GRADIENT_DESCENT does not work as expected

[mshafiei]

Hi,

I'm trying to get the gradient of local laplacian generator. When I add GRADIENT_DESCENT_OPT "GRADIENT_DESCENT" to python_bindings/apps/CMakeLists.txt:48 the local laplacian app breaks. The output image turns out to be pitch black. Also the gradient image is black (you need to add an additional buffer as the last parameter list of the generator in the local_laplacian_app.py). Is there any workaround for this issue?

[abadams]

Sounds like something is broken, but before investigating I wanted to point out that the gradient w.r.t. the input image is going to be zero for that algorithm anyway, because it starts with a nearest-neighbor-sampled lookup table.

[mshafiei]

I'm actually interested in the derivative w.r.t. alpha. I think that doesn't rely on the nearest neighbor look up. Is that correct? I also tried gradient on other generators e.g. blur, bilateral grid but none of them seem to work.

[mshafiei]

The issue happens for any generator (I've tried LLF, blur, and bilateral filtering). Follow the steps below to reproduce:

1. Add GRADIENT_DESCENT_OPT "GRADIENT_DESCENT" to add_halide_library in ${HALIDE_PATH}/python_bindings/apps/CMakeLists.txt.
2. Build Halide
3. Modify build_app.py, store the gradient buffer in an image using grad_buf_u8 = grad_buf.astype(np.uint8) and halide.imageio.imwrite(grad_path, grad_buf_u8)
4. Gradient image is black
5. Output image is also corrupted

[alexreinking]

Please consult the documentation for add_halide_library: https://github.com/halide/Halide/blob/main/doc/HalideCMakePackage.md#add_halide_library

The option is GRADIENT_DESCENT, no _OPT suffix, and it does not accept an argument. For example:

add_halide_library(
  lib FROM gen
  GRADIENT_DESCENT
)

[mshafiei]

Thanks @alexreinking. I think CMake still picked up GRADIENT_DESCENT and the unnecessary GRADIENT_DESCENT_OPT that I have specified was ignored. So the issue still exists. This code snippet reproduces it.

from blur import blur
import halide.imageio
import numpy as np
import sys
import timeit


if __name__ == "__main__":
    input_path = sys.argv[1]
    output_path = sys.argv[2]
    grad_path = sys.argv[3]
    timing_iterations = 10

    print("Reading from %s ..." % input_path)
    input_buf_u8 = halide.imageio.imread(input_path)[0,:,:]
    assert input_buf_u8.dtype == np.uint8
    
    # Convert to uint16... but remember that the blur() generator
    # is documented as only working on <= 14 bits of image; if
    # we use the upper two bits we'll get incorrect results.
    # We'll just leave it with 8 bits of useful data.
    input_buf = input_buf_u8.astype(np.uint16)
    output_buf = np.empty(input_buf.shape, dtype=input_buf.dtype)
    grad_buf = np.empty(input_buf.shape, dtype=input_buf.dtype)

    blur(input_buf, output_buf, grad_buf)

    output_buf_u8 = output_buf.astype(np.uint8)
    
    print("Saving to %s ..." % output_path)
    halide.imageio.imwrite(output_path, output_buf_u8)

    print("Saving to %s ..." % grad_path)
    grad_buf_u8 = grad_buf.astype(np.uint8)
    halide.imageio.imwrite(grad_path, grad_buf_u8)

[mshafiei]

@alexreinking @abadams Can you reproduce the issue with the code snippet above?

[alexreinking]

We're all out this week for Thanksgiving. I can take a look next week!

[abadams]

Bump. I don't feel qualified to look at this one, because it's probably something related to buffer management in the python layer.

[steven-johnson]

FWIW, enabling GRADIENT_DESCENT for the Python apps in this way seems to take ~forever to compile -- investigating

EDIT: building in verbose mode reveals Warning: Autoscheduling is not enabled in build_gradient_module(), so the resulting gradient module will be unscheduled; this is very unlikely to be what you want.

[steven-johnson]

The option is GRADIENT_DESCENT, no _OPT suffix

Why didn't this produce an error at CMake configure time? Surely unknown keywords should fail.