Try using load_and_convert_image instead of load_image

Thanks. I tested it and it worked faster than OpenCV. That's awesome :)

Just two more question:

1- Tools::save_image cannot save float output. I don't need to save. But want to see the output. Is there a way to convert Halide::Buffer to cv::Mat (OpenCV) and vice-versa without copying the data?

2- How can I see the Halide-generated C or C++ code?

Also, is there any macro to set the number of threads in Halide?
OpenCV's GaussianBlur works single-threaded. I want to know that if Halide just uses one core, would it be still faster than OpenCV?

I used this code to convert OpenCV's Mat to Halide::Buffer:

Halide::Buffer<float> in((float *)OpenCV_Mat.data, OpenCV_Mat.rows, OpenCV_Mat.cols, OpenCV_Mat.channels());

and used this do vice-versa:

Mat OpenCV_Output(Halide_Output.width(), Halide_Output.height(), CV_32F, Halide_Output.data());

But the Halide-blurred image seems strange! I didn't change the formula here.

Also, is there any macro to set the number of threads in Halide?
OpenCV's GaussianBlur works single-threaded. I want to know that if Halide just uses one core, would it be still faster than OpenCV?

call halide_set_num_threads(1);
before calling your function.

To my understanding, to boost up the performance, you might want to scheduler it with single-thread in mind.(without using parallel() )

I used this code to convert OpenCV's Mat to Halide::Buffer:

Halide::Buffer<float> in((float *)OpenCV_Mat.data, OpenCV_Mat.rows, OpenCV_Mat.cols, OpenCV_Mat.channels());

and used this do vice-versa:

Mat OpenCV_Output(Halide_Output.width(), Halide_Output.height(), CV_32F, Halide_Output.data());

But the Halide-blurred image seems strange! I didn't change the formula here.

check this post
https://stackoverflow.com/questions/39970361/how-to-convert-opencv-mat-to-halide-image-and-back

Thanks. I tested it and it worked faster than OpenCV. That's awesome :)

Just two more question:

1- Tools::save_image cannot save float output. I don't need to save. But want to see the output. Is there a way to convert Halide::Buffer to cv::Mat (OpenCV) and vice-versa without copying the data?

2- How can I see the Halide-generated C or C++ code?

Please refer to lesson-21
If you are using halide generator, add c_source to -e.
It will give you a xxx.halide_generated.cpp

./lesson_21_generate -o . -g auto_schedule_gen -f auto_schedule_true -e static_library,c_source,h,schedule target=host auto_schedule=true machine_params=32,16777216,40

I updated that stack overflow answer with the current way to do this. The issue is that opencv stores data interleaved by default. This is slow, unfortunately, but Halide::Runtime::Buffer represents this like so:

Halide::Runtime::Buffer<float>::make_interleaved(image.data, image.cols, image.rows, image.channels());

The catch is that the generated code will expect planar inputs by default. See tutorial lesson 16 for details of how to work with different formats.