Stream-based Kernels - Stream-based Kernels - 2026.1 English - UG1864

Stream-based kernels process data incrementally as it becomes available. They read from the input stream and write to the output stream sample-by-sample.

Partial data (fewer samples than the kernel loop):

• Input: 4 samples (kernel loops 8 times)
• Output: 4 samples
• Behavior: the kernel processes the available data and produces partial output immediately.

input  = [1, 2, 3, 4]
output = kernel.run(va.array(input, dtype=va.int32))
# output: [2, 4, 6, 8]

Exact data (matches the loop count):

• Input: 8 samples → output: 8 samples. The kernel completes one full processing cycle.

Oversized data (multiple of the loop count):

• Input: 16 samples → output: 16 samples. The kernel loop runs twice automatically.

input  = list(range(1, 17))
output = kernel.run(va.array(input, dtype=va.int32))
# output: [2, 4, 6, ..., 32]

Characteristics:

• Incremental processing
• Immediate output for partial data
• Suitable for streaming applications and pipelined dataflow
• Lower latency for partial results