Correlator

The Correlator is the compute block of a BLADE interferometry pipeline. It cross-correlates dual-polarization antenna voltages into baseline visibilities, producing the four polarization products XX, XY, YX, and YY for every antenna pair, including the autocorrelations. Successive input buffers can be accumulated into a single output with the integration rate, and the result pairs naturally with a stacked writer such as the UVH5 Writer.

How it works

Each compute cycle launches a runtime-compiled CUDA kernel. A thread block owns one channel and one contiguous slice of the time axis. It stages every antenna's voltages for that slice into shared memory, then each thread reduces one 2x2 tile of the upper-triangular baseline matrix over the whole slice, holding the four polarization products of its four baselines in registers. Input is therefore read from global memory exactly once per launch rather than once per baseline, and the only atomics are a single flush per baseline at the end of each block. The output is zeroed at the start of each integration window and the buffer is only emitted once integrationRate inputs have been accumulated. In between, the block skips downstream processing.

For eligible CI8 input the kernel packs four consecutive time samples into each 32-bit word and correlates them with dp4a, the four-way 8-bit dot-product instruction. Every product, and every sum of products over one input buffer, is an exact integer that fits in 32 bits, so this path evaluates the correlation exactly and calculationMode has no effect on the result. When packing is unavailable, CI8 falls back to the generic path and uses the configured calculation mode. A CF32 input always takes the generic path and supports single-precision or double-precision floating-point calculation. Integer calculation is rejected.

Floating-point addition depends on order. If several CUDA blocks added partial results to the same visibility, their completion order could change the lowest output bits between runs. The correlator avoids that behavior so identical input and configuration produce bit-reproducible visibilities.

The packed CI8 path can safely split the time axis because each block adds its partial result to an exact 32-bit integer accumulator. Integer addition does not depend on block completion order. After the complete input buffer has been reduced, a second kernel converts the result to CF32 and adds it to the integration output once. The generic path, used by CF32 and unpacked CI8, assigns the complete time axis to one block instead. This gives up time-axis parallelism but keeps the calculation order fixed. In both paths, each visibility receives one floating-point addition per input buffer, and the CUDA stream processes those buffers in order.

Both paths derive their launch geometry from the input shape. The antenna axis is padded to an even count, and the shared-memory stage is sized to the 48 KB static budget. The packed CI8 path may split the time axis across several blocks. It requires the time axis to tile by four and to be at most 65535 samples. Otherwise CI8 falls back to the generic path, which uses one time chunk.

Support matrix

Input condition	integer	single_precision_fp	double_precision_fp
CI8, packed eligible	Packed + accumulate: safe, CF32 rounding	Packed + accumulate: safe, CF32 rounding	Packed + accumulate: safe, CF32 rounding
CI8, packed unavailable and samples <= 65535	Generic: safe, CF32 rounding	Generic: safe, CF32 rounding	Generic: safe, CF32 rounding
CI8, samples > 65535	Rejected to prevent integer overflow	Generic: conditional on integration size	Generic: conditional on integration size
CF32	Rejected to prevent integer overflow	Generic: conditional on input magnitude	Generic: recommended, but CF32 output can still overflow

Packed eligibility requires a time dimension divisible by four and no greater than 65535 samples, as well as launch geometry that fits the shared-memory budget. Safe means overflow-safe, not numerically exact after conversion to the CF32 output.

Configuration

Parameter	Type	Default	Description
integrationRate	integer	1	Number of input buffers accumulated into each output buffer.
conjugateAntennaIndex	integer	1	Which antenna of the pair is conjugated, 0 for antenna A and 1 for antenna B.
calculationMode	string	double_precision_fp	Intermediate calculation precision. See the support matrix for available modes.

Input

Name	Description
buffer	Contiguous CI8 or CF32 tensor shaped [antennas, channels, samples, polarizations] with exactly two polarizations.

Output

Name	Description
buffer	CF32 tensor shaped [baselines, channels, 1, 4] where baselines is antennas * (antennas + 1) / 2.