GPU X-Engine

A CUDA reimplementation of the FPGA cross-correlation engine (X-engine) from the ISI Digital Spectrometer/Correlator, a mid-infrared astronomy signal processing system deployed at Mount Wilson Observatory in 2010. I built the original FPGA system and wanted to explore how the same DSP pipeline maps to a GPU architecture.

The F-engine simulator generates synthetic input matching the original FPGA output format: complex spectra from 3 antennas, channelized into 64 bins, and quantized to 4 bits. Two GPU kernels perform the auto- and cross-correlations:

1. Corner turn — transposes the data from [spectrum][antenna][channel] to [channel][antenna][spectrum] to group spectra by channel.
2. Cross-correlate + integrate — unpacks 4-bit samples, computes all baseline products per channel, and integrates over a configurable number of spectra.

All arithmetic is integer (int8 inputs, int32 accumulators), so we can validate the GPU results by exact match against the CPU results.

Results

The original Virtex-5 FPGA processed 45.0 M spectra/sec. This GPU implementation reaches 69.8 M spectra/sec on an A40, and is transfer-bound: H2D accounts for >50% of wall time.

Performance History

Version	GPU	Change	Spectra/sec	vs. FPGA
v0.1	A40	Naive implementation	32.5 M	72%
v0.2	A40	Move unpack from corner_turn kernel to correlation kernel	34.6 M	77%
v0.3	A40	Tile corner turn in shared memory with coalesced reads	39.7 M	88%
v0.4	A40	Coalesce writes in corner turn	56.1 M	125%
v0.5	A40	Pinned host memory for H2D transfer	80.3 M	178%
v0.6	A40	Generalize to arbitrary N_ANTENNAS	69.8 M	155%

Build

Requires CUDA toolkit (nvcc). Override SM_ARCH for target GPU:

make SM_ARCH=sm_86      # Default: sm_86

Run

./test_cpu              # Confirm correctness of CPU reference implementation
./test_gpu              # Validate GPU results against CPU reference results
./benchmark             # Measure throughput of pipeline stages

Source

File	Description
inc/common.h	System parameters, data types, utilities, macros
src/fengine_sim.cu	Simulated F-engine: generates packed 4-bit test data
src/xengine_cpu.cu	CPU cross-correlator for validation
src/corner_turn.cu	GPU kernel: transpose spectra
src/xengine_gpu.cu	GPU kernel: cross-correlate + integrate
src/test_cpu.cu	CPU test driver
src/test_gpu.cu	GPU test driver (validates against CPU reference)
src/benchmark.cu	Performance benchmarking (GPU vs original FPGA)