Calculation range fail about halide-hls

[xuanyoya]

In you original code, both input and output are 3 dimensional, especially you bound the 3rd dimension of output using output.bound(c, 0, 3); so please use 3 dimensions in run.cpp. If you want to reduce the 3rd dimension of input, please change

local_sum(x, y, c) += cast<uint16_t>(in(x+r.x, y+r.y, c)) * weight(r.x+2, r.y+2); to
local_sum(x, y) += cast<uint16_t>(in(x+r.x, y+r.y, r.z)) * weight(r.x+2, r.y+2);

And also change some other code accordingly. But in that case, it is not in the scope of this issue, so the current issue will be closed.

[xuanyoya]

There are several issues in your code. Main reason is your tile size is 1 x 1, and window size is 3 x 3, that makes the maximum size of your image tile -1 x -1. What you should change in your code is:

1. local_sum(x, y, c) += cast<uint16_t>(in(x+r.x, y+r.y, c)) * weight(r.x+2, r.y+2); change to local_sum(x, y, c) += cast<uint16_t>(in(x+r.x, y+r.y, c)) * weight(r.x+1, r.y+1); so that the weight is accessed at the right position.
2. hw_output.tile(x, y, xo, yo, xi, yi, 1, 1); change to hw_output.tile(x, y, xo, yo, xi, yi, 16, 16);, note 16 is just an example, but you should use a number that larger than the window size.
3. input_buf(x,y,c) = clamped(x,y,c); get removed, since it is not used.

[chingyi071]

Thank you @xuanyoya for detail description. However, the max(-1) problem still occur after the modifications. Moreover, I have tried to use different size of input buffer, but it didn't change anything. Is there any other potential problem?

[xuanyoya]

@chingyi071 I didn't read your run.cpp file before. But this time, I realized you also have another error in this file. Since you are doing a convolution with window 3 x 3, so the output image size will be smaller than the input image size, you should NOT specify your output image to be:

    BufferMinimal<uint8_t> out_native(input.width(),input.height());
    BufferMinimal<uint8_t> out_hls(input.width(), input.height());

Instead, change input.width() and input.height() to some numbers that smaller than or equal to input.width() - 2 and input.height() - 2 .

Let me know the result. If that passes, I will close this issue.

[chingyi071]

@xuanyoya I modify it but it still fail. But the error message become Error: Output buffer output is accessed at 2, which is beyond the max (-1) in dimension 2. The original error message is accessed at 0. I'm not sure whether this phenomenon can help you

[xuanyoya]

@chingyi071 Besides, for this kind of errors, you can just try the same program and same schedule with original C backend or with original Halide repo. If you see this error occurs with the original C backend or original Halide repo, then the fastest way for you debug is to post this question on stackoverflow or Halide developer mailing list.

[xuanyoya]

Since you are doing repeat_edge as clamped = BoundaryConditions::repeat_edge(input);, I don't think my previous suggestion is helpful. The change you should make is

    BufferMinimal<uint8_t> out_native(input.width(),input.height(), input.channels());
    BufferMinimal<uint8_t> out_hls(input.width(), input.height(), input.channels());

Besides, for this line pipeline_hls(input, weight, out_native);, it should be pipeline_hls(input, weight, out_hls);

[chingyi071]

The dimension of input image will be reduced from 3 to 2 in the function, so I remove input.channels() in output buffer declaration. I fix the input argument in pipeline_hls() but it still failed.

[xuanyoya]

I didn't see how you reduced the dimension from 3 to 2, and I have tried with my modification, it already passed.

[chingyi071]

@xuanyoya Sorry I just use reduced version of my function. This code does not reduce the dimension. And I pass the test. Thank you very much