blob: 9f8b3f3362012b3e15f362627ce781a99c609721 [file] [log] [blame] [edit]
/*
* Copyright (c) 2021, Alliance for Open Media. All rights reserved
*
* This source code is subject to the terms of the BSD 3-Clause Clear License
* and the Alliance for Open Media Patent License 1.0. If the BSD 3-Clause Clear
* License was not distributed with this source code in the LICENSE file, you
* can obtain it at aomedia.org/license/software-license/bsd-3-c-c/. If the
* Alliance for Open Media Patent License 1.0 was not distributed with this
* source code in the PATENTS file, you can obtain it at
* aomedia.org/license/patent-license/.
*/
#ifndef AOM_AV1_COMMON_MV_H_
#define AOM_AV1_COMMON_MV_H_
#include "av1/common/common.h"
#include "av1/common/common_data.h"
#include "aom_dsp/aom_filter.h"
#ifdef __cplusplus
extern "C" {
#endif
#define INVALID_MV 0x80008000
#define GET_MV_RAWPEL(x) (((x) + 3 + ((x) >= 0)) >> 3)
#define GET_MV_SUBPEL(x) ((x)*8)
#define MV_IN_USE_BITS 14
#define MV_UPP (1 << MV_IN_USE_BITS)
#define MV_LOW (-(1 << MV_IN_USE_BITS))
#define MARK_MV_INVALID(mv) \
do { \
((int_mv *)(mv))->as_int = INVALID_MV; \
} while (0);
#define CHECK_MV_EQUAL(x, y) (((x).row == (y).row) && ((x).col == (y).col))
// The motion vector in units of full pixel
typedef struct fullpel_mv {
int16_t row;
int16_t col;
} FULLPEL_MV;
// The motion vector in units of 1/8-pel
typedef struct mv {
int16_t row;
int16_t col;
} MV;
static const MV kZeroMv = { 0, 0 };
static const FULLPEL_MV kZeroFullMv = { 0, 0 };
typedef union int_mv {
uint32_t as_int;
MV as_mv;
FULLPEL_MV as_fullmv;
} int_mv; /* facilitates faster equality tests and copies */
typedef struct mv32 {
int32_t row;
int32_t col;
} MV32;
enum {
MV_PRECISION_8_PEL = 0,
MV_PRECISION_FOUR_PEL = 1,
MV_PRECISION_TWO_PEL = 2,
MV_PRECISION_ONE_PEL = 3,
MV_PRECISION_HALF_PEL = 4,
MV_PRECISION_QTR_PEL = 5,
MV_PRECISION_ONE_EIGHTH_PEL = 6,
NUM_MV_PRECISIONS,
} SENUM1BYTE(MvSubpelPrecision);
typedef struct {
uint8_t num_precisions;
MvSubpelPrecision precision[NUM_MV_PRECISIONS];
} PRECISION_SET;
static const PRECISION_SET av1_mv_precision_sets[2] = {
{ 4,
{ MV_PRECISION_FOUR_PEL, MV_PRECISION_ONE_PEL, MV_PRECISION_HALF_PEL,
MV_PRECISION_ONE_EIGHTH_PEL, NUM_MV_PRECISIONS, NUM_MV_PRECISIONS,
NUM_MV_PRECISIONS } },
{ 4,
{ MV_PRECISION_8_PEL, MV_PRECISION_FOUR_PEL, MV_PRECISION_ONE_PEL,
MV_PRECISION_QTR_PEL, NUM_MV_PRECISIONS, NUM_MV_PRECISIONS,
NUM_MV_PRECISIONS } },
};
#define MAX_NUM_OF_SUPPORTED_PRECISIONS 4
#define NUMBER_OF_PRECISION_SETS 1
#define MV_PREC_DOWN_CONTEXTS 2
#define FLEX_MV_COSTS_SIZE (MAX_NUM_OF_SUPPORTED_PRECISIONS - 1)
#define NUM_MV_PREC_MPP_CONTEXT 3
#define NUM_PB_FLEX_QUALIFIED_MAX_PREC \
((NUM_MV_PRECISIONS) - (MV_PRECISION_HALF_PEL))
// The mv limit for fullpel mvs
typedef struct {
int col_min;
int col_max;
int row_min;
int row_max;
} FullMvLimits;
// The mv limit for subpel mvs
typedef struct {
int col_min;
int col_max;
int row_min;
int row_max;
} SubpelMvLimits;
static AOM_INLINE FULLPEL_MV get_fullmv_from_mv(const MV *subpel_mv) {
const FULLPEL_MV full_mv = { (int16_t)GET_MV_RAWPEL(subpel_mv->row),
(int16_t)GET_MV_RAWPEL(subpel_mv->col) };
return full_mv;
}
#if CONFIG_C071_SUBBLK_WARPMV
static AOM_INLINE void get_phase_from_mv(MV ref_mv, MV *sub_mv_offset,
MvSubpelPrecision precision) {
sub_mv_offset->col = 0;
sub_mv_offset->row = 0;
int col_phase = ref_mv.col - GET_MV_SUBPEL(GET_MV_RAWPEL(ref_mv.col));
int row_phase = ref_mv.row - GET_MV_SUBPEL(GET_MV_RAWPEL(ref_mv.row));
if (precision == MV_PRECISION_QTR_PEL) {
sub_mv_offset->col = (col_phase & 1) ? col_phase : 0;
sub_mv_offset->row = (row_phase & 1) ? row_phase : 0;
} else if (precision == MV_PRECISION_HALF_PEL) {
sub_mv_offset->col = ((col_phase & 1) || (col_phase & 2)) ? col_phase : 0;
sub_mv_offset->row = ((row_phase & 1) || (row_phase & 2)) ? row_phase : 0;
} else if (precision == MV_PRECISION_ONE_PEL) {
sub_mv_offset->col = col_phase;
sub_mv_offset->row = row_phase;
} else {
assert(precision == MV_PRECISION_ONE_EIGHTH_PEL ||
precision < MV_PRECISION_ONE_PEL);
}
}
#endif // CONFIG_C071_SUBBLK_WARPMV
static AOM_INLINE MV get_mv_from_fullmv(const FULLPEL_MV *full_mv) {
const MV subpel_mv = { (int16_t)GET_MV_SUBPEL(full_mv->row),
(int16_t)GET_MV_SUBPEL(full_mv->col) };
return subpel_mv;
}
static AOM_INLINE void convert_fullmv_to_mv(int_mv *mv) {
mv->as_mv = get_mv_from_fullmv(&mv->as_fullmv);
}
#define ABS(x) (((x) >= 0) ? (x) : (-(x)))
// Reduce the precision of the MV to the target precision
// The parameter radix define the step size of the MV .
// For instance, radix = 1 for 1/8th pel, 2 for 1/4-th perl, 4 for 1/2 pel, 8
// for integer pel
static INLINE void lower_mv_precision(MV *mv, MvSubpelPrecision precision) {
const int radix = (1 << (MV_PRECISION_ONE_EIGHTH_PEL - precision));
if (radix == 1) return;
int mod = (mv->row % radix);
if (mod != 0) {
mv->row -= mod;
if (ABS(mod) > (radix >> 1)) {
if (mod > 0) {
mv->row += radix;
} else {
mv->row -= radix;
}
}
mv->row = clamp(mv->row, MV_LOW + radix, MV_UPP - radix);
}
mod = (mv->col % radix);
if (mod != 0) {
mv->col -= mod;
if (ABS(mod) > (radix >> 1)) {
if (mod > 0) {
mv->col += radix;
} else {
mv->col -= radix;
}
}
mv->col = clamp(mv->col, MV_LOW + radix, MV_UPP - radix);
}
}
static INLINE void full_pel_lower_mv_precision(FULLPEL_MV *full_pel_mv,
MvSubpelPrecision precision) {
if (precision >= MV_PRECISION_ONE_PEL) return;
const int radix = (1 << (MV_PRECISION_ONE_PEL - precision));
if (radix == 1) return;
int mod = (full_pel_mv->row % radix);
if (mod != 0) {
full_pel_mv->row -= mod;
if (ABS(mod) > radix / 2) {
if (mod > 0) {
full_pel_mv->row += radix;
} else {
full_pel_mv->row -= radix;
}
}
full_pel_mv->row = clamp(full_pel_mv->row, GET_MV_RAWPEL(MV_LOW) + radix,
GET_MV_RAWPEL(MV_UPP) - radix);
}
mod = (full_pel_mv->col % radix);
if (mod != 0) {
full_pel_mv->col -= mod;
if (ABS(mod) > radix / 2) {
if (mod > 0) {
full_pel_mv->col += radix;
} else {
full_pel_mv->col -= radix;
}
}
full_pel_mv->col = clamp(full_pel_mv->col, GET_MV_RAWPEL(MV_LOW) + radix,
GET_MV_RAWPEL(MV_UPP) - radix);
}
}
static INLINE void full_pel_lower_mv_precision_one_comp(
int *comp_value, MvSubpelPrecision precision, int is_max) {
if (precision >= MV_PRECISION_ONE_PEL) return;
const int radix = (1 << (MV_PRECISION_ONE_PEL - precision));
int value = *comp_value;
int mod = (value % radix);
if (mod != 0) {
if (mod < 0)
value -= mod;
else
value += (radix - ABS(mod));
if (is_max) {
value -= radix;
}
*comp_value = clamp(value, GET_MV_RAWPEL(MV_LOW) + radix,
GET_MV_RAWPEL(MV_UPP) - radix);
}
}
// Calculation precision for warp models
#define WARPEDMODEL_PREC_BITS 16
#define WARPEDMODEL_ROW3HOMO_PREC_BITS 16
#if CONFIG_EXTENDED_WARP_PREDICTION
// Storage precision for warp models
//
// Warp models are initially calculated using WARPEDMODEL_PREC_BITS fractional
// bits. This value is set quite high to reduce rounding error, especially
// during the least-squares process.
//
// However, this precision is far more than is needed for the warp filter and
// during storage, and excessive precision requires more hardware resources
// for little gain. So we reduce the parameters to a lower precision
// of (WARPEDMODEL_PREC_BITS - WARP_PARAM_REDUCE_BITS) after calculation.
//
// Note that the constraints in av1_get_shear_params() imply that the
// non-translational parameters are limited to a range a little wider than
// (-1/4, +1/4), but certainly narrower than (-1/2, +1/2). So they can be safely
// stored in (WARPEDMODEL_PREC_BITS - WARP_PARAM_REDUCE_BITS) bits, including
// the sign bit.
//
// In addition, the translational part of a warp model is clamped, to further
// limit the number of bits required for storage.
//
// The upshot of this is that, to store a single 6-parameter AFFINE warp model,
// hardware requires:
// * (WARPEDMODEL_PREC_BITS - WARP_PARAM_REDUCE_BITS) bits for each of the 4
// non-translational parameters
// * (WARPEDMODEL_PREC_BITS - WARP_PARAM_REDUCE_BITS + WARP_TRANS_INTEGER_BITS)
// bits for each of the 2 translational parameters
//
// for a total of 4 * 10 + 2 * 22 = 84 bits/model
#define WARP_PARAM_REDUCE_BITS 6
#define WARP_TRANS_INTEGER_BITS 12
#else
#define WARP_PARAM_REDUCE_BITS 6
#define WARP_TRANS_INTEGER_BITS 8
#endif // CONFIG_EXTENDED_WARP_PREDICTION
#define WARPEDMODEL_TRANS_CLAMP \
(1 << (WARPEDMODEL_PREC_BITS + WARP_TRANS_INTEGER_BITS - 1))
#define WARPEDMODEL_NONDIAGAFFINE_CLAMP (1 << (WARPEDMODEL_PREC_BITS - 3))
#define WARPEDMODEL_ROW3HOMO_CLAMP (1 << (WARPEDMODEL_PREC_BITS - 2))
// Shift required to convert between warp parameter and MV precision
#define WARPEDMODEL_TO_MV_SHIFT (WARPEDMODEL_PREC_BITS - 3)
// Bits of subpel precision for warped interpolation
#define WARPEDPIXEL_PREC_BITS 6
#define WARPEDPIXEL_PREC_SHIFTS (1 << WARPEDPIXEL_PREC_BITS)
#define WARPEDDIFF_PREC_BITS (WARPEDMODEL_PREC_BITS - WARPEDPIXEL_PREC_BITS)
/* clang-format off */
enum {
IDENTITY = 0, // identity transformation, 0-parameter
TRANSLATION = 1, // translational motion 2-parameter
ROTZOOM = 2, // simplified affine with rotation + zoom only, 4-parameter
AFFINE = 3, // affine, 6-parameter
TRANS_TYPES,
} UENUM1BYTE(TransformationType);
/* clang-format on */
// Number of types used for global motion (must be >= 3 and <= TRANS_TYPES)
// The following can be useful:
// GLOBAL_TRANS_TYPES 3 - up to rotation-zoom
// GLOBAL_TRANS_TYPES 4 - up to affine
// GLOBAL_TRANS_TYPES 6 - up to hor/ver trapezoids
// GLOBAL_TRANS_TYPES 7 - up to full homography
#define GLOBAL_TRANS_TYPES 4
typedef struct {
int global_warp_allowed;
int local_warp_allowed;
} WarpTypesAllowed;
// number of parameters used by each transformation in TransformationTypes
static const int trans_model_params[TRANS_TYPES] = { 0, 2, 4, 6 };
// The order of values in the wmmat matrix below is best described
// by the homography:
// [x' (m2 m3 m0 [x
// z . y' = m4 m5 m1 * y
// 1] m6 m7 1) 1]
typedef struct {
int32_t wmmat[8];
int16_t alpha, beta, gamma, delta;
TransformationType wmtype;
int8_t invalid;
#if CONFIG_EXT_WARP_FILTER
// Flag that indicates whether to use the affine warp filter
// (av1_highbd_warp_affine) or the translational warp filter
// (av1_ext_highbd_warp_affine)
bool use_affine_filter;
#endif // CONFIG_EXT_WARP_FILTER
} WarpedMotionParams;
/* clang-format off */
static const WarpedMotionParams default_warp_params = {
{ 0, 0, (1 << WARPEDMODEL_PREC_BITS), 0, 0, (1 << WARPEDMODEL_PREC_BITS), 0,
0 },
0, 0, 0, 0,
IDENTITY,
0,
#if CONFIG_EXT_WARP_FILTER
true
#endif // CONFIG_EXT_WARP_FILTER
};
/* clang-format on */
// The following constants describe the various precisions
// of different parameters in the global motion experiment.
//
// Given the general homography:
// [x' (a b c [x
// z . y' = d e f * y
// 1] g h i) 1]
//
// Constants using the name ALPHA here are related to parameters
// a, b, d, e. Constants using the name TRANS are related
// to parameters c and f.
//
// Anything ending in PREC_BITS is the number of bits of precision
// to maintain when converting from double to integer.
//
// The ABS parameters are used to create an upper and lower bound
// for each parameter. In other words, after a parameter is integerized
// it is clamped between -(1 << ABS_XXX_BITS) and (1 << ABS_XXX_BITS).
//
// XXX_PREC_DIFF and XXX_DECODE_FACTOR
// are computed once here to prevent repetitive
// computation on the decoder side. These are
// to allow the global motion parameters to be encoded in a lower
// precision than the warped model precision. This means that they
// need to be changed to warped precision when they are decoded.
//
// XX_MIN, XX_MAX are also computed to avoid repeated computation
#define SUBEXPFIN_K 3
#if CONFIG_EXTENDED_WARP_PREDICTION || CONFIG_IMPROVED_GLOBAL_MOTION
#define GM_TRANS_PREC_BITS 3
#define GM_ABS_TRANS_BITS 14
#define GM_ABS_TRANS_ONLY_BITS (GM_ABS_TRANS_BITS - GM_TRANS_PREC_BITS + 3)
#define GM_TRANS_PREC_DIFF (WARPEDMODEL_PREC_BITS - GM_TRANS_PREC_BITS)
#define GM_TRANS_ONLY_PREC_DIFF (WARPEDMODEL_PREC_BITS - 3)
#define GM_TRANS_DECODE_FACTOR (1 << GM_TRANS_PREC_DIFF)
#define GM_TRANS_ONLY_DECODE_FACTOR (1 << GM_TRANS_ONLY_PREC_DIFF)
#define GM_ALPHA_PREC_BITS 10
#if CONFIG_EXT_WARP_FILTER
#define GM_ABS_ALPHA_BITS 9
#else
#if CONFIG_IMPROVED_GLOBAL_MOTION
#define GM_ABS_ALPHA_BITS 8
#else
#define GM_ABS_ALPHA_BITS 7
#endif // CONFIG_IMPROVED_GLOBAL_MOTION
#endif // CONFIG_EXT_WARP_FILTER
#define GM_ALPHA_PREC_DIFF (WARPEDMODEL_PREC_BITS - GM_ALPHA_PREC_BITS)
#define GM_ALPHA_DECODE_FACTOR (1 << GM_ALPHA_PREC_DIFF)
#else
#define GM_TRANS_PREC_BITS 6
#define GM_ABS_TRANS_BITS 12
#define GM_ABS_TRANS_ONLY_BITS (GM_ABS_TRANS_BITS - GM_TRANS_PREC_BITS + 3)
#define GM_TRANS_PREC_DIFF (WARPEDMODEL_PREC_BITS - GM_TRANS_PREC_BITS)
#define GM_TRANS_ONLY_PREC_DIFF (WARPEDMODEL_PREC_BITS - 3)
#define GM_TRANS_DECODE_FACTOR (1 << GM_TRANS_PREC_DIFF)
#define GM_TRANS_ONLY_DECODE_FACTOR (1 << GM_TRANS_ONLY_PREC_DIFF)
#define GM_ALPHA_PREC_BITS 15
#define GM_ABS_ALPHA_BITS 12
#define GM_ALPHA_PREC_DIFF (WARPEDMODEL_PREC_BITS - GM_ALPHA_PREC_BITS)
#define GM_ALPHA_DECODE_FACTOR (1 << GM_ALPHA_PREC_DIFF)
#endif // CONFIG_EXTENDED_WARP_PREDICTION || CONFIG_IMPROVED_GLOBAL_MOTION
#define GM_ROW3HOMO_PREC_BITS 16
#define GM_ABS_ROW3HOMO_BITS 11
#define GM_ROW3HOMO_PREC_DIFF \
(WARPEDMODEL_ROW3HOMO_PREC_BITS - GM_ROW3HOMO_PREC_BITS)
#define GM_ROW3HOMO_DECODE_FACTOR (1 << GM_ROW3HOMO_PREC_DIFF)
#if CONFIG_IMPROVED_GLOBAL_MOTION
#define GM_TRANS_MAX ((1 << GM_ABS_TRANS_BITS) - 1)
#else
#define GM_TRANS_MAX (1 << GM_ABS_TRANS_BITS)
#endif // CONFIG_IMPROVED_GLOBAL_MOTION
#if CONFIG_EXT_WARP_FILTER
#define GM_ALPHA_MAX ((1 << GM_ABS_ALPHA_BITS) - 1)
#else
#define GM_ALPHA_MAX (1 << GM_ABS_ALPHA_BITS)
#endif // CONFIG_EXT_WARP_FILTER
#define GM_ROW3HOMO_MAX (1 << GM_ABS_ROW3HOMO_BITS)
#define GM_TRANS_MIN -GM_TRANS_MAX
#define GM_ALPHA_MIN -GM_ALPHA_MAX
#define GM_ROW3HOMO_MIN -GM_ROW3HOMO_MAX
static INLINE int block_center_x(int mi_col, BLOCK_SIZE bs) {
const int bw = block_size_wide[bs];
return mi_col * MI_SIZE + bw / 2 - 1;
}
static INLINE int block_center_y(int mi_row, BLOCK_SIZE bs) {
const int bh = block_size_high[bs];
return mi_row * MI_SIZE + bh / 2 - 1;
}
static INLINE int convert_to_trans_prec(MvSubpelPrecision precision, int coor) {
if (precision > MV_PRECISION_QTR_PEL)
return ROUND_POWER_OF_TWO_SIGNED(coor, WARPEDMODEL_PREC_BITS - 3);
else
return ROUND_POWER_OF_TWO_SIGNED(coor, WARPEDMODEL_PREC_BITS - 2) * 2;
}
// Returns how many bits do not need to be signaled relative to
// MV_PRECISION_ONE_EIGHTH_PEL
static INLINE int get_gm_precision_loss(MvSubpelPrecision precision) {
// NOTE: there is a bit of an anomaly in AV1 that the translation-only
// global parameters are sent only at 1/4 or 1/8 pel resolution depending
// on whether the allow_high_precision_mv flag is 0 or 1, but the
// cur_frame_force_integer_mv is ignored. Hence the AOMMIN(1, ...)
// below, but in here we correct that so that translation-
// only global parameters are sent at the MV resolution of the frame.
return AOMMIN(1, MV_PRECISION_ONE_EIGHTH_PEL - precision);
}
static INLINE TransformationType get_wmtype(const WarpedMotionParams *model) {
if (model->wmmat[5] == (1 << WARPEDMODEL_PREC_BITS) && !model->wmmat[4] &&
model->wmmat[2] == (1 << WARPEDMODEL_PREC_BITS) && !model->wmmat[3]) {
return ((!model->wmmat[1] && !model->wmmat[0]) ? IDENTITY : TRANSLATION);
}
if (model->wmmat[2] == model->wmmat[5] && model->wmmat[3] == -model->wmmat[4])
return ROTZOOM;
else
return AFFINE;
}
#if CONFIG_EXTENDED_WARP_PREDICTION
// Special value for row_offset and col_offset in the `CANDIDATE_MV` struct,
// to indicate that this motion vector did not come from spatial prediction
// (eg, temporal prediction, or a scaled MV from a nearby block which used
// a different ref frame)
//
// The special value is 0 because the spatial scan area consists of blocks
// both above and left of the current block. Thus valid offsets will always
// have at least one of row_offset and col_offset negative.
#define OFFSET_NONSPATIAL 0
#endif // CONFIG_EXTENDED_WARP_PREDICTION
typedef struct candidate_mv {
int_mv this_mv;
int_mv comp_mv;
#if CONFIG_EXTENDED_WARP_PREDICTION
// Position of the candidate block relative to the current block.
// This is used to decide whether to signal the WARP_EXTEND mode,
// and to fetch the corresponding warp model if that is used
//
// Note(rachelbarker):
// If these are both set to OFFSET_NONSPATIAL, then this is a non-spatial
// candidate, and so does not allow WARP_EXTEND
int row_offset;
int col_offset;
// Record the cwp index of the neighboring blocks
int8_t cwp_idx;
#endif // CONFIG_EXTENDED_WARP_PREDICTION
} CANDIDATE_MV;
#if CONFIG_EXTENDED_WARP_PREDICTION
// structure of the warp-reference-list (WRL)
// Each entry of the WRL contain warp parameter and projection type.
typedef struct warp_candidate {
WarpedMotionParams wm_params;
WarpProjectionType proj_type;
} WARP_CANDIDATE;
#endif // CONFIG_EXTENDED_WARP_PREDICTION
static INLINE int is_zero_mv(const MV *mv) {
return *((const uint32_t *)mv) == 0;
}
static INLINE int is_equal_mv(const MV *a, const MV *b) {
return *((const uint32_t *)a) == *((const uint32_t *)b);
}
static INLINE void clamp_mv(MV *mv, const SubpelMvLimits *mv_limits) {
mv->col = clamp(mv->col, mv_limits->col_min, mv_limits->col_max);
mv->row = clamp(mv->row, mv_limits->row_min, mv_limits->row_max);
}
static INLINE void clamp_fullmv(FULLPEL_MV *mv, const FullMvLimits *mv_limits) {
mv->col = clamp(mv->col, mv_limits->col_min, mv_limits->col_max);
mv->row = clamp(mv->row, mv_limits->row_min, mv_limits->row_max);
}
#ifdef __cplusplus
} // extern "C"
#endif
#endif // AOM_AV1_COMMON_MV_H_