wrapper/FFmpeg
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

lavc: VP9 decoder

Browse source 
Originally written by Ronald S. Bultje <rsbultje@gmail.com> and
Clément Bœsch <u@pkh.me>

Further contributions by:
Anton Khirnov <anton@khirnov.net>
Diego Biurrun <diego@biurrun.de>
Luca Barbato <lu_zero@gentoo.org>
Martin Storsjö <martin@martin.st>

Signed-off-by: Luca Barbato <lu_zero@gentoo.org>
Signed-off-by: Anton Khirnov <anton@khirnov.net>
This commit is contained in:
Ronald S. Bultje 2013-09-30 23:03:30 -04:00 • committed by Luca Barbato
parent 458446acfa
commit 72ca830f51
221 changed files with 10716 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

2
libavcodec/Makefile
View file

@ -392,6 +392,8 @@ OBJS-$(CONFIG_VP5_DECODER) += vp5.o vp56.o vp56data.o vp56dsp.o \
OBJS-$(CONFIG_VP6_DECODER) += vp6.o vp56.o vp56data.o vp56dsp.o \
vp6dsp.o vp56rac.o
OBJS-$(CONFIG_VP8_DECODER) += vp8.o vp8dsp.o vp56rac.o
OBJS-$(CONFIG_VP9_DECODER) += vp9.o vp9data.o vp9dsp.o \
vp9block.o vp9prob.o vp9mvs.o vp56rac.o
OBJS-$(CONFIG_VQA_DECODER) += vqavideo.o
OBJS-$(CONFIG_WAVPACK_DECODER) += wavpack.o
OBJS-$(CONFIG_WEBP_DECODER) += webp.o

1
libavcodec/allcodecs.c
View file

@ -257,6 +257,7 @@ void avcodec_register_all(void)
REGISTER_DECODER(VP6A, vp6a);
REGISTER_DECODER(VP6F, vp6f);
REGISTER_DECODER(VP8, vp8);
REGISTER_DECODER(VP9, vp9);
REGISTER_DECODER(VQA, vqa);
REGISTER_DECODER(WEBP, webp);
REGISTER_ENCDEC (WMV1, wmv1);

2
libavcodec/version.h
View file

@ -27,7 +27,7 @@
*/
#define LIBAVCODEC_VERSION_MAJOR 55
#define LIBAVCODEC_VERSION_MINOR 27
#define LIBAVCODEC_VERSION_MINOR 28
#define LIBAVCODEC_VERSION_MICRO 0
#define LIBAVCODEC_VERSION_INT AV_VERSION_INT(LIBAVCODEC_VERSION_MAJOR, \

1270
libavcodec/vp9.c Normal file
View file

File diff suppressed because it is too large Load diff

419
libavcodec/vp9.h Normal file
View file

@ -0,0 +1,419 @@
/*
* VP9 compatible video decoder
*
* Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>
* Copyright (C) 2013 Clément Bœsch <u pkh me>
*
* This file is part of Libav.
*
* Libav is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* Libav is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with Libav; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#ifndef AVCODEC_VP9_H
#define AVCODEC_VP9_H
#include <stddef.h>
#include <stdint.h>
#include "libavutil/internal.h"
#include "avcodec.h"
#include "vp56.h"
enum TxfmMode {
TX_4X4,
TX_8X8,
TX_16X16,
TX_32X32,
N_TXFM_SIZES,
TX_SWITCHABLE = N_TXFM_SIZES,
N_TXFM_MODES
};
enum TxfmType {
DCT_DCT,
DCT_ADST,
ADST_DCT,
ADST_ADST,
N_TXFM_TYPES
};
enum IntraPredMode {
VERT_PRED,
HOR_PRED,
DC_PRED,
DIAG_DOWN_LEFT_PRED,
DIAG_DOWN_RIGHT_PRED,
VERT_RIGHT_PRED,
HOR_DOWN_PRED,
VERT_LEFT_PRED,
HOR_UP_PRED,
TM_VP8_PRED,
LEFT_DC_PRED,
TOP_DC_PRED,
DC_128_PRED,
DC_127_PRED,
DC_129_PRED,
N_INTRA_PRED_MODES
};
enum FilterMode {
FILTER_8TAP_SMOOTH,
FILTER_8TAP_REGULAR,
FILTER_8TAP_SHARP,
FILTER_BILINEAR,
FILTER_SWITCHABLE,
};
enum BlockPartition {
PARTITION_NONE, // [ ] <-.
PARTITION_H, // [-] |
PARTITION_V, // [|] |
PARTITION_SPLIT, // [+] --'
};
enum InterPredMode {
NEARESTMV = 10,
NEARMV = 11,
ZEROMV = 12,
NEWMV = 13,
};
enum MVJoint {
MV_JOINT_ZERO,
MV_JOINT_H,
MV_JOINT_V,
MV_JOINT_HV,
};
typedef struct ProbContext {
uint8_t y_mode[4][9];
uint8_t uv_mode[10][9];
uint8_t filter[4][2];
uint8_t mv_mode[7][3];
uint8_t intra[4];
uint8_t comp[5];
uint8_t single_ref[5][2];
uint8_t comp_ref[5];
uint8_t tx32p[2][3];
uint8_t tx16p[2][2];
uint8_t tx8p[2];
uint8_t skip[3];
uint8_t mv_joint[3];
struct {
uint8_t sign;
uint8_t classes[10];
uint8_t class0;
uint8_t bits[10];
uint8_t class0_fp[2][3];
uint8_t fp[3];
uint8_t class0_hp;
uint8_t hp;
} mv_comp[2];
uint8_t partition[4][4][3];
} ProbContext;
typedef void (*vp9_mc_func)(uint8_t *dst, const uint8_t *ref,
ptrdiff_t dst_stride,
ptrdiff_t ref_stride,
int h, int mx, int my);
typedef struct VP9DSPContext {
/*
* dimension 1: 0=4x4, 1=8x8, 2=16x16, 3=32x32
* dimension 2: intra prediction modes
*
* dst/left/top is aligned by transform-size (i.e. 4, 8, 16 or 32 pixels)
* stride is aligned by 16 pixels
* top[-1] is top/left; top[4,7] is top-right for 4x4
*/
// FIXME(rbultje) maybe replace left/top pointers with HAVE_TOP/
// HAVE_LEFT/HAVE_TOPRIGHT flags instead, and then handle it in-place?
// also needs to fit in with what h264/vp8/etc do
void (*intra_pred[N_TXFM_SIZES][N_INTRA_PRED_MODES])(uint8_t *dst,
ptrdiff_t stride,
const uint8_t *left,
const uint8_t *top);
/*
* dimension 1: 0=4x4, 1=8x8, 2=16x16, 3=32x32, 4=lossless (3-4=dct only)
* dimension 2: 0=dct/dct, 1=dct/adst, 2=adst/dct, 3=adst/adst
*
* dst is aligned by transform-size (i.e. 4, 8, 16 or 32 pixels)
* stride is aligned by 16 pixels
* block is 16-byte aligned
* eob indicates the position (+1) of the last non-zero coefficient,
* in scan-order. This can be used to write faster versions, e.g. a
* dc-only 4x4/8x8/16x16/32x32, or a 4x4-only (eob<10) 8x8/16x16/32x32,
* etc.
*/
// FIXME also write idct_add_block() versions for whole (inter) pred
// blocks, so we can do 2 4x4s at once
void (*itxfm_add[N_TXFM_SIZES + 1][N_TXFM_TYPES])(uint8_t *dst,
ptrdiff_t stride,
int16_t *block, int eob);
/*
* dimension 1: width of filter (0=4, 1=8, 2=16)
* dimension 2: 0=col-edge filter (h), 1=row-edge filter (v)
*
* dst/stride are aligned by 8
*/
void (*loop_filter_8[3][2])(uint8_t *dst, ptrdiff_t stride,
int mb_lim, int lim, int hev_thr);
/*
* dimension 1: 0=col-edge filter (h), 1=row-edge filter (v)
*
* The width of filter is assumed to be 16; dst/stride are aligned by 16
*/
void (*loop_filter_16[2])(uint8_t *dst, ptrdiff_t stride,
int mb_lim, int lim, int hev_thr);
/*
* dimension 1/2: width of filter (0=4, 1=8) for each filter half
* dimension 3: 0=col-edge filter (h), 1=row-edge filter (v)
*
* dst/stride are aligned by operation size
* this basically calls loop_filter[d1][d3][0](), followed by
* loop_filter[d2][d3][0]() on the next 8 pixels
* mb_lim/lim/hev_thr contain two values in the lowest two bytes of the
* integer.
*/
// FIXME perhaps a mix4 that operates on 32px (for AVX2)
void (*loop_filter_mix2[2][2][2])(uint8_t *dst, ptrdiff_t stride,
int mb_lim, int lim, int hev_thr);
/*
* dimension 1: hsize (0: 64, 1: 32, 2: 16, 3: 8, 4: 4)
* dimension 2: filter type (0: smooth, 1: regular, 2: sharp, 3: bilin)
* dimension 3: averaging type (0: put, 1: avg)
* dimension 4: x subpel interpolation (0: none, 1: 8tap/bilin)
* dimension 5: y subpel interpolation (1: none, 1: 8tap/bilin)
*
* dst/stride are aligned by hsize
*/
vp9_mc_func mc[5][4][2][2][2];
} VP9DSPContext;
enum CompPredMode {
PRED_SINGLEREF,
PRED_COMPREF,
PRED_SWITCHABLE,
};
typedef struct VP9MVRefPair {
VP56mv mv[2];
int8_t ref[2];
} VP9MVRefPair;
typedef struct VP9Filter {
uint8_t level[8 * 8];
uint8_t /* bit=col */ mask[2 /* 0=y, 1=uv */][2 /* 0=col, 1=row */]
[8 /* rows */][4 /* 0=16, 1=8, 2=4, 3=inner4 */];
} VP9Filter;
enum BlockLevel {
BL_64X64,
BL_32X32,
BL_16X16,
BL_8X8,
};
enum BlockSize {
BS_64x64,
BS_64x32,
BS_32x64,
BS_32x32,
BS_32x16,
BS_16x32,
BS_16x16,
BS_16x8,
BS_8x16,
BS_8x8,
BS_8x4,
BS_4x8,
BS_4x4,
N_BS_SIZES,
};
typedef struct VP9Block {
uint8_t seg_id, intra, comp, ref[2], mode[4], uvmode, skip;
enum FilterMode filter;
VP56mv mv[4 /* b_idx */][2 /* ref */];
enum BlockSize bs;
enum TxfmMode tx, uvtx;
int row, row7, col, col7;
uint8_t *dst[3];
ptrdiff_t y_stride, uv_stride;
} VP9Block;
typedef struct VP9Context {
VP9DSPContext dsp;
VideoDSPContext vdsp;
GetBitContext gb;
VP56RangeCoder c;
VP56RangeCoder *c_b;
unsigned c_b_size;
VP9Block b;
// bitstream header
uint8_t profile;
uint8_t keyframe, last_keyframe;
uint8_t invisible;
uint8_t use_last_frame_mvs;
uint8_t errorres;
uint8_t colorspace;
uint8_t fullrange;
uint8_t intraonly;
uint8_t resetctx;
uint8_t refreshrefmask;
uint8_t highprecisionmvs;
enum FilterMode filtermode;
uint8_t allowcompinter;
uint8_t fixcompref;
uint8_t refreshctx;
uint8_t parallelmode;
uint8_t framectxid;
uint8_t refidx[3];
uint8_t signbias[3];
uint8_t varcompref[2];
AVFrame *refs[8];
AVFrame *cur_frame;
struct {
uint8_t level;
int8_t sharpness;
uint8_t lim_lut[64];
uint8_t mblim_lut[64];
} filter;
struct {
uint8_t enabled;
int8_t mode[2];
int8_t ref[4];
} lf_delta;
uint8_t yac_qi;
int8_t ydc_qdelta, uvdc_qdelta, uvac_qdelta;
uint8_t lossless;
struct {
uint8_t enabled;
uint8_t temporal;
uint8_t absolute_vals;
uint8_t update_map;
struct {
uint8_t q_enabled;
uint8_t lf_enabled;
uint8_t ref_enabled;
uint8_t skip_enabled;
uint8_t ref_val;
int16_t q_val;
int8_t lf_val;
int16_t qmul[2][2];
uint8_t lflvl[4][2];
} feat[8];
} segmentation;
struct {
unsigned log2_tile_cols, log2_tile_rows;
unsigned tile_cols, tile_rows;
unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;
} tiling;
unsigned sb_cols, sb_rows, rows, cols;
struct {
ProbContext p;
uint8_t coef[4][2][2][6][6][3];
} prob_ctx[4];
struct {
ProbContext p;
uint8_t coef[4][2][2][6][6][11];
uint8_t seg[7];
uint8_t segpred[3];
} prob;
struct {
unsigned y_mode[4][10];
unsigned uv_mode[10][10];
unsigned filter[4][3];
unsigned mv_mode[7][4];
unsigned intra[4][2];
unsigned comp[5][2];
unsigned single_ref[5][2][2];
unsigned comp_ref[5][2];
unsigned tx32p[2][4];
unsigned tx16p[2][3];
unsigned tx8p[2][2];
unsigned skip[3][2];
unsigned mv_joint[4];
struct {
unsigned sign[2];
unsigned classes[11];
unsigned class0[2];
unsigned bits[10][2];
unsigned class0_fp[2][4];
unsigned fp[4];
unsigned class0_hp[2];
unsigned hp[2];
} mv_comp[2];
unsigned partition[4][4][4];
unsigned coef[4][2][2][6][6][3];
unsigned eob[4][2][2][6][6][2];
} counts;
enum TxfmMode txfmmode;
enum CompPredMode comppredmode;
// contextual (left/above) cache
uint8_t left_partition_ctx[8], *above_partition_ctx;
uint8_t left_mode_ctx[16], *above_mode_ctx;
// FIXME maybe merge some of the below in a flags field?
uint8_t left_y_nnz_ctx[16], *above_y_nnz_ctx;
uint8_t left_uv_nnz_ctx[2][8], *above_uv_nnz_ctx[2];
uint8_t left_skip_ctx[8], *above_skip_ctx; // 1bit
uint8_t left_txfm_ctx[8], *above_txfm_ctx; // 2bit
uint8_t left_segpred_ctx[8], *above_segpred_ctx; // 1bit
uint8_t left_intra_ctx[8], *above_intra_ctx; // 1bit
uint8_t left_comp_ctx[8], *above_comp_ctx; // 1bit
uint8_t left_ref_ctx[8], *above_ref_ctx; // 2bit
uint8_t left_filter_ctx[8], *above_filter_ctx;
VP56mv left_mv_ctx[16][2], (*above_mv_ctx)[2];
// whole-frame cache
uint8_t *intra_pred_data[3];
uint8_t *segmentation_map;
VP9MVRefPair *mv[2];
VP9Filter *lflvl;
DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[71 * 80];
// block reconstruction intermediates
DECLARE_ALIGNED(32, int16_t, block)[4096];
DECLARE_ALIGNED(32, int16_t, uvblock)[2][1024];
uint8_t eob[256];
uint8_t uveob[2][64];
VP56mv min_mv, max_mv;
DECLARE_ALIGNED(32, uint8_t, tmp_y)[64 * 64];
DECLARE_ALIGNED(32, uint8_t, tmp_uv)[2][32 * 32];
} VP9Context;
void ff_vp9dsp_init(VP9DSPContext *dsp);
void ff_vp9dsp_init_x86(VP9DSPContext *dsp);
void ff_vp9_fill_mv(VP9Context *s, VP56mv *mv, int mode, int sb);
void ff_vp9_adapt_probs(VP9Context *s);
int ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,
VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,
enum BlockLevel bl, enum BlockPartition bp);
#endif /* AVCODEC_VP9_H */

1684
libavcodec/vp9block.c Normal file
View file

File diff suppressed because it is too large Load diff

2133
libavcodec/vp9data.c Normal file
View file

File diff suppressed because it is too large Load diff

70
libavcodec/vp9data.h Normal file
View file

@ -0,0 +1,70 @@
/*
* Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>
* Copyright (C) 2013 Clément Bœsch <u pkh me>
*
* This file is part of Libav.
*
* Libav is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* Libav is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with Libav; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#ifndef AVCODEC_VP9DATA_H
#define AVCODEC_VP9DATA_H
#include <stdint.h>
#include "vp9.h"
extern const int8_t ff_vp9_partition_tree[3][2];
extern const uint8_t ff_vp9_default_kf_partition_probs[4][4][3];
extern const int8_t ff_vp9_segmentation_tree[7][2];
extern const int8_t ff_vp9_intramode_tree[9][2];
extern const uint8_t ff_vp9_default_kf_ymode_probs[10][10][9];
extern const uint8_t ff_vp9_default_kf_uvmode_probs[10][9];
extern const int8_t ff_vp9_inter_mode_tree[3][2];
extern const int8_t ff_vp9_filter_tree[2][2];
extern const enum FilterMode ff_vp9_filter_lut[3];
extern const int16_t ff_vp9_dc_qlookup[256];
extern const int16_t ff_vp9_ac_qlookup[256];
extern const enum TxfmType ff_vp9_intra_txfm_type[14];
extern const int16_t ff_vp9_default_scan_4x4[16];
extern const int16_t ff_vp9_col_scan_4x4[16];
extern const int16_t ff_vp9_row_scan_4x4[16];
extern const int16_t ff_vp9_default_scan_8x8[64];
extern const int16_t ff_vp9_col_scan_8x8[64];
extern const int16_t ff_vp9_row_scan_8x8[64];
extern const int16_t ff_vp9_default_scan_16x16[256];
extern const int16_t ff_vp9_col_scan_16x16[256];
extern const int16_t ff_vp9_row_scan_16x16[256];
extern const int16_t ff_vp9_default_scan_32x32[1024];
extern const int16_t *ff_vp9_scans[5][4];
extern const int16_t ff_vp9_default_scan_4x4_nb[16][2];
extern const int16_t ff_vp9_col_scan_4x4_nb[16][2];
extern const int16_t ff_vp9_row_scan_4x4_nb[16][2];
extern const int16_t ff_vp9_default_scan_8x8_nb[64][2];
extern const int16_t ff_vp9_col_scan_8x8_nb[64][2];
extern const int16_t ff_vp9_row_scan_8x8_nb[64][2];
extern const int16_t ff_vp9_default_scan_16x16_nb[256][2];
extern const int16_t ff_vp9_col_scan_16x16_nb[256][2];
extern const int16_t ff_vp9_row_scan_16x16_nb[256][2];
extern const int16_t ff_vp9_default_scan_32x32_nb[1024][2];
extern const int16_t (*ff_vp9_scans_nb[5][4])[2];
extern const uint8_t ff_vp9_model_pareto8[256][8];
extern const ProbContext ff_vp9_default_probs;
extern const uint8_t ff_vp9_default_coef_probs[4][2][2][6][6][3];
extern const int8_t ff_vp9_mv_joint_tree[3][2];
extern const int8_t ff_vp9_mv_class_tree[10][2];
extern const int8_t ff_vp9_mv_fp_tree[3][2];
#endif /* AVCODEC_VP9DATA_H */

2174
libavcodec/vp9dsp.c Normal file
View file

File diff suppressed because it is too large Load diff

344
libavcodec/vp9mvs.c Normal file
View file

@ -0,0 +1,344 @@
/*
* VP9 compatible video decoder
*
* Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>
* Copyright (C) 2013 Clément Bœsch <u pkh me>
*
* This file is part of Libav.
*
* Libav is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* Libav is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with Libav; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "internal.h"
#include "vp56.h"
#include "vp9.h"
#include "vp9data.h"
static av_always_inline void clamp_mv(VP56mv *dst, const VP56mv *src,
VP9Context *s)
{
dst->x = av_clip(src->x, s->min_mv.x, s->max_mv.x);
dst->y = av_clip(src->y, s->min_mv.y, s->max_mv.y);
}
static void find_ref_mvs(VP9Context *s,
VP56mv *pmv, int ref, int z, int idx, int sb)
{
static const int8_t mv_ref_blk_off[N_BS_SIZES][8][2] = {
[BS_64x64] = { { 3, -1 }, { -1, 3 }, { 4, -1 }, { -1, 4 },
{ -1, -1 }, { 0, -1 }, { -1, 0 }, { 6, -1 } },
[BS_64x32] = { { 0, -1 }, { -1, 0 }, { 4, -1 }, { -1, 2 },
{ -1, -1 }, { 0, -3 }, { -3, 0 }, { 2, -1 } },
[BS_32x64] = { { -1, 0 }, { 0, -1 }, { -1, 4 }, { 2, -1 },
{ -1, -1 }, { -3, 0 }, { 0, -3 }, { -1, 2 } },
[BS_32x32] = { { 1, -1 }, { -1, 1 }, { 2, -1 }, { -1, 2 },
{ -1, -1 }, { 0, -3 }, { -3, 0 }, { -3, -3 } },
[BS_32x16] = { { 0, -1 }, { -1, 0 }, { 2, -1 }, { -1, -1 },
{ -1, 1 }, { 0, -3 }, { -3, 0 }, { -3, -3 } },
[BS_16x32] = { { -1, 0 }, { 0, -1 }, { -1, 2 }, { -1, -1 },
{ 1, -1 }, { -3, 0 }, { 0, -3 }, { -3, -3 } },
[BS_16x16] = { { 0, -1 }, { -1, 0 }, { 1, -1 }, { -1, 1 },
{ -1, -1 }, { 0, -3 }, { -3, 0 }, { -3, -3 } },
[BS_16x8] = { { 0, -1 }, { -1, 0 }, { 1, -1 }, { -1, -1 },
{ 0, -2 }, { -2, 0 }, { -2, -1 }, { -1, -2 } },
[BS_8x16] = { { -1, 0 }, { 0, -1 }, { -1, 1 }, { -1, -1 },
{ -2, 0 }, { 0, -2 }, { -1, -2 }, { -2, -1 } },
[BS_8x8] = { { 0, -1 }, { -1, 0 }, { -1, -1 }, { 0, -2 },
{ -2, 0 }, { -1, -2 }, { -2, -1 }, { -2, -2 } },
[BS_8x4] = { { 0, -1 }, { -1, 0 }, { -1, -1 }, { 0, -2 },
{ -2, 0 }, { -1, -2 }, { -2, -1 }, { -2, -2 } },
[BS_4x8] = { { 0, -1 }, { -1, 0 }, { -1, -1 }, { 0, -2 },
{ -2, 0 }, { -1, -2 }, { -2, -1 }, { -2, -2 } },
[BS_4x4] = { { 0, -1 }, { -1, 0 }, { -1, -1 }, { 0, -2 },
{ -2, 0 }, { -1, -2 }, { -2, -1 }, { -2, -2 } },
};
VP9Block *const b = &s->b;
int row = b->row, col = b->col, row7 = b->row7;
const int8_t (*p)[2] = mv_ref_blk_off[b->bs];
#define INVALID_MV 0x80008000U
uint32_t mem = INVALID_MV;
int i;
#define RETURN_DIRECT_MV(mv) \
do { \
uint32_t m = AV_RN32A(&mv); \
if (!idx) { \
AV_WN32A(pmv, m); \
return; \
} else if (mem == INVALID_MV) { \
mem = m; \
} else if (m != mem) { \
AV_WN32A(pmv, m); \
return; \
} \
} while (0)
if (sb >= 0) {
if (sb == 2 || sb == 1) {
RETURN_DIRECT_MV(b->mv[0][z]);
} else if (sb == 3) {
RETURN_DIRECT_MV(b->mv[2][z]);
RETURN_DIRECT_MV(b->mv[1][z]);
RETURN_DIRECT_MV(b->mv[0][z]);
}
#define RETURN_MV(mv) \
do { \
if (sb > 0) { \
VP56mv tmp; \
uint32_t m; \
clamp_mv(&tmp, &mv, s); \
m = AV_RN32A(&tmp); \
if (!idx) { \
AV_WN32A(pmv, m); \
return; \
} else if (mem == INVALID_MV) { \
mem = m; \
} else if (m != mem) { \
AV_WN32A(pmv, m); \
return; \
} \
} else { \
uint32_t m = AV_RN32A(&mv); \
if (!idx) { \
clamp_mv(pmv, &mv, s); \
return; \
} else if (mem == INVALID_MV) { \
mem = m; \
} else if (m != mem) { \
clamp_mv(pmv, &mv, s); \
return; \
} \
} \
} while (0)
if (row > 0) {
VP9MVRefPair *mv = &s->mv[0][(row - 1) * s->sb_cols * 8 + col];
if (mv->ref[0] == ref)
RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);
else if (mv->ref[1] == ref)
RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);
}
if (col > s->tiling.tile_col_start) {
VP9MVRefPair *mv = &s->mv[0][row * s->sb_cols * 8 + col - 1];
if (mv->ref[0] == ref)
RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);
else if (mv->ref[1] == ref)
RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);
}
i = 2;
} else {
i = 0;
}
// previously coded MVs in the neighborhood, using same reference frame
for (; i < 8; i++) {
int c = p[i][0] + col, r = p[i][1] + row;
if (c >= s->tiling.tile_col_start && c < s->cols &&
r >= 0 && r < s->rows) {
VP9MVRefPair *mv = &s->mv[0][r * s->sb_cols * 8 + c];
if (mv->ref[0] == ref)
RETURN_MV(mv->mv[0]);
else if (mv->ref[1] == ref)
RETURN_MV(mv->mv[1]);
}
}
// MV at this position in previous frame, using same reference frame
if (s->use_last_frame_mvs) {
VP9MVRefPair *mv = &s->mv[1][row * s->sb_cols * 8 + col];
if (mv->ref[0] == ref)
RETURN_MV(mv->mv[0]);
else if (mv->ref[1] == ref)
RETURN_MV(mv->mv[1]);
}
#define RETURN_SCALE_MV(mv, scale) \
do { \
if (scale) { \
VP56mv mv_temp = { -mv.x, -mv.y }; \
RETURN_MV(mv_temp); \
} else { \
RETURN_MV(mv); \
} \
} while (0)
// previously coded MVs in the neighborhood, using different reference frame
for (i = 0; i < 8; i++) {
int c = p[i][0] + col, r = p[i][1] + row;
if (c >= s->tiling.tile_col_start && c < s->cols &&
r >= 0 && r < s->rows) {
VP9MVRefPair *mv = &s->mv[0][r * s->sb_cols * 8 + c];
if (mv->ref[0] != ref && mv->ref[0] >= 0)
RETURN_SCALE_MV(mv->mv[0],
s->signbias[mv->ref[0]] != s->signbias[ref]);
if (mv->ref[1] != ref && mv->ref[1] >= 0)
RETURN_SCALE_MV(mv->mv[1],
s->signbias[mv->ref[1]] != s->signbias[ref]);
}
}
// MV at this position in previous frame, using different reference frame
if (s->use_last_frame_mvs) {
VP9MVRefPair *mv = &s->mv[1][row * s->sb_cols * 8 + col];
if (mv->ref[0] != ref && mv->ref[0] >= 0)
RETURN_SCALE_MV(mv->mv[0],
s->signbias[mv->ref[0]] != s->signbias[ref]);
if (mv->ref[1] != ref && mv->ref[1] >= 0)
RETURN_SCALE_MV(mv->mv[1],
s->signbias[mv->ref[1]] != s->signbias[ref]);
}
AV_ZERO32(pmv);
#undef INVALID_MV
#undef RETURN_MV
#undef RETURN_SCALE_MV
}
static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)
{
int bit, sign = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].sign);
int n, c = vp8_rac_get_tree(&s->c, ff_vp9_mv_class_tree,
s->prob.p.mv_comp[idx].classes);
s->counts.mv_comp[idx].sign[sign]++;
s->counts.mv_comp[idx].classes[c]++;
if (c) {
int m;
for (n = 0, m = 0; m < c; m++) {
bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);
n |= bit << m;
s->counts.mv_comp[idx].bits[m][bit]++;
}
n <<= 3;
bit = vp8_rac_get_tree(&s->c, ff_vp9_mv_fp_tree,
s->prob.p.mv_comp[idx].fp);
n |= bit << 1;
s->counts.mv_comp[idx].fp[bit]++;
if (hp) {
bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);
s->counts.mv_comp[idx].hp[bit]++;
n |= bit;
} else {
n |= 1;
// bug in libvpx - we count for bw entropy purposes even if the
// bit wasn't coded
s->counts.mv_comp[idx].hp[1]++;
}
n += 8 << c;
} else {
n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);
s->counts.mv_comp[idx].class0[n]++;
bit = vp8_rac_get_tree(&s->c, ff_vp9_mv_fp_tree,
s->prob.p.mv_comp[idx].class0_fp[n]);
s->counts.mv_comp[idx].class0_fp[n][bit]++;
n = (n << 3) | (bit << 1);
if (hp) {
bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);
s->counts.mv_comp[idx].class0_hp[bit]++;
n |= bit;
} else {
n |= 1;
// bug in libvpx - we count for bw entropy purposes even if the
// bit wasn't coded
s->counts.mv_comp[idx].class0_hp[1]++;
}
}
return sign ? -(n + 1) : (n + 1);
}
void ff_vp9_fill_mv(VP9Context *s, VP56mv *mv, int mode, int sb)
{
VP9Block *const b = &s->b;
if (mode == ZEROMV) {
memset(mv, 0, sizeof(*mv) * 2);
} else {
int hp;
// FIXME cache this value and reuse for other subblocks
find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,
mode == NEWMV ? -1 : sb);
// FIXME maybe move this code into find_ref_mvs()
if ((mode == NEWMV || sb == -1) &&
!(hp = s->highprecisionmvs &&
abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {
if (mv[0].y & 1) {
if (mv[0].y < 0)
mv[0].y++;
else
mv[0].y--;
}
if (mv[0].x & 1) {
if (mv[0].x < 0)
mv[0].x++;
else
mv[0].x--;
}
}
if (mode == NEWMV) {
enum MVJoint j = vp8_rac_get_tree(&s->c, ff_vp9_mv_joint_tree,
s->prob.p.mv_joint);
s->counts.mv_joint[j]++;
if (j >= MV_JOINT_V)
mv[0].y += read_mv_component(s, 0, hp);
if (j & 1)
mv[0].x += read_mv_component(s, 1, hp);
}
if (b->comp) {
// FIXME cache this value and reuse for other subblocks
find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,
mode == NEWMV ? -1 : sb);
if ((mode == NEWMV || sb == -1) &&
!(hp = s->highprecisionmvs &&
abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {
if (mv[1].y & 1) {
if (mv[1].y < 0)
mv[1].y++;
else
mv[1].y--;
}
if (mv[1].x & 1) {
if (mv[1].x < 0)
mv[1].x++;
else
mv[1].x--;
}
}
if (mode == NEWMV) {
enum MVJoint j = vp8_rac_get_tree(&s->c, ff_vp9_mv_joint_tree,
s->prob.p.mv_joint);
s->counts.mv_joint[j]++;
if (j >= MV_JOINT_V)
mv[1].y += read_mv_component(s, 0, hp);
if (j & 1)
mv[1].x += read_mv_component(s, 1, hp);
}
}
}
}

274
libavcodec/vp9prob.c Normal file
View file

@ -0,0 +1,274 @@
/*
* VP9 compatible video decoder
*
* Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>
* Copyright (C) 2013 Clément Bœsch <u pkh me>
*
* This file is part of Libav.
*
* Libav is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* Libav is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with Libav; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "vp56.h"
#include "vp9.h"
#include "vp9data.h"
static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,
int max_count, int update_factor)
{
unsigned ct = ct0 + ct1, p2, p1;
if (!ct)
return;
p1 = *p;
p2 = ((ct0 << 8) + (ct >> 1)) / ct;
p2 = av_clip(p2, 1, 255);
ct = FFMIN(ct, max_count);
update_factor = FASTDIV(update_factor * ct, max_count);
// (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8
*p = p1 + (((p2 - p1) * update_factor + 128) >> 8);
}
void ff_vp9_adapt_probs(VP9Context *s)
{
int i, j, k, l, m;
ProbContext *p = &s->prob_ctx[s->framectxid].p;
int uf = (s->keyframe || s->intraonly || !s->last_keyframe) ? 112 : 128;
// coefficients
for (i = 0; i < 4; i++)
for (j = 0; j < 2; j++)
for (k = 0; k < 2; k++)
for (l = 0; l < 6; l++)
for (m = 0; m < 6; m++) {
uint8_t *pp = s->prob_ctx[s->framectxid].coef[i][j][k][l][m];
unsigned *e = s->counts.eob[i][j][k][l][m];
unsigned *c = s->counts.coef[i][j][k][l][m];
if (l == 0 && m >= 3) // dc only has 3 pt
break;
adapt_prob(&pp[0], e[0], e[1], 24, uf);
adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);
adapt_prob(&pp[2], c[1], c[2], 24, uf);
}
if (s->keyframe || s->intraonly) {
memcpy(p->skip, s->prob.p.skip, sizeof(p->skip));
memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));
memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));
memcpy(p->tx8p, s->prob.p.tx8p, sizeof(p->tx8p));
return;
}
// skip flag
for (i = 0; i < 3; i++)
adapt_prob(&p->skip[i], s->counts.skip[i][0],
s->counts.skip[i][1], 20, 128);
// intra/inter flag
for (i = 0; i < 4; i++)
adapt_prob(&p->intra[i], s->counts.intra[i][0],
s->counts.intra[i][1], 20, 128);
// comppred flag
if (s->comppredmode == PRED_SWITCHABLE) {
for (i = 0; i < 5; i++)
adapt_prob(&p->comp[i], s->counts.comp[i][0],
s->counts.comp[i][1], 20, 128);
}
// reference frames
if (s->comppredmode != PRED_SINGLEREF) {
for (i = 0; i < 5; i++)
adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],
s->counts.comp_ref[i][1], 20, 128);
}
if (s->comppredmode != PRED_COMPREF) {
for (i = 0; i < 5; i++) {
uint8_t *pp = p->single_ref[i];
unsigned (*c)[2] = s->counts.single_ref[i];
adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);
adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);
}
}
// block partitioning
for (i = 0; i < 4; i++)
for (j = 0; j < 4; j++) {
uint8_t *pp = p->partition[i][j];
unsigned *c = s->counts.partition[i][j];
adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);
adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);
adapt_prob(&pp[2], c[2], c[3], 20, 128);
}
// tx size
if (s->txfmmode == TX_SWITCHABLE) {
for (i = 0; i < 2; i++) {
unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];
adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0],
s->counts.tx8p[i][1], 20, 128);
adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);
adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);
adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);
adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);
adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);
}
}
// interpolation filter
if (s->filtermode == FILTER_SWITCHABLE) {
for (i = 0; i < 4; i++) {
uint8_t *pp = p->filter[i];
unsigned *c = s->counts.filter[i];
adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);
adapt_prob(&pp[1], c[1], c[2], 20, 128);
}
}
// inter modes
for (i = 0; i < 7; i++) {
uint8_t *pp = p->mv_mode[i];
unsigned *c = s->counts.mv_mode[i];
adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);
adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);
adapt_prob(&pp[2], c[1], c[3], 20, 128);
}
// mv joints
{
uint8_t *pp = p->mv_joint;
unsigned *c = s->counts.mv_joint;
adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);
adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);
adapt_prob(&pp[2], c[2], c[3], 20, 128);
}
// mv components
for (i = 0; i < 2; i++) {
uint8_t *pp;
unsigned *c, (*c2)[2], sum;
adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],
s->counts.mv_comp[i].sign[1], 20, 128);
pp = p->mv_comp[i].classes;
c = s->counts.mv_comp[i].classes;
sum = c[1] + c[2] + c[3] + c[4] + c[5] +
c[6] + c[7] + c[8] + c[9] + c[10];
adapt_prob(&pp[0], c[0], sum, 20, 128);
sum -= c[1];
adapt_prob(&pp[1], c[1], sum, 20, 128);
sum -= c[2] + c[3];
adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);
adapt_prob(&pp[3], c[2], c[3], 20, 128);
sum -= c[4] + c[5];
adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);
adapt_prob(&pp[5], c[4], c[5], 20, 128);
sum -= c[6];
adapt_prob(&pp[6], c[6], sum, 20, 128);
adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);
adapt_prob(&pp[8], c[7], c[8], 20, 128);
adapt_prob(&pp[9], c[9], c[10], 20, 128);
adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],
s->counts.mv_comp[i].class0[1], 20, 128);
pp = p->mv_comp[i].bits;
c2 = s->counts.mv_comp[i].bits;
for (j = 0; j < 10; j++)
adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);
for (j = 0; j < 2; j++) {
pp = p->mv_comp[i].class0_fp[j];
c = s->counts.mv_comp[i].class0_fp[j];
adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);
adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);
adapt_prob(&pp[2], c[2], c[3], 20, 128);
}
pp = p->mv_comp[i].fp;
c = s->counts.mv_comp[i].fp;
adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);
adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);
adapt_prob(&pp[2], c[2], c[3], 20, 128);
if (s->highprecisionmvs) {
adapt_prob(&p->mv_comp[i].class0_hp,
s->counts.mv_comp[i].class0_hp[0],
s->counts.mv_comp[i].class0_hp[1], 20, 128);
adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],
s->counts.mv_comp[i].hp[1], 20, 128);
}
}
// y intra modes
for (i = 0; i < 4; i++) {
uint8_t *pp = p->y_mode[i];
unsigned *c = s->counts.y_mode[i], sum, s2;
sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];
adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);
sum -= c[TM_VP8_PRED];
adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);
sum -= c[VERT_PRED];
adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);
s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];
sum -= s2;
adapt_prob(&pp[3], s2, sum, 20, 128);
s2 -= c[HOR_PRED];
adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);
adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED],
20, 128);
sum -= c[DIAG_DOWN_LEFT_PRED];
adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);
sum -= c[VERT_LEFT_PRED];
adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);
adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);
}
// uv intra modes
for (i = 0; i < 10; i++) {
uint8_t *pp = p->uv_mode[i];
unsigned *c = s->counts.uv_mode[i], sum, s2;
sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];
adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);
sum -= c[TM_VP8_PRED];
adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);
sum -= c[VERT_PRED];
adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);
s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];
sum -= s2;
adapt_prob(&pp[3], s2, sum, 20, 128);
s2 -= c[HOR_PRED];
adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);
adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED],
20, 128);
sum -= c[DIAG_DOWN_LEFT_PRED];
adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);
sum -= c[VERT_LEFT_PRED];
adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);
adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);
}
}

2
libavcodec/x86/Makefile
View file

@ -35,6 +35,7 @@ OBJS-$(CONFIG_VORBIS_DECODER) += x86/vorbisdsp_init.o
OBJS-$(CONFIG_VP3DSP) += x86/vp3dsp_init.o
OBJS-$(CONFIG_VP6_DECODER) += x86/vp6dsp_init.o
OBJS-$(CONFIG_VP8_DECODER) += x86/vp8dsp_init.o
OBJS-$(CONFIG_VP9_DECODER) += x86/vp9dsp_init.o
OBJS-$(CONFIG_XMM_CLOBBER_TEST) += x86/w64xmmtest.o
MMX-OBJS-$(CONFIG_DSPUTIL) += x86/dsputil_mmx.o \
@ -90,3 +91,4 @@ YASM-OBJS-$(CONFIG_VP3DSP) += x86/vp3dsp.o
YASM-OBJS-$(CONFIG_VP6_DECODER) += x86/vp6dsp.o
YASM-OBJS-$(CONFIG_VP8_DECODER) += x86/vp8dsp.o \
x86/vp8dsp_loopfilter.o
YASM-OBJS-$(CONFIG_VP9_DECODER) += x86/vp9dsp.o

277
libavcodec/x86/vp9dsp.asm Normal file
View file

@ -0,0 +1,277 @@
;******************************************************************************
;* VP9 SIMD optimizations
;*
;* Copyright (c) 2013 Ronald S. Bultje <rsbultje gmail com>
;*
;* This file is part of Libav.
;*
;* Libav is free software; you can redistribute it and/or
;* modify it under the terms of the GNU Lesser General Public
;* License as published by the Free Software Foundation; either
;* version 2.1 of the License, or (at your option) any later version.
;*
;* Libav is distributed in the hope that it will be useful,
;* but WITHOUT ANY WARRANTY; without even the implied warranty of
;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
;* Lesser General Public License for more details.
;*
;* You should have received a copy of the GNU Lesser General Public
;* License along with Libav; if not, write to the Free Software
;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
;******************************************************************************
%include "libavutil/x86/x86util.asm"
SECTION_RODATA
; FIXME share with vp8dsp.asm
pw_256: times 8 dw 256
%macro F8_TAPS 8
times 8 db %1, %2
times 8 db %3, %4
times 8 db %5, %6
times 8 db %7, %8
%endmacro
; int8_t ff_filters_ssse3[3][15][4][16]
const filters_ssse3 ; smooth
F8_TAPS -3, -1, 32, 64, 38, 1, -3, 0
F8_TAPS -2, -2, 29, 63, 41, 2, -3, 0
F8_TAPS -2, -2, 26, 63, 43, 4, -4, 0
F8_TAPS -2, -3, 24, 62, 46, 5, -4, 0
F8_TAPS -2, -3, 21, 60, 49, 7, -4, 0
F8_TAPS -1, -4, 18, 59, 51, 9, -4, 0
F8_TAPS -1, -4, 16, 57, 53, 12, -4, -1
F8_TAPS -1, -4, 14, 55, 55, 14, -4, -1
F8_TAPS -1, -4, 12, 53, 57, 16, -4, -1
F8_TAPS 0, -4, 9, 51, 59, 18, -4, -1
F8_TAPS 0, -4, 7, 49, 60, 21, -3, -2
F8_TAPS 0, -4, 5, 46, 62, 24, -3, -2
F8_TAPS 0, -4, 4, 43, 63, 26, -2, -2
F8_TAPS 0, -3, 2, 41, 63, 29, -2, -2
F8_TAPS 0, -3, 1, 38, 64, 32, -1, -3
; regular
F8_TAPS 0, 1, -5, 126, 8, -3, 1, 0
F8_TAPS -1, 3, -10, 122, 18, -6, 2, 0
F8_TAPS -1, 4, -13, 118, 27, -9, 3, -1
F8_TAPS -1, 4, -16, 112, 37, -11, 4, -1
F8_TAPS -1, 5, -18, 105, 48, -14, 4, -1
F8_TAPS -1, 5, -19, 97, 58, -16, 5, -1
F8_TAPS -1, 6, -19, 88, 68, -18, 5, -1
F8_TAPS -1, 6, -19, 78, 78, -19, 6, -1
F8_TAPS -1, 5, -18, 68, 88, -19, 6, -1
F8_TAPS -1, 5, -16, 58, 97, -19, 5, -1
F8_TAPS -1, 4, -14, 48, 105, -18, 5, -1
F8_TAPS -1, 4, -11, 37, 112, -16, 4, -1
F8_TAPS -1, 3, -9, 27, 118, -13, 4, -1
F8_TAPS 0, 2, -6, 18, 122, -10, 3, -1
F8_TAPS 0, 1, -3, 8, 126, -5, 1, 0
; sharp
F8_TAPS -1, 3, -7, 127, 8, -3, 1, 0
F8_TAPS -2, 5, -13, 125, 17, -6, 3, -1
F8_TAPS -3, 7, -17, 121, 27, -10, 5, -2
F8_TAPS -4, 9, -20, 115, 37, -13, 6, -2
F8_TAPS -4, 10, -23, 108, 48, -16, 8, -3
F8_TAPS -4, 10, -24, 100, 59, -19, 9, -3
F8_TAPS -4, 11, -24, 90, 70, -21, 10, -4
F8_TAPS -4, 11, -23, 80, 80, -23, 11, -4
F8_TAPS -4, 10, -21, 70, 90, -24, 11, -4
F8_TAPS -3, 9, -19, 59, 100, -24, 10, -4
F8_TAPS -3, 8, -16, 48, 108, -23, 10, -4
F8_TAPS -2, 6, -13, 37, 115, -20, 9, -4
F8_TAPS -2, 5, -10, 27, 121, -17, 7, -3
F8_TAPS -1, 3, -6, 17, 125, -13, 5, -2
F8_TAPS 0, 1, -3, 8, 127, -7, 3, -1
SECTION .text
%macro filter_h_fn 1
%assign %%px mmsize/2
cglobal %1_8tap_1d_h_ %+ %%px, 6, 6, 11, dst, src, dstride, sstride, h, filtery
mova m6, [pw_256]
mova m7, [filteryq+ 0]
%if ARCH_X86_64 && mmsize > 8
mova m8, [filteryq+16]
mova m9, [filteryq+32]
mova m10, [filteryq+48]
%endif
.loop:
movh m0, [srcq-3]
movh m1, [srcq-2]
movh m2, [srcq-1]
movh m3, [srcq+0]
movh m4, [srcq+1]
movh m5, [srcq+2]
punpcklbw m0, m1
punpcklbw m2, m3
movh m1, [srcq+3]
movh m3, [srcq+4]
add srcq, sstrideq
punpcklbw m4, m5
punpcklbw m1, m3
pmaddubsw m0, m7
%if ARCH_X86_64 && mmsize > 8
pmaddubsw m2, m8
pmaddubsw m4, m9
pmaddubsw m1, m10
%else
pmaddubsw m2, [filteryq+16]
pmaddubsw m4, [filteryq+32]
pmaddubsw m1, [filteryq+48]
%endif
paddw m0, m2
paddw m4, m1
paddsw m0, m4
pmulhrsw m0, m6
%ifidn %1, avg
movh m1, [dstq]
%endif
packuswb m0, m0
%ifidn %1, avg
pavgb m0, m1
%endif
movh [dstq], m0
add dstq, dstrideq
dec hd
jg .loop
RET
%endmacro
INIT_MMX ssse3
filter_h_fn put
filter_h_fn avg
INIT_XMM ssse3
filter_h_fn put
filter_h_fn avg
%macro filter_v_fn 1
%assign %%px mmsize/2
%if ARCH_X86_64
cglobal %1_8tap_1d_v_ %+ %%px, 6, 8, 11, dst, src, dstride, sstride, h, filtery, src4, sstride3
%else
cglobal %1_8tap_1d_v_ %+ %%px, 4, 7, 11, dst, src, dstride, sstride, filtery, src4, sstride3
mov filteryq, r5mp
%define hd r4mp
%endif
sub srcq, sstrideq
lea sstride3q, [sstrideq*3]
sub srcq, sstrideq
mova m6, [pw_256]
sub srcq, sstrideq
mova m7, [filteryq+ 0]
lea src4q, [srcq+sstrideq*4]
%if ARCH_X86_64 && mmsize > 8
mova m8, [filteryq+16]
mova m9, [filteryq+32]
mova m10, [filteryq+48]
%endif
.loop:
; FIXME maybe reuse loads from previous rows, or just more generally
; unroll this to prevent multiple loads of the same data?
movh m0, [srcq]
movh m1, [srcq+sstrideq]
movh m2, [srcq+sstrideq*2]
movh m3, [srcq+sstride3q]
movh m4, [src4q]
movh m5, [src4q+sstrideq]
punpcklbw m0, m1
punpcklbw m2, m3
movh m1, [src4q+sstrideq*2]
movh m3, [src4q+sstride3q]
add srcq, sstrideq
add src4q, sstrideq
punpcklbw m4, m5
punpcklbw m1, m3
pmaddubsw m0, m7
%if ARCH_X86_64 && mmsize > 8
pmaddubsw m2, m8
pmaddubsw m4, m9
pmaddubsw m1, m10
%else
pmaddubsw m2, [filteryq+16]
pmaddubsw m4, [filteryq+32]
pmaddubsw m1, [filteryq+48]
%endif
paddw m0, m2
paddw m4, m1
paddsw m0, m4
pmulhrsw m0, m6
%ifidn %1, avg
movh m1, [dstq]
%endif
packuswb m0, m0
%ifidn %1, avg
pavgb m0, m1
%endif
movh [dstq], m0
add dstq, dstrideq
dec hd
jg .loop
RET
%endmacro
INIT_MMX ssse3
filter_v_fn put
filter_v_fn avg
INIT_XMM ssse3
filter_v_fn put
filter_v_fn avg
%macro fpel_fn 6
%if %2 == 4
%define %%srcfn movh
%define %%dstfn movh
%else
%define %%srcfn movu
%define %%dstfn mova
%endif
%if %2 <= 16
cglobal %1%2, 5, 7, 4, dst, src, dstride, sstride, h, dstride3, sstride3
lea sstride3q, [sstrideq*3]
lea dstride3q, [dstrideq*3]
%else
cglobal %1%2, 5, 5, 4, dst, src, dstride, sstride, h
%endif
.loop:
%%srcfn m0, [srcq]
%%srcfn m1, [srcq+s%3]
%%srcfn m2, [srcq+s%4]
%%srcfn m3, [srcq+s%5]
lea srcq, [srcq+sstrideq*%6]
%ifidn %1, avg
pavgb m0, [dstq]
pavgb m1, [dstq+d%3]
pavgb m2, [dstq+d%4]
pavgb m3, [dstq+d%5]
%endif
%%dstfn [dstq], m0
%%dstfn [dstq+d%3], m1
%%dstfn [dstq+d%4], m2
%%dstfn [dstq+d%5], m3
lea dstq, [dstq+dstrideq*%6]
sub hd, %6
jnz .loop
RET
%endmacro
%define d16 16
%define s16 16
INIT_MMX mmx
fpel_fn put, 4, strideq, strideq*2, stride3q, 4
fpel_fn put, 8, strideq, strideq*2, stride3q, 4
INIT_MMX sse
fpel_fn avg, 4, strideq, strideq*2, stride3q, 4
fpel_fn avg, 8, strideq, strideq*2, stride3q, 4
INIT_XMM sse
fpel_fn put, 16, strideq, strideq*2, stride3q, 4
fpel_fn put, 32, mmsize, strideq, strideq+mmsize, 2
fpel_fn put, 64, mmsize, mmsize*2, mmsize*3, 1
INIT_XMM sse2
fpel_fn avg, 16, strideq, strideq*2, stride3q, 4
fpel_fn avg, 32, mmsize, strideq, strideq+mmsize, 2
fpel_fn avg, 64, mmsize, mmsize*2, mmsize*3, 1
%undef s16
%undef d16

245
libavcodec/x86/vp9dsp_init.c Normal file
View file

@ -0,0 +1,245 @@
/*
* VP9 SIMD optimizations
*
* Copyright (c) 2013 Ronald S. Bultje <rsbultje@gmail.com>
*
* This file is part of Libav.
*
* Libav is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* Libav is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with Libav; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/cpu.h"
#include "libavutil/internal.h"
#include "libavutil/mem.h"
#include "libavutil/x86/asm.h"
#include "libavutil/x86/cpu.h"
#include "libavcodec/vp9.h"
#if HAVE_YASM
#define fpel_func(avg, sz, opt) \
void ff_ ## avg ## sz ## _ ## opt(uint8_t *dst, const uint8_t *src, \
ptrdiff_t dst_stride, \
ptrdiff_t src_stride, \
int h, int mx, int my)
fpel_func(put, 4, mmx);
fpel_func(put, 8, mmx);
fpel_func(put, 16, sse);
fpel_func(put, 32, sse);
fpel_func(put, 64, sse);
fpel_func(avg, 4, sse);
fpel_func(avg, 8, sse);
fpel_func(avg, 16, sse2);
fpel_func(avg, 32, sse2);
fpel_func(avg, 64, sse2);
#undef fpel_func
#define mc_func(avg, sz, dir, opt) \
void \
ff_ ## avg ## _8tap_1d_ ## dir ## _ ## sz ## _ ## opt(uint8_t *dst, \
const uint8_t *src, \
ptrdiff_t dst_stride, \
ptrdiff_t src_stride, \
int h, \
const int8_t (*filter)[16])
#define mc_funcs(sz) \
mc_func(put, sz, h, ssse3); \
mc_func(avg, sz, h, ssse3); \
mc_func(put, sz, v, ssse3); \
mc_func(avg, sz, v, ssse3)
mc_funcs(4);
mc_funcs(8);
#undef mc_funcs
#undef mc_func
#define mc_rep_func(avg, sz, hsz, dir, opt) \
static av_always_inline void \
ff_ ## avg ## _8tap_1d_ ## dir ## _ ## sz ## _ ## opt(uint8_t *dst, \
const uint8_t *src, \
ptrdiff_t dst_stride, \
ptrdiff_t src_stride, \
int h, \
const int8_t (*filter)[16]) \
{ \
ff_ ## avg ## _8tap_1d_ ## dir ## _ ## hsz ## _ ## opt(dst, src, \
dst_stride, \
src_stride, \
h, \
filter); \
ff_ ## avg ## _8tap_1d_ ## dir ## _ ## hsz ## _ ## opt(dst + hsz, \
src + hsz, \
dst_stride, \
src_stride, \
h, filter); \
}
#define mc_rep_funcs(sz, hsz) \
mc_rep_func(put, sz, hsz, h, ssse3); \
mc_rep_func(avg, sz, hsz, h, ssse3); \
mc_rep_func(put, sz, hsz, v, ssse3); \
mc_rep_func(avg, sz, hsz, v, ssse3)
mc_rep_funcs(16, 8);
mc_rep_funcs(32, 16);
mc_rep_funcs(64, 32);
#undef mc_rep_funcs
#undef mc_rep_func
extern const int8_t ff_filters_ssse3[3][15][4][16];
#define filter_8tap_2d_fn(op, sz, f, fname) \
static void \
op ## _8tap_ ## fname ## _ ## sz ## hv_ssse3(uint8_t *dst, \
const uint8_t *src, \
ptrdiff_t dst_stride, \
ptrdiff_t src_stride, \
int h, int mx, int my) \
{ \
LOCAL_ALIGNED_16(uint8_t, temp, [71 * 64]); \
ff_put_8tap_1d_h_ ## sz ## _ssse3(temp, src - 3 * src_stride, \
64, src_stride, \
h + 7, \
ff_filters_ssse3[f][mx - 1]); \
ff_ ## op ## _8tap_1d_v_ ## sz ## _ssse3(dst, temp + 3 * 64, \
dst_stride, 64, \
h, \
ff_filters_ssse3[f][my - 1]); \
}
#define filters_8tap_2d_fn(op, sz) \
filter_8tap_2d_fn(op, sz, FILTER_8TAP_REGULAR, regular) \
filter_8tap_2d_fn(op, sz, FILTER_8TAP_SHARP, sharp) \
filter_8tap_2d_fn(op, sz, FILTER_8TAP_SMOOTH, smooth)
#define filters_8tap_2d_fn2(op) \
filters_8tap_2d_fn(op, 64) \
filters_8tap_2d_fn(op, 32) \
filters_8tap_2d_fn(op, 16) \
filters_8tap_2d_fn(op, 8) \
filters_8tap_2d_fn(op, 4)
filters_8tap_2d_fn2(put)
filters_8tap_2d_fn2(avg)
#undef filters_8tap_2d_fn2
#undef filters_8tap_2d_fn
#undef filter_8tap_2d_fn
#define filter_8tap_1d_fn(op, sz, f, fname, dir, dvar) \
static void \
op ## _8tap_ ## fname ## _ ## sz ## dir ## _ssse3(uint8_t *dst, \
const uint8_t *src, \
ptrdiff_t dst_stride, \
ptrdiff_t src_stride, \
int h, int mx, \
int my) \
{ \
ff_ ## op ## _8tap_1d_ ## dir ## _ ## sz ## _ssse3(dst, src, \
dst_stride, \
src_stride, h, \
ff_filters_ssse3[f][dvar - 1]); \
}
#define filters_8tap_1d_fn(op, sz, dir, dvar) \
filter_8tap_1d_fn(op, sz, FILTER_8TAP_REGULAR, regular, dir, dvar) \
filter_8tap_1d_fn(op, sz, FILTER_8TAP_SHARP, sharp, dir, dvar) \
filter_8tap_1d_fn(op, sz, FILTER_8TAP_SMOOTH, smooth, dir, dvar)
#define filters_8tap_1d_fn2(op, sz) \
filters_8tap_1d_fn(op, sz, h, mx) \
filters_8tap_1d_fn(op, sz, v, my)
#define filters_8tap_1d_fn3(op) \
filters_8tap_1d_fn2(op, 64) \
filters_8tap_1d_fn2(op, 32) \
filters_8tap_1d_fn2(op, 16) \
filters_8tap_1d_fn2(op, 8) \
filters_8tap_1d_fn2(op, 4)
filters_8tap_1d_fn3(put)
filters_8tap_1d_fn3(avg)
#undef filters_8tap_1d_fn
#undef filters_8tap_1d_fn2
#undef filters_8tap_1d_fn3
#undef filter_8tap_1d_fn
#endif /* HAVE_YASM */
av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp)
{
#if HAVE_YASM
int cpu_flags = av_get_cpu_flags();
#define init_fpel(idx1, idx2, sz, type, opt) \
dsp->mc[idx1][FILTER_8TAP_SMOOTH ][idx2][0][0] = \
dsp->mc[idx1][FILTER_8TAP_REGULAR][idx2][0][0] = \
dsp->mc[idx1][FILTER_8TAP_SHARP ][idx2][0][0] = \
dsp->mc[idx1][FILTER_BILINEAR ][idx2][0][0] = ff_ ## type ## sz ## _ ## opt
#define init_subpel1(idx1, idx2, idxh, idxv, sz, dir, type, opt) \
dsp->mc[idx1][FILTER_8TAP_SMOOTH][idx2][idxh][idxv] = type ## _8tap_smooth_ ## sz ## dir ## _ ## opt; \
dsp->mc[idx1][FILTER_8TAP_REGULAR][idx2][idxh][idxv] = type ## _8tap_regular_ ## sz ## dir ## _ ## opt; \
dsp->mc[idx1][FILTER_8TAP_SHARP][idx2][idxh][idxv] = type ## _8tap_sharp_ ## sz ## dir ## _ ## opt
#define init_subpel2(idx, idxh, idxv, dir, type, opt) \
init_subpel1(0, idx, idxh, idxv, 64, dir, type, opt); \
init_subpel1(1, idx, idxh, idxv, 32, dir, type, opt); \
init_subpel1(2, idx, idxh, idxv, 16, dir, type, opt); \
init_subpel1(3, idx, idxh, idxv, 8, dir, type, opt); \
init_subpel1(4, idx, idxh, idxv, 4, dir, type, opt)
#define init_subpel3(idx, type, opt) \
init_subpel2(idx, 1, 1, hv, type, opt); \
init_subpel2(idx, 0, 1, v, type, opt); \
init_subpel2(idx, 1, 0, h, type, opt)
if (EXTERNAL_MMX(cpu_flags)) {
init_fpel(4, 0, 4, put, mmx);
init_fpel(3, 0, 8, put, mmx);
}
if (EXTERNAL_SSE(cpu_flags)) {
init_fpel(2, 0, 16, put, sse);
init_fpel(1, 0, 32, put, sse);
init_fpel(0, 0, 64, put, sse);
init_fpel(4, 1, 4, avg, sse);
init_fpel(3, 1, 8, avg, sse);
}
if (EXTERNAL_SSE2(cpu_flags)) {
init_fpel(2, 1, 16, avg, sse2);
init_fpel(1, 1, 32, avg, sse2);
init_fpel(0, 1, 64, avg, sse2);
}
if (EXTERNAL_SSSE3(cpu_flags)) {
init_subpel3(0, put, ssse3);
init_subpel3(1, avg, ssse3);
}
#undef init_fpel
#undef init_subpel1
#undef init_subpel2
#undef init_subpel3
#endif /* HAVE_YASM */
}