Merge pull request #33527 from clayjohn/GLES2-bufferdata_optimization

Improve glBufferSubData usage where safe

7 files changed, 49 insertions, 9 deletions

diff --git a/drivers/gles2/rasterizer_canvas_gles2.cpp b/drivers/gles2/rasterizer_canvas_gles2.cpp
index e34705f7b7..325f2d6707 100644
--- a/drivers/gles2/rasterizer_canvas_gles2.cpp
+++ b/drivers/gles2/rasterizer_canvas_gles2.cpp
@@ -291,6 +291,10 @@ RasterizerStorageGLES2::Texture *RasterizerCanvasGLES2::_bind_canvas_texture(con
 void RasterizerCanvasGLES2::_draw_polygon(const int *p_indices, int p_index_count, int p_vertex_count, const Vector2 *p_vertices, const Vector2 *p_uvs, const Color *p_colors, bool p_singlecolor, const float *p_weights, const int *p_bones) {
 
 	glBindBuffer(GL_ARRAY_BUFFER, data.polygon_buffer);
+#ifndef GLES_OVER_GL
+	// Orphan the buffer to avoid CPU/GPU sync points caused by glBufferSubData
+	glBufferData(GL_ARRAY_BUFFER, data.polygon_buffer_size, NULL, GL_DYNAMIC_DRAW);
+#endif
 
 	uint32_t buffer_ofs = 0;
 
@@ -339,6 +343,11 @@ void RasterizerCanvasGLES2::_draw_polygon(const int *p_indices, int p_index_coun
 	}
 
 	glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, data.polygon_index_buffer);
+#ifndef GLES_OVER_GL
+	// Orphan the buffer to avoid CPU/GPU sync points caused by glBufferSubData
+	glBufferData(GL_ELEMENT_ARRAY_BUFFER, data.polygon_index_buffer_size, NULL, GL_DYNAMIC_DRAW);
+#endif
+
 	if (storage->config.support_32_bits_indices) { //should check for
 		glBufferSubData(GL_ELEMENT_ARRAY_BUFFER, 0, sizeof(int) * p_index_count, p_indices);
 		glDrawElements(GL_TRIANGLES, p_index_count, GL_UNSIGNED_INT, 0);
@@ -358,6 +367,10 @@ void RasterizerCanvasGLES2::_draw_polygon(const int *p_indices, int p_index_coun
 void RasterizerCanvasGLES2::_draw_generic(GLuint p_primitive, int p_vertex_count, const Vector2 *p_vertices, const Vector2 *p_uvs, const Color *p_colors, bool p_singlecolor) {
 
 	glBindBuffer(GL_ARRAY_BUFFER, data.polygon_buffer);
+#ifndef GLES_OVER_GL
+	// Orphan the buffer to avoid CPU/GPU sync points caused by glBufferSubData
+	glBufferData(GL_ARRAY_BUFFER, data.polygon_buffer_size, NULL, GL_DYNAMIC_DRAW);
+#endif
 
 	uint32_t buffer_ofs = 0;
 
@@ -435,6 +448,10 @@ void RasterizerCanvasGLES2::_draw_gui_primitive(int p_points, const Vector2 *p_v
 	}
 
 	glBindBuffer(GL_ARRAY_BUFFER, data.polygon_buffer);
+#ifndef GLES_OVER_GL
+	// Orphan the buffer to avoid CPU/GPU sync points caused by glBufferSubData
+	glBufferData(GL_ARRAY_BUFFER, data.polygon_buffer_size, NULL, GL_DYNAMIC_DRAW);
+#endif
 	glBufferSubData(GL_ARRAY_BUFFER, 0, p_points * stride * 4 * sizeof(float), buffer_data);
 
 	glVertexAttribPointer(VS::ARRAY_VERTEX, 2, GL_FLOAT, GL_FALSE, stride * sizeof(float), NULL);
@@ -883,7 +900,7 @@ void RasterizerCanvasGLES2::_canvas_item_render_commands(Item *p_item, Item *cur
 				}
 
 				glBindBuffer(GL_ARRAY_BUFFER, data.ninepatch_vertices);
-				glBufferSubData(GL_ARRAY_BUFFER, 0, sizeof(float) * (16 + 16) * 2, buffer);
+				glBufferData(GL_ARRAY_BUFFER, sizeof(float) * (16 + 16) * 2, buffer, GL_DYNAMIC_DRAW);
 
 				glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, data.ninepatch_elements);
 
@@ -2034,6 +2051,8 @@ void RasterizerCanvasGLES2::initialize() {
 		glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, data.polygon_index_buffer);
 		glBufferData(GL_ELEMENT_ARRAY_BUFFER, index_size, NULL, GL_DYNAMIC_DRAW);
 		glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, 0);
+
+		data.polygon_index_buffer_size = index_size;
 	}
 
 	// ninepatch buffers
diff --git a/drivers/gles2/rasterizer_canvas_gles2.h b/drivers/gles2/rasterizer_canvas_gles2.h
index ab636dca71..ba636a9763 100644
--- a/drivers/gles2/rasterizer_canvas_gles2.h
+++ b/drivers/gles2/rasterizer_canvas_gles2.h
@@ -65,6 +65,7 @@ public:
 		GLuint polygon_index_buffer;
 
 		uint32_t polygon_buffer_size;
+		uint32_t polygon_index_buffer_size;
 
 		GLuint ninepatch_vertices;
 		GLuint ninepatch_elements;
diff --git a/drivers/gles2/rasterizer_scene_gles2.cpp b/drivers/gles2/rasterizer_scene_gles2.cpp
index 1deecbf3cc..bc47fb46f5 100644
--- a/drivers/gles2/rasterizer_scene_gles2.cpp
+++ b/drivers/gles2/rasterizer_scene_gles2.cpp
@@ -2702,7 +2702,7 @@ void RasterizerSceneGLES2::_draw_sky(RasterizerStorageGLES2::Sky *p_sky, const C
 	}
 
 	glBindBuffer(GL_ARRAY_BUFFER, state.sky_verts);
-	glBufferSubData(GL_ARRAY_BUFFER, 0, sizeof(Vector3) * 8, vertices);
+	glBufferData(GL_ARRAY_BUFFER, sizeof(Vector3) * 8, vertices, GL_DYNAMIC_DRAW);
 
 	// bind sky vertex array....
 	glVertexAttribPointer(VS::ARRAY_VERTEX, 3, GL_FLOAT, GL_FALSE, sizeof(Vector3) * 2, 0);
diff --git a/drivers/gles3/rasterizer_canvas_gles3.cpp b/drivers/gles3/rasterizer_canvas_gles3.cpp
index edffe852a2..55d16a6c2e 100644
--- a/drivers/gles3/rasterizer_canvas_gles3.cpp
+++ b/drivers/gles3/rasterizer_canvas_gles3.cpp
@@ -128,7 +128,7 @@ void RasterizerCanvasGLES3::light_internal_update(RID p_rid, Light *p_light) {
 	li->ubo_data.shadow_distance_mult = (p_light->radius_cache * 1.1);
 
 	glBindBuffer(GL_UNIFORM_BUFFER, li->ubo);
-	glBufferSubData(GL_UNIFORM_BUFFER, 0, sizeof(LightInternal::UBOData), &li->ubo_data);
+	glBufferData(GL_UNIFORM_BUFFER, sizeof(LightInternal::UBOData), &li->ubo_data, GL_DYNAMIC_DRAW);
 	glBindBuffer(GL_UNIFORM_BUFFER, 0);
 }
 
@@ -326,6 +326,12 @@ void RasterizerCanvasGLES3::_draw_polygon(const int *p_indices, int p_index_coun
 	glBindVertexArray(data.polygon_buffer_pointer_array);
 	glBindBuffer(GL_ARRAY_BUFFER, data.polygon_buffer);
 
+#ifndef GLES_OVER_GL
+	// Orphan the buffers to avoid CPU/GPU sync points caused by glBufferSubData
+	glBufferData(GL_ARRAY_BUFFER, data.polygon_buffer_size, NULL, GL_DYNAMIC_DRAW);
+	glBufferData(GL_ELEMENT_ARRAY_BUFFER, data.polygon_index_buffer_size, NULL, GL_DYNAMIC_DRAW);
+#endif
+
 	uint32_t buffer_ofs = 0;
 
 	//vertex
@@ -417,6 +423,12 @@ void RasterizerCanvasGLES3::_draw_generic(GLuint p_primitive, int p_vertex_count
 	glBindVertexArray(data.polygon_buffer_pointer_array);
 	glBindBuffer(GL_ARRAY_BUFFER, data.polygon_buffer);
 
+#ifndef GLES_OVER_GL
+	// Orphan the buffers to avoid CPU/GPU sync points caused by glBufferSubData
+	glBufferData(GL_ARRAY_BUFFER, data.polygon_buffer_size, NULL, GL_DYNAMIC_DRAW);
+	glBufferData(GL_ELEMENT_ARRAY_BUFFER, data.polygon_index_buffer_size, NULL, GL_DYNAMIC_DRAW);
+#endif
+
 	uint32_t buffer_ofs = 0;
 
 	//vertex
@@ -508,6 +520,11 @@ void RasterizerCanvasGLES3::_draw_gui_primitive(int p_points, const Vector2 *p_v
 	}
 
 	glBindBuffer(GL_ARRAY_BUFFER, data.polygon_buffer);
+#ifndef GLES_OVER_GL
+	// Orphan the buffer to avoid CPU/GPU sync points caused by glBufferSubData
+	glBufferData(GL_ARRAY_BUFFER, data.polygon_buffer_size, NULL, GL_DYNAMIC_DRAW);
+#endif
+	//TODO the below call may need to be replaced with: glBufferSubData(GL_ARRAY_BUFFER, 0, p_points * stride * 4 * sizeof(float), &b[0]);
 	glBufferSubData(GL_ARRAY_BUFFER, 0, p_points * stride * 4, &b[0]);
 	glBindVertexArray(data.polygon_buffer_quad_arrays[version]);
 	glDrawArrays(prim[p_points], 0, p_points);
@@ -1882,7 +1899,7 @@ void RasterizerCanvasGLES3::reset_canvas() {
 	state.canvas_item_ubo_data.time = storage->frame.time[0];
 
 	glBindBuffer(GL_UNIFORM_BUFFER, state.canvas_item_ubo);
-	glBufferSubData(GL_UNIFORM_BUFFER, 0, sizeof(CanvasItemUBO), &state.canvas_item_ubo_data);
+	glBufferData(GL_UNIFORM_BUFFER, sizeof(CanvasItemUBO), &state.canvas_item_ubo_data, GL_DYNAMIC_DRAW);
 	glBindBuffer(GL_UNIFORM_BUFFER, 0);
 
 	state.canvas_texscreen_used = false;
@@ -2109,6 +2126,8 @@ void RasterizerCanvasGLES3::initialize() {
 		glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, data.polygon_index_buffer);
 		glBufferData(GL_ELEMENT_ARRAY_BUFFER, index_size, NULL, GL_DYNAMIC_DRAW); //allocate max size
 		glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, 0);
+
+		data.polygon_index_buffer_size = index_size;
 	}
 
 	store_transform(Transform(), state.canvas_item_ubo_data.projection_matrix);
diff --git a/drivers/gles3/rasterizer_canvas_gles3.h b/drivers/gles3/rasterizer_canvas_gles3.h
index bf5ef30820..382b1e5640 100644
--- a/drivers/gles3/rasterizer_canvas_gles3.h
+++ b/drivers/gles3/rasterizer_canvas_gles3.h
@@ -64,6 +64,7 @@ public:
 		GLuint particle_quad_array;
 
 		uint32_t polygon_buffer_size;
+		uint32_t polygon_index_buffer_size;
 
 	} data;
 
diff --git a/drivers/gles3/rasterizer_scene_gles3.cpp b/drivers/gles3/rasterizer_scene_gles3.cpp
index 519fdf2b3b..0820997529 100644
--- a/drivers/gles3/rasterizer_scene_gles3.cpp
+++ b/drivers/gles3/rasterizer_scene_gles3.cpp
@@ -2550,7 +2550,7 @@ void RasterizerSceneGLES3::_draw_sky(RasterizerStorageGLES3::Sky *p_sky, const C
 	}
 
 	glBindBuffer(GL_ARRAY_BUFFER, state.sky_verts);
-	glBufferSubData(GL_ARRAY_BUFFER, 0, sizeof(Vector3) * 8, vertices);
+	glBufferData(GL_ARRAY_BUFFER, sizeof(Vector3) * 8, vertices, GL_DYNAMIC_DRAW);
 	glBindBuffer(GL_ARRAY_BUFFER, 0); //unbind
 
 	glBindVertexArray(state.sky_array);
@@ -2690,7 +2690,7 @@ void RasterizerSceneGLES3::_setup_environment(Environment *env, const CameraMatr
 	}
 
 	glBindBuffer(GL_UNIFORM_BUFFER, state.scene_ubo);
-	glBufferSubData(GL_UNIFORM_BUFFER, 0, sizeof(State::SceneDataUBO), &state.ubo_data);
+	glBufferData(GL_UNIFORM_BUFFER, sizeof(State::SceneDataUBO), &state.ubo_data, GL_DYNAMIC_DRAW);
 	glBindBuffer(GL_UNIFORM_BUFFER, 0);
 
 	//fill up environment
@@ -2698,7 +2698,7 @@ void RasterizerSceneGLES3::_setup_environment(Environment *env, const CameraMatr
 	store_transform(sky_orientation * p_cam_transform, state.env_radiance_data.transform);
 
 	glBindBuffer(GL_UNIFORM_BUFFER, state.env_radiance_ubo);
-	glBufferSubData(GL_UNIFORM_BUFFER, 0, sizeof(State::EnvironmentRadianceUBO), &state.env_radiance_data);
+	glBufferData(GL_UNIFORM_BUFFER, sizeof(State::EnvironmentRadianceUBO), &state.env_radiance_data, GL_DYNAMIC_DRAW);
 	glBindBuffer(GL_UNIFORM_BUFFER, 0);
 }
 
@@ -2808,7 +2808,7 @@ void RasterizerSceneGLES3::_setup_directional_light(int p_index, const Transform
 	}
 
 	glBindBuffer(GL_UNIFORM_BUFFER, state.directional_ubo);
-	glBufferSubData(GL_UNIFORM_BUFFER, 0, sizeof(LightDataUBO), &ubo_data);
+	glBufferData(GL_UNIFORM_BUFFER, sizeof(LightDataUBO), &ubo_data, GL_DYNAMIC_DRAW);
 	glBindBuffer(GL_UNIFORM_BUFFER, 0);
 
 	directional_light = li;
diff --git a/drivers/gles3/rasterizer_storage_gles3.cpp b/drivers/gles3/rasterizer_storage_gles3.cpp
index 71737426a9..625bfb4554 100644
--- a/drivers/gles3/rasterizer_storage_gles3.cpp
+++ b/drivers/gles3/rasterizer_storage_gles3.cpp
@@ -3099,7 +3099,7 @@ void RasterizerStorageGLES3::_update_material(Material *material) {
 		}
 
 		glBindBuffer(GL_UNIFORM_BUFFER, material->ubo_id);
-		glBufferSubData(GL_UNIFORM_BUFFER, 0, material->ubo_size, local_ubo);
+		glBufferData(GL_UNIFORM_BUFFER, material->ubo_size, local_ubo, GL_STATIC_DRAW);
 		glBindBuffer(GL_UNIFORM_BUFFER, 0);
 	}