Demonstrates how to use specialized functions to select a level of detail based on dynamic conditions.
- iOS 11.0+
- tvOS 11.0+
- macOS 10.13+
- Xcode 11.0+Beta
A high-quality gaming experience has to manage trade-offs between great graphics and great performance. High-quality models look great, but their complexity requires a significant amount of processing power. By increasing or decreasing the level of detail (LOD) of a model, games can selectively manage both graphics and performance.
Instead of selecting a fixed LOD at build time, games can dynamically select between a range of LODs at runtime based on certain model-view conditions. For example, a focal-point foreground model could have a high LOD, whereas a fast-moving background model could have a low LOD.
This sample demonstrates dynamic LOD selection for a fire truck model, based on its distance from the scene’s camera. When the model is closer to the camera, the renderer uses a higher LOD; when the model is further from the camera, the renderer uses a lower LOD.
The Xcode project contains schemes for running the sample on macOS, iOS, or tvOS. Metal is not supported in the iOS or tvOS Simulator, so the iOS and tvOS schemes require a physical device to run the sample. The default scheme is macOS, which runs the sample as is on your Mac.
GPU Branch Statements
Unlike CPU code, graphics processing unit (GPU) branch statements like
else are very expensive. The massively parallel architecture of GPUs isn’t particularly well suited to handling GPU functions that have many branches. More branches result in more register allocations and thus decrease the number of GPU threads that can execute concurrently. Nevertheless, branch statements are useful programming constructs, particularly for functions that share a lot of code. In fact, a common problem for graphics functions that share code is how to handle a branch condition that differs only between draw calls, not between individual threads executing within a single draw call.
Traditionally, branches that differ between draw calls are mitigated in one of these ways:
Writing per-branch functions. Each branch is written as a complete and separate function, and the render loop determines which function to use at runtime. This approach greatly increases code duplication because all possible outcomes of each branch condition require their own standalone function. For example, a single
ifstatement requires one function for the
trueoutcome and another function for the
Using preprocessor directives. Instead of using a regular
ifstatement, functions can use the
#ifpreprocessor directive that selectively compiles a function after evaluating its branch conditions. This approach avoids code duplication but reduces the performance benefits of precompiled Metal shading language code. Because the branch conditions can only be evaluated at runtime, the functions can’t be precompiled at build time.
Metal’s function specialization feature reduces branch performance costs, avoids code duplication, and leverages build time compilation. Function specialization allows you to create multiple executable versions of a single source function. You create specialized functions by declaring function constants in your Metal shading language code and setting their values at runtime. Doing so allows the front-end compiler to precompile your source function at build time and the back-end compiler to compile the specialized function at runtime, when the pipeline is created.
Define Your LOD Selection Criteria
This sample demonstrates function specialization by creating different render pipelines for different LODs. All of the pipelines share the same source function, but function constants determine LOD-specific paths and inputs for each pipeline. Specifically, the sample demonstrates dynamic LOD selection for a fire truck model, based on its distance from the scene’s camera. When the fire truck is close to the camera, it occupies more pixels on the screen; therefore, the sample uses a high-quality render pipeline. When the fire truck is far from the camera, it occupies fewer pixels on the screen; therefore, the sample uses a low-quality render pipeline.
The fire truck model in this sample uses many types of textures, such as albedo, normal, metallic, roughness, ambient occlusion, and irradiance. It’s too wasteful to sample from each of these textures when the model is far from the camera because the detail provided by the full combination of textures isn’t seen. The sample uses various function constant values to create specialized functions that sample from more or fewer textures, depending on the selected LOD. Additionally, specialized functions that sample from fewer textures also perform less complex computations and result in a faster render pipeline.
is method controls whether a material property is set by sampling from a texture or by reading a constant value.
Implement Specialized Functions
The sample uses six function constants to control the various inputs available to the
fragment fragment function.
The sample also declares a derived function constant,
has, that’s used in the
vertex vertex function. This value determines whether the render pipeline requires the vertex function to output a texture coordinate to the
Color return value.
When the value of
false, the vertex function does not write a value to the
The function constants control the source of a parameter to the lighting computation in the
calculate function. When you use the
[[function attribute, this function can determine whether it should sample from a texture. The function only samples from a texture if the attribute indicates that a texture parameter is present; otherwise, it reads a uniform value from the
The corresponding inputs to the fragment function also use the same function constants.
Create Different Pipelines
This sample uses three different
MTLRender objects, each representing a different LOD. Specializing functions and building pipelines is expensive, so the sample performs these tasks asynchronously before starting the render loop. When the
AAPLRenderer object is initialized, each LOD pipeline is created asynchronously by using dispatch groups, completion handlers, and notification blocks.
The sample creates six specialized functions overall: one vertex and one fragment function for each of the three LODs. This task is monitored by the
specialization dispatch group, and each function is specialized by calling the
notify block builds the three render pipelines. This task is monitored by the
_pipeline dispatch group, and each pipeline is built by calling the
Render with a Specific LOD
At the beginning of the render loop, for each frame, the sample calls the
_calculate method to update the
_current value. This value defines the LOD for the frame based on the distance between the model and the camera. The
_calculate method also sets a
_global value that creates a smooth transition between LOD boundaries.
_current value is used to set the corresponding
MTLRender object for the frame.
_global value is used to interpolate between quality levels and prevent abrupt LOD transitions.
Finally, the render loop draws each submesh in the model with the specific LOD pipeline.