vo_gpu: attempt to avoid UBOs for dynamic variables

This makes the radeon driver shut up about frequently updating
STATIC_DRAW UBOs (--opengl-debug), and also reduces the amount of
synchronization necessary for vulkan uniform buffers.

Also add some extra debugging/tracing code paths. I went with a
flags-based approach in case we ever want to extend this.

video/out/gpu/shader_cache.c

@@ -76,6 +76,7 @@ struct gl_shader_cache {
     // Next binding point (texture unit, image unit, buffer binding, etc.)
     // In OpenGL these are separate for each input type
     int next_binding[RA_VARTYPE_COUNT];
+    bool next_uniform_dynamic;
 
     struct ra_renderpass_params params;
 
@@ -135,6 +136,7 @@ void gl_sc_reset(struct gl_shader_cache *sc)
     sc->pushc_size = 0;
     for (int i = 0; i < RA_VARTYPE_COUNT; i++)
         sc->next_binding[i] = 0;
+    sc->next_uniform_dynamic = false;
     sc->current_shader = NULL;
     sc->params = (struct ra_renderpass_params){0};
     sc->needs_reset = false;
@@ -259,14 +261,22 @@ static int gl_sc_next_binding(struct gl_shader_cache *sc, enum ra_vartype type)
     }
 }
 
+void gl_sc_uniform_dynamic(struct gl_shader_cache *sc)
+{
+    sc->next_uniform_dynamic = true;
+}
+
 // Updates the metadata for the given sc_uniform. Assumes sc_uniform->input
 // and glsl_type/buffer_format are already set.
 static void update_uniform_params(struct gl_shader_cache *sc, struct sc_uniform *u)
 {
+    bool dynamic = sc->next_uniform_dynamic;
+    sc->next_uniform_dynamic = false;
+
     // Try not using push constants for "large" values like matrices, since
     // this is likely to both exceed the VGPR budget as well as the pushc size
     // budget
-    bool try_pushc = u->input.dim_m == 1;
+    bool try_pushc = u->input.dim_m == 1 || dynamic;
 
     // Attempt using push constants first
     if (try_pushc && sc->ra->glsl_vulkan && sc->ra->max_pushc_size) {
@@ -287,7 +297,10 @@ static void update_uniform_params(struct gl_shader_cache *sc, struct sc_uniform
     // to explicit offsets on UBO entries. In theory we could leave away
     // the offsets and support UBOs for older GL as well, but this is a nice
     // safety net for driver bugs (and also rules out potentially buggy drivers)
-    if (sc->ra->glsl_version >= 440 && (sc->ra->caps & RA_CAP_BUF_RO)) {
+    // Also avoid UBOs for highly dynamic stuff since that requires synchronizing
+    // the UBO writes every frame
+    bool try_ubo = !(sc->ra->caps & RA_CAP_GLOBAL_UNIFORM) || !dynamic;
+    if (try_ubo && sc->ra->glsl_version >= 440 && (sc->ra->caps & RA_CAP_BUF_RO)) {
         u->type = SC_UNIFORM_TYPE_UBO;
         u->layout = sc->ra->fns->uniform_layout(&u->input);
         u->offset = MP_ALIGN_UP(sc->ubo_size, u->layout.align);
@@ -514,6 +527,13 @@ static void update_uniform(struct gl_shader_cache *sc, struct sc_entry *e,
     un->v = u->v;
     un->set = true;
 
+    static const char *desc[] = {
+        [SC_UNIFORM_TYPE_UBO]    = "UBO",
+        [SC_UNIFORM_TYPE_PUSHC]  = "PC",
+        [SC_UNIFORM_TYPE_GLOBAL] = "global",
+    };
+    MP_TRACE(sc, "Updating %s uniform '%s'\n", desc[u->type], u->input.name);
+
     switch (u->type) {
     case SC_UNIFORM_TYPE_GLOBAL: {
         struct ra_renderpass_input_val value = {

video/out/gpu/shader_cache.h

@@ -25,6 +25,10 @@ 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_paddf(struct gl_shader_cache *sc, const char *textf, ...)
     PRINTF_ATTRIBUTE(2, 3);
+
+// A hint that the next data-type (i.e. non-binding) uniform is expected to
+// change frequently. This refers to the _f, _i, _vecN etc. uniform types.
+void gl_sc_uniform_dynamic(struct gl_shader_cache *sc);
 void gl_sc_uniform_texture(struct gl_shader_cache *sc, char *name,
                            struct ra_tex *tex);
 void gl_sc_uniform_image2D_wo(struct gl_shader_cache *sc, const char *name,

video/out/gpu/video.c

@@ -1500,7 +1500,9 @@ found:
 static void load_shader(struct gl_video *p, struct bstr body)
 {
     gl_sc_hadd_bstr(p->sc, body);
+    gl_sc_uniform_dynamic(p->sc);
     gl_sc_uniform_f(p->sc, "random", (double)av_lfg_get(&p->lfg) / UINT32_MAX);
+    gl_sc_uniform_dynamic(p->sc);
     gl_sc_uniform_i(p->sc, "frame", p->frames_uploaded);
     gl_sc_uniform_vec2(p->sc, "input_size",
                        (float[]){(p->src_rect.x1 - p->src_rect.x0) *
@@ -2579,6 +2581,7 @@ static void pass_dither(struct gl_video *p)
 
         float matrix[2][2] = {{cos(r),     -sin(r)    },
                               {sin(r) * m,  cos(r) * m}};
+        gl_sc_uniform_dynamic(p->sc);
         gl_sc_uniform_mat2(p->sc, "dither_trafo", true, &matrix[0][0]);
 
         GLSL(dither_pos = dither_trafo * dither_pos;)
@@ -2957,11 +2960,13 @@ static void gl_video_interpolate_frame(struct gl_video *p, struct vo_frame *t,
 
         // Blend the frames together
         if (oversample || linear) {
+            gl_sc_uniform_dynamic(p->sc);
             gl_sc_uniform_f(p->sc, "inter_coeff", mix);
             GLSL(color = mix(texture(texture0, texcoord0),
                              texture(texture1, texcoord1),
                              inter_coeff);)
         } else {
+            gl_sc_uniform_dynamic(p->sc);
             gl_sc_uniform_f(p->sc, "fcoord", mix);
             pass_sample_separated_gen(p->sc, tscale, 0, 0);
         }

video/out/gpu/video_shaders.c

@@ -768,6 +768,7 @@ static void prng_init(struct gl_shader_cache *sc, AVLFG *lfg)
     // Initialize the PRNG by hashing the position + a random uniform
     GLSL(vec3 _m = vec3(HOOKED_pos, random) + vec3(1.0);)
     GLSL(float h = permute(permute(permute(_m.x)+_m.y)+_m.z);)
+    gl_sc_uniform_dynamic(sc);
     gl_sc_uniform_f(sc, "random", (double)av_lfg_get(lfg) / UINT32_MAX);
 }