yading@10: /*
yading@10:  * Copyright (c) 2010 Stefano Sabatini
yading@10:  * Copyright (c) 2010 Baptiste Coudurier
yading@10:  * Copyright (c) 2007 Bobby Bingham
yading@10:  *
yading@10:  * This file is part of FFmpeg.
yading@10:  *
yading@10:  * FFmpeg is free software; you can redistribute it and/or
yading@10:  * modify it under the terms of the GNU Lesser General Public
yading@10:  * License as published by the Free Software Foundation; either
yading@10:  * version 2.1 of the License, or (at your option) any later version.
yading@10:  *
yading@10:  * FFmpeg is distributed in the hope that it will be useful,
yading@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
yading@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
yading@10:  * Lesser General Public License for more details.
yading@10:  *
yading@10:  * You should have received a copy of the GNU Lesser General Public
yading@10:  * License along with FFmpeg; if not, write to the Free Software
yading@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
yading@10:  */
yading@10: 
yading@10: /**
yading@10:  * @file
yading@10:  * overlay one video on top of another
yading@10:  */
yading@10: 
yading@10: /* #define DEBUG */
yading@10: 
yading@10: #include "avfilter.h"
yading@10: #include "formats.h"
yading@10: #include "libavutil/common.h"
yading@10: #include "libavutil/eval.h"
yading@10: #include "libavutil/avstring.h"
yading@10: #include "libavutil/opt.h"
yading@10: #include "libavutil/pixdesc.h"
yading@10: #include "libavutil/imgutils.h"
yading@10: #include "libavutil/mathematics.h"
yading@10: #include "libavutil/opt.h"
yading@10: #include "libavutil/timestamp.h"
yading@10: #include "internal.h"
yading@10: #include "bufferqueue.h"
yading@10: #include "drawutils.h"
yading@10: #include "video.h"
yading@10: 
yading@10: static const char *const var_names[] = {
yading@10:     "main_w",    "W", ///< width  of the main    video
yading@10:     "main_h",    "H", ///< height of the main    video
yading@10:     "overlay_w", "w", ///< width  of the overlay video
yading@10:     "overlay_h", "h", ///< height of the overlay video
yading@10:     "hsub",
yading@10:     "vsub",
yading@10:     "x",
yading@10:     "y",
yading@10:     "n",            ///< number of frame
yading@10:     "pos",          ///< position in the file
yading@10:     "t",            ///< timestamp expressed in seconds
yading@10:     NULL
yading@10: };
yading@10: 
yading@10: enum var_name {
yading@10:     VAR_MAIN_W,    VAR_MW,
yading@10:     VAR_MAIN_H,    VAR_MH,
yading@10:     VAR_OVERLAY_W, VAR_OW,
yading@10:     VAR_OVERLAY_H, VAR_OH,
yading@10:     VAR_HSUB,
yading@10:     VAR_VSUB,
yading@10:     VAR_X,
yading@10:     VAR_Y,
yading@10:     VAR_N,
yading@10:     VAR_POS,
yading@10:     VAR_T,
yading@10:     VAR_VARS_NB
yading@10: };
yading@10: 
yading@10: #define MAIN    0
yading@10: #define OVERLAY 1
yading@10: 
yading@10: #define R 0
yading@10: #define G 1
yading@10: #define B 2
yading@10: #define A 3
yading@10: 
yading@10: #define Y 0
yading@10: #define U 1
yading@10: #define V 2
yading@10: 
yading@10: typedef struct {
yading@10:     const AVClass *class;
yading@10:     int x, y;                   ///< position of overlayed picture
yading@10:     double enable;              ///< tells if blending is enabled
yading@10: 
yading@10:     int allow_packed_rgb;
yading@10:     uint8_t frame_requested;
yading@10:     uint8_t overlay_eof;
yading@10:     uint8_t main_is_packed_rgb;
yading@10:     uint8_t main_rgba_map[4];
yading@10:     uint8_t main_has_alpha;
yading@10:     uint8_t overlay_is_packed_rgb;
yading@10:     uint8_t overlay_rgba_map[4];
yading@10:     uint8_t overlay_has_alpha;
yading@10:     enum OverlayFormat { OVERLAY_FORMAT_YUV420, OVERLAY_FORMAT_YUV444, OVERLAY_FORMAT_RGB, OVERLAY_FORMAT_NB} format;
yading@10:     enum EvalMode { EVAL_MODE_INIT, EVAL_MODE_FRAME, EVAL_MODE_NB } eval_mode;
yading@10: 
yading@10:     AVFrame *overpicref;
yading@10:     struct FFBufQueue queue_main;
yading@10:     struct FFBufQueue queue_over;
yading@10: 
yading@10:     int main_pix_step[4];       ///< steps per pixel for each plane of the main output
yading@10:     int overlay_pix_step[4];    ///< steps per pixel for each plane of the overlay
yading@10:     int hsub, vsub;             ///< chroma subsampling values
yading@10:     int shortest;               ///< terminate stream when the shortest input terminates
yading@10:     int repeatlast;             ///< repeat last overlay frame
yading@10: 
yading@10:     double var_values[VAR_VARS_NB];
yading@10:     char *x_expr, *y_expr;
yading@10:     char *enable_expr;
yading@10:     AVExpr *x_pexpr, *y_pexpr, *enable_pexpr;
yading@10: } OverlayContext;
yading@10: 
yading@10: static av_cold int init(AVFilterContext *ctx)
yading@10: {
yading@10:     OverlayContext *over = ctx->priv;
yading@10: 
yading@10:     if (over->allow_packed_rgb) {
yading@10:         av_log(ctx, AV_LOG_WARNING,
yading@10:                "The rgb option is deprecated and is overriding the format option, use format instead\n");
yading@10:         over->format = OVERLAY_FORMAT_RGB;
yading@10:     }
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: static av_cold void uninit(AVFilterContext *ctx)
yading@10: {
yading@10:     OverlayContext *over = ctx->priv;
yading@10: 
yading@10:     av_frame_free(&over->overpicref);
yading@10:     ff_bufqueue_discard_all(&over->queue_main);
yading@10:     ff_bufqueue_discard_all(&over->queue_over);
yading@10:     av_expr_free(over->x_pexpr); over->x_pexpr = NULL;
yading@10:     av_expr_free(over->y_pexpr); over->y_pexpr = NULL;
yading@10:     av_expr_free(over->enable_pexpr); over->enable_pexpr = NULL;
yading@10: }
yading@10: 
yading@10: static inline int normalize_xy(double d, int chroma_sub)
yading@10: {
yading@10:     if (isnan(d))
yading@10:         return INT_MAX;
yading@10:     return (int)d & ~((1 << chroma_sub) - 1);
yading@10: }
yading@10: 
yading@10: enum EvalTarget { EVAL_XY, EVAL_ENABLE, EVAL_ALL };
yading@10: 
yading@10: static void eval_expr(AVFilterContext *ctx, enum EvalTarget eval_tgt)
yading@10: {
yading@10:     OverlayContext  *over = ctx->priv;
yading@10: 
yading@10:     if (eval_tgt == EVAL_XY || eval_tgt == EVAL_ALL) {
yading@10:         over->var_values[VAR_X] = av_expr_eval(over->x_pexpr, over->var_values, NULL);
yading@10:         over->var_values[VAR_Y] = av_expr_eval(over->y_pexpr, over->var_values, NULL);
yading@10:         over->var_values[VAR_X] = av_expr_eval(over->x_pexpr, over->var_values, NULL);
yading@10:         over->x = normalize_xy(over->var_values[VAR_X], over->hsub);
yading@10:         over->y = normalize_xy(over->var_values[VAR_Y], over->vsub);
yading@10:     }
yading@10:     if (eval_tgt == EVAL_ENABLE || eval_tgt == EVAL_ALL) {
yading@10:         over->enable = av_expr_eval(over->enable_pexpr, over->var_values, NULL);
yading@10:     }
yading@10: }
yading@10: 
yading@10: static int set_expr(AVExpr **pexpr, const char *expr, const char *option, void *log_ctx)
yading@10: {
yading@10:     int ret;
yading@10:     AVExpr *old = NULL;
yading@10: 
yading@10:     if (*pexpr)
yading@10:         old = *pexpr;
yading@10:     ret = av_expr_parse(pexpr, expr, var_names,
yading@10:                         NULL, NULL, NULL, NULL, 0, log_ctx);
yading@10:     if (ret < 0) {
yading@10:         av_log(log_ctx, AV_LOG_ERROR,
yading@10:                "Error when evaluating the expression '%s' for %s\n",
yading@10:                expr, option);
yading@10:         *pexpr = old;
yading@10:         return ret;
yading@10:     }
yading@10: 
yading@10:     av_expr_free(old);
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: static int process_command(AVFilterContext *ctx, const char *cmd, const char *args,
yading@10:                            char *res, int res_len, int flags)
yading@10: {
yading@10:     OverlayContext *over = ctx->priv;
yading@10:     int ret;
yading@10: 
yading@10:     if      (!strcmp(cmd, "x"))
yading@10:         ret = set_expr(&over->x_pexpr, args, cmd, ctx);
yading@10:     else if (!strcmp(cmd, "y"))
yading@10:         ret = set_expr(&over->y_pexpr, args, cmd, ctx);
yading@10:     else if (!strcmp(cmd, "enable"))
yading@10:         ret = set_expr(&over->enable_pexpr, args, cmd, ctx);
yading@10:     else
yading@10:         ret = AVERROR(ENOSYS);
yading@10: 
yading@10:     if (ret < 0)
yading@10:         return ret;
yading@10: 
yading@10:     if (over->eval_mode == EVAL_MODE_INIT) {
yading@10:         eval_expr(ctx, EVAL_ALL);
yading@10:         av_log(ctx, AV_LOG_VERBOSE, "x:%f xi:%d y:%f yi:%d enable:%f\n",
yading@10:                over->var_values[VAR_X], over->x,
yading@10:                over->var_values[VAR_Y], over->y,
yading@10:                over->enable);
yading@10:     }
yading@10:     return ret;
yading@10: }
yading@10: 
yading@10: static int query_formats(AVFilterContext *ctx)
yading@10: {
yading@10:     OverlayContext *over = ctx->priv;
yading@10: 
yading@10:     /* overlay formats contains alpha, for avoiding conversion with alpha information loss */
yading@10:     static const enum AVPixelFormat main_pix_fmts_yuv420[] = {
yading@10:         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUVA420P, AV_PIX_FMT_NONE
yading@10:     };
yading@10:     static const enum AVPixelFormat overlay_pix_fmts_yuv420[] = {
yading@10:         AV_PIX_FMT_YUVA420P, AV_PIX_FMT_NONE
yading@10:     };
yading@10: 
yading@10:     static const enum AVPixelFormat main_pix_fmts_yuv444[] = {
yading@10:         AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUVA444P, AV_PIX_FMT_NONE
yading@10:     };
yading@10:     static const enum AVPixelFormat overlay_pix_fmts_yuv444[] = {
yading@10:         AV_PIX_FMT_YUVA444P, AV_PIX_FMT_NONE
yading@10:     };
yading@10: 
yading@10:     static const enum AVPixelFormat main_pix_fmts_rgb[] = {
yading@10:         AV_PIX_FMT_ARGB,  AV_PIX_FMT_RGBA,
yading@10:         AV_PIX_FMT_ABGR,  AV_PIX_FMT_BGRA,
yading@10:         AV_PIX_FMT_RGB24, AV_PIX_FMT_BGR24,
yading@10:         AV_PIX_FMT_NONE
yading@10:     };
yading@10:     static const enum AVPixelFormat overlay_pix_fmts_rgb[] = {
yading@10:         AV_PIX_FMT_ARGB,  AV_PIX_FMT_RGBA,
yading@10:         AV_PIX_FMT_ABGR,  AV_PIX_FMT_BGRA,
yading@10:         AV_PIX_FMT_NONE
yading@10:     };
yading@10: 
yading@10:     AVFilterFormats *main_formats;
yading@10:     AVFilterFormats *overlay_formats;
yading@10: 
yading@10:     switch (over->format) {
yading@10:     case OVERLAY_FORMAT_YUV420:
yading@10:         main_formats    = ff_make_format_list(main_pix_fmts_yuv420);
yading@10:         overlay_formats = ff_make_format_list(overlay_pix_fmts_yuv420);
yading@10:         break;
yading@10:     case OVERLAY_FORMAT_YUV444:
yading@10:         main_formats    = ff_make_format_list(main_pix_fmts_yuv444);
yading@10:         overlay_formats = ff_make_format_list(overlay_pix_fmts_yuv444);
yading@10:         break;
yading@10:     case OVERLAY_FORMAT_RGB:
yading@10:         main_formats    = ff_make_format_list(main_pix_fmts_rgb);
yading@10:         overlay_formats = ff_make_format_list(overlay_pix_fmts_rgb);
yading@10:         break;
yading@10:     default:
yading@10:         av_assert0(0);
yading@10:     }
yading@10: 
yading@10:     ff_formats_ref(main_formats,    &ctx->inputs [MAIN   ]->out_formats);
yading@10:     ff_formats_ref(overlay_formats, &ctx->inputs [OVERLAY]->out_formats);
yading@10:     ff_formats_ref(main_formats,    &ctx->outputs[MAIN   ]->in_formats );
yading@10: 
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: static const enum AVPixelFormat alpha_pix_fmts[] = {
yading@10:     AV_PIX_FMT_YUVA420P, AV_PIX_FMT_YUVA444P,
yading@10:     AV_PIX_FMT_ARGB, AV_PIX_FMT_ABGR, AV_PIX_FMT_RGBA,
yading@10:     AV_PIX_FMT_BGRA, AV_PIX_FMT_NONE
yading@10: };
yading@10: 
yading@10: static int config_input_main(AVFilterLink *inlink)
yading@10: {
yading@10:     OverlayContext *over = inlink->dst->priv;
yading@10:     const AVPixFmtDescriptor *pix_desc = av_pix_fmt_desc_get(inlink->format);
yading@10: 
yading@10:     av_image_fill_max_pixsteps(over->main_pix_step,    NULL, pix_desc);
yading@10: 
yading@10:     over->hsub = pix_desc->log2_chroma_w;
yading@10:     over->vsub = pix_desc->log2_chroma_h;
yading@10: 
yading@10:     over->main_is_packed_rgb =
yading@10:         ff_fill_rgba_map(over->main_rgba_map, inlink->format) >= 0;
yading@10:     over->main_has_alpha = ff_fmt_is_in(inlink->format, alpha_pix_fmts);
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: static int config_input_overlay(AVFilterLink *inlink)
yading@10: {
yading@10:     AVFilterContext *ctx  = inlink->dst;
yading@10:     OverlayContext  *over = inlink->dst->priv;
yading@10:     int ret;
yading@10:     const AVPixFmtDescriptor *pix_desc = av_pix_fmt_desc_get(inlink->format);
yading@10: 
yading@10:     av_image_fill_max_pixsteps(over->overlay_pix_step, NULL, pix_desc);
yading@10: 
yading@10:     /* Finish the configuration by evaluating the expressions
yading@10:        now when both inputs are configured. */
yading@10:     over->var_values[VAR_MAIN_W   ] = over->var_values[VAR_MW] = ctx->inputs[MAIN   ]->w;
yading@10:     over->var_values[VAR_MAIN_H   ] = over->var_values[VAR_MH] = ctx->inputs[MAIN   ]->h;
yading@10:     over->var_values[VAR_OVERLAY_W] = over->var_values[VAR_OW] = ctx->inputs[OVERLAY]->w;
yading@10:     over->var_values[VAR_OVERLAY_H] = over->var_values[VAR_OH] = ctx->inputs[OVERLAY]->h;
yading@10:     over->var_values[VAR_HSUB]  = 1<<pix_desc->log2_chroma_w;
yading@10:     over->var_values[VAR_VSUB]  = 1<<pix_desc->log2_chroma_h;
yading@10:     over->var_values[VAR_X]     = NAN;
yading@10:     over->var_values[VAR_Y]     = NAN;
yading@10:     over->var_values[VAR_N]     = 0;
yading@10:     over->var_values[VAR_T]     = NAN;
yading@10:     over->var_values[VAR_POS]   = NAN;
yading@10: 
yading@10:     if ((ret = set_expr(&over->x_pexpr,      over->x_expr,      "x",      ctx)) < 0 ||
yading@10:         (ret = set_expr(&over->y_pexpr,      over->y_expr,      "y",      ctx)) < 0 ||
yading@10:         (ret = set_expr(&over->enable_pexpr, over->enable_expr, "enable", ctx)) < 0)
yading@10:         return ret;
yading@10: 
yading@10:     over->overlay_is_packed_rgb =
yading@10:         ff_fill_rgba_map(over->overlay_rgba_map, inlink->format) >= 0;
yading@10:     over->overlay_has_alpha = ff_fmt_is_in(inlink->format, alpha_pix_fmts);
yading@10: 
yading@10:     if (over->eval_mode == EVAL_MODE_INIT) {
yading@10:         eval_expr(ctx, EVAL_ALL);
yading@10:         av_log(ctx, AV_LOG_VERBOSE, "x:%f xi:%d y:%f yi:%d enable:%f\n",
yading@10:                over->var_values[VAR_X], over->x,
yading@10:                over->var_values[VAR_Y], over->y,
yading@10:                over->enable);
yading@10:     }
yading@10: 
yading@10:     av_log(ctx, AV_LOG_VERBOSE,
yading@10:            "main w:%d h:%d fmt:%s overlay w:%d h:%d fmt:%s\n",
yading@10:            ctx->inputs[MAIN]->w, ctx->inputs[MAIN]->h,
yading@10:            av_get_pix_fmt_name(ctx->inputs[MAIN]->format),
yading@10:            ctx->inputs[OVERLAY]->w, ctx->inputs[OVERLAY]->h,
yading@10:            av_get_pix_fmt_name(ctx->inputs[OVERLAY]->format));
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: static int config_output(AVFilterLink *outlink)
yading@10: {
yading@10:     AVFilterContext *ctx = outlink->src;
yading@10: 
yading@10:     outlink->w = ctx->inputs[MAIN]->w;
yading@10:     outlink->h = ctx->inputs[MAIN]->h;
yading@10:     outlink->time_base = ctx->inputs[MAIN]->time_base;
yading@10: 
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: // divide by 255 and round to nearest
yading@10: // apply a fast variant: (X+127)/255 = ((X+127)*257+257)>>16 = ((X+128)*257)>>16
yading@10: #define FAST_DIV255(x) ((((x) + 128) * 257) >> 16)
yading@10: 
yading@10: // calculate the unpremultiplied alpha, applying the general equation:
yading@10: // alpha = alpha_overlay / ( (alpha_main + alpha_overlay) - (alpha_main * alpha_overlay) )
yading@10: // (((x) << 16) - ((x) << 9) + (x)) is a faster version of: 255 * 255 * x
yading@10: // ((((x) + (y)) << 8) - ((x) + (y)) - (y) * (x)) is a faster version of: 255 * (x + y)
yading@10: #define UNPREMULTIPLY_ALPHA(x, y) ((((x) << 16) - ((x) << 9) + (x)) / ((((x) + (y)) << 8) - ((x) + (y)) - (y) * (x)))
yading@10: 
yading@10: /**
yading@10:  * Blend image in src to destination buffer dst at position (x, y).
yading@10:  */
yading@10: static void blend_image(AVFilterContext *ctx,
yading@10:                         AVFrame *dst, AVFrame *src,
yading@10:                         int x, int y)
yading@10: {
yading@10:     OverlayContext *over = ctx->priv;
yading@10:     int i, imax, j, jmax, k, kmax;
yading@10:     const int src_w = src->width;
yading@10:     const int src_h = src->height;
yading@10:     const int dst_w = dst->width;
yading@10:     const int dst_h = dst->height;
yading@10: 
yading@10:     if (x >= dst_w || x+dst_w  < 0 ||
yading@10:         y >= dst_h || y+dst_h < 0)
yading@10:         return; /* no intersection */
yading@10: 
yading@10:     if (over->main_is_packed_rgb) {
yading@10:         uint8_t alpha;          ///< the amount of overlay to blend on to main
yading@10:         const int dr = over->main_rgba_map[R];
yading@10:         const int dg = over->main_rgba_map[G];
yading@10:         const int db = over->main_rgba_map[B];
yading@10:         const int da = over->main_rgba_map[A];
yading@10:         const int dstep = over->main_pix_step[0];
yading@10:         const int sr = over->overlay_rgba_map[R];
yading@10:         const int sg = over->overlay_rgba_map[G];
yading@10:         const int sb = over->overlay_rgba_map[B];
yading@10:         const int sa = over->overlay_rgba_map[A];
yading@10:         const int sstep = over->overlay_pix_step[0];
yading@10:         const int main_has_alpha = over->main_has_alpha;
yading@10:         uint8_t *s, *sp, *d, *dp;
yading@10: 
yading@10:         i = FFMAX(-y, 0);
yading@10:         sp = src->data[0] + i     * src->linesize[0];
yading@10:         dp = dst->data[0] + (y+i) * dst->linesize[0];
yading@10: 
yading@10:         for (imax = FFMIN(-y + dst_h, src_h); i < imax; i++) {
yading@10:             j = FFMAX(-x, 0);
yading@10:             s = sp + j     * sstep;
yading@10:             d = dp + (x+j) * dstep;
yading@10: 
yading@10:             for (jmax = FFMIN(-x + dst_w, src_w); j < jmax; j++) {
yading@10:                 alpha = s[sa];
yading@10: 
yading@10:                 // if the main channel has an alpha channel, alpha has to be calculated
yading@10:                 // to create an un-premultiplied (straight) alpha value
yading@10:                 if (main_has_alpha && alpha != 0 && alpha != 255) {
yading@10:                     uint8_t alpha_d = d[da];
yading@10:                     alpha = UNPREMULTIPLY_ALPHA(alpha, alpha_d);
yading@10:                 }
yading@10: 
yading@10:                 switch (alpha) {
yading@10:                 case 0:
yading@10:                     break;
yading@10:                 case 255:
yading@10:                     d[dr] = s[sr];
yading@10:                     d[dg] = s[sg];
yading@10:                     d[db] = s[sb];
yading@10:                     break;
yading@10:                 default:
yading@10:                     // main_value = main_value * (1 - alpha) + overlay_value * alpha
yading@10:                     // since alpha is in the range 0-255, the result must divided by 255
yading@10:                     d[dr] = FAST_DIV255(d[dr] * (255 - alpha) + s[sr] * alpha);
yading@10:                     d[dg] = FAST_DIV255(d[dg] * (255 - alpha) + s[sg] * alpha);
yading@10:                     d[db] = FAST_DIV255(d[db] * (255 - alpha) + s[sb] * alpha);
yading@10:                 }
yading@10:                 if (main_has_alpha) {
yading@10:                     switch (alpha) {
yading@10:                     case 0:
yading@10:                         break;
yading@10:                     case 255:
yading@10:                         d[da] = s[sa];
yading@10:                         break;
yading@10:                     default:
yading@10:                         // apply alpha compositing: main_alpha += (1-main_alpha) * overlay_alpha
yading@10:                         d[da] += FAST_DIV255((255 - d[da]) * s[sa]);
yading@10:                     }
yading@10:                 }
yading@10:                 d += dstep;
yading@10:                 s += sstep;
yading@10:             }
yading@10:             dp += dst->linesize[0];
yading@10:             sp += src->linesize[0];
yading@10:         }
yading@10:     } else {
yading@10:         const int main_has_alpha = over->main_has_alpha;
yading@10:         if (main_has_alpha) {
yading@10:             uint8_t alpha;          ///< the amount of overlay to blend on to main
yading@10:             uint8_t *s, *sa, *d, *da;
yading@10: 
yading@10:             i = FFMAX(-y, 0);
yading@10:             sa = src->data[3] + i     * src->linesize[3];
yading@10:             da = dst->data[3] + (y+i) * dst->linesize[3];
yading@10: 
yading@10:             for (imax = FFMIN(-y + dst_h, src_h); i < imax; i++) {
yading@10:                 j = FFMAX(-x, 0);
yading@10:                 s = sa + j;
yading@10:                 d = da + x+j;
yading@10: 
yading@10:                 for (jmax = FFMIN(-x + dst_w, src_w); j < jmax; j++) {
yading@10:                     alpha = *s;
yading@10:                     if (alpha != 0 && alpha != 255) {
yading@10:                         uint8_t alpha_d = *d;
yading@10:                         alpha = UNPREMULTIPLY_ALPHA(alpha, alpha_d);
yading@10:                     }
yading@10:                     switch (alpha) {
yading@10:                     case 0:
yading@10:                         break;
yading@10:                     case 255:
yading@10:                         *d = *s;
yading@10:                         break;
yading@10:                     default:
yading@10:                         // apply alpha compositing: main_alpha += (1-main_alpha) * overlay_alpha
yading@10:                         *d += FAST_DIV255((255 - *d) * *s);
yading@10:                     }
yading@10:                     d += 1;
yading@10:                     s += 1;
yading@10:                 }
yading@10:                 da += dst->linesize[3];
yading@10:                 sa += src->linesize[3];
yading@10:             }
yading@10:         }
yading@10:         for (i = 0; i < 3; i++) {
yading@10:             int hsub = i ? over->hsub : 0;
yading@10:             int vsub = i ? over->vsub : 0;
yading@10:             int src_wp = FFALIGN(src_w, 1<<hsub) >> hsub;
yading@10:             int src_hp = FFALIGN(src_h, 1<<vsub) >> vsub;
yading@10:             int dst_wp = FFALIGN(dst_w, 1<<hsub) >> hsub;
yading@10:             int dst_hp = FFALIGN(dst_h, 1<<vsub) >> vsub;
yading@10:             int yp = y>>vsub;
yading@10:             int xp = x>>hsub;
yading@10:             uint8_t *s, *sp, *d, *dp, *a, *ap;
yading@10: 
yading@10:             j = FFMAX(-yp, 0);
yading@10:             sp = src->data[i] + j         * src->linesize[i];
yading@10:             dp = dst->data[i] + (yp+j)    * dst->linesize[i];
yading@10:             ap = src->data[3] + (j<<vsub) * src->linesize[3];
yading@10: 
yading@10:             for (jmax = FFMIN(-yp + dst_hp, src_hp); j < jmax; j++) {
yading@10:                 k = FFMAX(-xp, 0);
yading@10:                 d = dp + xp+k;
yading@10:                 s = sp + k;
yading@10:                 a = ap + (k<<hsub);
yading@10: 
yading@10:                 for (kmax = FFMIN(-xp + dst_wp, src_wp); k < kmax; k++) {
yading@10:                     int alpha_v, alpha_h, alpha;
yading@10: 
yading@10:                     // average alpha for color components, improve quality
yading@10:                     if (hsub && vsub && j+1 < src_hp && k+1 < src_wp) {
yading@10:                         alpha = (a[0] + a[src->linesize[3]] +
yading@10:                                  a[1] + a[src->linesize[3]+1]) >> 2;
yading@10:                     } else if (hsub || vsub) {
yading@10:                         alpha_h = hsub && k+1 < src_wp ?
yading@10:                             (a[0] + a[1]) >> 1 : a[0];
yading@10:                         alpha_v = vsub && j+1 < src_hp ?
yading@10:                             (a[0] + a[src->linesize[3]]) >> 1 : a[0];
yading@10:                         alpha = (alpha_v + alpha_h) >> 1;
yading@10:                     } else
yading@10:                         alpha = a[0];
yading@10:                     // if the main channel has an alpha channel, alpha has to be calculated
yading@10:                     // to create an un-premultiplied (straight) alpha value
yading@10:                     if (main_has_alpha && alpha != 0 && alpha != 255) {
yading@10:                         // average alpha for color components, improve quality
yading@10:                         uint8_t alpha_d;
yading@10:                         if (hsub && vsub && j+1 < src_hp && k+1 < src_wp) {
yading@10:                             alpha_d = (d[0] + d[src->linesize[3]] +
yading@10:                                        d[1] + d[src->linesize[3]+1]) >> 2;
yading@10:                         } else if (hsub || vsub) {
yading@10:                             alpha_h = hsub && k+1 < src_wp ?
yading@10:                                 (d[0] + d[1]) >> 1 : d[0];
yading@10:                             alpha_v = vsub && j+1 < src_hp ?
yading@10:                                 (d[0] + d[src->linesize[3]]) >> 1 : d[0];
yading@10:                             alpha_d = (alpha_v + alpha_h) >> 1;
yading@10:                         } else
yading@10:                             alpha_d = d[0];
yading@10:                         alpha = UNPREMULTIPLY_ALPHA(alpha, alpha_d);
yading@10:                     }
yading@10:                     *d = FAST_DIV255(*d * (255 - alpha) + *s * alpha);
yading@10:                     s++;
yading@10:                     d++;
yading@10:                     a += 1 << hsub;
yading@10:                 }
yading@10:                 dp += dst->linesize[i];
yading@10:                 sp += src->linesize[i];
yading@10:                 ap += (1 << vsub) * src->linesize[3];
yading@10:             }
yading@10:         }
yading@10:     }
yading@10: }
yading@10: 
yading@10: static int try_filter_frame(AVFilterContext *ctx, AVFrame *mainpic)
yading@10: {
yading@10:     OverlayContext *over = ctx->priv;
yading@10:     AVFilterLink *inlink = ctx->inputs[0];
yading@10:     AVFrame *next_overpic;
yading@10:     int ret;
yading@10: 
yading@10:     /* Discard obsolete overlay frames: if there is a next overlay frame with pts
yading@10:      * before the main frame, we can drop the current overlay. */
yading@10:     while (1) {
yading@10:         next_overpic = ff_bufqueue_peek(&over->queue_over, 0);
yading@10:         if (!next_overpic && over->overlay_eof && !over->repeatlast) {
yading@10:             av_frame_free(&over->overpicref);
yading@10:             break;
yading@10:         }
yading@10:         if (!next_overpic || av_compare_ts(next_overpic->pts, ctx->inputs[OVERLAY]->time_base,
yading@10:                                            mainpic->pts     , ctx->inputs[MAIN]->time_base) > 0)
yading@10:             break;
yading@10:         ff_bufqueue_get(&over->queue_over);
yading@10:         av_frame_free(&over->overpicref);
yading@10:         over->overpicref = next_overpic;
yading@10:     }
yading@10: 
yading@10:     /* If there is no next frame and no EOF and the overlay frame is before
yading@10:      * the main frame, we can not know yet if it will be superseded. */
yading@10:     if (!over->queue_over.available && !over->overlay_eof &&
yading@10:         (!over->overpicref || av_compare_ts(over->overpicref->pts, ctx->inputs[OVERLAY]->time_base,
yading@10:                                             mainpic->pts         , ctx->inputs[MAIN]->time_base) < 0))
yading@10:         return AVERROR(EAGAIN);
yading@10: 
yading@10:     /* At this point, we know that the current overlay frame extends to the
yading@10:      * time of the main frame. */
yading@10:     av_dlog(ctx, "main_pts:%s main_pts_time:%s",
yading@10:             av_ts2str(mainpic->pts), av_ts2timestr(mainpic->pts, &ctx->inputs[MAIN]->time_base));
yading@10:     if (over->overpicref)
yading@10:         av_dlog(ctx, " over_pts:%s over_pts_time:%s",
yading@10:                 av_ts2str(over->overpicref->pts), av_ts2timestr(over->overpicref->pts, &ctx->inputs[OVERLAY]->time_base));
yading@10:     av_dlog(ctx, "\n");
yading@10: 
yading@10:     if (over->overpicref) {
yading@10:         if (over->eval_mode == EVAL_MODE_FRAME) {
yading@10:             int64_t pos = av_frame_get_pkt_pos(mainpic);
yading@10: 
yading@10:             over->var_values[VAR_T] = mainpic->pts == AV_NOPTS_VALUE ?
yading@10:                 NAN : mainpic->pts * av_q2d(inlink->time_base);
yading@10:             over->var_values[VAR_POS] = pos == -1 ? NAN : pos;
yading@10: 
yading@10:             eval_expr(ctx, EVAL_ALL);
yading@10:             av_log(ctx, AV_LOG_DEBUG, "n:%f t:%f pos:%f x:%f xi:%d y:%f yi:%d enable:%f\n",
yading@10:                    over->var_values[VAR_N], over->var_values[VAR_T], over->var_values[VAR_POS],
yading@10:                    over->var_values[VAR_X], over->x,
yading@10:                    over->var_values[VAR_Y], over->y,
yading@10:                    over->enable);
yading@10:         }
yading@10:         if (over->enable)
yading@10:             blend_image(ctx, mainpic, over->overpicref, over->x, over->y);
yading@10: 
yading@10:         over->var_values[VAR_N] += 1.0;
yading@10:     }
yading@10:     ret = ff_filter_frame(ctx->outputs[0], mainpic);
yading@10:     av_assert1(ret != AVERROR(EAGAIN));
yading@10:     over->frame_requested = 0;
yading@10:     return ret;
yading@10: }
yading@10: 
yading@10: static int try_filter_next_frame(AVFilterContext *ctx)
yading@10: {
yading@10:     OverlayContext *over = ctx->priv;
yading@10:     AVFrame *next_mainpic = ff_bufqueue_peek(&over->queue_main, 0);
yading@10:     int ret;
yading@10: 
yading@10:     if (!next_mainpic)
yading@10:         return AVERROR(EAGAIN);
yading@10:     if ((ret = try_filter_frame(ctx, next_mainpic)) == AVERROR(EAGAIN))
yading@10:         return ret;
yading@10:     ff_bufqueue_get(&over->queue_main);
yading@10:     return ret;
yading@10: }
yading@10: 
yading@10: static int flush_frames(AVFilterContext *ctx)
yading@10: {
yading@10:     int ret;
yading@10: 
yading@10:     while (!(ret = try_filter_next_frame(ctx)));
yading@10:     return ret == AVERROR(EAGAIN) ? 0 : ret;
yading@10: }
yading@10: 
yading@10: static int filter_frame_main(AVFilterLink *inlink, AVFrame *inpicref)
yading@10: {
yading@10:     AVFilterContext *ctx = inlink->dst;
yading@10:     OverlayContext *over = ctx->priv;
yading@10:     int ret;
yading@10: 
yading@10:     if ((ret = flush_frames(ctx)) < 0)
yading@10:         return ret;
yading@10:     if ((ret = try_filter_frame(ctx, inpicref)) < 0) {
yading@10:         if (ret != AVERROR(EAGAIN))
yading@10:             return ret;
yading@10:         ff_bufqueue_add(ctx, &over->queue_main, inpicref);
yading@10:     }
yading@10: 
yading@10:     if (!over->overpicref)
yading@10:         return 0;
yading@10:     flush_frames(ctx);
yading@10: 
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: static int filter_frame_over(AVFilterLink *inlink, AVFrame *inpicref)
yading@10: {
yading@10:     AVFilterContext *ctx = inlink->dst;
yading@10:     OverlayContext *over = ctx->priv;
yading@10:     int ret;
yading@10: 
yading@10:     if ((ret = flush_frames(ctx)) < 0)
yading@10:         return ret;
yading@10:     ff_bufqueue_add(ctx, &over->queue_over, inpicref);
yading@10:     ret = try_filter_next_frame(ctx);
yading@10:     return ret == AVERROR(EAGAIN) ? 0 : ret;
yading@10: }
yading@10: 
yading@10: static int request_frame(AVFilterLink *outlink)
yading@10: {
yading@10:     AVFilterContext *ctx = outlink->src;
yading@10:     OverlayContext *over = ctx->priv;
yading@10:     int input, ret;
yading@10: 
yading@10:     if (!try_filter_next_frame(ctx))
yading@10:         return 0;
yading@10:     over->frame_requested = 1;
yading@10:     while (over->frame_requested) {
yading@10:         /* TODO if we had a frame duration, we could guess more accurately */
yading@10:         input = !over->overlay_eof && (over->queue_main.available ||
yading@10:                                        over->queue_over.available < 2) ?
yading@10:                 OVERLAY : MAIN;
yading@10:         ret = ff_request_frame(ctx->inputs[input]);
yading@10:         /* EOF on main is reported immediately */
yading@10:         if (ret == AVERROR_EOF && input == OVERLAY) {
yading@10:             over->overlay_eof = 1;
yading@10:             if (over->shortest)
yading@10:                 return ret;
yading@10:             if ((ret = try_filter_next_frame(ctx)) != AVERROR(EAGAIN))
yading@10:                 return ret;
yading@10:             ret = 0; /* continue requesting frames on main */
yading@10:         }
yading@10:         if (ret < 0)
yading@10:             return ret;
yading@10:     }
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: #define OFFSET(x) offsetof(OverlayContext, x)
yading@10: #define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
yading@10: 
yading@10: static const AVOption overlay_options[] = {
yading@10:     { "x", "set the x expression", OFFSET(x_expr), AV_OPT_TYPE_STRING, {.str = "0"}, CHAR_MIN, CHAR_MAX, FLAGS },
yading@10:     { "y", "set the y expression", OFFSET(y_expr), AV_OPT_TYPE_STRING, {.str = "0"}, CHAR_MIN, CHAR_MAX, FLAGS },
yading@10:     { "enable", "set expression which enables overlay", OFFSET(enable_expr), AV_OPT_TYPE_STRING, {.str = "1"}, .flags = FLAGS },
yading@10:     { "eval", "specify when to evaluate expressions", OFFSET(eval_mode), AV_OPT_TYPE_INT, {.i64 = EVAL_MODE_FRAME}, 0, EVAL_MODE_NB-1, FLAGS, "eval" },
yading@10:          { "init",  "eval expressions once during initialization", 0, AV_OPT_TYPE_CONST, {.i64=EVAL_MODE_INIT},  .flags = FLAGS, .unit = "eval" },
yading@10:          { "frame", "eval expressions per-frame",                  0, AV_OPT_TYPE_CONST, {.i64=EVAL_MODE_FRAME}, .flags = FLAGS, .unit = "eval" },
yading@10:     { "rgb", "force packed RGB in input and output (deprecated)", OFFSET(allow_packed_rgb), AV_OPT_TYPE_INT, {.i64=0}, 0, 1, FLAGS },
yading@10:     { "shortest", "force termination when the shortest input terminates", OFFSET(shortest), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, FLAGS },
yading@10:     { "format", "set output format", OFFSET(format), AV_OPT_TYPE_INT, {.i64=OVERLAY_FORMAT_YUV420}, 0, OVERLAY_FORMAT_NB-1, FLAGS, "format" },
yading@10:         { "yuv420", "", 0, AV_OPT_TYPE_CONST, {.i64=OVERLAY_FORMAT_YUV420}, .flags = FLAGS, .unit = "format" },
yading@10:         { "yuv444", "", 0, AV_OPT_TYPE_CONST, {.i64=OVERLAY_FORMAT_YUV444}, .flags = FLAGS, .unit = "format" },
yading@10:         { "rgb",    "", 0, AV_OPT_TYPE_CONST, {.i64=OVERLAY_FORMAT_RGB},    .flags = FLAGS, .unit = "format" },
yading@10:     { "repeatlast", "repeat overlay of the last overlay frame", OFFSET(repeatlast), AV_OPT_TYPE_INT, {.i64=1}, 0, 1, FLAGS },
yading@10:     { NULL }
yading@10: };
yading@10: 
yading@10: AVFILTER_DEFINE_CLASS(overlay);
yading@10: 
yading@10: static const AVFilterPad avfilter_vf_overlay_inputs[] = {
yading@10:     {
yading@10:         .name         = "main",
yading@10:         .type         = AVMEDIA_TYPE_VIDEO,
yading@10:         .get_video_buffer = ff_null_get_video_buffer,
yading@10:         .config_props = config_input_main,
yading@10:         .filter_frame = filter_frame_main,
yading@10:         .needs_writable = 1,
yading@10:     },
yading@10:     {
yading@10:         .name         = "overlay",
yading@10:         .type         = AVMEDIA_TYPE_VIDEO,
yading@10:         .config_props = config_input_overlay,
yading@10:         .filter_frame = filter_frame_over,
yading@10:     },
yading@10:     { NULL }
yading@10: };
yading@10: 
yading@10: static const AVFilterPad avfilter_vf_overlay_outputs[] = {
yading@10:     {
yading@10:         .name          = "default",
yading@10:         .type          = AVMEDIA_TYPE_VIDEO,
yading@10:         .config_props  = config_output,
yading@10:         .request_frame = request_frame,
yading@10:     },
yading@10:     { NULL }
yading@10: };
yading@10: 
yading@10: AVFilter avfilter_vf_overlay = {
yading@10:     .name      = "overlay",
yading@10:     .description = NULL_IF_CONFIG_SMALL("Overlay a video source on top of the input."),
yading@10: 
yading@10:     .init      = init,
yading@10:     .uninit    = uninit,
yading@10: 
yading@10:     .priv_size = sizeof(OverlayContext),
yading@10:     .priv_class = &overlay_class,
yading@10: 
yading@10:     .query_formats = query_formats,
yading@10:     .process_command = process_command,
yading@10: 
yading@10:     .inputs    = avfilter_vf_overlay_inputs,
yading@10:     .outputs   = avfilter_vf_overlay_outputs,
yading@10: };