Merge pull request #42734 from lawnjelly/mac_orphan2

Fix buffer orphaning on desktop

drivers/gles2/rasterizer_canvas_base_gles2.cpp

@@ -430,14 +430,10 @@ void RasterizerCanvasBaseGLES2::_copy_texscreen(const Rect2 &p_rect) {
 void RasterizerCanvasBaseGLES2::_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;
+	storage->buffer_orphan_and_upload(data.polygon_buffer_size, 0, sizeof(Vector2) * p_vertex_count, p_vertices);
 
-	glBufferSubData(GL_ARRAY_BUFFER, 0, sizeof(Vector2) * p_vertex_count, p_vertices);
 	glEnableVertexAttribArray(VS::ARRAY_VERTEX);
 	glVertexAttribPointer(VS::ARRAY_VERTEX, 2, GL_FLOAT, GL_FALSE, sizeof(Vector2), NULL);
 	buffer_ofs += sizeof(Vector2) * p_vertex_count;
@@ -482,13 +478,9 @@ void RasterizerCanvasBaseGLES2::_draw_polygon(const int *p_indices, int p_index_
 	}
 
 	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);
+		storage->buffer_orphan_and_upload(data.polygon_index_buffer_size, 0, sizeof(int) * p_index_count, p_indices, GL_ELEMENT_ARRAY_BUFFER);
 		glDrawElements(GL_TRIANGLES, p_index_count, GL_UNSIGNED_INT, 0);
 		storage->info.render._2d_draw_call_count++;
 	} else {
@@ -496,7 +488,7 @@ void RasterizerCanvasBaseGLES2::_draw_polygon(const int *p_indices, int p_index_
 		for (int i = 0; i < p_index_count; i++) {
 			index16[i] = uint16_t(p_indices[i]);
 		}
-		glBufferSubData(GL_ELEMENT_ARRAY_BUFFER, 0, sizeof(uint16_t) * p_index_count, index16);
+		storage->buffer_orphan_and_upload(data.polygon_index_buffer_size, 0, sizeof(uint16_t) * p_index_count, index16, GL_ELEMENT_ARRAY_BUFFER);
 		glDrawElements(GL_TRIANGLES, p_index_count, GL_UNSIGNED_SHORT, 0);
 		storage->info.render._2d_draw_call_count++;
 	}
@@ -508,14 +500,10 @@ void RasterizerCanvasBaseGLES2::_draw_polygon(const int *p_indices, int p_index_
 void RasterizerCanvasBaseGLES2::_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;
+	storage->buffer_orphan_and_upload(data.polygon_buffer_size, 0, sizeof(Vector2) * p_vertex_count, p_vertices);
 
-	glBufferSubData(GL_ARRAY_BUFFER, 0, sizeof(Vector2) * p_vertex_count, p_vertices);
 	glEnableVertexAttribArray(VS::ARRAY_VERTEX);
 	glVertexAttribPointer(VS::ARRAY_VERTEX, 2, GL_FLOAT, GL_FALSE, sizeof(Vector2), NULL);
 	buffer_ofs += sizeof(Vector2) * p_vertex_count;
@@ -551,14 +539,10 @@ void RasterizerCanvasBaseGLES2::_draw_generic(GLuint p_primitive, int p_vertex_c
 void RasterizerCanvasBaseGLES2::_draw_generic_indices(GLuint p_primitive, 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) {
 
 	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;
+	storage->buffer_orphan_and_upload(data.polygon_buffer_size, 0, sizeof(Vector2) * p_vertex_count, p_vertices);
 
-	glBufferSubData(GL_ARRAY_BUFFER, 0, sizeof(Vector2) * p_vertex_count, p_vertices);
 	glEnableVertexAttribArray(VS::ARRAY_VERTEX);
 	glVertexAttribPointer(VS::ARRAY_VERTEX, 2, GL_FLOAT, GL_FALSE, sizeof(Vector2), NULL);
 	buffer_ofs += sizeof(Vector2) * p_vertex_count;
@@ -587,13 +571,9 @@ void RasterizerCanvasBaseGLES2::_draw_generic_indices(GLuint p_primitive, const
 	}
 
 	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);
+		storage->buffer_orphan_and_upload(data.polygon_index_buffer_size, 0, sizeof(int) * p_index_count, p_indices, GL_ELEMENT_ARRAY_BUFFER);
 		glDrawElements(p_primitive, p_index_count, GL_UNSIGNED_INT, 0);
 		storage->info.render._2d_draw_call_count++;
 	} else {
@@ -601,7 +581,7 @@ void RasterizerCanvasBaseGLES2::_draw_generic_indices(GLuint p_primitive, const
 		for (int i = 0; i < p_index_count; i++) {
 			index16[i] = uint16_t(p_indices[i]);
 		}
-		glBufferSubData(GL_ELEMENT_ARRAY_BUFFER, 0, sizeof(uint16_t) * p_index_count, index16);
+		storage->buffer_orphan_and_upload(data.polygon_index_buffer_size, 0, sizeof(uint16_t) * p_index_count, index16, GL_ELEMENT_ARRAY_BUFFER);
 		glDrawElements(p_primitive, p_index_count, GL_UNSIGNED_SHORT, 0);
 		storage->info.render._2d_draw_call_count++;
 	}
@@ -664,11 +644,7 @@ void RasterizerCanvasBaseGLES2::_draw_gui_primitive(int p_points, const Vector2
 	}
 
 	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);
+	storage->buffer_orphan_and_upload(data.polygon_buffer_size, 0, p_points * stride * 4 * sizeof(float), buffer_data);
 
 	glVertexAttribPointer(VS::ARRAY_VERTEX, 2, GL_FLOAT, GL_FALSE, stride * sizeof(float), NULL);
 

drivers/gles2/rasterizer_storage_gles2.h

@@ -1333,7 +1333,18 @@ public:
 	virtual String get_video_adapter_name() const;
 	virtual String get_video_adapter_vendor() const;
 
+	void buffer_orphan_and_upload(unsigned int p_buffer_size, unsigned int p_offset, unsigned int p_data_size, const void *p_data, GLenum p_target = GL_ARRAY_BUFFER);
+
 	RasterizerStorageGLES2();
 };
 
+// standardize the orphan / upload in one place so it can be changed per platform as necessary, and avoid future
+// bugs causing pipeline stalls
+inline void RasterizerStorageGLES2::buffer_orphan_and_upload(unsigned int p_buffer_size, unsigned int p_offset, unsigned int p_data_size, const void *p_data, GLenum p_target) {
+	// Orphan the buffer to avoid CPU/GPU sync points caused by glBufferSubData
+	// Was previously #ifndef GLES_OVER_GL however this causes stalls on desktop mac also (and possibly other)
+	glBufferData(p_target, p_buffer_size, NULL, GL_DYNAMIC_DRAW);
+	glBufferSubData(p_target, p_offset, p_data_size, p_data);
+}
+
 #endif // RASTERIZERSTORAGEGLES2_H

drivers/gles3/rasterizer_canvas_gles3.cpp

@@ -331,15 +331,9 @@ 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 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;
+	storage->buffer_orphan_and_upload(data.polygon_buffer_size, buffer_ofs, sizeof(Vector2) * p_vertex_count, p_vertices);
 
-	//vertex
-	glBufferSubData(GL_ARRAY_BUFFER, buffer_ofs, sizeof(Vector2) * p_vertex_count, p_vertices);
 	glEnableVertexAttribArray(VS::ARRAY_VERTEX);
 	glVertexAttribPointer(VS::ARRAY_VERTEX, 2, GL_FLOAT, false, sizeof(Vector2), CAST_INT_TO_UCHAR_PTR(buffer_ofs));
 	buffer_ofs += sizeof(Vector2) * p_vertex_count;
@@ -406,11 +400,7 @@ void RasterizerCanvasGLES3::_draw_polygon(const int *p_indices, int p_index_coun
 
 	//bind the indices buffer.
 	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
-	glBufferSubData(GL_ELEMENT_ARRAY_BUFFER, 0, sizeof(int) * p_index_count, p_indices);
+	storage->buffer_orphan_and_upload(data.polygon_index_buffer_size, 0, sizeof(int) * p_index_count, p_indices, GL_ELEMENT_ARRAY_BUFFER);
 
 	//draw the triangles.
 	glDrawElements(GL_TRIANGLES, p_index_count, GL_UNSIGNED_INT, 0);
@@ -432,15 +422,9 @@ 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 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;
+	storage->buffer_orphan_and_upload(data.polygon_buffer_size, buffer_ofs, sizeof(Vector2) * p_vertex_count, p_vertices);
 
-	//vertex
-	glBufferSubData(GL_ARRAY_BUFFER, buffer_ofs, sizeof(Vector2) * p_vertex_count, p_vertices);
 	glEnableVertexAttribArray(VS::ARRAY_VERTEX);
 	glVertexAttribPointer(VS::ARRAY_VERTEX, 2, GL_FLOAT, false, sizeof(Vector2), CAST_INT_TO_UCHAR_PTR(buffer_ofs));
 	buffer_ofs += sizeof(Vector2) * p_vertex_count;
@@ -485,15 +469,9 @@ void RasterizerCanvasGLES3::_draw_generic_indices(GLuint p_primitive, const int
 	glBindVertexArray(data.polygon_buffer_pointer_array);
 	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;
+	storage->buffer_orphan_and_upload(data.polygon_buffer_size, buffer_ofs, sizeof(Vector2) * p_vertex_count, p_vertices);
 
-	//vertex
-	glBufferSubData(GL_ARRAY_BUFFER, buffer_ofs, sizeof(Vector2) * p_vertex_count, p_vertices);
 	glEnableVertexAttribArray(VS::ARRAY_VERTEX);
 	glVertexAttribPointer(VS::ARRAY_VERTEX, 2, GL_FLOAT, false, sizeof(Vector2), CAST_INT_TO_UCHAR_PTR(buffer_ofs));
 	buffer_ofs += sizeof(Vector2) * p_vertex_count;
@@ -538,11 +516,8 @@ void RasterizerCanvasGLES3::_draw_generic_indices(GLuint p_primitive, const int
 
 	//bind the indices buffer.
 	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
-	glBufferSubData(GL_ELEMENT_ARRAY_BUFFER, 0, sizeof(int) * p_index_count, p_indices);
+
+	storage->buffer_orphan_and_upload(data.polygon_index_buffer_size, 0, sizeof(int) * p_index_count, p_indices, GL_ELEMENT_ARRAY_BUFFER);
 
 	//draw the triangles.
 	glDrawElements(p_primitive, p_index_count, GL_UNSIGNED_INT, 0);
@@ -616,12 +591,9 @@ 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]);
+	//TODO the below call may need to be replaced with: (p_points * stride * 4 * sizeof(float), &b[0]);
+	storage->buffer_orphan_and_upload(data.polygon_buffer_size, 0, p_points * stride * 4, &b[0]);
+
 	glBindVertexArray(data.polygon_buffer_quad_arrays[version]);
 	glDrawArrays(prim[p_points], 0, p_points);
 	glBindVertexArray(0);

drivers/gles3/rasterizer_storage_gles3.h

@@ -1481,7 +1481,18 @@ public:
 	virtual String get_video_adapter_name() const;
 	virtual String get_video_adapter_vendor() const;
 
+	void buffer_orphan_and_upload(unsigned int p_buffer_size, unsigned int p_offset, unsigned int p_data_size, const void *p_data, GLenum p_target = GL_ARRAY_BUFFER);
+
 	RasterizerStorageGLES3();
 };
 
+// standardize the orphan / upload in one place so it can be changed per platform as necessary, and avoid future
+// bugs causing pipeline stalls
+inline void RasterizerStorageGLES3::buffer_orphan_and_upload(unsigned int p_buffer_size, unsigned int p_offset, unsigned int p_data_size, const void *p_data, GLenum p_target) {
+	// Orphan the buffer to avoid CPU/GPU sync points caused by glBufferSubData
+	// Was previously #ifndef GLES_OVER_GL however this causes stalls on desktop mac also (and possibly other)
+	glBufferData(p_target, p_buffer_size, NULL, GL_DYNAMIC_DRAW);
+	glBufferSubData(p_target, p_offset, p_data_size, p_data);
+}
+
 #endif // RASTERIZERSTORAGEGLES3_H