Merge branch 'wip/chergert/fewer-uniforms' into 'master'

ngl: apply fewer uniforms

See merge request GNOME/gtk!3289

gsk/ngl/gsknglcommandqueue.c

@@ -731,98 +731,6 @@ gsk_ngl_command_queue_delete_program (GskNglCommandQueue *self,
   glDeleteProgram (program);
 }
 
-static inline void
-apply_uniform (gconstpointer     dataptr,
-               GskNglUniformInfo info,
-               guint             location)
-{
-  g_assert (dataptr != NULL);
-  g_assert (info.format > 0);
-  g_assert (location < GL_MAX_UNIFORM_LOCATIONS);
-
-  switch (info.format)
-    {
-    case GSK_NGL_UNIFORM_FORMAT_1F:
-      glUniform1fv (location, 1, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_2F:
-      glUniform2fv (location, 1, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_3F:
-      glUniform3fv (location, 1, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_4F:
-      glUniform4fv (location, 1, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_1FV:
-      glUniform1fv (location, info.array_count, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_2FV:
-      glUniform2fv (location, info.array_count, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_3FV:
-      glUniform3fv (location, info.array_count, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_4FV:
-      glUniform4fv (location, info.array_count, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_1I:
-    case GSK_NGL_UNIFORM_FORMAT_TEXTURE:
-      glUniform1iv (location, 1, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_2I:
-      glUniform2iv (location, 1, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_3I:
-      glUniform3iv (location, 1, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_4I:
-      glUniform4iv (location, 1, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_1UI:
-      glUniform1uiv (location, 1, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_MATRIX: {
-      float mat[16];
-      graphene_matrix_to_float (dataptr, mat);
-      glUniformMatrix4fv (location, 1, GL_FALSE, mat);
-#if 0
-      /* TODO: If Graphene can give us a peek here on platforms
-       * where the format is float[16] (most/all x86_64?) then
-       * We can avoid the SIMD operation to convert the format.
-       */
-      G_STATIC_ASSERT (sizeof (graphene_matrix_t) == 16*4);
-      glUniformMatrix4fv (location, 1, GL_FALSE, dataptr);
-#endif
-    }
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_COLOR:
-      glUniform4fv (location, 1, dataptr);
-    break;
-
-    case GSK_NGL_UNIFORM_FORMAT_ROUNDED_RECT:
-      glUniform4fv (location, 3, dataptr);
-    break;
-
-    default:
-      g_assert_not_reached ();
-    }
-}
-
 static inline void
 apply_viewport (guint *current_width,
                 guint *current_height,
@@ -1193,8 +1101,7 @@ gsk_ngl_command_queue_execute (GskNglCommandQueue   *self,
               const GskNglCommandUniform *u = &self->batch_uniforms.items[batch->draw.uniform_offset];
 
               for (guint i = 0; i < batch->draw.uniform_count; i++, u++)
-                apply_uniform (gsk_ngl_uniform_state_get_uniform_data (self->uniforms, u->info.offset),
+                gsk_ngl_uniform_state_apply (self->uniforms, program, u->location, u->info);
-                               u->info, u->location);
 
               n_uniforms += batch->draw.uniform_count;
             }

gsk/ngl/gskngluniformstate.c

@@ -65,6 +65,8 @@ gsk_ngl_uniform_state_new (void)
   state->values_pos = 0;
   state->values_buf = g_malloc (4096);
 
+  memset (state->apply_hash, 0, sizeof state->apply_hash);
+
   return g_steal_pointer (&state);
 }
 
@@ -221,6 +223,8 @@ gsk_ngl_uniform_state_end_frame (GskNglUniformState *state)
   state->values_pos = allocator;
 
   g_assert (allocator <= state->values_len);
+
+  memset (state->apply_hash, 0, sizeof state->apply_hash);
 }
 
 gsize

gsk/ngl/gskngluniformstateprivate.h

@@ -83,6 +83,7 @@ typedef struct _GskNglUniformState
   guint8 *values_buf;
   guint values_pos;
   guint values_len;
+  GskNglUniformInfo apply_hash[512];
 } GskNglUniformState;
 
 /**
@@ -272,12 +273,15 @@ gsk_ngl_uniform_state_set1f (GskNglUniformState   *state,
   g_assert (program != 0);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_1F, 1, location, stamp, &info)))
+    {
+      if (info->info.initial || u->v0 != value0)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform1f , 1);
           u->v0 = value0;
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set2f (GskNglUniformState   *state,
@@ -294,6 +298,8 @@ gsk_ngl_uniform_state_set2f (GskNglUniformState   *state,
   g_assert (program != NULL);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_2F, 1, location, stamp, &info)))
+    {
+      if (info->info.initial || u->v0 != value0 || u->v1 != value1)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform2f, 1);
           u->v0 = value0;
@@ -301,6 +307,7 @@ gsk_ngl_uniform_state_set2f (GskNglUniformState   *state,
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set3f (GskNglUniformState   *state,
@@ -318,6 +325,8 @@ gsk_ngl_uniform_state_set3f (GskNglUniformState   *state,
   g_assert (program != NULL);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_3F, 1, location, stamp, &info)))
+    {
+      if (info->info.initial || u->v0 != value0 || u->v1 != value1 || u->v2 != value2)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform3f, 1);
           u->v0 = value0;
@@ -326,6 +335,7 @@ gsk_ngl_uniform_state_set3f (GskNglUniformState   *state,
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set4f (GskNglUniformState   *state,
@@ -344,6 +354,8 @@ gsk_ngl_uniform_state_set4f (GskNglUniformState   *state,
   g_assert (program != NULL);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_4F, 1, location, stamp, &info)))
+    {
+      if (info->info.initial || u->v0 != value0 || u->v1 != value1 || u->v2 != value2 || u->v3 != value3)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform4f, 1);
           u->v0 = value0;
@@ -353,6 +365,7 @@ gsk_ngl_uniform_state_set4f (GskNglUniformState   *state,
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set1ui (GskNglUniformState   *state,
@@ -368,12 +381,15 @@ gsk_ngl_uniform_state_set1ui (GskNglUniformState   *state,
   g_assert (program != NULL);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_1UI, 1, location, stamp, &info)))
+    {
+      if (info->info.initial || u->v0 != value0)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform1ui, 1);
           u->v0 = value0;
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set1i (GskNglUniformState   *state,
@@ -389,12 +405,15 @@ gsk_ngl_uniform_state_set1i (GskNglUniformState   *state,
   g_assert (program != NULL);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_1I, 1, location, stamp, &info)))
+    {
+      if (info->info.initial || u->v0 != value0)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform1i, 1);
           u->v0 = value0;
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set2i (GskNglUniformState   *state,
@@ -411,6 +430,8 @@ gsk_ngl_uniform_state_set2i (GskNglUniformState   *state,
   g_assert (program != NULL);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_2I, 1, location, stamp, &info)))
+    {
+      if (info->info.initial || u->v0 != value0 || u->v1 != value1)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform2i, 1);
           u->v0 = value0;
@@ -418,6 +439,7 @@ gsk_ngl_uniform_state_set2i (GskNglUniformState   *state,
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set3i (GskNglUniformState   *state,
@@ -435,6 +457,8 @@ gsk_ngl_uniform_state_set3i (GskNglUniformState   *state,
   g_assert (program != NULL);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_3I, 1, location, stamp, &info)))
+    {
+      if (info->info.initial || u->v0 != value0 || u->v1 != value1 || u->v2 != value2)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform3i, 1);
           u->v0 = value0;
@@ -443,6 +467,7 @@ gsk_ngl_uniform_state_set3i (GskNglUniformState   *state,
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set4i (GskNglUniformState   *state,
@@ -461,6 +486,8 @@ gsk_ngl_uniform_state_set4i (GskNglUniformState   *state,
   g_assert (program != NULL);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_4I, 1, location, stamp, &info)))
+    {
+      if (info->info.initial || u->v0 != value0 || u->v1 != value1 || u->v2 != value2 || u->v3 != value3)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform4i, 1);
           u->v0 = value0;
@@ -470,6 +497,7 @@ gsk_ngl_uniform_state_set4i (GskNglUniformState   *state,
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set_rounded_rect (GskNglUniformState   *state,
@@ -486,12 +514,15 @@ gsk_ngl_uniform_state_set_rounded_rect (GskNglUniformState   *state,
   g_assert (rounded_rect != NULL);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_ROUNDED_RECT, 1, location, stamp, &info)))
+    {
+      if (info->info.initial || memcmp (u, rounded_rect, sizeof *u) != 0)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, GskRoundedRect, 1);
           memcpy (u, rounded_rect, sizeof *rounded_rect);
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set_matrix (GskNglUniformState      *state,
@@ -508,12 +539,15 @@ gsk_ngl_uniform_state_set_matrix (GskNglUniformState      *state,
   g_assert (matrix != NULL);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_MATRIX, 1, location, stamp, &info)))
+    {
+      if (info->info.initial || memcmp (u, matrix, sizeof *u) != 0)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, graphene_matrix_t, 1);
           memcpy (u, matrix, sizeof *matrix);
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 /**
  * gsk_ngl_uniform_state_set_texture:
@@ -545,12 +579,15 @@ gsk_ngl_uniform_state_set_texture (GskNglUniformState   *state,
   texture_slot -= GL_TEXTURE0;
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_TEXTURE, 1, location, stamp, &info)))
+    {
+      if (info->info.initial || *u != texture_slot)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, guint, 1);
           *u = texture_slot;
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 /**
  * gsk_ngl_uniform_state_set_color:
@@ -582,11 +619,14 @@ gsk_ngl_uniform_state_set_color (GskNglUniformState   *state,
       if (color == NULL)
         color = &transparent;
 
+      if (info->info.initial || memcmp (color, u, sizeof *u) != 0)
+        {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, GdkRGBA, 1);
           memcpy (u, color, sizeof *color);
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set1fv (GskNglUniformState   *state,
@@ -604,12 +644,15 @@ gsk_ngl_uniform_state_set1fv (GskNglUniformState   *state,
   g_assert (count > 0);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_1FV, count, location, stamp, &info)))
+    {
+      if (info->info.initial || count != info->info.array_count || memcmp (u, value, sizeof *u * count) != 0)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform1f, count);
           memcpy (u, value, sizeof (Uniform1f) * count);
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set2fv (GskNglUniformState   *state,
@@ -627,12 +670,15 @@ gsk_ngl_uniform_state_set2fv (GskNglUniformState   *state,
   g_assert (count > 0);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_2FV, count, location, stamp, &info)))
+    {
+      if (info->info.initial || count != info->info.array_count || memcmp (u, value, sizeof *u * count) != 0)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform2f, count);
           memcpy (u, value, sizeof (Uniform2f) * count);
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set3fv (GskNglUniformState   *state,
@@ -650,12 +696,15 @@ gsk_ngl_uniform_state_set3fv (GskNglUniformState   *state,
   g_assert (count > 0);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_3FV, count, location, stamp, &info)))
+    {
+      if (info->info.initial || count != info->info.array_count || memcmp (u, value, sizeof *u * count) != 0)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform3f, count);
           memcpy (u, value, sizeof (Uniform3f) * count);
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
 
 static inline void
 gsk_ngl_uniform_state_set4fv (GskNglUniformState   *state,
@@ -673,12 +722,146 @@ gsk_ngl_uniform_state_set4fv (GskNglUniformState   *state,
   g_assert (count > 0);
 
   if ((u = gsk_ngl_uniform_state_get_value (state, program, GSK_NGL_UNIFORM_FORMAT_4FV, count, location, stamp, &info)))
+    {
+      if (info->info.initial || count != info->info.array_count || memcmp (u, value, sizeof *u * count) != 0)
         {
           GSK_NGL_UNIFORM_STATE_REPLACE (info, u, Uniform4f, count);
           memcpy (u, value, sizeof (Uniform4f) * count);
           gsk_ngl_uniform_info_changed (info, location, stamp);
         }
     }
+}
+
+static inline guint
+gsk_ngl_uniform_state_fmix (guint program,
+                            guint location)
+{
+  guint h = (program << 16) | location;
+
+   h ^= h >> 16;
+   h *= 0x85ebca6b;
+   h ^= h >> 13;
+   h *= 0xc2b2ae35;
+   h ^= h >> 16;
+
+   return h;
+}
+
+/*
+ * gsk_ngl_uniform_state_apply:
+ * @state: the uniform state
+ * @program: the program id
+ * @location: the location of the uniform
+ * @offset: the offset of the data within the buffer
+ * @info: the uniform info
+ *
+ * This function can be used to apply state that was previously recorded
+ * by the #GskNglUniformState.
+ *
+ * It is specifically useful from the GskNglCommandQueue to execute uniform
+ * changes but only when they have changed from the current value.
+ */
+static inline void
+gsk_ngl_uniform_state_apply (GskNglUniformState *state,
+                             guint               program,
+                             guint               location,
+                             GskNglUniformInfo   info)
+{
+  guint index = gsk_ngl_uniform_state_fmix (program, location) % G_N_ELEMENTS (state->apply_hash);
+  gconstpointer dataptr = GSK_NGL_UNIFORM_VALUE (state->values_buf, info.offset);
+
+  /* aligned, can treat as unsigned */
+  if (*(guint *)&info == *(guint *)&state->apply_hash[index])
+    return;
+
+  state->apply_hash[index] = info;
+
+  /* TODO: We could do additional comparisons here to make sure we are
+   *       changing state.
+   */
+
+  switch (info.format)
+    {
+    case GSK_NGL_UNIFORM_FORMAT_1F:
+      glUniform1fv (location, 1, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_2F:
+      glUniform2fv (location, 1, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_3F:
+      glUniform3fv (location, 1, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_4F:
+      glUniform4fv (location, 1, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_1FV:
+      glUniform1fv (location, info.array_count, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_2FV:
+      glUniform2fv (location, info.array_count, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_3FV:
+      glUniform3fv (location, info.array_count, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_4FV:
+      glUniform4fv (location, info.array_count, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_1I:
+    case GSK_NGL_UNIFORM_FORMAT_TEXTURE:
+      glUniform1iv (location, 1, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_2I:
+      glUniform2iv (location, 1, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_3I:
+      glUniform3iv (location, 1, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_4I:
+      glUniform4iv (location, 1, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_1UI:
+      glUniform1uiv (location, 1, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_MATRIX: {
+      float mat[16];
+      graphene_matrix_to_float (dataptr, mat);
+      glUniformMatrix4fv (location, 1, GL_FALSE, mat);
+#if 0
+      /* TODO: If Graphene can give us a peek here on platforms
+       * where the format is float[16] (most/all x86_64?) then
+       * We can avoid the SIMD operation to convert the format.
+       */
+      G_STATIC_ASSERT (sizeof (graphene_matrix_t) == 16*4);
+      glUniformMatrix4fv (location, 1, GL_FALSE, dataptr);
+#endif
+    }
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_COLOR:
+      glUniform4fv (location, 1, dataptr);
+    break;
+
+    case GSK_NGL_UNIFORM_FORMAT_ROUNDED_RECT:
+      glUniform4fv (location, 3, dataptr);
+    break;
+
+    default:
+      g_assert_not_reached ();
+    }
+}
 
 G_END_DECLS