vo_gpu: invalidate fbotex before drawing

Don't discard the OSD or pass_draw_to_screen passes though. Could be
faster on some hardware.

4 files changed, 11 insertions, 10 deletions

diff --git a/video/out/gpu/osd.c b/video/out/gpu/osd.c
index 317deb6e4d..75f69f0c9b 100644
--- a/video/out/gpu/osd.c
+++ b/video/out/gpu/osd.c
@@ -314,7 +314,7 @@ void mpgl_osd_draw_finish(struct mpgl_osd *ctx, int index,
     const int *factors = &blend_factors[part->format][0];
     gl_sc_blend(sc, factors[0], factors[1], factors[2], factors[3]);
 
-    gl_sc_dispatch_draw(sc, fbo.tex, vertex_vao, MP_ARRAY_SIZE(vertex_vao),
+    gl_sc_dispatch_draw(sc, fbo.tex, false, vertex_vao, MP_ARRAY_SIZE(vertex_vao),
                         sizeof(struct vertex), part->vertices, part->num_vertices);
 }
 
diff --git a/video/out/gpu/shader_cache.c b/video/out/gpu/shader_cache.c
index 6d0f370568..780bf7f5d7 100644
--- a/video/out/gpu/shader_cache.c
+++ b/video/out/gpu/shader_cache.c
@@ -965,13 +965,14 @@ static void gl_sc_generate(struct gl_shader_cache *sc,
 }
 
 struct mp_pass_perf gl_sc_dispatch_draw(struct gl_shader_cache *sc,
-                                        struct ra_tex *target,
+                                        struct ra_tex *target, bool discard,
                                         const struct ra_renderpass_input *vao,
                                         int vao_len, size_t vertex_stride,
                                         void *vertices, size_t num_vertices)
 {
     struct timer_pool *timer = NULL;
 
+    sc->params.invalidate_target = discard;
     gl_sc_generate(sc, RA_RENDERPASS_TYPE_RASTER, target->params.format,
                    vao, vao_len, vertex_stride);
     if (!sc->current_shader)
diff --git a/video/out/gpu/shader_cache.h b/video/out/gpu/shader_cache.h
index 2fe7dcfb9d..547c6b6307 100644
--- a/video/out/gpu/shader_cache.h
+++ b/video/out/gpu/shader_cache.h
@@ -50,7 +50,7 @@ void gl_sc_blend(struct gl_shader_cache *sc,
                  enum ra_blend blend_dst_alpha);
 void gl_sc_enable_extension(struct gl_shader_cache *sc, char *name);
 struct mp_pass_perf gl_sc_dispatch_draw(struct gl_shader_cache *sc,
-                                        struct ra_tex *target,
+                                        struct ra_tex *target, bool discard,
                                         const struct ra_renderpass_input *vao,
                                         int vao_len, size_t vertex_stride,
                                         void *ptr, size_t num);
diff --git a/video/out/gpu/video.c b/video/out/gpu/video.c
index f80d63a8b1..3f0959931d 100644
--- a/video/out/gpu/video.c
+++ b/video/out/gpu/video.c
@@ -1153,7 +1153,7 @@ static void dispatch_compute(struct gl_video *p, int w, int h,
 }
 
 static struct mp_pass_perf render_pass_quad(struct gl_video *p,
-                                            struct ra_fbo fbo,
+                                            struct ra_fbo fbo, bool discard,
                                             const struct mp_rect *dst)
 {
     // The first element is reserved for `vec2 position`
@@ -1211,15 +1211,15 @@ static struct mp_pass_perf render_pass_quad(struct gl_video *p,
             &p->tmp_vertex[num_vertex_attribs * 1],
             vertex_stride);
 
-    return gl_sc_dispatch_draw(p->sc, fbo.tex, p->vao, num_vertex_attribs,
+    return gl_sc_dispatch_draw(p->sc, fbo.tex, discard, p->vao, num_vertex_attribs,
                                vertex_stride, p->tmp_vertex, num_vertices);
 }
 
 static void finish_pass_fbo(struct gl_video *p, struct ra_fbo fbo,
-                            const struct mp_rect *dst)
+                            bool discard, const struct mp_rect *dst)
 {
     pass_prepare_src_tex(p);
-    pass_record(p, render_pass_quad(p, fbo, dst));
+    pass_record(p, render_pass_quad(p, fbo, discard, dst));
     debug_check_gl(p, "after rendering");
     cleanup_binds(p);
 }
@@ -1248,7 +1248,7 @@ static void finish_pass_tex(struct gl_video *p, struct ra_tex **dst_tex,
         debug_check_gl(p, "after dispatching compute shader");
     } else {
         struct ra_fbo fbo = { .tex = *dst_tex, };
-        finish_pass_fbo(p, fbo, &(struct mp_rect){0, 0, w, h});
+        finish_pass_fbo(p, fbo, true, &(struct mp_rect){0, 0, w, h});
     }
 }
 
@@ -2799,7 +2799,7 @@ static void pass_draw_to_screen(struct gl_video *p, struct ra_fbo fbo)
 
     pass_dither(p);
     pass_describe(p, "output to screen");
-    finish_pass_fbo(p, fbo, &p->dst_rect);
+    finish_pass_fbo(p, fbo, false, &p->dst_rect);
 }
 
 static bool update_surface(struct gl_video *p, struct mp_image *mpi,
@@ -3205,7 +3205,7 @@ static void reinterleave_vdpau(struct gl_video *p,
         const struct ra_format *fmt = ra_find_unorm_format(p->ra, 1, comps);
         ra_tex_resize(p->ra, p->log, tex, w, h * 2, fmt);
         struct ra_fbo fbo = { *tex };
-        finish_pass_fbo(p, fbo, &(struct mp_rect){0, 0, w, h * 2});
+        finish_pass_fbo(p, fbo, true, &(struct mp_rect){0, 0, w, h * 2});
 
         output[n] = *tex;
     }