level][orientation];
const uint8_t *len_lut = &s->coef_lut_len[q_idx*COEF_LUT_TAB]; const uint64_t q_m = ((uint64_t)s->qmagic_lut[q_idx][0]) << 2;
const int qfactor = ff_dirac_qscale_tab[q_idx]; const uint64_t q_a = s->qmagic_lut[q_idx][1];
const int q_s = av_log2(ff_dirac_qscale_tab[q_idx]) + 32;
const int left = b->width * slice->x / s->num_x; const int left = b->width * slice->x / s->num_x;
const int right = b->width *(slice->x+1) / s->num_x; const int right = b->width *(slice->x+1) / s->num_x;
@@ -631,14 +589,9 @@ static int count_hq_slice(SliceArgs *slice, int quant_idx)
for (y = top; y < bottom; y++) { for (y = top; y < bottom; y++) {
for (x = left; x < right; x++) { for (x = left; x < right; x++) {
uint32_t c_abs = FFABS(buf[x]); uint32_t c_abs = QUANT(FFABS(buf[x]), q_m, q_a, q_s);
if (c_abs < COEF_LUT_TAB) { bits += count_vc2_ue_uint(c_abs);
bits += len_lut[c_abs]; bits += !!c_abs;
} else {
c_abs = QUANT(c_abs, qfactor);
bits += count_vc2_ue_uint(c_abs);
bits += !!c_abs;
}
} }
buf += b->stride; buf += b->stride;
} }
@@ -1059,8 +1012,6 @@ static av_cold int vc2_encode_end(AVCodecContext *avctx)
} }
av_freep(&s->slice_args); av_freep(&s->slice_args);
av_freep(&s->coef_lut_len);
av_freep(&s->coef_lut_val);
return 0; return 0;
} }
@@ -1069,7 +1020,7 @@ static av_cold int vc2_encode_init(AVCodecContext *avctx)
{ {
Plane *p; Plane *p;
SubBand *b; SubBand *b;
int i, j, level, o, shift, ret; int i, level, o, shift, ret;
const AVPixFmtDescriptor *fmt = av_pix_fmt_desc_get(avctx->pix_fmt); const AVPixFmtDescriptor *fmt = av_pix_fmt_desc_get(avctx->pix_fmt);
const int depth = fmt->comp[0].depth; const int depth = fmt->comp[0].depth;
VC2EncContext *s = avctx->priv_data; VC2EncContext *s = avctx->priv_data;
@@ -1211,27 +1162,20 @@ static av_cold int vc2_encode_init(AVCodecContext *avctx)
if (!s->slice_args) if (!s->slice_args)
goto alloc_fail; goto alloc_fail;
/* Lookup tables */ for (i = 0; i < 116; i++) {
s->coef_lut_len = av_malloc(COEF_LUT_TAB*(s->q_ceil+1)*sizeof(*s->coef_lut_len)); const uint32_t qf = ff_dirac_qscale_tab[i];
if (!s->coef_lut_len) const int m = av_log2(qf);
goto alloc_fail; const uint32_t t = (1UL << (m + 32)) / qf;
const uint32_t r = (t*qf + qf) & ((1UL << 32) - 1);
s->coef_lut_val = av_malloc(COEF_LUT_TAB*(s->q_ceil+1)*sizeof(*s->coef_lut_val)); if (!(qf & (qf - 1))) {
if (!s->coef_lut_val) s->qmagic_lut[i][0] = 0xFFFFFFFF;
goto alloc_fail; s->qmagic_lut[i][1] = 0xFFFFFFFF;
} else if (r <= 1UL << m) {
for (i = 0; i < s->q_ceil; i++) { s->qmagic_lut[i][0] = t + 1;
uint8_t *len_lut = &s->coef_lut_len[i*COEF_LUT_TAB]; s->qmagic_lut[i][1] = 0;
uint32_t *val_lut = &s->coef_lut_val[i*COEF_LUT_TAB]; } else {
for (j = 0; j < COEF_LUT_TAB; j++) { s->qmagic_lut[i][0] = t;
get_vc2_ue_uint(QUANT(j, ff_dirac_qscale_tab[i]), s->qmagic_lut[i][1] = t;
&len_lut[j], &val_lut[j]);
if (len_lut[j] != 1) {
len_lut[j] += 1;
val_lut[j] <<= 1;
} else {
val_lut[j] = 1;
}
} }
} }