3x3 blur in Halide

blur.gen.cpp

#include <Halide.h>
using namespace Halide;

class Blur3x3 : public Generator<Blur3x3> {
public:
    Input<Buffer<uint8_t>> _input{"input", 2};
    Output<Buffer<uint8_t>> blurV{"blurV", 2};

    Func blurH{"blurH"};
    Var x, y, xi, yi;

    void generate() {
        Func input;
        input(x, y) = cast<uint16_t>(BoundaryConditions::repeat_edge(_input)(x, y));
        blurH(x, y) = (input(x - 1, y) + input(x, y) + input(x + 1, y)) / 3;
        blurV(x, y) = cast<uint8_t>((blurH(x, y - 1) + blurH(x, y) + blurH(x, y + 1)) / 3);
    }

    void schedule() {
        blurV.tile(x, y, xi, yi, 256, 32)
            .vectorize(xi, 8)
            .parallel(y);
        blurH.compute_at(blurV, x)
            .vectorize(x, 8);
    }
};

HALIDE_REGISTER_GENERATOR(Blur3x3, blur3x3)

CMakeLists.txt

cmake_minimum_required(VERSION 3.14)
project(blur3x3)

# Set up language settings
set(CMAKE_CXX_STANDARD 11)
set(CMAKE_CXX_STANDARD_REQUIRED YES)
set(CMAKE_CXX_EXTENSIONS NO)

# Find Halide
find_package(Halide REQUIRED)

add_executable(blur_generator blur.gen.cpp)
target_link_libraries(blur_generator PRIVATE Halide::Generator)

add_halide_library(blur3x3 FROM blur_generator
                   REGISTRATION REG_CPP)

add_executable(run_blur ${REG_CPP})
target_link_libraries(run_blur PRIVATE Halide::RunGenMain blur3x3)

Note that this requires the forthcoming CMake support in Halide's cmake-modern branch.