This chapter is from the book
This chapter is from the book
This chapter is from the book
This chapter is from the book
What’s New in OpenGL ES 3.0
OpenGL ES 2.0 ushered in the era of programmable shaders for handheld devices and has been wildly successful in powering games, applications, and user interfaces across a wide range of devices. OpenGL ES 3.0 extends OpenGL ES 2.0 to support many new rendering techniques, optimizations, and visual quality enhancements. The following sections provide a categorized overview of the major new features that have been added to OpenGL ES 3.0. Each of these features will be described in detail later in the book.
Texturing
OpenGL ES 3.0 introduces many new features related to texturing:
- sRGB textures and framebuffers—Allow the application to perform gamma-correct rendering. Textures can be stored in gamma-corrected sRGB space, uncorrected to linear space upon being fetched in the shader, and then converted back to sRGB gamma-corrected space on output to the framebuffer. This enables potentially higher visual fidelity by properly computing lighting and other calculations in linear space.
- 2D texture arrays—A texture target that stores an array of 2D textures. Such arrays might, for example, be used to perform texture animation. Prior to 2D texture arrays, such animation was typically done by tiling the frames of an animation in a single 2D texture and modifying the texture coordinates to change animation frames. With 2D texture arrays, each frame of the animation can be specified in a 2D slice of the array.
- 3D textures—While some OpenGL ES 2.0 implementations supported 3D textures through an extension, OpenGL ES 3.0 has made this a mandatory feature. 3D textures are essential in many medical imaging applications, such as those that perform direct volume rendering of 3D voxel data (e.g., CT, MRI, or PET data).
- Depth textures and shadow comparison—Enable the depth buffer to be stored in a texture. The most common use for depth textures is in rendering shadows, where a depth buffer is rendered from the viewpoint of the light source and then used for comparison when rendering the scene to determine whether a fragment is in shadow. In addition to depth textures, OpenGL ES 3.0 allows the comparison against the depth texture to be done at the time of fetch, thereby allowing bilinear filtering to be done on depth textures (also known as percentage closest filtering [PCF]).
- Seamless cubemaps—In OpenGL ES 2.0, rendering with cubemaps could produce artifacts at the boundaries between cubemap faces. In OpenGL ES 3.0, cubemaps can be sampled such that filtering uses data from adjacent faces and removes the seaming artifact.
- Floating-point textures—OpenGL ES 3.0 greatly expands on the texture formats supported. Floating-point half-float (16-bit) textures are supported and can be filtered, whereas full-float (32-bit) textures are supported but not filterable. The ability to access floating-point texture data has many applications, including high dynamic range texturing to general-purpose computation.
- ETC2/EAC texture compression—While several OpenGL ES 2.0 implementations provided support for vendor-specific compressed texture formats (e.g., ATC by Qualcomm, PVRTC by Imagination Technologies, and Ericsson Texture Compression by Sony Ericsson), there was no standard compression format that developers could rely on. In OpenGL ES 3.0, support for ETC2/EAC is mandatory. The ETC2/EAC formats provide compression for RGB888, RGBA8888, and one- and two-channel signed/unsigned texture data. Texture compression offers several advantages, including better performance (due to better utilization of the texture cache) as well as a reduction in GPU memory utilization.
- Integer textures—OpenGL ES 3.0 introduces the capability to render to and fetch from textures stored as unnormalized signed or unsigned 8-bit, 16-bit, and 32-bit integer textures.
- Additional texture formats—In addition to those formats already mentioned, OpenGL ES 3.0 includes support for 11-11-10 RGB floating-point textures, shared exponent RGB 9-9-9-5 textures, 10-10-10-2 integer textures, and 8-bit-per-component signed normalized textures.
- Non-power-of-2 textures (NPOT)—Textures can now be specified with non-power-of-2 dimensions. This is useful in many situations, such as when texturing from a video or camera feed that is captured/recorded at a non-power-of-2 dimension.
- Texture level of detail (LOD) features—The texture LOD parameter used to determine which mipmap to fetch from can now be clamped. Additionally, the base and maximum mipmap level can be clamped. These two features, in combination, make it possible to stream mipmaps. As larger mipmap levels become available, the base level can be increased and the LOD value can be smoothly increased to provide smooth-looking streaming textures. This is very useful, for example, when downloading texture mipmap data over a network connection.
- Texture swizzles—A new texture object state was introduced to allow independent control of where each channel (R, G, B, and A) of texture data is mapped to in the shader.
- Immutable textures—Provide a mechanism for the application to specify the format and size of a texture before loading it with data. In doing so, the texture format becomes immutable and the OpenGL ES driver can perform all consistency and memory checks up-front. This can improve performance by allowing the driver to skip consistency checks at draw time.
- Increased minimum sizes—All OpenGL ES 3.0 implementations are required to support much larger texture resources than OpenGL ES 2.0. For example, the minimum supported 2D texture dimension in OpenGL ES 2.0 was 64 but was increased to 2048 in OpenGL ES 3.0.
Shaders
OpenGL ES 3.0 includes a major update to the OpenGL ES Shading Language (ESSL; to v3.00) and new API features to support new shader features:
- Program binaries—In OpenGL ES 2.0, it was possible to store shaders in a binary format, but it was still required to link them into program at runtime. In OpenGL ES 3.0, the entire linked program binary (containing the vertex and fragment shader) can be stored in an offline binary format with no link step required at runtime. This can potentially help reduce the load time of applications. Additionally, OpenGL ES 3.0 provides an interface to retrieve the program binary from the driver so no offline tools are required to use program binaries.
- Mandatory online compiler—OpenGL ES 2.0 made it optional whether the driver would support online compilation of shaders. The intent was to reduce the memory requirements of the driver, but this achievement came at a major cost to developers in terms of having to rely on vendor-specific tools to generate shaders. In OpenGL ES 3.0, all implementations will have an online shader compiler.
- Non-square matrices—New matrix types other than square matrices are supported, and associated uniform calls were added to the API to support loading them. Non-square matrices can reduce the instruction count required for performing transformations. For example, if performing an affine transformation, a 4 × 3 matrix can be used in place of a 4 × 4 where the last row is (0, 0, 0, 1), thus reducing the instructions required to perform the transformation.
- Full integer support—Integer (and unsigned integer) scalar and vector types, along with full integer operations, are supported in ESSL 3.00. There are various built-in functions such as conversion from int to float, and from float to int, as well as the ability to read integer values from textures and output integer values to integer color buffers.
- Centroid sampling—To avoid rendering artifacts when multisampling, the output variables from the vertex shader (and inputs to the fragment shader) can be declared with centroid sampling.
- Flat/smooth interpolators—In OpenGL ES 2.0, all interpolators were implicitly linearly interpolated across the primitive. In ESSL 3.00, interpolators (vertex shader outputs/fragment shader inputs) can be explicitly declared to have either smooth or flat shading.
- Uniform blocks—Uniform values can be grouped together into uniform blocks. Uniform blocks can be loaded more efficiently and also shared across multiple shader programs.
- Layout qualifiers—Vertex shader inputs can be declared with layout qualifiers to explicitly bind the location in the shader source without requiring making API calls. Layout qualifiers can also be used for fragment shader outputs to bind the outputs to each target when rendering to multiple render targets. Further, layout qualifiers can be used to control the memory layout for uniform blocks.
- Instance and vertex ID—The vertex index is now accessible in the vertex shader as well as the instance ID if using instance rendering.
- Fragment depth—The fragment shader can explicitly control the depth value for the current fragment rather than relying on the interpolation of its depth value.
- New built-in functions—ESSL 3.00 introduces many new built-in functions to support new texture features, fragment derivatives, half-float data conversion, and matrix and math operations.
- Relaxed limitations—ESSL 3.0 greatly relaxes the restrictions on shaders. Shaders are no longer limited in terms of instruction length, fully support looping and branching on variables, and support indexing on arrays.
Geometry
OpenGL ES 3.0 introduces several new features related to geometry specification and control of primitive rendering:
- Transform feedback—Allows the output of the vertex shader to be captured in a buffer object. This is useful for a wide range of techniques that perform animation on the GPU without any CPU intervention—for example, particle animation or physics simulation using render-to-vertex-buffer.
- Boolean occlusion queries—Enable the application to query whether any pixels of a draw call (or a set of draw calls) passes the depth test. This feature can be used within a variety of techniques, such as visibility determination for a lens flare effect as well as optimization to avoid performing geometry processing on objects whose bounding volume is obscured.
- Instanced rendering—Efficiently renders objects that contain similar geometry but differ by attributes (such as transformation matrix, color, or size). This feature is useful in rendering large quantities of similar objects, such as for crowd rendering.
- Primitive restart—When using triangle strips in OpenGL ES 2.0 for a new primitive, the application would have to insert indices into the index buffer to represent a degenerate triangle. In OpenGL ES 3.0, a special index value can be used that indicates the beginning of a new primitive. This obviates the need for generating degenerate triangles when using triangle strips.
- New vertex formats—New vertex formats, including 10-10-10-2 signed and unsigned normalized vertex attributes; 8-bit, 16-bit, and 32-bit integer attributes; and 16-bit half-float, are supported in OpenGL ES 3.0.
Buffer Objects
OpenGL ES 3.0 introduces many new buffer objects to increase the efficiency and flexibility of specifying data to various parts of the graphics pipeline:
- Uniform buffer objects—Provide an efficient method for storing/binding large blocks of uniforms. Uniform buffer objects can be used to reduce the performance cost of binding uniform values to shaders, which is a common bottleneck in OpenGL ES 2.0 applications.
- Vertex array objects—Provide an efficient method for binding and switching between vertex array states. Vertex array objects are essentially container objects for vertex array states. Using them allows an application to switch the vertex array state in a single API call rather than making several calls.
- Sampler objects—Separate the sampler state (texture wrap mode and filtering) from the texture object. This provides a more efficient method of sharing the sampler state across textures.
- Sync objects—Provide a mechanism for the application to check on whether a set of OpenGL ES operations has finished executing on the GPU. A related new feature is a fence, which provides a way for the application to inform the GPU that it should wait until a set of OpenGL ES operations has finished executing before queuing up more operations for execution.
- Pixel buffer objects—Enable the application to perform asynchronous transfer of data to pixel operations and texture transfer operations. This optimization is primarily intended to provide faster transfer of data between the CPU and the GPU, where the application can continue doing work during the transfer operation.
- Buffer subrange mapping—Allows the application to map a subregion of a buffer for access by the CPU. This can provide better performance than traditional buffer mapping, in which the whole buffer needs to be available to the client.
- Buffer object to buffer object copies—Provide a mechanism to efficiently transfer data from one buffer object to another without intervention on the CPU.
Framebuffer
OpenGL ES 3.0 adds many new features related to off-screen rendering to framebuffer objects:
- Multiple render targets (MRTs)—Allow the application to render simultaneously to several color buffers at one time. With MRTs, the fragment shader outputs several colors, one for each attached color buffer. MRTs are used in many advanced rendering algorithms, such as deferred shading.
- Multisample renderbuffers—Enable the application to render to off-screen framebuffers with multisample anti-aliasing. The multisample renderbuffers cannot be directly bound to textures, but they can be resolved to single-sample textures using the newly introduced framebuffer blit.
- Framebuffer invalidation hints—Many implementations of OpenGL ES 3.0 are based on GPUs that use tile-based rendering (TBR; explained in the Framebuffer Invalidation section in Chapter 12). It is often the case that TBR incurs a significant performance cost when having to unnecessarily restore the contents of the tiles for further rendering to a framebuffer. Framebuffer invalidation gives the application a mechanism to inform the driver that the contents of the framebuffer are no longer needed. This allows the driver to take optimization steps to skip unnecessary restore operations on the tiles. Such functionality is very important to achieve peak performance in many applications, especially those that do significant amounts of off-screen rendering.
- New blend equations—The min/max functions are supported in OpenGL ES 3.0 as a blend equation.