Awesome

starch - a framework for selecting architecture-specific code at runtime

starch helps generates glue code to select architecture-specific versions of code depending on the hardware detected at runtime.

It arranges for code to be built multiple times with different compiler options. At runtime, user code calls a dispatcher entry point which selects the best compiled version of the versions that can safely run on the hardware used at runtime.

It tries to be agnostic about the details of the code being generated and the details of the hardware.

Caution caution work in progress

This documentation isn't very complete. You'll need to look at the example and the code itself.

Design notes

Architecture-independent generated output; the generated outputs can be generated during development and committed as part of the main source code, and at build time starch does not need to be re-run.
Doesn't care about the details of the functions you call; they can have any signature.
Can automatically generate benchmarking code given a benchmarking helper that sets up inputs to the function.
Does not do any hardware detection itself, and does not care about the hardware details; for each combination of compiler flags, the user code provides a test function to be called at runtime to determine if it is safe to run code compiled with those flags.
Allows the same generic code to be compiled multiple times with different compile flags to take advantage of compile auto-vectorization that requires additional instruction set features (AVX, NEON, ..) being enabled.
Emits makefile fragments to be included into a larger makefile structure

License

The generator script and templates are licensed under a BSD 2-clause license, see the LICENSE file.

No copyright claim is made on generated code.

Prerequisites

At generation time (results can be committed to version control):

Python 3
Mako

At build time:

a C compiler
make

Quickstart

Look in example/ for a full example.

Concepts

A function is the user-visible API to starch-generated code. It just looks like a C function pointer. Initially, this pointer points to a dispatcher routine which will select an appropriate implementation at runtime and call it. For subsequent calls, the dispatcher updates the function pointer to point directly to the selected implementation.

A function impl is one particular way of implementing a function. All impls should produce the same results given the same inputs to avoid confusing user code. There may be different impls with different performance characteristics - for example, different degrees of manual loop unrolling, or an impl that takes advantage of a particular instruction set (NEON, AVX, etc). Each impl has a unique-within-the-function "variant" name that identifies it.

Function impls may be conditionally compiled depending on build features (see below). This is useful for impls that cannot always be compiled e.g. they depend on the availability of a particular instruction set.

A build flavor is a particular way of building the function impl. It consists of a set of compiler flags to use, plus an associated test function that determines at runtime if it is safe to run the code. For example, a flavor may enable use of specific instructions that may or may not be available at runtime via -mavx, -march=..., and similar flags. Each flavor declares that it provides zero or more features.

A feature is a characteristic of the build flavor compiler flags that allows certain impls to be compiled. For example, an impl that uses NEON intrinsics can only be compiled if the compiler is building for an ARM instruction set that supports NEON. Features are defined in the build flavor, and are advertised at compile time by the presence of a STARCH_FEATURE_x macro; implementations may conditionally compile on this macro and should use STARCH_IMPL_REQUIRES to indicate they will only be emitted when a given feature is present.

A build mix is a combination of build flavors that can coexist in the same binary. For example, an "x86" mix might include build flavors that build for generic x86, x86-with-AVX, and x86-with-AVX2; but it would not include a build flavor for ARM, because ARM and x86 object code can't be linked together into a single binary.

Alignment

A function can optionally include an aligned version; this is a version of the function with an independent call point and wisdom, which assumes that data passed to the function is already aligned. Each flavor has an associated alignment in bytes, but otherwise it is up to the implementations to decide what exactly is aligned. Implementations for an aligned function on a flavor that specifies an alignment (>1 byte) will be compiled twice, once with an alignment of 1 and once with the flavor's alignment, to generate two different compiled versions.

starch provides macros to help with alignment:

STARCH_ALIGNMENT, in implementations, is the alignment (in bytes) that implementations can assume.
STARCH_MIX_ALIGNMENT, defined in the generated header file, is the required alignment (in bytes) for callers of the _aligned version of a function. It is the largest alignment of all flavors in the mix.
STARCH_ALIGNED(ptr) in implementations evaluates to ptr while hinting to the compiler that the data is aligned according to STARCH_ALIGNMENT. This maps to gcc's __builtin_assume_aligned builtin.

Benchmarks

Functions can optionally provide a benchmark helper by defining a (no args, void return typer) function using the STARCH_BENCHMARK macro. This macro is only present when benchmark code is being compiled.

The benchmark helper should set up function inputs for benchmarking and then use the STARCH_BENCHMARK_RUN macro. This macro expands to code that will benchmark each possible impl in turn with the provided arguments.

If the benchmark needs to allocated possibly-aligned buffers, two macros STARCH_BENCHMARK_ALLOC and STARCH_BENCHMARK_FREE will allocate suitably aligned buffers for the current STARCH_ALIGNMENT value. STARCH_BENCHMARK_ALLOC(count,type) will allocate count elements of type type, aligned to either STARCH_ALIGNMENT or the required alignment for type, whichever is larger. STARCH_BENCHMARK_FREE(ptr) will free a buffer previously allocated by STARCH_BENCHMARK_ALLOC.

See example/benchmark/subtract_n_benchmark.c for examples.

Gotchas

Files added by scan_file are #include-d into surrounding support files. Multiple files may be included into the same compilation unit. You should ensure that you don't pollute the global namespace (macros, static functions names, etc) for subsequent files that will follow.

Files added by scan_file will be compiled multiple times. You should ensure that any symbols other than those handled by STARCH_IMPL / STARCH_IMPL_REQUIRES are either static or use the STARCH_SYMBOL macro to get a unique name for this compilation pass.

You probably want to separate out benchmark-support code into separate files to avoid an extra version of any impls present in the same file from being emitted.

Wisdom

There is partial support for a wisdom implementation. Wisdom is a priori information about the preferred code to use for a given function, for example as the result of benchmarking to find the fastest version. It is simply the order in which compiled impls are tried until one that is supported is found.

To set wisdom, there are two options:

Provide a wisdom ordering for the function when defining a build mix. This controls the order in which the compiled impls are included in the generated registry that is searched at runtime.
Call starch_<function>_set_wisdom at runtime. This accepts an array of function variants, terminated by NULL. When called, the registry is re-sorted to prefer the listed variants in the order provided (and the function pointer is reset to the dispatcher so the chosen code will be re-selected on the next call). This could be used to load install-specific wisdom during program startup.