1.5 Using and Understanding the Code
The pbrt source code distribution is available from pbrt.org. The website also includes additional documentation, images rendered with pbrt, example scenes, errata, and links to a bug reporting system. We encourage you to visit the website and subscribe to the pbrt mailing list.
pbrt is written in C++, but we have tried to make it accessible to non-C++ experts by limiting the use of esoteric features of the language. Staying close to the core language features also helps with the system’s portability. We make use of C++’s extensive standard library whenever it is applicable but will not discuss the semantics of calls to standard library functions in the text. Our expectation is that the reader will consult documentation of the standard library as necessary.
We will occasionally omit short sections of pbrt’s source code from the book. For example, when there are a number of cases to be handled, all with nearly identical code, we will present one case and note that the code for the remaining cases has been omitted from the text. Default class constructors are generally not shown, and the text also does not include details like the various #include directives at the start of each source file. All the omitted code can be found in the pbrt source code distribution.
1.5.1 Source Code Organization
The source code used for building pbrt is under the src directory in the pbrt distribution. In that directory are src/ext, which has the source code for various third-party libraries that are used by pbrt, and src/pbrt, which contains pbrt’s source code. We will not discuss the third-party libraries’ implementations in the book.
The source files in the src/pbrt directory mostly consist of implementations of the various interface types. For example, shapes.h and shapes.cpp have implementations of the Shape interface, materials.h and materials.cpp have materials, and so forth. That directory also holds the source code for parsing pbrt’s scene description files.
The pbrt.h header file in src/pbrt is the first file that is included by all other source files in the system. It contains a few macros and widely useful forward declarations, though we have tried to keep it short and to minimize the number of other headers that it includes in the interests of compile time efficiency.
The src/pbrt directory also contains a number of subdirectories. They have the following roles:
- base: Header files defining the interfaces for 12 of the common interface types listed in Table 1.1 (Primitive and Integrator are CPU-only and so are defined in files in the cpu directory).
- cmd: Source files containing the main() functions for the executables that are built for pbrt. (Others besides the pbrt executable include imgtool, which performs various image processing operations, and pbrt_test, which contains unit tests.)
- cpu: CPU-specific code, including Integrator implementations.
- gpu: GPU-specific source code, including functions for allocating memory and launching work on the GPU.
- util: Lower-level utility code, most of it not specific to rendering.
- wavefront: Implementation of the WavefrontPathIntegrator, which is introduced in Chapter 15. This integrator runs on both CPUs and GPUs.
1.5.2 Naming Conventions
Functions and classes are generally named using Camel case, with the first letter of each word capitalized and no delineation for spaces. One exception is some methods of container classes, which follow the naming convention of the C++ standard library when they have matching functionality (e.g., size() and begin() and end() for iterators). Variables also use Camel case, though with the first letter lowercase, except for a few global variables.
We also try to match mathematical notation in naming: for example, we use variables like p for points and w for directions . We will occasionally add a p to the end of a variable to denote a primed symbol: wp for . Underscores are used to indicate subscripts in equations: theta_o for , for example.
Our use of underscores is not perfectly consistent, however. Short variable names often omit the underscore—we use wi for and we have already seen the use of Li for . We also occasionally use an underscore to separate a word from a lowercase mathematical symbol. For example, we use Sample_f for a method that samples a function rather than Samplef, which would be more difficult to read, or SampleF, which would obscure the connection to the function (“where was the function defined?”).
1.5.3 Pointer or Reference?
C++ provides two different mechanisms for passing an object to a function or method by reference: pointers and references. If a function argument is not intended as an output variable, either can be used to save the expense of passing the entire structure on the stack. The convention in pbrt is to use a pointer when the argument will be completely changed by the function or method, a reference when some of its internal state will be changed but it will not be fully reinitialized, and const references when it will not be changed at all. One important exception to this rule is that we will always use a pointer when we want to be able to pass nullptr to indicate that a parameter is not available or should not be used.
1.5.4 Abstraction versus Efficiency
One of the primary tensions when designing interfaces for software systems is making a reasonable trade-off between abstraction and efficiency. For example, many programmers religiously make all data in all classes private and provide methods to obtain or modify the values of the data items. For simple classes (e.g., Vector3f), we believe that approach needlessly hides a basic property of the implementation—that the class holds three floating-point coordinates—that we can reasonably expect to never change. Of course, using no information hiding and exposing all details of all classes’ internals leads to a code maintenance nightmare, but we believe that there is nothing wrong with judiciously exposing basic design decisions throughout the system. For example, the fact that a Ray is represented with a point, a vector, a time, and the medium it is in is a decision that does not need to be hidden behind a layer of abstraction. Code elsewhere is shorter and easier to understand when details like these are exposed.
An important thing to keep in mind when writing a software system and making these sorts of trade-offs is the expected final size of the system. pbrt is roughly 70,000 lines of code and it is never going to grow to be a million lines of code; this fact should be reflected in the amount of information hiding used in the system. It would be a waste of programmer time (and likely a source of runtime inefficiency) to design the interfaces to accommodate a system of a much higher level of complexity.
We have reimplemented a subset of the C++ standard library in the pstd namespace; this was necessary in order to use those parts of it interchangeably on the CPU and on the GPU. For the purposes of reading pbrt’s source code, anything in pstd provides the same functionality with the same type and methods as the corresponding entity in std. We will therefore not document usage of pstd in the text here.
Almost all dynamic memory allocation for the objects that represent the scene in pbrt is performed using an instance of an Allocator that is provided to the object creation methods. In pbrt, Allocator is shorthand for the C++ standard library’s pmr::polymorphic_allocator type. Its definition is in pbrt.h so that it is available to all other source files.
std::pmr::polymorphic_allocator implementations provide a few methods for allocating and freeing objects. These three are used widely in pbrt:
The first, allocate_bytes(), allocates the specified number of bytes of memory. Next, allocate_object() allocates an array of n objects of the specified type T, initializing each one with its default constructor. The final method, new_object(), allocates a single object of type T and calls its constructor with the provided arguments. There are corresponding methods for freeing each type of allocation: deallocate_bytes(), deallocate_object(), and delete_object().
A tricky detail related to the use of allocators with data structures from the C++ standard library is that a container’s allocator is fixed once its constructor has run. Thus, if one container is assigned to another, the target container’s allocator is unchanged even though all the values it stores are updated. (This is the case even with C++’s move semantics.) Therefore, it is common to see objects’ constructors in pbrt passing along an allocator in member initializer lists for containers that they store even if they are not yet ready to set the values stored in them.
Using an explicit memory allocator rather than direct calls to new and delete has a few advantages. Not only does it make it easy to do things like track the total amount of memory that has been allocated, but it also makes it easy to substitute allocators that are optimized for many small allocations, as is useful when building acceleration structures in Chapter 7. Using allocators in this way also makes it easy to store the scene objects in memory that is visible to the GPU when GPU rendering is being used.
1.5.7 Dynamic Dispatch
As mentioned in Section 1.3, virtual functions are generally not used for dynamic dispatch with polymorphic types in pbrt (the main exception being the Integrators). Instead, the TaggedPointer class is used to represent a pointer to one of a specified set of types; it includes machinery for runtime type identification and thence dynamic dispatch. (Its implementation can be found in Appendix B.4.4.) Two considerations motivate its use.
First, in C++, an instance of an object that inherits from an abstract base class includes a hidden virtual function table pointer that is used to resolve virtual function calls. On most modern systems, this pointer uses eight bytes of memory. While eight bytes may not seem like much, we have found that when rendering complex scenes with previous versions of pbrt, a substantial amount of memory would be used just for virtual function pointers for shapes and primitives. With the TaggedPointer class, there is no incremental storage cost for type information.
The other problem with virtual function tables is that they store function pointers that point to executable code. Of course, that’s what they are supposed to do, but this characteristic means that a virtual function table can be valid for method calls from either the CPU or from the GPU, but not from both simultaneously, since the executable code for the different processors is stored at different memory locations. When using the GPU for rendering, it is useful to be able to call methods from both processors, however.
For all the code that just calls methods of polymorphic objects, the use of pbrt’s TaggedPointer in place of virtual functions makes no difference other than the fact that method calls are made using the . operator, just as would be used for a C++ reference. Section 4.5.1, which introduces Spectrum, the first class based on TaggedPointer that occurs in the book, has more details about how pbrt’s dynamic dispatch scheme is implemented.
1.5.8 Code Optimization
We have tried to make pbrt efficient through the use of well-chosen algorithms rather than through local micro-optimizations, so that the system can be more easily understood. However, efficiency is an integral part of rendering, and so we discuss performance issues throughout the book.
For both CPUs and GPUs, processing performance continues to grow more quickly than the speed at which data can be loaded from main memory into the processor. This means that waiting for values to be fetched from memory can be a major performance limitation. The most important optimizations that we discuss relate to minimizing unnecessary memory access and organizing algorithms and data structures in ways that lead to coherent access patterns; paying attention to these issues can speed up program execution much more than reducing the total number of instructions executed.
1.5.9 Debugging and Logging
Debugging a renderer can be challenging, especially in cases where the result is correct most of the time but not always. pbrt includes a number of facilities to ease debugging.
One of the most important is a suite of unit tests. We have found unit testing to be invaluable in the development of pbrt for the reassurance it gives that the tested functionality is very likely to be correct. Having this assurance relieves the concern behind questions during debugging such as “am I sure that the hash table that is being used here is not itself the source of my bug?” Alternatively, a failing unit test is almost always easier to debug than an incorrect image generated by the renderer; many of the tests have been added along the way as we have debugged pbrt. Unit tests for a file code.cpp are found in code_tests.cpp. All the unit tests are executed by an invocation of the pbrt_test executable and specific ones can be selected via command-line options.
There are many assertions throughout the pbrt codebase, most of them not included in the book text. These check conditions that should never be true and issue an error and exit immediately if they are found to be true at runtime. (See Section B.3.6 for the definitions of the assertion macros used in pbrt.) A failed assertion gives a first hint about the source of an error; like a unit test, an assertion helps focus debugging, at least with a starting point. Some of the more computationally expensive assertions in pbrt are only enabled for debug builds; if the renderer is crashing or otherwise producing incorrect output, it is worthwhile to try running a debug build to see if one of those additional assertions fails and yields a clue.
We have also endeavored to make the execution of pbrt at a given pixel sample deterministic. One challenge with debugging a renderer is a crash that only happens after minutes or hours of rendering computation. With deterministic execution, rendering can be restarted at a single pixel sample in order to more quickly return to the point of a crash. Furthermore, upon a crash pbrt will print a message such as “Rendering failed at pixel (16, 27) sample 821. Debug with --debugstart 16,27,821”. The values printed after “debugstart” depend on the integrator being used, but are sufficient to restart its computation close to the point of a crash.
Finally, it is often useful to print out the values stored in a data structure during the course of debugging. We have implemented ToString() methods for nearly all of pbrt’s classes. They return a std::string representation of them so that it is easy to print their full object state during program execution. Furthermore, pbrt’s custom Printf() and StringPrintf() functions (Section B.3.3) automatically use the string returned by ToString() for an object when a %s specifier is found in the formatting string.
1.5.10 Parallelism and Thread Safety
In pbrt (as is the case for most ray tracers), the vast majority of data at rendering time is read only (e.g., the scene description and texture images). Much of the parsing of the scene file and creation of the scene representation in memory is done with a single thread of execution, so there are few synchronization issues during that phase of execution. During rendering, concurrent read access to all the read-only data by multiple threads works with no problems on both the CPU and the GPU; we only need to be concerned with situations where data in memory is being modified.
As a general rule, the low-level classes and structures in the system are not thread-safe. For example, the Point3f class, which stores three float values to represent a point in 3D space, is not safe for multiple threads to call methods that modify it at the same time. (Multiple threads can use Point3fs as read-only data simultaneously, of course.) The runtime overhead to make Point3f thread-safe would have a substantial effect on performance with little benefit in return.
The same is true for classes like Vector3f, Normal3f, SampledSpectrum, Transform, Quaternion, and SurfaceInteraction. These classes are usually either created at scene construction time and then used as read-only data or allocated on the stack during rendering and used only by a single thread.
The utility classes ScratchBuffer (used for high-performance temporary memory allocation) and RNG (pseudo-random number generation) are also not safe for use by multiple threads; these classes store state that is modified when their methods are called, and the overhead from protecting modification to their state with mutual exclusion would be excessive relative to the amount of computation they perform. Consequently, in code like the ImageTileIntegrator::Render() method earlier, pbrt allocates per-thread instances of these classes on the stack.
With two exceptions, implementations of the base types listed in Table 1.1 are safe for multiple threads to use simultaneously. With a little care, it is usually straightforward to implement new instances of these base classes so they do not modify any shared state in their methods.
The first exceptions are the Light Preprocess() method implementations. These are called by the system during scene construction, and implementations of them generally modify shared state in their objects. Therefore, it is helpful to allow the implementer to assume that only a single thread will call into these methods. (This is a separate issue from the consideration that implementations of these methods that are computationally intensive may use ParallelFor() to parallelize their computation.)
The second exception is Sampler class implementations; their methods are also not expected to be thread-safe. This is another instance where this requirement would impose an excessive performance and scalability impact; many threads simultaneously trying to get samples from a single Sampler would limit the system’s overall performance. Therefore, as described in Section 1.3.4, a unique Sampler is created for each rendering thread using Sampler::Clone().
All stand-alone functions in pbrt are thread-safe (as long as multiple threads do not pass pointers to the same data to them).
1.5.11 Extending the System
One of our goals in writing this book and building the pbrt system was to make it easier for developers and researchers to experiment with new (or old!) ideas in rendering. One of the great joys in computer graphics is writing new software that makes a new image; even small changes to the system can be fun to experiment with. The exercises throughout the book suggest many changes to make to the system, ranging from small tweaks to major open-ended research projects. Section C.4 in Appendix C has more information about the mechanics of adding new implementations of the interfaces listed in Table 1.1.
Although we made every effort to make pbrt as correct as possible through extensive testing, it is inevitable that some bugs are still present.
If you believe you have found a bug in the system, please do the following:
- Reproduce the bug with an unmodified copy of the latest version of pbrt.
- Check the online discussion forum and the bug-tracking system at pbrt.org. Your issue may be a known bug, or it may be a commonly misunderstood feature.
- Try to find the simplest possible test case that demonstrates the bug. Many bugs can be demonstrated by scene description files that are just a few lines long, and debugging is much easier with a simple scene than a complex one.
- Submit a detailed bug report using our online bug-tracking system. Make sure that you include the scene file that demonstrates the bug and a detailed description of why you think pbrt is not behaving correctly with the scene. If you can provide a patch that fixes the bug, all the better!
We will periodically update the pbrt source code repository with bug fixes and minor enhancements. (Be aware that we often let bug reports accumulate for a few months before going through them; do not take this as an indication that we do not value them!) However, we will not make major changes to the pbrt source code so that it does not diverge from the system described here in the book.