Use the Stencil Backend

Author: Yuze Chi (Blaok)

In this tutorial, we show how to use the stencil backend in HeteroCL and generate HLS C++ code as the result.

import numpy as np

import heterocl as hcl

Stencil Comptuation

Stencil kernels compute output based on a sliding window over the input. The following shows an example. It computes the average over a 5-point window.

def jacobi(input_image, output_image):
    def jacobi_kernel(y, x):
        return (input_image[y+1, x-1] +
                input_image[y  , x  ] +
                input_image[y+1, x  ] +
                input_image[y+1, x+1] +
                input_image[y+2, x  ]) / 5

    return hcl.update(output_image, jacobi_kernel, name=output_image.name)

Use the Stencil Backend

HeteroCL provides a special backend for stencil computation kernels. It can be used via the target argument when building a program.

dtype = hcl.Float()
input_image = hcl.placeholder((480, 640), name="input", dtype=dtype)
output_image = hcl.placeholder((480, 640), name="output", dtype=dtype)
soda_schedule = hcl.create_schedule([input_image, output_image], jacobi)
print(hcl.build(soda_schedule, target='soda'))

Increase Parallelism

The above program is written in the SODA DSL, which provides advanced optimizations to stencil kernels. One of the optimizations is to provide scalable parallelism. To increase parallelism, one can unroll the inner-most stencil loop, as follows. The same SODA DSL will be generated, except the unroll factor will become 8.

tensor = jacobi.output
axis = tensor.axis
soda_schedule[tensor].unroll(axis[len(axis) - 1], factor=8)
print(hcl.build(soda_schedule, target='soda'))

Generatel HLS C++ Code

The SODA DSL certainly does not compile directly. It needs to be passed to the SODA Compiler. HeteroCL provides a built-in target that generates HLS C++ code from the intermediate SODA code directly. The generated C++ code is valid HLS code and can be passed to HLS vendor tools without modifications.

print(hcl.build(soda_schedule, target='soda_xhls'))