sgnl-cpu-interp

Fast polyphase resampling for multichannel data with multi-architecture SIMD support

---

Features

• Multi-Architecture SIMD: Automatic runtime CPU detection with optimized kernels for:
  • x86_64: AVX-512, AVX2+FMA, AVX, SSE4.1, SSE2
  • ARM64: NEON (Apple Silicon, AWS Graviton, etc.)
  • Fallback: Optimized scalar implementation
• No External Dependencies: Only requires NumPy (removed GSL and FFTW dependencies)
• High Performance: ~5x faster than GSL-based implementations
• Multichannel: Optimized for processing many channels simultaneously (tested with 1024+ channels)
• Quality: Lanczos-windowed sinc interpolation for high-quality upsampling
• Two Memory Layouts: Supports both (time, channels) and (channels, time) layouts
• Simple API: Easy-to-use NumPy-based interface

Installation

From PyPI (when available)

pip install sgnl-cpu-interp

Binary wheels are built in CI for:

Platform	Architecture	Minimum version
Linux	x86_64, aarch64	glibc 2.28 / manylinux_2_28 (e.g. RHEL/Rocky 8, Debian 10, Ubuntu 18.10)
macOS	x86_64	macOS 13 (Ventura)
macOS	arm64	macOS 15 (Sequoia)

From source

git clone https://git.ligo.org/greg/fast-resample-cpu.git
cd fast-resample-cpu
pip install .

No external dependencies are required beyond NumPy and a C compiler; the build backend (scikit-build-core) provisions CMake and Ninja automatically if they are not already installed. The build detects your CPU architecture and compiles the appropriate SIMD kernels; the best implementation is selected at runtime.

Quick Start

import numpy as np
from sgnl_cpu_interp import upsample, get_simd_info

# Check which SIMD implementation is being used
print(get_simd_info())
# {'implementation': 'NEON', 'available': ['NEON', 'Scalar'], 'cpu_features': 'NEON+FMA'}

# Upsample a 50 Hz sine wave from 128 Hz to 2048 Hz
fs_in = 128
fs_out = 2048
factor = fs_out // fs_in  # 16x upsampling

# Generate test signal
t = np.arange(0, 0.5, 1/fs_in)
signal = np.sin(2 * np.pi * 50 * t).astype(np.float32)

# Upsample
upsampled = upsample(signal, factor=factor, half_length=8)
print(f"Input: {len(signal)} samples at {fs_in} Hz")
print(f"Output: {len(upsampled)} samples at {fs_out} Hz")

Usage Examples

Single channel upsampling

import numpy as np
from sgnl_cpu_interp import upsample

# 1D signal (single channel)
signal = np.random.randn(1024).astype(np.float32)
upsampled = upsample(signal, factor=2)

Multichannel upsampling

# 2D array: (n_samples, n_channels)
n_samples, n_channels = 1024, 128
data = np.random.randn(n_samples, n_channels).astype(np.float32)

# Upsample by factor of 2
upsampled = upsample(data, factor=2)
print(upsampled.shape)  # (2016, 128) - note: loses 2*half_length samples

# Upsample by factor of 4 with longer kernel for better quality
upsampled = upsample(data, factor=4, half_length=16)
print(upsampled.shape)  # (3972, 128)

Transposed layout

from sgnl_cpu_interp import upsample_transposed

# Transposed layout: (n_channels, n_samples)
data = np.random.randn(128, 1024).astype(np.float32)
upsampled = upsample_transposed(data, factor=2)
print(upsampled.shape)  # (128, 2016)

API Reference

upsample(data, factor=2, half_length=8)

Upsample multichannel data using polyphase filtering (standard layout).

Parameters:

• data (ndarray): Input array of shape (n_samples,) for single channel or (n_samples, n_channels) for multichannel. float32, float64, complex64, and complex128 are processed natively and preserved in the output; any other dtype is converted to float32. Complex data is resampled by filtering the real and imaginary parts independently with the same real kernel.
• factor (int, optional): Upsampling factor (default: 2). Must be >= 2.
• half_length (int, optional): Half-length of the sinc kernel (default: 8). Larger values provide better quality but are slower. Total kernel length = 2 * half_length + 1.

Returns:

• output (ndarray): Upsampled array of shape ((n_samples - kernel_len + 1) * factor,) or ((n_samples - kernel_len + 1) * factor, n_channels) where kernel_len = 2 * half_length + 1. Same dtype as the (possibly converted) input.

upsample_transposed(data, factor=2, half_length=8)

Upsample multichannel data using polyphase filtering (transposed layout).

Same as upsample() but expects input in (n_channels, n_samples) layout. Use this when your data is already in channels-first format to avoid transpose overhead.

Parameters:

• data (ndarray): Input array of shape (n_channels, n_samples). Must be 2D; float32, float64, complex64, and complex128 are processed natively and preserved, other dtypes are converted to float32.
• factor (int, optional): Upsampling factor (default: 2). Must be >= 2.
• half_length (int, optional): Half-length of the sinc kernel (default: 8).

Returns:

• output (ndarray): Upsampled array of shape (n_channels, (n_samples - kernel_len + 1) * factor).

downsample(data, factor=2, half_length=8)

Downsample multichannel data with an anti-aliasing FIR filter (standard layout).

The signal is filtered with a Lanczos-windowed sinc lowpass (cutoff at the output Nyquist, unit DC gain) and decimated in a single pass — only every factor-th output is computed. Any integer factor >= 2 is supported (e.g. factor=1024 for 16384 Hz -> 16 Hz).

Parameters:

• data (ndarray): Input array of shape (n_samples,) or (n_samples, n_channels). float32, float64, complex64, and complex128 are processed natively and preserved; other dtypes are converted to float32.
• factor (int, optional): Downsampling factor (default: 2). Must be >= 2.
• half_length (int, optional): Filter half-length parameter (default: 8). The filter has 2 * half_length * factor + 1 taps.

Returns:

• output (ndarray): Downsampled array of length (n_samples - kernel_len) // factor + 1 per channel, where kernel_len = 2 * half_length * factor + 1. The group delay is exactly half_length * factor input samples.

downsample_transposed(data, factor=2, half_length=8)

Same as downsample() but expects (n_channels, n_samples) layout.

get_simd_info()

Get information about the current SIMD implementation.

Returns:

• dict with keys:
  • implementation: Name of current implementation (e.g., 'AVX2+FMA', 'NEON', 'Scalar')
  • available: List of all available implementations for this CPU
  • cpu_features: Detected CPU SIMD features

set_implementation(name)

Manually select a SIMD implementation. Useful for testing and benchmarking.

Parameters:

• name (str): Implementation name from get_simd_info()['available']

Can also be set via the SGNL_CPU_IMPL environment variable:

SGNL_CPU_IMPL=Scalar python my_script.py

Important Notes

• Edge loss: The convolution loses kernel_len - 1 samples from the edges. For half_length=8, you lose 16 input samples.
• Time alignment: The output has a delay of (kernel_len - 1) / 2 samples at the input sample rate.
• Minimum length: Input must have at least kernel_len samples.
• Uses Lanczos-windowed sinc kernel: h(x) = sinc(x/factor) * sinc(x/kernel_length)

Performance

Benchmark on 1024 channels, 1024 samples:

Platform	Implementation	Time	Notes
Apple Silicon (M-series)	NEON	~0.4 ms	Auto-selected
Apple Silicon	Scalar	~0.4 ms	Compiler auto-vectorizes well
x86_64 (Haswell+)	AVX2+FMA	~0.3 ms	Expected
x86_64 (older)	SSE2	~0.8 ms	Baseline x86_64

Comparison with previous GSL-based implementation:

Implementation	Time	Speedup
GSL BLAS (old)	9.8 ms	1.0x
This package	0.4 ms	~25x

Architecture

The package automatically detects CPU features at module load time and selects the best available implementation:

┌─────────────────────────────────────────────────────────┐
│                    Python API                           │
│         upsample() / upsample_transposed()              │
└─────────────────────────────────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────┐
│                  Runtime Dispatch                        │
│         cpu_detect() → select best implementation        │
└─────────────────────────────────────────────────────────┘
                           │
          ┌────────────────┼────────────────┐
          ▼                ▼                ▼
    ┌──────────┐    ┌──────────┐    ┌──────────┐
    │ AVX-512  │    │   NEON   │    │  Scalar  │
    │  AVX2    │    │  (ARM)   │    │(fallback)│
    │   AVX    │    └──────────┘    └──────────┘
    │  SSE4.1  │
    │  SSE2    │
    │  (x86)   │
    └──────────┘

Development

Environment

Any C toolchain and Python >= 3.10 will do:

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Alternatively, a nix flake provides a self-contained development environment (C toolchain, CMake, Ninja, Python, uv):

nix develop
uv venv --seed
uv pip install -e ".[dev]"

Building and testing

make build         # rebuild the C extension and reinstall (editable, incremental)
make test          # run the test suite
make lint          # linter
make format        # code formatter
make type-check    # static type checker
make bench         # downsample benchmark vs scipy (needs pip install ".[bench]")

The C extension is built with scikit-build-core + CMake; per-kernel SIMD flags live in CMakeLists.txt.

Project structure

fast-resample-cpu/
├── csrc/
│   ├── cpu_detect.c      # Runtime CPU feature detection
│   ├── dispatch.c        # Function pointer dispatch table
│   ├── resample_ext_simd.c  # Python extension wrapper
│   └── kernels/
│       ├── convolve_scalar.c   # Baseline implementation
│       ├── convolve_sse2.c     # x86 SSE2
│       ├── convolve_sse4.c     # x86 SSE4.1
│       ├── convolve_avx.c      # x86 AVX
│       ├── convolve_avx2.c     # x86 AVX2+FMA
│       ├── convolve_avx512.c   # x86 AVX-512
│       └── convolve_neon.c     # ARM NEON
├── src/
│   └── sgnl_cpu_interp/  # Python API
├── CMakeLists.txt        # C extension build configuration
├── pyproject.toml        # Package metadata and build backend
└── tests/                # Test suite

Adding a new SIMD implementation

1. Create csrc/kernels/convolve_<name>.c implementing convolve_<name>() and convolve_transposed_<name>()
2. Add the implementation to the dispatch table in csrc/dispatch.c
3. Add per-file compile flags in CMakeLists.txt (see the existing set_source_files_properties calls)
4. Add CPU feature detection if needed in csrc/cpu_detect.c

Algorithm

This implementation uses polyphase filtering for efficient upsampling:

1. Kernel generation: Creates a Lanczos-windowed sinc kernel and splits it into factor polyphase components
2. SIMD convolution: Vectorized dot product across channels (standard layout) or time samples (transposed layout)
3. Phase-blocked upsampling: Processes all output samples with the same phase together to maximize kernel data reuse in cache

The approach is specifically optimized for:

• Many channels (100+)
• Small to moderate upsampling factors (2-16x)
• Short to medium input lengths (100s to 1000s of samples)

License

MIT License - see LICENSE file for details.

Contributing

Contributions welcome! Please open an issue or merge request on git.ligo.org.

Citation

If you use this in research, please cite:

@software{sgnl_cpu_interp,
  title = {sgnl-cpu-interp: Fast polyphase resampling with multi-architecture SIMD},
  url = {https://git.ligo.org/greg/fast-resample-cpu},
  year = {2025}
}