RepoMirrors/mpv
1
0
Fork 0
mirror of https://github.com/mpv-player/mpv synced 2024-12-21 06:14:32 +00:00
Code Issues Releases Wiki Activity

vo_opengl: adjust interpolation code for the new video-sync mechanism

Browse Source 
This should make interpolation work much better in general, although
there still might be some side effects for unusual framerates (eg. 35 Hz
or 48 Hz). Most of the common framerates are tested and working fine.
(24 Hz, 30 Hz, 60 Hz)

The new code doesn't have support for oversample yet, so it's been
removed (and will most likely be reimplemented in a cleaner way if
there's enough demand). I would recommend using something like robidoux
or mitchell instead of oversample, though - they're much
smoother for the common cases.
This commit is contained in:
Niklas Haas 2015-06-26 10:59:57 +02:00 committed by wm4
parent c5e3613bba
commit f166d12985
5 changed files with 117 additions and 115 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
DOCS/man/vo.rst
View File

@ -495,12 +495,11 @@ Available video output drivers are:
The filter used for interpolating the temporal axis (frames). This is
only used if ``interpolation`` is enabled. The only valid choices
for ``tscale`` are separable convolution filters (use ``tscale=help``
to get a list). The default is ``oversample``.
to get a list). The default is ``robidoux``.
Note that the maximum supported filter radius is currently 3, and that
using filters with larger radius may introduce issues when pausing or
framestepping, proportional to the radius used. It is recommended to
stick to a radius of 1 or 2.
Note that the maximum supported filter radius is currently 3, due to
limitations in the number of video textures that can be loaded
simultaneously.
``dscale-radius``, ``cscale-radius``, ``tscale-radius``, etc.
Set filter parameters for ``dscale``, ``cscale`` and ``tscale``,

203
video/out/gl_video.c
View File

@ -64,7 +64,7 @@ static const char *const fixed_scale_filters[] = {
NULL
};
static const char *const fixed_tscale_filters[] = {
"oversample",
//"oversample",
NULL
};
@ -130,8 +130,7 @@ struct scaler {
struct fbosurface {
struct fbotex fbotex;
int64_t pts;
double vpts; // used for synchronizing subtitles only
double pts;
};
#define FBOSURFACES_MAX 10
@ -340,7 +339,7 @@ const struct gl_video_opts gl_video_opts_def = {
{{"bilinear", .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // scale
{{NULL, .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // dscale
{{"bilinear", .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // cscale
{{"oversample", .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // tscale
{{"robidoux", .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // tscale
},
.alpha_mode = 2,
.background = {0, 0, 0, 255},
@ -359,7 +358,7 @@ const struct gl_video_opts gl_video_opts_hq_def = {
{{"spline36", .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // scale
{{"mitchell", .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // dscale
{{"spline36", .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // cscale
{{"oversample", .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // tscale
{{"robidoux", .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // tscale
},
.alpha_mode = 2,
.background = {0, 0, 0, 255},
@ -523,8 +522,7 @@ void gl_video_set_debug(struct gl_video *p, bool enable)
static void gl_video_reset_surfaces(struct gl_video *p)
{
for (int i = 0; i < FBOSURFACES_MAX; i++) {
p->surfaces[i].pts = 0;
p->surfaces[i].vpts = MP_NOPTS_VALUE;
p->surfaces[i].pts = MP_NOPTS_VALUE;
}
p->surface_idx = 0;
p->surface_now = 0;
@ -1981,8 +1979,13 @@ static void pass_draw_osd(struct gl_video *p, int draw_flags, double pts,
// The main rendering function, takes care of everything up to and including
// upscaling
static void pass_render_frame(struct gl_video *p)
static void pass_render_frame(struct gl_video *p, struct mp_image *img)
{
if (img) {
gl_video_set_image(p, img);
gl_video_upload_image(p);
}
p->use_linear = p->opts.linear_scaling || p->opts.sigmoid_upscaling;
p->use_indirect = false; // set to true as needed by pass_*
pass_read_video(p);
@ -2068,49 +2071,68 @@ static void gl_video_interpolate_frame(struct gl_video *p, int fbo,
// First of all, figure out if we have a frame availble at all, and draw
// it manually + reset the queue if not
if (!p->surfaces[p->surface_now].pts) {
pass_render_frame(p);
if (p->surfaces[p->surface_now].pts == MP_NOPTS_VALUE) {
pass_render_frame(p, t->frame);
finish_pass_fbo(p, &p->surfaces[p->surface_now].fbotex,
vp_w, vp_h, 0, FBOTEX_FUZZY);
p->surfaces[p->surface_now].pts = t ? t->pts : 0;
p->surfaces[p->surface_now].vpts = p->image.mpi->pts;
p->surfaces[p->surface_now].pts = p->image.mpi->pts;
p->surface_idx = p->surface_now;
}
// Find the right frame for this instant
if (t->frame && t->frame->pts != MP_NOPTS_VALUE) {
int next = fbosurface_wrap(p->surface_now + 1);
while (p->surfaces[next].pts != MP_NOPTS_VALUE &&
p->surfaces[next].pts > p->surfaces[p->surface_now].pts &&
p->surfaces[p->surface_now].pts < t->frame->pts)
{
p->surface_now = next;
next = fbosurface_wrap(next + 1);
}
}
// Figure out the queue size. For illustration, a filter radius of 2 would
// look like this: _ A [B] C D _
// A is surface_bse, B is surface_now, C is surface_nxt and D is
// surface_end.
struct scaler *tscale = &p->scaler[3];
reinit_scaler(p, tscale, &p->opts.scaler[3], 1, tscale_sizes);
bool oversample = strcmp(tscale->conf.kernel.name, "oversample") == 0;
int size;
if (oversample) {
size = 2;
} else {
assert(tscale->kernel && !tscale->kernel->polar);
size = ceil(tscale->kernel->size);
assert(size <= TEXUNIT_VIDEO_NUM);
}
int radius = size/2;
assert(tscale->kernel && !tscale->kernel->polar);
size = ceil(tscale->kernel->size);
assert(size <= TEXUNIT_VIDEO_NUM);
int radius = size/2;
int surface_now = p->surface_now;
int surface_nxt = fbosurface_wrap(surface_now + 1);
int surface_bse = fbosurface_wrap(surface_now - (radius-1));
int surface_end = fbosurface_wrap(surface_now + radius);
assert(fbosurface_wrap(surface_bse + size-1) == surface_end);
// Render a new frame if it came in and there's room in the queue
// Render new frames while there's room in the queue. Note that technically,
// this should be done before the step where we find the right frame, but
// it only barely matters at the very beginning of playback, and this way
// makes the code much more linear.
int surface_dst = fbosurface_wrap(p->surface_idx+1);
if (t && surface_dst != surface_bse &&
p->surfaces[p->surface_idx].pts < t->pts) {
MP_STATS(p, "new-pts");
pass_render_frame(p);
finish_pass_fbo(p, &p->surfaces[surface_dst].fbotex,
vp_w, vp_h, 0, FBOTEX_FUZZY);
p->surfaces[surface_dst].pts = t->pts;
p->surfaces[surface_dst].vpts = p->image.mpi->pts;
p->surface_idx = surface_dst;
for (int i = -1; i < t->num_future_frames; i++) {
// Avoid overwriting data we might still need
if (surface_dst == surface_bse - 1)
break;
struct mp_image *f = i < 0 ? t->frame : t->future_frames[i];
if (!f || f->pts == MP_NOPTS_VALUE)
continue;
if (f->pts > p->surfaces[p->surface_idx].pts) {
MP_STATS(p, "new-pts");
pass_render_frame(p, f);
finish_pass_fbo(p, &p->surfaces[surface_dst].fbotex,
vp_w, vp_h, 0, FBOTEX_FUZZY);
p->surfaces[surface_dst].pts = f->pts;
p->surface_idx = surface_dst;
surface_dst = fbosurface_wrap(surface_dst+1);
}
}
// Figure out whether the queue is "valid". A queue is invalid if the
@ -2121,7 +2143,9 @@ static void gl_video_interpolate_frame(struct gl_video *p, int fbo,
bool valid = true;
for (int i = surface_bse, ii; valid && i != surface_end; i = ii) {
ii = fbosurface_wrap(i+1);
if (!p->surfaces[i].pts || !p->surfaces[ii].pts) {
if (p->surfaces[i].pts == MP_NOPTS_VALUE ||
p->surfaces[ii].pts == MP_NOPTS_VALUE)
{
valid = false;
} else if (p->surfaces[ii].pts < p->surfaces[i].pts) {
valid = false;
@ -2130,81 +2154,63 @@ static void gl_video_interpolate_frame(struct gl_video *p, int fbo,
}
// Update OSD PTS to synchronize subtitles with the displayed frame
if (t) {
double vpts_now = p->surfaces[surface_now].vpts,
vpts_nxt = p->surfaces[surface_nxt].vpts,
vpts_new = p->image.mpi->pts;
if (vpts_now != MP_NOPTS_VALUE &&
vpts_nxt != MP_NOPTS_VALUE &&
vpts_new != MP_NOPTS_VALUE)
{
// Round to nearest neighbour
double vpts_vsync = (t->next_vsync - t->pts)/1e6 + vpts_new;
p->osd_pts = fabs(vpts_vsync-vpts_now) < fabs(vpts_vsync-vpts_nxt)
? vpts_now : vpts_nxt;
}
}
p->osd_pts = p->surfaces[surface_now].pts;
// Finally, draw the right mix of frames to the screen.
if (!t || !valid) {
if (!valid) {
// surface_now is guaranteed to be valid, so we can safely use it.
pass_load_fbotex(p, &p->surfaces[surface_now].fbotex, 0, vp_w, vp_h);
GLSL(vec4 color = texture(texture0, texcoord0);)
p->is_interpolated = false;
} else {
int64_t pts_now = p->surfaces[surface_now].pts,
pts_nxt = p->surfaces[surface_nxt].pts;
double fscale = pts_nxt - pts_now, mix;
if (oversample) {
double vsync_interval = t->next_vsync - t->prev_vsync,
threshold = tscale->conf.kernel.params[0];
threshold = isnan(threshold) ? 0.0 : threshold;
mix = (pts_nxt - t->next_vsync) / vsync_interval;
mix = mix <= 0 + threshold ? 0 : mix;
mix = mix >= 1 - threshold ? 1 : mix;
mix = 1 - mix;
gl_sc_uniform_f(p->sc, "inter_coeff", mix);
GLSL(vec4 color = mix(texture(texture0, texcoord0),
texture(texture1, texcoord1),
inter_coeff);)
} else {
mix = (t->next_vsync - pts_now) / fscale;
gl_sc_uniform_f(p->sc, "fcoord", mix);
pass_sample_separated_gen(p, tscale, 0, 0);
double pts_now = p->surfaces[surface_now].pts,
pts_nxt = p->surfaces[surface_nxt].pts;
double mix = (t->vsync_offset / 1e6) / (pts_nxt - pts_now);
// The scaler code always wants the fcoord to be between 0 and 1,
// so we try to adjust by using the previous set of N frames instead
// (which requires some extra checking to make sure it's valid)
if (mix < 0.0) {
int prev = fbosurface_wrap(surface_bse - 1);
if (p->surfaces[prev].pts != MP_NOPTS_VALUE &&
p->surfaces[prev].pts < p->surfaces[surface_bse].pts)
{
mix += 1.0;
surface_bse = prev;
} else {
mix = 0.0; // at least don't blow up, this should only
// ever happen at the start of playback
}
}
// Non-oversample case
gl_sc_uniform_f(p->sc, "fcoord", mix);
pass_sample_separated_gen(p, tscale, 0, 0);
// Load all the required frames
for (int i = 0; i < size; i++) {
pass_load_fbotex(p, &p->surfaces[fbosurface_wrap(surface_bse+i)].fbotex,
i, vp_w, vp_h);
}
MP_STATS(p, "frame-mix");
MP_DBG(p, "inter frame ppts: %lld, pts: %lld, vsync: %lld, mix: %f\n",
(long long)pts_now, (long long)pts_nxt,
(long long)t->next_vsync, mix);
MP_DBG(p, "inter frame pts: %lld, vsync: %lld, mix: %f\n",
(long long)t->pts, (long long)t->next_vsync, mix);
p->is_interpolated = true;
}
pass_draw_to_screen(p, fbo);
// Dequeue frames if necessary
if (t) {
int64_t vsync_interval = t->next_vsync - t->prev_vsync;
int64_t vsync_guess = t->next_vsync + vsync_interval;
if (p->surfaces[surface_nxt].pts > p->surfaces[p->surface_now].pts &&
p->surfaces[surface_nxt].pts < vsync_guess)
{
p->surface_now = surface_nxt;
}
}
}
// (fbo==0 makes BindFramebuffer select the screen backbuffer)
void gl_video_render_frame(struct gl_video *p, int fbo, struct frame_timing *t)
void gl_video_render_frame(struct gl_video *p, struct mp_image *mpi, int fbo,
struct frame_timing *t)
{
GL *gl = p->gl;
struct video_image *vimg = &p->image;
gl->BindFramebuffer(GL_FRAMEBUFFER, fbo);
if (!vimg->mpi || p->dst_rect.x0 > 0 || p->dst_rect.y0 > 0 ||
if ((!mpi && !vimg->mpi) || p->dst_rect.x0 > 0 || p->dst_rect.y0 > 0 ||
p->dst_rect.x1 < p->vp_w || p->dst_rect.y1 < abs(p->vp_h))
{
struct m_color c = p->opts.background;
@ -2212,22 +2218,18 @@ void gl_video_render_frame(struct gl_video *p, int fbo, struct frame_timing *t)
gl->Clear(GL_COLOR_BUFFER_BIT);
}
if (vimg->mpi) {
gl_video_upload_image(p);
gl_sc_set_vao(p->sc, &p->vao);
gl_sc_set_vao(p->sc, &p->vao);
if (p->opts.interpolation) {
gl_video_interpolate_frame(p, fbo, t);
} else {
// Skip interpolation if there's nothing to be done
pass_render_frame(p);
pass_draw_to_screen(p, fbo);
}
debug_check_gl(p, "after video rendering");
if (p->opts.interpolation && t) {
gl_video_interpolate_frame(p, fbo, t);
} else {
// Skip interpolation if there's nothing to be done
pass_render_frame(p, mpi);
pass_draw_to_screen(p, fbo);
}
debug_check_gl(p, "after video rendering");
gl->BindFramebuffer(GL_FRAMEBUFFER, fbo);
if (p->osd) {
@ -2297,9 +2299,12 @@ void gl_video_set_image(struct gl_video *p, struct mp_image *mpi)
struct video_image *vimg = &p->image;
talloc_free(vimg->mpi);
vimg->mpi = mpi;
vimg->mpi = mp_image_new_ref(mpi);
vimg->needs_upload = true;
if (!vimg->mpi)
abort();
p->osd_pts = mpi->pts;
}
@ -2793,16 +2798,14 @@ void gl_video_set_options(struct gl_video *p, struct gl_video_opts *opts,
}
// Figure out an adequate size for the interpolation queue. The larger
// the radius, the earlier we need to queue frames. This rough heuristic
// seems to work for now, but ideally we want to rework the pause/unpause
// logic to make larger queue sizes the default.
// the radius, the earlier we need to queue frames.
if (queue_size && p->opts.interpolation) {
const struct filter_kernel *kernel =
mp_find_filter_kernel(p->opts.scaler[3].kernel.name);
if (kernel) {
double radius = kernel->f.radius;
radius = radius > 0 ? radius : p->opts.scaler[3].radius;
*queue_size = 50e3 * ceil(radius);
*queue_size = 1 + ceil(radius);
}
}

3
video/out/gl_video.h
View File

@ -88,7 +88,8 @@ void gl_video_config(struct gl_video *p, struct mp_image_params *params);
void gl_video_set_output_depth(struct gl_video *p, int r, int g, int b);
void gl_video_set_lut3d(struct gl_video *p, struct lut3d *lut3d);
void gl_video_set_image(struct gl_video *p, struct mp_image *img);
void gl_video_render_frame(struct gl_video *p, int fbo, struct frame_timing *t);
void gl_video_render_frame(struct gl_video *p, struct mp_image *img, int fbo,
struct frame_timing *t);
void gl_video_resize(struct gl_video *p, int vp_w, int vp_h,
struct mp_rect *src, struct mp_rect *dst,
struct mp_osd_res *osd);

15
video/out/vo_opengl.c
View File

@ -170,14 +170,11 @@ static void draw_image_timed(struct vo *vo, mp_image_t *mpi,
struct gl_priv *p = vo->priv;
GL *gl = p->gl;
if (mpi)
gl_video_set_image(p->renderer, mpi);
if (p->glctx->start_frame && !p->glctx->start_frame(p->glctx))
return;
p->frame_started = true;
gl_video_render_frame(p->renderer, 0, t);
gl_video_render_frame(p->renderer, mpi, 0, t);
// The playloop calls this last before waiting some time until it decides
// to call flip_page(). Tell OpenGL to start execution of the GPU commands
@ -186,6 +183,8 @@ static void draw_image_timed(struct vo *vo, mp_image_t *mpi,
if (p->use_glFinish)
gl->Finish();
talloc_free(mpi);
}
static void draw_image(struct vo *vo, mp_image_t *mpi)
@ -302,9 +301,9 @@ static bool reparse_cmdline(struct gl_priv *p, char *args)
}
if (r >= 0) {
int queue = 0;
int queue = 1;
gl_video_set_options(p->renderer, opts->renderer_opts, &queue);
vo_set_queue_params(p->vo, queue, opts->renderer_opts->interpolation, 1);
vo_set_queue_params(p->vo, 0, opts->renderer_opts->interpolation, queue);
p->vo->want_redraw = true;
}
@ -359,7 +358,7 @@ static int control(struct vo *vo, uint32_t request, void *data)
case VOCTRL_REDRAW_FRAME:
if (!(p->glctx->start_frame && !p->glctx->start_frame(p->glctx))) {
p->frame_started = true;
gl_video_render_frame(p->renderer, 0, NULL);
gl_video_render_frame(p->renderer, NULL, 0, NULL);
}
return true;
case VOCTRL_SET_COMMAND_LINE: {
@ -443,7 +442,7 @@ static int preinit(struct vo *vo)
p->glctx->depth_b);
int queue = 0;
gl_video_set_options(p->renderer, p->renderer_opts, &queue);
vo_set_queue_params(p->vo, queue, p->renderer_opts->interpolation, 0);
vo_set_queue_params(p->vo, 0, p->renderer_opts->interpolation, queue);
p->cms = gl_lcms_init(p, vo->log, vo->global);
if (!p->cms)

2
video/out/vo_opengl_cb.c
View File

@ -372,7 +372,7 @@ int mpv_opengl_cb_draw(mpv_opengl_cb_context *ctx, int fbo, int vp_w, int vp_h)
if (mpi)
gl_video_set_image(ctx->renderer, mpi);
gl_video_render_frame(ctx->renderer, fbo, timing.pts ? &timing : NULL);
gl_video_render_frame(ctx->renderer, mpi, fbo, timing.pts ? &timing : NULL);
gl_video_unset_gl_state(ctx->renderer);