Name QCOM_tiled_rendering Name Strings GL_QCOM_tiled_rendering Contributors Colin Sharp Jeff Leger Contacts Chuck Smith, Qualcomm (chucks 'at' qualcomm.com) Maurice Ribble, Qualcomm (mribble 'at' qualcomm.com) Notice Copyright Qualcomm 2009. IP Status Qualcomm Proprietary. Status Complete. Version Last Modified Date: August 20, 2009 Revision: #1.6 Number OpenGL ES Extension #70 Dependencies OpenGL ES 1.0 or higher is required. This extension interacts with QCOM_write_only_rendering. This extension is written based on the wording of the OpenGL ES 2.0 specification. Overview In the handheld graphics space, a typical challenge is achieving efficient rendering performance given the different characteristics of the various types of graphics memory. Some types of memory ("slow" memory) are less expensive but have low bandwidth, higher latency, and/or higher power consumption, while other types ("fast" memory) are more expensive but have higher bandwidth, lower latency, and/or lower power consumption. In many cases, it is more efficient for a graphics processing unit (GPU) to render directly to fast memory, but at most common display resolutions it is not practical for a device to contain enough fast memory to accommodate both the full color and depth/stencil buffers (the frame buffer). In some devices, this problem can be addressed by providing both types of memory; a large amount of slow memory that is sufficient to store the entire frame buffer, and a small, dedicated amount of fast memory that allows the GPU to render with optimal performance. The challenge lies in finding a way for the GPU to render to fast memory when it is not large enough to contain the actual frame buffer. One approach to solving this problem is to design the GPU and/or driver using a tiled rendering architecture. With this approach the render target is subdivided into a number of individual tiles, which are sized to fit within the available amount of fast memory. Under normal operation, the entire scene will be rendered to each individual tile using a multi-pass technique, in which primitives that lie entirely outside of the tile being rendered are trivially discarded. After each tile has been rendered, its contents are saved out to the actual frame buffer in slow memory (a process referred to as the "resolve"). The resolve introduces significant overhead, both for the CPU and the GPU. However, even with this additional overhead, rendering using this method is usually more efficient than rendering directly to slow memory. This extension allows the application to specify a rectangular tile rendering area and have full control over the resolves for that area. The information given to the driver through this API can be used to perform various optimizations in the driver and hardware. One example optimization is being able to reduce the size or number of the resolves. Another optimization might be to reduce the number of passes needed in the tiling approach mentioned above. Even traditional rendering GPUs that don't use tiles may benefit from this extension depending on their implemention of certain common GPU operations. One typical use case could involve an application only rendering to select portions of the render target using this technique (which shall be referred to as "application tiling"), leaving all other portions of the render target untouched. Therefore, in order to preserve the contents of the untouched portions of the render target, the application must request an EGL (or other context management API) configuration with a non-destructive swap. A destructive swap may only be used safely if the application renders to the entire area of the render target during each frame (otherwise the contents of the untouched portions of the frame buffer will be undefined). Additionally, care must be taken to avoid the cost of mixing rendering with and without application tiling within a single frame. Rendering without application tiling ("normal" rendering) is most efficient when all of the rendering for the entire scene can be encompassed within a single resolve. If any portions of the scene are rendered prior to that resolve (such as via a prior resolve, or via application tiling), then that resolve becomes much more heavyweight. When this occurs, prior to rendering each tile the fast memory must be populated with the existing contents of the frame buffer region corresponding to that tile. This operation can double the cost of resolves, so it is recommended that applications avoid mixing application tiling and normal rendering within a single frame. If both rendering methods must be used in the same frame, then the most efficient approach is to perform all normal rendering first, followed by rendering done with application tiling. An implicit resolve will occur (if needed) at the start of application tiling, so any pending normal rendering operations will be flushed at the time application tiling is initiated. This extension provides interfaces for the application to communicate to the driver whether or not rendering done with application tiling depends on the existing contents of the specified tile, and whether or not the rendered contents of the specified tile need to be preserved upon completion. This mechanism can be used to obtain optimal performance, e.g. when the application knows that every pixel in a tile will be completely rendered or when the resulting contents of the depth/stencil buffers do not need to be preserved. Issues (1) How do Viewport and Scissor interact with this extension? RESOLVED: They don't. When application tiling is used, the viewport and scissor retain their existing values, relative to the render target, not the specified tile. Therefore, all rendering commands issued between StartTilingQCOM and EndTilingQCOM will be subject to the same scissor, and will undergo the same viewport transformation, as normal rendering commands. (2) How do Flush and Finish interact with this extension? RESOLVED: When Flush or Finish is called while application tiling is active, the behavior will be as if EndTilingQCOM was called, except that the application tiling state will remain unchanged (meaning the active tile will not be reset). This means that any pending rendering commands will be performed to the active tile, and application tiling will continue to be active for any following rendering commands. (3) How does SwapBuffers interact with this extension? RESOLVED: It doesn't. If SwapBuffers is called while application tiling is active, the contents of the entire back buffer will be copied to the visible window, ignoring the active tile. SwapBuffers will have no effect on the application tiling state. (4) What happens if the render target is changed while application tiling is active? RESOLVED: If the current render target is changed, either by binding a new framebuffer object or changing the write surface of the active framebuffer (either explicitly or by deleting the currently bound framebuffer or write surface), an implicit EndTilingQCOM will occur. The active tile will be reset and application tiling will be deactivated. This is necessary because the active tile may not be valid for the new render target. (5) Should this extension provide a query mechanism for determining things such as tile offset, alignment, and size requirements so a developer can intelligently choose tile regions? RESOLVED: No. This information is very device-dependent and difficult to present in an easily understood manner. Instead, this extension will let developers specify an arbitrary rectangular tile region and all these requirements, including subdividing the given tile into multiple tiles if necessary, will be handled by the driver and hardware. (6) Should this extension allow multiple tiles? RESOLVED: No. While earlier versions of this extension allowed for this, after support for arbitrary tile sizes was added the benefit of multiple tiles became negligible. Allowing multiple tiles complicated the API and made it much more difficult for traditional rendering and some tile-based rendering GPUs to support this extension. (7) Should multiple render targets be supported? They are not supported by either the OpenGL ES core specification or any existing OpenGL ES extensions. Support could be added with some new bitmasks for the parameter. Should this be added now, or deferred for inclusion in any possible future MRT extension? RESOLVED: Yes. It is not difficult to add now and doing it now makes supporting MRTs in the future easier. New Procedures and Functions void StartTilingQCOM(uint x, uint y, uint width, uint height, bitfield preserveMask); void EndTilingQCOM(bitfield preserveMask); New Tokens Accepted by the parameter of StartTilingQCOM and EndTilingQCOM GL_COLOR_BUFFER_BIT0_QCOM 0x00000001 GL_COLOR_BUFFER_BIT1_QCOM 0x00000002 GL_COLOR_BUFFER_BIT2_QCOM 0x00000004 GL_COLOR_BUFFER_BIT3_QCOM 0x00000008 GL_COLOR_BUFFER_BIT4_QCOM 0x00000010 GL_COLOR_BUFFER_BIT5_QCOM 0x00000020 GL_COLOR_BUFFER_BIT6_QCOM 0x00000040 GL_COLOR_BUFFER_BIT7_QCOM 0x00000080 GL_DEPTH_BUFFER_BIT0_QCOM 0x00000100 GL_DEPTH_BUFFER_BIT1_QCOM 0x00000200 GL_DEPTH_BUFFER_BIT2_QCOM 0x00000400 GL_DEPTH_BUFFER_BIT3_QCOM 0x00000800 GL_DEPTH_BUFFER_BIT4_QCOM 0x00001000 GL_DEPTH_BUFFER_BIT5_QCOM 0x00002000 GL_DEPTH_BUFFER_BIT6_QCOM 0x00004000 GL_DEPTH_BUFFER_BIT7_QCOM 0x00008000 GL_STENCIL_BUFFER_BIT0_QCOM 0x00010000 GL_STENCIL_BUFFER_BIT1_QCOM 0x00020000 GL_STENCIL_BUFFER_BIT2_QCOM 0x00040000 GL_STENCIL_BUFFER_BIT3_QCOM 0x00080000 GL_STENCIL_BUFFER_BIT4_QCOM 0x00100000 GL_STENCIL_BUFFER_BIT5_QCOM 0x00200000 GL_STENCIL_BUFFER_BIT6_QCOM 0x00400000 GL_STENCIL_BUFFER_BIT7_QCOM 0x00800000 GL_MULTISAMPLE_BUFFER_BIT0_QCOM 0x01000000 GL_MULTISAMPLE_BUFFER_BIT1_QCOM 0x02000000 GL_MULTISAMPLE_BUFFER_BIT2_QCOM 0x04000000 GL_MULTISAMPLE_BUFFER_BIT3_QCOM 0x08000000 GL_MULTISAMPLE_BUFFER_BIT4_QCOM 0x10000000 GL_MULTISAMPLE_BUFFER_BIT5_QCOM 0x20000000 GL_MULTISAMPLE_BUFFER_BIT6_QCOM 0x40000000 GL_MULTISAMPLE_BUFFER_BIT7_QCOM 0x80000000 Additions to Chapter 2 of the OpenGL ES 2.0 Specification (OpenGL Operation) Add a new section "Rendering with Application Tiling" after section 2.13: "2.14 Rendering with Application Tiling The application may specify an arbitrary rectangular region (a 'tile') to which rendering commands should be restricted. The command void StartTilingQCOM(uint x, uint y, uint width, uint height, bitfield preserveMask); specifies the tile described by , , , . Until the next call to EndTilingQCOM, all rendering commands (including clears) will only update the contents of the render target defined by the extents of this tile. The parameters and specify the screen-space origin of the tile, and and specify the screen-space width and height of the tile. The tile origin is located at the lower left corner of the tile. If the size of the tile is too large for the fast memory on the device then it will be internally subdivided into multiple tiles. The parameter is the bitwise OR of a number of values indicating which buffers need to be initialized with the existing contents of the frame buffer region corresponding to the specified tile prior to rendering, or the single value NONE. The values allowed are COLOR_BUFFER_BIT*_QCOM, DEPTH_BUFFER_BIT*_QCOM, STENCIL_BUFFER_BIT*_QCOM, and MULTISAMPLE_BUFFER_BIT*_QCOM. These indicate the color buffer, the depth buffer, the stencil buffer, and a multisample buffer modifier, respectively. The multisample bits are different since they modify the meaning of the color, depth, and stencil bits if the active surface is a multisample surface. If a multisample bit is set then the corresponding color, depth, and/or stencil bit will cause all the samples to be copied across the memory bus in devices that are using fast tiled memory, but if the multisample bit is not set then only a single resolved sample is copied across the bus. In practice, not setting the multisample bit when rendering to a multisample buffer can greatly improve performance, but could cause small rendering artifacts in some multiple-pass rendering algorithms. The 0-7 number is to specify which render target is being used. If multiple render targets are not being used then 0 should be specified. Any buffers specifed in that do not exist in the current rendering state will be silently ignored (simlilar to the behavior of Clear). If NONE is specified, then no buffers will be initialized. For any buffers not initialized in this manner, the initial contents will be undefined. The values of , , and are silently clamped to the extents of the render target. The command void EndTilingQCOM(bitfield preserveMask); notifies the driver that the application has completed all desired rendering to the tile specified by StartTilingQCOM. This allows the driver to flush the contents of the specified tile to the corresponding region of the render target, and disables application tiling (resuming normal rendering). The parameter is specified using the same values as the equivalent argument of StartTilingQCOM, but indicates which buffers need to be preserved upon completion of all rendering commands issued with application tiling. For any buffers not preserved in this manner, the resulting contents of the buffer regions corresponding to the active tile will be undefined. GLX Protocol None. Errors INVALID_OPERATION error is generated if StartTilingQCOM is called while WRITEONLY_RENDERING_QCOM is enabled or the current framebuffer is not framebuffer complete INVALID_OPERATION error is generated if EndTilingQCOM is called without a corresponding call to StartTilingQCOM INVALID_OPERATION error is generated if StartTilingQCOM is called after calling StartTilingQCOM without a corresponding call to EndTilingQCOM INVALID_OPERATION error is generated if Enable(WRITEONLY_RENDERING_QCOM) is called between StartTilingQCOM and EndTilingQCOM New State None. Sample Usage GLboolean renderTiledTriangle(GLuint x, GLuint y, GLuint width, GLuint height) { // set the active tile and initialize the color and depth buffers with // the existing contents glStartTilingQCOM(x, y, width, height, GL_COLOR_BUFFER_BIT0_QCOM | GL_DEPTH_BUFFER_BIT0_QCOM); // draw the triangle glDrawArrays(GL_TRIANGLES, 0, 3); // finished with this tile -- preserve the color buffer glEndTilingQCOM(GL_COLOR_BUFFER_BIT0_QCOM); // return success return GL_TRUE; } Revision History #09 08/20/2009 Chuck Smith Cosmetic changes #08 08/19/2009 Maurice Ribble Add support for multiple render targets #07 07/28/2009 Maurice Ribble Clean up spec Remove multiple tile support #06 07/23/2009 Maurice Ribble Updated overview to match latest spec #05 07/15/2009 Maurice Ribble Changed from spec to subdivide tiles instead of returning out of memory #04 07/06/2009 Maurice Ribble Update due to the AMD->Qualcomm move; general extension cleanup. #03 11/17/2008 Chuck Smith Clarified the results of EndTilingQCOM for unpreserved buffers. #02 11/10/2008 Chuck Smith Updates to clarify behavior; additions to the Issues section. #01 11/04/2008 Chuck Smith First draft.