TLDR: To target all current generation Graviton instances (Graviton2 and Graviton3), use -march=arm8.2-a.

C and C++ code can be built for Graviton with a variety of flags, depending on the goal. If the goal is to get the best performance for a specific generation, select a flag from the table column "performance". If the goal is to get a good compromise of feature availability and performance balanced across generations, select the flag from the "balanced" column. If you want to target multiple generations of Graviton, select the "balanced" flag for the oldest generation planned for deployment, since code built for a newer generation may not run on an older generation. On arm64 -mcpu= acts as both specifying the appropriate architecture and tuning and it's generally better to use that vs -march if you're building for a specific CPU.

¹ Requires GCC-9 or later (or GCC-7 for Amazon Linux 2); otherwise we suggest using -mcpu=cortex-a72

² If your compiler doesn't support neoverse-512tvb, please use the Graviton2 tuning.

For some applications, it may be necessary to support a broad range of Arm64 targets while still making use of more advanced features such as LSE (Large System Extensions) or SVE (Scalable Vector Extension). For this case choose a more conservative build flag, such as -march=armv8-a and make use of runtime CPU support detection of features such as SVE. You can enable runtime detection and use of LSE atomics instructions by adding the additional compiler flag, -moutline-atomics.

Newer compilers provide better support and optimizations for Graviton processors. We have seen 15% better performance on Graviton2 when using gcc-10 instead of Amazon Linux 2 system's compiler gcc-7. When possible please use the latest compiler version available on your system. The table shows GCC and LLVM compiler versions available in Linux distributions. Starred version marks the default system compiler.

The Graviton2 and Graviton3(E) processors have support for the Large-System Extensions (LSE) which was first introduced in vArmv8.1. LSE provides low-cost atomic operations which can improve system throughput for CPU-to-CPU communication, locks, and mutexes. The improvement can be up to an order of magnitude when using LSE instead of load/store exclusives.

POSIX threads library needs LSE atomic instructions. LSE is important for locking and thread synchronization routines. The following systems distribute a libc compiled with LSE instructions:

• Amazon Linux 2,
• Amazon Linux 2022,
• Ubuntu 18.04 (needs apt install libc6-lse),
• Ubuntu 20.04,
• Ubuntu 22.04.

The compiler needs to generate LSE instructions for applications that use atomic operations. For example, the code of databases like PostgreSQL contain atomic constructs; c++11 code with std::atomic statements translate into atomic operations. GCC's -march=armv8.2-a flag enables all instructions supported by Graviton2, including LSE. To confirm that LSE instructions are created, the output of objdump command line utility should contain LSE instructions:

$ objdump -d app | grep -i 'cas\|casp\|swp\|ldadd\|stadd\|ldclr\|stclr\|ldeor\|steor\|ldset\|stset\|ldsmax\|stsmax\|ldsmin\|stsmin\|ldumax\|stumax\|ldumin\|stumin' | wc -l

To check whether the application binary contains load and store exclusives:

$ objdump -d app | grep -i 'ldxr\|ldaxr\|stxr\|stlxr' | wc -l

GCC's -moutline-atomics flag produces a binary that runs on both Graviton and Graviton2. Supporting both platforms with the same binary comes at a small extra cost: one load and one branch. To check that an application has been compiled with -moutline-atomics, nm command line utility displays the name of functions and global variables in an application binary. The boolean variable that GCC uses to check for LSE hardware capability is __aarch64_have_lse_atomics and it should appear in the list of symbols:

$ nm app | grep __aarch64_have_lse_atomics | wc -l
# the output should be 1 if app has been compiled with -moutline-atomics

GCC 10.1+ enables -moutline-atomics by default as does the version of GCC 7 used by Amazon Linux 2.

When programs contain code with x64 intrinsics, the following procedure can help to quickly obtain a working program on Arm, assess the performance of the program running on Graviton processors, profile hot paths, and improve the quality of code on the hot paths.

To quickly get a prototype running on Arm, one can use https://github.com/DLTcollab/sse2neon a translator of x64 intrinsics to NEON. sse2neon provides a quick starting point in porting performance critical codes to Arm. It shortens the time needed to get an Arm working program that then can be used to extract profiles and to identify hot paths in the code. A header file sse2neon.h contains several of the functions provided by standard x64 include files like xmmintrin.h, only implemented with NEON instructions to produce the exact semantics of the x64 intrinsic. Once a profile is established, the hot paths can be rewritten directly with NEON intrinsics to avoid the overhead of the generic sse2neon translation.

The C standard doesn't specify the signedness of char. On x86 char is signed by default while on Arm it is unsigned by default. This can be addressed by using standard int types that explicitly specify the signedness (e.g. uint8_t and int8_t) or compile with -fsigned-char. When using the getchar function, instead of the commonly used but incorrect:

char c;
while((c = getchar()) != EOF) {
    // Do something with the character c
}
// Assume we have reached the end of file here

you should use an int type and the standard function feof and ferror to check for the end of file, as follows:

int c;

while ((c = getchar()) != EOF) {
    // Do something with the character c
}
// Once we get EOF, we should check if it is actually an EOF or an error

if (feof(stdin)) {
    // End of file has been reached
} else if (ferror(stdin)) {
    // Handle the error (check errno, etc)
}

Graviton2 processors been optimized for performance and power efficient machine learning by enabling Arm dot-product instructions commonly used for Machine Learning (quantized) inference workloads, and enabling Half precision floating point - _float16 to double the number of operations per second, reducing the memory footprint compared to single precision floating point (_float32), while still enjoying large dynamic range.

The scalable vector extensions (SVE) require both a new enough tool-chain to auto-vectorize to SVE (GCC 11+, LLVM 14+) and a 4.15+ kernel that supports SVE. One notable exception is that Amazon Linux 2 with a 4.14 kernel doesn't support SVE; please upgrade to a 5.4+ AL2 kernel.

The Arm instruction set includes instructions that can be used to speedup common code sequences. The table below lists common operations and links to code sequences: