/src/ffmpeg/libswscale/uops.h

Source
/**
 * Copyright (C) 2025 Niklas Haas
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg 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.
 *
 * FFmpeg 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 FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

#ifndef SWSCALE_UOPS_H
#define SWSCALE_UOPS_H

#include <assert.h>
#include <stdbool.h>
#include <stdint.h>

/***************************************************************************
 * Note: This header must be usable at build time, to generate asm sources *
 ***************************************************************************/

#include "libavutil/attributes.h"

typedef struct SwsContext       SwsContext;
typedef struct SwsFilterWeights SwsFilterWeights;
typedef struct SwsLut3D         SwsLut3D;
typedef struct SwsOpList        SwsOpList;

typedef enum SwsPixelType {
    SWS_PIXEL_NONE = 0,
    SWS_PIXEL_U8,
    SWS_PIXEL_U16,
    SWS_PIXEL_U32,
    SWS_PIXEL_F32,
    SWS_PIXEL_TYPE_NB
} SwsPixelType;

const char *ff_sws_pixel_type_name(SwsPixelType type);

static inline av_const int ff_sws_pixel_type_size(SwsPixelType type)
{
    switch (type) {
    case SWS_PIXEL_U8:  return sizeof(uint8_t);
    case SWS_PIXEL_U16: return sizeof(uint16_t);
    case SWS_PIXEL_U32: return sizeof(uint32_t);
    case SWS_PIXEL_F32: return sizeof(float);
    case SWS_PIXEL_NONE: break;
    case SWS_PIXEL_TYPE_NB: break;
    }
    return 0;
}

static inline av_const bool ff_sws_pixel_type_is_int(SwsPixelType type)
{
    switch (type) {
    case SWS_PIXEL_U8:
    case SWS_PIXEL_U16:
    case SWS_PIXEL_U32:
        return true;
    case SWS_PIXEL_F32:
        return false;
    case SWS_PIXEL_NONE:
    case SWS_PIXEL_TYPE_NB: break;
    }
    return false;
}

typedef union SwsPixel {
    char data[4];

    uint8_t  u8;
    uint16_t u16;
    uint32_t u32;
    float    f32;
} SwsPixel;

/* Ensures (SwsPixel) {0} is properly initialized to all zeros */
static_assert(sizeof(SwsPixel) == sizeof(char[4]), "SwsPixel size mismatch");

/**
 * Bit-mask of components. Exact meaning depends on the usage context.
 */
typedef uint8_t SwsCompMask;
enum {
    SWS_COMP_NONE = 0,
    SWS_COMP_ALL  = 0xF,
#define SWS_COMP(X) (1 << (X))
#define SWS_COMP_TEST(mask, X) (!!((mask) & SWS_COMP(X)))
#define SWS_COMP_INV(mask) ((mask) ^ SWS_COMP_ALL)
#define SWS_COMP_ELEMS(N) ((1 << (N)) - 1)
#define SWS_COMP_COUNT(mask) (av_popcount((mask) & SWS_COMP_ALL))
#define SWS_COMP_MASK(X, Y, Z, W)   \
    (((X) ? SWS_COMP(0) : 0) |      \
     ((Y) ? SWS_COMP(1) : 0) |      \
     ((Z) ? SWS_COMP(2) : 0) |      \
     ((W) ? SWS_COMP(3) : 0))
};


#define ff_sws_comp_mask_str(mask) ff_sws_comp_mask_print(mask, (char[5]){0})
static inline char *ff_sws_comp_mask_print(SwsCompMask mask, char buf[5])
{
    char *ptr = buf;
    for (int c = 0; c < 4; c++) {
        if (SWS_COMP_TEST(mask, c))
            *ptr++ = "xyzw"[c];
    }
    *ptr = '\0';
    return buf;
}

typedef uint32_t SwsUOpFlags;
typedef enum SwsUOpFlagBits {
    SWS_UOP_FLAG_NONE   = 0,
    SWS_UOP_FLAG_FMA    = (1 << 0), /* platform supports FMA ops */
    SWS_UOP_FLAG_PSHUFB = (1 << 1), /* platform supports pshufb equivalent */
} SwsUOpFlagBits;

typedef enum SwsUOpType {
    SWS_UOP_INVALID = 0,

    /* Read/write uops; mask = components to read/write */
    SWS_UOP_READ_PLANAR,     /* simple planar byte-aligned read */
    SWS_UOP_READ_PLANAR_FH,  /* planar read with horizontal filter */
    SWS_UOP_READ_PLANAR_FV,  /* planar read with vertical filter */
    SWS_UOP_READ_PLANAR_FV_FMA,
    SWS_UOP_READ_PACKED,     /* simple packed byte-aligned read */
    SWS_UOP_READ_NIBBLE,     /* fractional read (4 bits) from single plane */
    SWS_UOP_READ_BIT,        /* fractional read (1 bit) from single plane */
    SWS_UOP_READ_PALETTE,    /* indexed read from palette in plane 1 */

    SWS_UOP_WRITE_PLANAR,    /* simple planar byte-aligned write */
    SWS_UOP_WRITE_PACKED,    /* simple packed byte-aligned write */
    SWS_UOP_WRITE_NIBBLE,    /* fractional write (4 bits) to single plane */
    SWS_UOP_WRITE_BIT,       /* fractional write (1 bit) to single plane */

    /* Packed shuffle / gather uops */
    SWS_UOP_RW_SHUFFLE,      /* in-place (packed) indexed shuffle/gather */

    /* Data rearrangement uops; mask = needed or trivial components */
    SWS_UOP_PERMUTE,         /* permute pointers (no duplicates) */
    SWS_UOP_COPY,            /* permute data (may contain duplicates) */

    /* Data conversion / manipulation uops; mask = affected components */
    SWS_UOP_SWAP_BYTES,      /* swap byte order in components */
    SWS_UOP_EXPAND_BIT,      /* expand low-order bit to all bits in type */
    SWS_UOP_EXPAND_PAIR,     /* expand bytes in pairs (16 bit) */
    SWS_UOP_EXPAND_QUAD,     /* expand bytes in quads (32 bit) */
    SWS_UOP_TO_U8,           /* cast pixel values to SWS_PIXEL_U8  */
    SWS_UOP_TO_U16,          /* cast pixel values to SWS_PIXEL_U16 */
    SWS_UOP_TO_U32,          /* cast pixel values to SWS_PIXEL_U32 */
    SWS_UOP_TO_F32,          /* cast pixel values to SWS_PIXEL_F32 */

    /* Arithmetic uops */
    SWS_UOP_SCALE,           /* multiply masked components by scalar */
    SWS_UOP_ADD,             /* add vec4 to masked components */
    SWS_UOP_MIN,             /* min(x, vec4) on masked components */
    SWS_UOP_MAX,             /* max(x, vec4) on masked components */

    /* Identical to corresponding SwsOpType */
    SWS_UOP_UNPACK,          /* mask = nonzero components in pack pattern */
    SWS_UOP_PACK,            /* mask = nonzero components in pack pattern */
    SWS_UOP_LSHIFT,          /* mask = components to shift */
    SWS_UOP_RSHIFT,          /* mask = components to shift */
    SWS_UOP_CLEAR,           /* mask = components to clear */
    SWS_UOP_LINEAR,          /* mask = non-trivial output rows */
    SWS_UOP_LINEAR_FMA,      /* with SWS_UOP_FLAG_FMA */
    SWS_UOP_DITHER,          /* mask = components to dither */
    SWS_UOP_LUT_3D,          /* mask = needed output components */

    /* Platform-specific uops would go here */
    SWS_UOP_TYPE_NB,
} SwsUOpType;

typedef struct SwsShuffleUOp {
    uint8_t clear_value; /* value to clear elements with negative indices to */
    uint8_t read_size;   /* input bytes per iteration */
    uint8_t write_size;  /* output bytes per iteration */
} SwsShuffleUOp;

typedef struct SwsShuffleMask {
    int8_t mask[16];    /* shuffle index mask, or -1 to clear bytes (to `clear_value`) */
    uint8_t pixels;     /* number of pixels per iteration */
} SwsShuffleMask;

typedef struct SwsFilterUOp {
    SwsPixelType type; /* pixel type to store result as */
} SwsFilterUOp;

typedef struct SwsShiftUOp {
    uint8_t amount;
} SwsShiftUOp;

typedef struct SwsMoveUOp {
    /* The worst case number of moves (for two independent cycles) */
    #define SWS_UOP_MOVE_MAX 6
    int num_moves;

    /* This may involve a temporary register (index -1) */
    int8_t dst[SWS_UOP_MOVE_MAX]; /* destination register index */
    int8_t src[SWS_UOP_MOVE_MAX]; /* source register index */
} SwsMoveUOp;

typedef struct SwsPackUOp {
    uint8_t pattern[4]; /* bit depth pattern, from MSB to LSB */
} SwsPackUOp;

typedef struct SwsClearUOp {
    SwsCompMask one;  /* mask of coefficients equal to all 1s */
    SwsCompMask zero; /* mask of coefficients equal to all 0s */
} SwsClearUOp;

typedef struct SwsLinearUOp {
    uint32_t one;  /* mask of coefficients equal to one */
    uint32_t zero; /* mask of coefficients equal to zero */

    /* for SWS_UOP_LINEAR_FMA only */
    uint32_t exact; /* mask of coefficients whose product is exact */
} SwsLinearUOp;

#define SWS_MASK(I, J)  (1 << (5 * (I) + (J)))
#define SWS_MASK_OFF(I) SWS_MASK(I, 4)
#define SWS_MASK_ROW(I) (0x1F << (5 * (I)))
#define SWS_MASK_COL(J) (0x8421 << J)
#define SWS_MASK_DIAG4  (0x41041)

typedef struct SwsDitherUOp {
    uint8_t y_offset[4];
    uint8_t size_log2;
} SwsDitherUOp;

typedef struct SwsLut3DUOp {
    int dynamic;
} SwsLut3DUOp;

/**
 * Computes (1 << size_log2) + MAX(y_offset). The dither matrix attached to
 * the SwsUOp is always pre-padded to this number of lines.
 */
int ff_sws_dither_height(const SwsDitherUOp *dither);

typedef union SwsUOpParams {
    SwsShuffleUOp   shuffle; /* for SWS_UOP_RW_SHUFFLE */
    SwsFilterUOp    filter;  /* for SWS_UOP_READ_*_FV/FH */
    SwsShiftUOp     shift;
    SwsMoveUOp      move; /* for SWS_UOP_PERMUTE and SWS_UOP_COPY */
    SwsPackUOp      pack;
    SwsClearUOp     clear;
    SwsLinearUOp    lin;
    SwsDitherUOp    dither;
    SwsLut3DUOp     lut3d;
} SwsUOpParams;

typedef struct SwsUOp {
    /* These fields uniquely identify the uop implementation */
    SwsPixelType type;
    SwsUOpType uop;
    SwsCompMask mask;
    SwsUOpParams par;

    /* Constant data for this uop; not part of the unique identifier */
    union {
        SwsFilterWeights *kernel;   /* refstruct */
        SwsPixel *ptr;              /* refstruct */
        SwsPixel scalar;
        SwsPixel vec4[4];
        SwsPixel mat4[4][5];        /* row major */
        SwsShuffleMask shuffle;     /* for SWS_UOP_RW_SHUFFLE */
        const SwsLut3D *lut3d;      /* for SWS_UOP_LUT_3D; refstruct */
        void *opaque;               /* reserved for internal use */
    } data;
} SwsUOp;

/**
 * Compare two SwsUOps for equality (excluding constant data).
 */
int ff_sws_uop_cmp(const SwsUOp *a, const SwsUOp *b);

static inline int ff_sws_uop_cmp_v(const void *a, const void *b)
{
    return ff_sws_uop_cmp(a, b);
}

/**
 * Generate a unique name for a SwsUOp.
 */
#define SWS_UOP_NAME_MAX 64
void ff_sws_uop_name(const SwsUOp *op, char buf[SWS_UOP_NAME_MAX]);

typedef struct SwsUOpList {
    SwsUOp *ops;
    int num_ops;

    /* Additional metadata for implementations */
    SwsCompMask planes_in;  /* mask of planes read from */
    SwsCompMask planes_out; /* mask of planes written to */
    int pixel_size_max;     /* size of largest pixel type seen in any uop */
} SwsUOpList;

SwsUOpList *ff_sws_uop_list_alloc(void);
void ff_sws_uop_list_free(SwsUOpList **ops);
void ff_sws_uop_list_remove_at(SwsUOpList *uops, int index, int count);

/* Takes over ownership of `uop` and sets it to {0}, even on failure. */
int ff_sws_uop_list_append(SwsUOpList *uops, SwsUOp *uop);

/**
 * Called internally by ff_sws_ops_translate().
 */
int ff_sws_uop_list_optimize(SwsContext *ctx, SwsUOpFlags flags, SwsUOpList *uops);

/**
 * Translate a list of operations down to micro-ops, which can be further
 * optimized and then directly executed by backends.
 *
 * Return 0 or a negative error code.
 */
int ff_sws_ops_translate(SwsContext *ctx, const SwsOpList *ops,
                         SwsUOpFlags flags, SwsUOpList *uops);

/**
 * Compute a shuffle mask for `pshufb`-style ASM functions, by repeating
 * the shuffle pattern for as many groups as will fit.
 *
 * @param uop         An operation of type SWS_UOP_RW_SHUFFLE.
 * @param shuffle     The output shuffle index mask (or -1 to clear bytes).
 * @param size        The maximum size (in bytes) of the output shuffle mask.
 *
 * @return the number of groups on success, or a negative error code.
 *
 * @note The shuffle mask is already pre-expanded to fill up to 16 bytes,
 *       so this is only needed for larger shuffle instructions (e.g. vpermb).
 */
int ff_sws_shuffle_mask(const SwsUOp *uop, int8_t shuffle[], int size);

#endif

Coverage Report

Created: 2026-08-17 07:50

Line	Count	Source
1		/**
2		* Copyright (C) 2025 Niklas Haas
3		*
4		* This file is part of FFmpeg.
5		*
6		* FFmpeg is free software; you can redistribute it and/or
7		* modify it under the terms of the GNU Lesser General Public
8		* License as published by the Free Software Foundation; either
9		* version 2.1 of the License, or (at your option) any later version.
10		*
11		* FFmpeg is distributed in the hope that it will be useful,
12		* but WITHOUT ANY WARRANTY; without even the implied warranty of
13		* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14		* Lesser General Public License for more details.
15		*
16		* You should have received a copy of the GNU Lesser General Public
17		* License along with FFmpeg; if not, write to the Free Software
18		* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19		*/
20
21		#ifndef SWSCALE_UOPS_H
22		#define SWSCALE_UOPS_H
23
24		#include <assert.h>
25		#include <stdbool.h>
26		#include <stdint.h>
27
28		/***************************************************************************
29		* Note: This header must be usable at build time, to generate asm sources *
30		***************************************************************************/
31
32		#include "libavutil/attributes.h"
33
34		typedef struct SwsContext SwsContext;
35		typedef struct SwsFilterWeights SwsFilterWeights;
36		typedef struct SwsLut3D SwsLut3D;
37		typedef struct SwsOpList SwsOpList;
38
39		typedef enum SwsPixelType {
40		SWS_PIXEL_NONE = 0,
41		SWS_PIXEL_U8,
42		SWS_PIXEL_U16,
43		SWS_PIXEL_U32,
44		SWS_PIXEL_F32,
45		SWS_PIXEL_TYPE_NB
46		} SwsPixelType;
47
48		const char *ff_sws_pixel_type_name(SwsPixelType type);
49
50		static inline av_const int ff_sws_pixel_type_size(SwsPixelType type)
51	0	{
52	0	switch (type) {
53	0	case SWS_PIXEL_U8: return sizeof(uint8_t);
54	0	case SWS_PIXEL_U16: return sizeof(uint16_t);
55	0	case SWS_PIXEL_U32: return sizeof(uint32_t);
56	0	case SWS_PIXEL_F32: return sizeof(float);
57	0	case SWS_PIXEL_NONE: break;
58	0	case SWS_PIXEL_TYPE_NB: break;
59	0	}
60	0	return 0;
61	0	} Unexecuted instantiation: format.c:ff_sws_pixel_type_size Unexecuted instantiation: graph.c:ff_sws_pixel_type_size Unexecuted instantiation: ops.c:ff_sws_pixel_type_size Unexecuted instantiation: ops_dispatch.c:ff_sws_pixel_type_size Unexecuted instantiation: ops_memcpy.c:ff_sws_pixel_type_size Unexecuted instantiation: ops_optimizer.c:ff_sws_pixel_type_size Unexecuted instantiation: uops.c:ff_sws_pixel_type_size Unexecuted instantiation: uops_backend.c:ff_sws_pixel_type_size Unexecuted instantiation: ops_chain.c:ff_sws_pixel_type_size
62
63		static inline av_const bool ff_sws_pixel_type_is_int(SwsPixelType type)
64	0	{
65	0	switch (type) {
66	0	case SWS_PIXEL_U8:
67	0	case SWS_PIXEL_U16:
68	0	case SWS_PIXEL_U32:
69	0	return true;
70	0	case SWS_PIXEL_F32:
71	0	return false;
72	0	case SWS_PIXEL_NONE:
73	0	case SWS_PIXEL_TYPE_NB: break;
74	0	}
75	0	return false;
76	0	} Unexecuted instantiation: format.c:ff_sws_pixel_type_is_int Unexecuted instantiation: graph.c:ff_sws_pixel_type_is_int Unexecuted instantiation: ops.c:ff_sws_pixel_type_is_int Unexecuted instantiation: ops_dispatch.c:ff_sws_pixel_type_is_int Unexecuted instantiation: ops_memcpy.c:ff_sws_pixel_type_is_int Unexecuted instantiation: ops_optimizer.c:ff_sws_pixel_type_is_int Unexecuted instantiation: uops.c:ff_sws_pixel_type_is_int Unexecuted instantiation: uops_backend.c:ff_sws_pixel_type_is_int Unexecuted instantiation: ops_chain.c:ff_sws_pixel_type_is_int
77
78		typedef union SwsPixel {
79		char data[4];
80
81		uint8_t u8;
82		uint16_t u16;
83		uint32_t u32;
84		float f32;
85		} SwsPixel;
86
87		/* Ensures (SwsPixel) {0} is properly initialized to all zeros */
88		static_assert(sizeof(SwsPixel) == sizeof(char[4]), "SwsPixel size mismatch");
89
90		/**
91		* Bit-mask of components. Exact meaning depends on the usage context.
92		*/
93		typedef uint8_t SwsCompMask;
94		enum {
95		SWS_COMP_NONE = 0,
96		SWS_COMP_ALL = 0xF,
97	0	#define SWS_COMP(X) (1 << (X))
98	0	#define SWS_COMP_TEST(mask, X) (!!((mask) & SWS_COMP(X)))
99		#define SWS_COMP_INV(mask) ((mask) ^ SWS_COMP_ALL)
100	0	#define SWS_COMP_ELEMS(N) ((1 << (N)) - 1)
101		#define SWS_COMP_COUNT(mask) (av_popcount((mask) & SWS_COMP_ALL))
102		#define SWS_COMP_MASK(X, Y, Z, W) \
103	0	(((X) ? SWS_COMP(0) : 0) \| \
104	0	((Y) ? SWS_COMP(1) : 0) \| \
105	0	((Z) ? SWS_COMP(2) : 0) \| \
106	0	((W) ? SWS_COMP(3) : 0))
107		};
108
109
110	0	#define ff_sws_comp_mask_str(mask) ff_sws_comp_mask_print(mask, (char[5]){0})
111		static inline char *ff_sws_comp_mask_print(SwsCompMask mask, char buf[5])
112	0	{
113	0	char *ptr = buf;
114	0	for (int c = 0; c < 4; c++) {
115	0	if (SWS_COMP_TEST(mask, c))
116	0	*ptr++ = "xyzw"[c];
117	0	}
118	0	*ptr = '\0';
119	0	return buf;
120	0	} Unexecuted instantiation: format.c:ff_sws_comp_mask_print Unexecuted instantiation: graph.c:ff_sws_comp_mask_print Unexecuted instantiation: ops.c:ff_sws_comp_mask_print Unexecuted instantiation: ops_dispatch.c:ff_sws_comp_mask_print Unexecuted instantiation: ops_memcpy.c:ff_sws_comp_mask_print Unexecuted instantiation: ops_optimizer.c:ff_sws_comp_mask_print Unexecuted instantiation: uops.c:ff_sws_comp_mask_print Unexecuted instantiation: uops_backend.c:ff_sws_comp_mask_print Unexecuted instantiation: ops_chain.c:ff_sws_comp_mask_print
121
122		typedef uint32_t SwsUOpFlags;
123		typedef enum SwsUOpFlagBits {
124		SWS_UOP_FLAG_NONE = 0,
125		SWS_UOP_FLAG_FMA = (1 << 0), /* platform supports FMA ops */
126		SWS_UOP_FLAG_PSHUFB = (1 << 1), /* platform supports pshufb equivalent */
127		} SwsUOpFlagBits;
128
129		typedef enum SwsUOpType {
130		SWS_UOP_INVALID = 0,
131
132		/* Read/write uops; mask = components to read/write */
133		SWS_UOP_READ_PLANAR, /* simple planar byte-aligned read */
134		SWS_UOP_READ_PLANAR_FH, /* planar read with horizontal filter */
135		SWS_UOP_READ_PLANAR_FV, /* planar read with vertical filter */
136		SWS_UOP_READ_PLANAR_FV_FMA,
137		SWS_UOP_READ_PACKED, /* simple packed byte-aligned read */
138		SWS_UOP_READ_NIBBLE, /* fractional read (4 bits) from single plane */
139		SWS_UOP_READ_BIT, /* fractional read (1 bit) from single plane */
140		SWS_UOP_READ_PALETTE, /* indexed read from palette in plane 1 */
141
142		SWS_UOP_WRITE_PLANAR, /* simple planar byte-aligned write */
143		SWS_UOP_WRITE_PACKED, /* simple packed byte-aligned write */
144		SWS_UOP_WRITE_NIBBLE, /* fractional write (4 bits) to single plane */
145		SWS_UOP_WRITE_BIT, /* fractional write (1 bit) to single plane */
146
147		/* Packed shuffle / gather uops */
148		SWS_UOP_RW_SHUFFLE, /* in-place (packed) indexed shuffle/gather */
149
150		/* Data rearrangement uops; mask = needed or trivial components */
151		SWS_UOP_PERMUTE, /* permute pointers (no duplicates) */
152		SWS_UOP_COPY, /* permute data (may contain duplicates) */
153
154		/* Data conversion / manipulation uops; mask = affected components */
155		SWS_UOP_SWAP_BYTES, /* swap byte order in components */
156		SWS_UOP_EXPAND_BIT, /* expand low-order bit to all bits in type */
157		SWS_UOP_EXPAND_PAIR, /* expand bytes in pairs (16 bit) */
158		SWS_UOP_EXPAND_QUAD, /* expand bytes in quads (32 bit) */
159		SWS_UOP_TO_U8, /* cast pixel values to SWS_PIXEL_U8 */
160		SWS_UOP_TO_U16, /* cast pixel values to SWS_PIXEL_U16 */
161		SWS_UOP_TO_U32, /* cast pixel values to SWS_PIXEL_U32 */
162		SWS_UOP_TO_F32, /* cast pixel values to SWS_PIXEL_F32 */
163
164		/* Arithmetic uops */
165		SWS_UOP_SCALE, /* multiply masked components by scalar */
166		SWS_UOP_ADD, /* add vec4 to masked components */
167		SWS_UOP_MIN, /* min(x, vec4) on masked components */
168		SWS_UOP_MAX, /* max(x, vec4) on masked components */
169
170		/* Identical to corresponding SwsOpType */
171		SWS_UOP_UNPACK, /* mask = nonzero components in pack pattern */
172		SWS_UOP_PACK, /* mask = nonzero components in pack pattern */
173		SWS_UOP_LSHIFT, /* mask = components to shift */
174		SWS_UOP_RSHIFT, /* mask = components to shift */
175		SWS_UOP_CLEAR, /* mask = components to clear */
176		SWS_UOP_LINEAR, /* mask = non-trivial output rows */
177		SWS_UOP_LINEAR_FMA, /* with SWS_UOP_FLAG_FMA */
178		SWS_UOP_DITHER, /* mask = components to dither */
179		SWS_UOP_LUT_3D, /* mask = needed output components */
180
181		/* Platform-specific uops would go here */
182		SWS_UOP_TYPE_NB,
183		} SwsUOpType;
184
185		typedef struct SwsShuffleUOp {
186		uint8_t clear_value; /* value to clear elements with negative indices to */
187		uint8_t read_size; /* input bytes per iteration */
188		uint8_t write_size; /* output bytes per iteration */
189		} SwsShuffleUOp;
190
191		typedef struct SwsShuffleMask {
192		int8_t mask[16]; /* shuffle index mask, or -1 to clear bytes (to `clear_value`) */
193		uint8_t pixels; /* number of pixels per iteration */
194		} SwsShuffleMask;
195
196		typedef struct SwsFilterUOp {
197		SwsPixelType type; /* pixel type to store result as */
198		} SwsFilterUOp;
199
200		typedef struct SwsShiftUOp {
201		uint8_t amount;
202		} SwsShiftUOp;
203
204		typedef struct SwsMoveUOp {
205		/* The worst case number of moves (for two independent cycles) */
206		#define SWS_UOP_MOVE_MAX 6
207		int num_moves;
208
209		/* This may involve a temporary register (index -1) */
210		int8_t dst[SWS_UOP_MOVE_MAX]; /* destination register index */
211		int8_t src[SWS_UOP_MOVE_MAX]; /* source register index */
212		} SwsMoveUOp;
213
214		typedef struct SwsPackUOp {
215		uint8_t pattern[4]; /* bit depth pattern, from MSB to LSB */
216		} SwsPackUOp;
217
218		typedef struct SwsClearUOp {
219		SwsCompMask one; /* mask of coefficients equal to all 1s */
220		SwsCompMask zero; /* mask of coefficients equal to all 0s */
221		} SwsClearUOp;
222
223		typedef struct SwsLinearUOp {
224		uint32_t one; /* mask of coefficients equal to one */
225		uint32_t zero; /* mask of coefficients equal to zero */
226
227		/* for SWS_UOP_LINEAR_FMA only */
228		uint32_t exact; /* mask of coefficients whose product is exact */
229		} SwsLinearUOp;
230
231	0	#define SWS_MASK(I, J) (1 << (5 * (I) + (J)))
232		#define SWS_MASK_OFF(I) SWS_MASK(I, 4)
233	0	#define SWS_MASK_ROW(I) (0x1F << (5 * (I)))
234	0	#define SWS_MASK_COL(J) (0x8421 << J)
235	0	#define SWS_MASK_DIAG4 (0x41041)
236
237		typedef struct SwsDitherUOp {
238		uint8_t y_offset[4];
239		uint8_t size_log2;
240		} SwsDitherUOp;
241
242		typedef struct SwsLut3DUOp {
243		int dynamic;
244		} SwsLut3DUOp;
245
246		/**
247		* Computes (1 << size_log2) + MAX(y_offset). The dither matrix attached to
248		* the SwsUOp is always pre-padded to this number of lines.
249		*/
250		int ff_sws_dither_height(const SwsDitherUOp *dither);
251
252		typedef union SwsUOpParams {
253		SwsShuffleUOp shuffle; /* for SWS_UOP_RW_SHUFFLE */
254		SwsFilterUOp filter; /* for SWS_UOP_READ__FV/FH /
255		SwsShiftUOp shift;
256		SwsMoveUOp move; /* for SWS_UOP_PERMUTE and SWS_UOP_COPY */
257		SwsPackUOp pack;
258		SwsClearUOp clear;
259		SwsLinearUOp lin;
260		SwsDitherUOp dither;
261		SwsLut3DUOp lut3d;
262		} SwsUOpParams;
263
264		typedef struct SwsUOp {
265		/* These fields uniquely identify the uop implementation */
266		SwsPixelType type;
267		SwsUOpType uop;
268		SwsCompMask mask;
269		SwsUOpParams par;
270
271		/* Constant data for this uop; not part of the unique identifier */
272		union {
273		SwsFilterWeights kernel; / refstruct */
274		SwsPixel ptr; / refstruct */
275		SwsPixel scalar;
276		SwsPixel vec4[4];
277		SwsPixel mat4[4][5]; /* row major */
278		SwsShuffleMask shuffle; /* for SWS_UOP_RW_SHUFFLE */
279		const SwsLut3D lut3d; / for SWS_UOP_LUT_3D; refstruct */
280		void opaque; / reserved for internal use */
281		} data;
282		} SwsUOp;
283
284		/**
285		* Compare two SwsUOps for equality (excluding constant data).
286		*/
287		int ff_sws_uop_cmp(const SwsUOp a, const SwsUOp b);
288
289		static inline int ff_sws_uop_cmp_v(const void a, const void b)
290	0	{
291	0	return ff_sws_uop_cmp(a, b);
292	0	} Unexecuted instantiation: format.c:ff_sws_uop_cmp_v Unexecuted instantiation: graph.c:ff_sws_uop_cmp_v Unexecuted instantiation: ops.c:ff_sws_uop_cmp_v Unexecuted instantiation: ops_dispatch.c:ff_sws_uop_cmp_v Unexecuted instantiation: ops_memcpy.c:ff_sws_uop_cmp_v Unexecuted instantiation: ops_optimizer.c:ff_sws_uop_cmp_v Unexecuted instantiation: uops.c:ff_sws_uop_cmp_v Unexecuted instantiation: uops_backend.c:ff_sws_uop_cmp_v Unexecuted instantiation: ops_chain.c:ff_sws_uop_cmp_v
293
294		/**
295		* Generate a unique name for a SwsUOp.
296		*/
297	0	#define SWS_UOP_NAME_MAX 64
298		void ff_sws_uop_name(const SwsUOp *op, char buf[SWS_UOP_NAME_MAX]);
299
300		typedef struct SwsUOpList {
301		SwsUOp *ops;
302		int num_ops;
303
304		/* Additional metadata for implementations */
305		SwsCompMask planes_in; /* mask of planes read from */
306		SwsCompMask planes_out; /* mask of planes written to */
307		int pixel_size_max; /* size of largest pixel type seen in any uop */
308		} SwsUOpList;
309
310		SwsUOpList *ff_sws_uop_list_alloc(void);
311		void ff_sws_uop_list_free(SwsUOpList **ops);
312		void ff_sws_uop_list_remove_at(SwsUOpList *uops, int index, int count);
313
314		/* Takes over ownership of `uop` and sets it to {0}, even on failure. */
315		int ff_sws_uop_list_append(SwsUOpList uops, SwsUOp uop);
316
317		/**
318		* Called internally by ff_sws_ops_translate().
319		*/
320		int ff_sws_uop_list_optimize(SwsContext ctx, SwsUOpFlags flags, SwsUOpList uops);
321
322		/**
323		* Translate a list of operations down to micro-ops, which can be further
324		* optimized and then directly executed by backends.
325		*
326		* Return 0 or a negative error code.
327		*/
328		int ff_sws_ops_translate(SwsContext ctx, const SwsOpList ops,
329		SwsUOpFlags flags, SwsUOpList *uops);
330
331		/**
332		* Compute a shuffle mask for `pshufb`-style ASM functions, by repeating
333		* the shuffle pattern for as many groups as will fit.
334		*
335		* @param uop An operation of type SWS_UOP_RW_SHUFFLE.
336		* @param shuffle The output shuffle index mask (or -1 to clear bytes).
337		* @param size The maximum size (in bytes) of the output shuffle mask.
338		*
339		* @return the number of groups on success, or a negative error code.
340		*
341		* @note The shuffle mask is already pre-expanded to fill up to 16 bytes,
342		* so this is only needed for larger shuffle instructions (e.g. vpermb).
343		*/
344		int ff_sws_shuffle_mask(const SwsUOp *uop, int8_t shuffle[], int size);
345
346		#endif