ARM NEON SIMD-Optimized Sine and Cosine Library for Embedded Systems.
sincos_acc is a high-performance C library designed to accelerate trigonometric calculations on ARM NEON architectures. It provides a vectorized implementation of sine and cosine functions that process 4 elements simultaneously in a single operation, achieving significant speedups over standard C library implementations while maintaining high precision.
- SIMD Acceleration: Leverages ARM NEON intrinsics to process 4 single-precision floats at once.
- High Performance: Achieved 3.05x speedup compared to standard
sinf/cosf(tested on ARM64). - Exceptional Precision: Max Absolute Error (MAE) of 7.52e-08 (passing the 1e-05 target).
- Cache-Aware Processing: Optimized for both small (65x65) and large (260x257) grids.
- Simple API: One function call handles arbitrary element counts with automatic vectorization and remainder handling.
Tested with a grid of 66,820 elements (approx. 260x257) across a range of
| Implementation | Execution Time | Throughput | Precision (MAE) |
|---|---|---|---|
Scalar (sinf/cosf) |
0.000121 s | 554.16 M/s | Reference |
NEON (sincos_ps) |
0.000040 s | 1689.57 M/s | 7.52e-08 |
Speedup Ratio: 3.05x
- An ARM compiler (e.g.,
gccorclang) with NEON support.
-
Clone the repository:
git clone https://github.com/huntkao/sincos_acc.git cd sincos_acc -
Compile the benchmark:
make
-
Run the analysis:
./perf_test
include/sincos_neon.h: Public API definition.src/sincos_neon.c: Core implementation logic.extern/neon_mathfun.h: SIMD math engine (optimized minimax polynomial).benchmarks/perf_test.c: Comprehensive analysis and verification suite.
The core SIMD engine uses an optimized port of neon_mathfun.h, originally by Julien Pommier.
Distributed under the zlib license. See extern/neon_mathfun.h for copyright details.