From 1c8131167334183ad24b90cbbf4fd86d3d5406cc Mon Sep 17 00:00:00 2001
From: wm4 <wm4@nowhere>
Date: Sat, 22 Jul 2017 21:46:59 +0200
Subject: vo_opengl: osd: refactor and simplify

Reduce this to 1 draw call per OSD pass. This removes the need for some
annoying special handling regarding 3D video support (we supported
duplicating the OSD/subtitles for side-by-side 3D output etc.).

Remove the unneeded texture sampler uniform thing.
---
 video/out/opengl/osd.c   | 58 ++++++++++++++++++++++++++----------------------
 video/out/opengl/osd.h   |  1 +
 video/out/opengl/utils.c | 13 -----------
 video/out/opengl/utils.h |  2 --
 video/out/opengl/video.c |  3 ++-
 5 files changed, 34 insertions(+), 43 deletions(-)

diff --git a/video/out/opengl/osd.c b/video/out/opengl/osd.c
index 5df5bb199a..64b08a67fc 100644
--- a/video/out/opengl/osd.c
+++ b/video/out/opengl/osd.c
@@ -55,6 +55,7 @@ struct mpgl_osd_part {
     int num_subparts;
     int prev_num_subparts;
     struct sub_bitmap *subparts;
+    int num_vertices;
     struct vertex *vertices;
 };
 
@@ -226,14 +227,14 @@ static void write_quad(struct vertex *va, struct gl_transform t,
 #undef COLOR_INIT
 }
 
-static int generate_verts(struct mpgl_osd_part *part, struct gl_transform t)
+static void generate_verts(struct mpgl_osd_part *part, struct gl_transform t)
 {
     int num_vertices = part->num_subparts * 6;
-    MP_TARRAY_GROW(part, part->vertices, num_vertices);
+    MP_TARRAY_GROW(part, part->vertices, part->num_vertices + num_vertices);
 
     for (int n = 0; n < part->num_subparts; n++) {
         struct sub_bitmap *b = &part->subparts[n];
-        struct vertex *va = part->vertices;
+        struct vertex *va = &part->vertices[part->num_vertices];
 
         // NOTE: the blend color is used with SUBBITMAP_LIBASS only, so it
         //       doesn't matter that we upload garbage for the other formats
@@ -247,28 +248,7 @@ static int generate_verts(struct mpgl_osd_part *part, struct gl_transform t)
                    part->w, part->h, color);
     }
 
-    return num_vertices;
-}
-
-static void draw_part(struct mpgl_osd *ctx, int index, struct gl_transform t)
-{
-    GL *gl = ctx->gl;
-    struct mpgl_osd_part *part = ctx->parts[index];
-
-    int num_vertices = generate_verts(part, t);
-    if (!num_vertices)
-        return;
-
-    gl->Enable(GL_BLEND);
-    gl->BindTexture(GL_TEXTURE_2D, part->texture);
-
-    const int *factors = &blend_factors[part->format][0];
-    gl->BlendFuncSeparate(factors[0], factors[1], factors[2], factors[3]);
-
-    gl_vao_draw_data(&ctx->vao, GL_TRIANGLES, part->vertices, num_vertices);
-
-    gl->BindTexture(GL_TEXTURE_2D, 0);
-    gl->Disable(GL_BLEND);
+    part->num_vertices += num_vertices;
 }
 
 // number of screen divisions per axis (x=0, y=1) for the current 3D mode
@@ -285,10 +265,13 @@ static void get_3d_side_by_side(int stereo_mode, int div[2])
 
 void mpgl_osd_draw_part(struct mpgl_osd *ctx, int vp_w, int vp_h, int index)
 {
+    GL *gl = ctx->gl;
+    struct mpgl_osd_part *part = ctx->parts[index];
+
     int div[2];
     get_3d_side_by_side(ctx->stereo_mode, div);
 
-    ctx->gl->Viewport(0, 0, vp_w, abs(vp_h));
+    part->num_vertices = 0;
 
     for (int x = 0; x < div[0]; x++) {
         for (int y = 0; y < div[1]; y++) {
@@ -300,9 +283,24 @@ void mpgl_osd_draw_part(struct mpgl_osd *ctx, int vp_w, int vp_h, int index)
             t.t[0] += a_x * t.m[0][0] + a_y * t.m[1][0];
             t.t[1] += a_x * t.m[0][1] + a_y * t.m[1][1];
 
-            draw_part(ctx, index, t);
+            generate_verts(part, t);
         }
     }
+
+    if (!part->num_vertices)
+        return;
+
+    gl->Enable(GL_BLEND);
+
+    const int *factors = &blend_factors[part->format][0];
+    gl->BlendFuncSeparate(factors[0], factors[1], factors[2], factors[3]);
+
+    ctx->gl->Viewport(0, 0, vp_w, abs(vp_h));
+
+    gl_vao_draw_data(&ctx->vao, GL_TRIANGLES, part->vertices, part->num_vertices);
+
+    gl->BindTexture(GL_TEXTURE_2D, 0);
+    gl->Disable(GL_BLEND);
 }
 
 enum sub_bitmap_format mpgl_osd_get_part_format(struct mpgl_osd *ctx, int index)
@@ -311,6 +309,12 @@ enum sub_bitmap_format mpgl_osd_get_part_format(struct mpgl_osd *ctx, int index)
     return ctx->parts[index]->format;
 }
 
+GLuint mpgl_osd_get_part_texture(struct mpgl_osd *ctx, int index)
+{
+    assert(index >= 0 && index < MAX_OSD_PARTS);
+    return ctx->parts[index]->texture;
+}
+
 struct gl_vao *mpgl_osd_get_vao(struct mpgl_osd *ctx)
 {
     return &ctx->vao;
diff --git a/video/out/opengl/osd.h b/video/out/opengl/osd.h
index 34cbf2b383..197bb72501 100644
--- a/video/out/opengl/osd.h
+++ b/video/out/opengl/osd.h
@@ -16,6 +16,7 @@ void mpgl_osd_generate(struct mpgl_osd *ctx, struct mp_osd_res res, double pts,
                        int stereo_mode, int draw_flags);
 void mpgl_osd_resize(struct mpgl_osd *ctx, struct mp_osd_res res, int stereo_mode);
 enum sub_bitmap_format mpgl_osd_get_part_format(struct mpgl_osd *ctx, int index);
+GLuint mpgl_osd_get_part_texture(struct mpgl_osd *ctx, int index);
 struct gl_vao *mpgl_osd_get_vao(struct mpgl_osd *ctx);
 void mpgl_osd_draw_part(struct mpgl_osd *ctx, int vp_w, int vp_h, int index);
 int64_t mpgl_get_change_counter(struct mpgl_osd *ctx);
diff --git a/video/out/opengl/utils.c b/video/out/opengl/utils.c
index 4c453b07e8..d3066d3ccb 100644
--- a/video/out/opengl/utils.c
+++ b/video/out/opengl/utils.c
@@ -648,19 +648,6 @@ const char *mp_sampler_type(GLenum texture_target)
     }
 }
 
-// gl_sc_uniform_tex() should be preferred.
-void gl_sc_uniform_sampler(struct gl_shader_cache *sc, char *name, GLenum target,
-                           int unit)
-{
-    struct sc_uniform *u = find_uniform(sc, name);
-    u->type = UT_i;
-    u->size = 1;
-    u->glsl_type = mp_sampler_type(target);
-    u->v.i[0] = unit;
-    u->tex_target = 0;
-    u->tex_handle = 0;
-}
-
 void gl_sc_uniform_tex(struct gl_shader_cache *sc, char *name, GLenum target,
                        GLuint texture)
 {
diff --git a/video/out/opengl/utils.h b/video/out/opengl/utils.h
index c55670b8d6..191565dd93 100644
--- a/video/out/opengl/utils.h
+++ b/video/out/opengl/utils.h
@@ -154,8 +154,6 @@ void gl_sc_addf(struct gl_shader_cache *sc, const char *textf, ...);
 void gl_sc_hadd(struct gl_shader_cache *sc, const char *text);
 void gl_sc_haddf(struct gl_shader_cache *sc, const char *textf, ...);
 void gl_sc_hadd_bstr(struct gl_shader_cache *sc, struct bstr text);
-void gl_sc_uniform_sampler(struct gl_shader_cache *sc, char *name, GLenum target,
-                           int unit);
 void gl_sc_uniform_tex(struct gl_shader_cache *sc, char *name, GLenum target,
                        GLuint texture);
 void gl_sc_uniform_tex_ui(struct gl_shader_cache *sc, char *name, GLuint texture);
diff --git a/video/out/opengl/video.c b/video/out/opengl/video.c
index 742c36850a..4b2c8ecc56 100644
--- a/video/out/opengl/video.c
+++ b/video/out/opengl/video.c
@@ -2411,7 +2411,8 @@ static void pass_draw_osd(struct gl_video *p, int draw_flags, double pts,
         enum sub_bitmap_format fmt = mpgl_osd_get_part_format(p->osd, n);
         if (!fmt)
             continue;
-        gl_sc_uniform_sampler(p->sc, "osdtex", GL_TEXTURE_2D, 0);
+        gl_sc_uniform_tex(p->sc, "osdtex", GL_TEXTURE_2D,
+                          mpgl_osd_get_part_texture(p->osd, n));
         switch (fmt) {
         case SUBBITMAP_RGBA: {
             pass_describe(p, "drawing osd (rgba)");
-- 
cgit v1.2.3