Not able to save image in png format while using Halide - png

Question

I tried running the following program on a Windows machine using Visual Studio:
#include <Halide.h>
#include "halide_image_io.h"
#include "png.h"
using namespace Halide;
using namespace Halide::Tools;
int main(int argc, char** argv)
{
Buffer<uint8_t> in = load_image("images/rgb.png");
Func blurx, out;
Var x, y,c, xi, yi;
printf("width : %d, height: %d and channels: %d",in.width(),in.height(),in.channels());
//width = 768, height = 1280
blurx(x, y,c) = (in(x, y,c) + in(x, y,c) + in(x, y,c)) / 3.0f;
out(x, y,c) = (blurx(x, y,c) + blurx(x, y,c) + blurx(x, y,c)) / 3.0f;
out.tile(x, y, xi, yi, 256, 32).vectorize(xi, 8).parallel(y);
Buffer<uint8_t> result = out.realize(in.width(), in.height(),in.channels());
save_image(result, "output/output.png");
return 0;
}
I am getting the error "Image cannot be saved in this format". The error does not occur when I remove the '/3.0f'. So, division may be causing some pixel values to be an invalid format. Hence, I am not able to save it in .png format. How can I solve this?
Please note: the formula should have been (in(x-1, y,c) + in(x, y,c) + in(x+1, y,c)) / 3.0f; .... but, that gives me error"access out of boundaries of input buffer".... I am trying to resolve the division error first, so for the time being, I modified the formula .. which helped me to catch this error.

Answer 1

Adding the division by 3.0f changes the inferred type of out (and blurx) from uint8_t to float. I believe the issue is that the save_image helper doesn't allow saving floating point buffers as PNGs. To get back to uint8_t try wrapping the whole expression defining out in a cast<uint8_t>(...).
(Note: you also probably want to upcast the input values at least uint16_t to avoid quickly overflowing when you add 3 of them together.)

Answer 2

I made the following changes and it works now:
#include <Halide.h>
#include "halide_image_io.h"
#include "png.h"
using namespace Halide;
using namespace Halide::Tools;
int main(int argc, char** argv)
{
Buffer<uint8_t> in = load_image("images/rgb.png");
Func blurx, out;
Var x, y,c, xi, yi;
Func in_bounded = BoundaryConditions::repeat_edge(in);
Func input_16;
input_16(x, y) = cast<uint16_t>(in_bounded(x, y));
blurx(x, y,c) = (input_16(x, y,c) + input_16(x, y,c) + input_16(x, y,c)) / 3;
out(x, y,c) = cast<uint8_t>(blurx(x, y,c) + blurx(x, y,c) + blurx(x, y,c)) / 3);
out.tile(x, y, xi, yi, 256, 32).vectorize(xi, 8).parallel(y);
Buffer<uint8_t> result = out.realize(in.width(), in.height(),in.channels());
save_image(result, "output/output.png");
return 0;
}