TST: Calculate RMS and diff image in C++ #29102
Conversation
The current implementation is not slow, but uses a lot of memory per image. In `compare_images`, we have: - one actual and one expected image as uint8 (2×image) - both converted to int16 (though original is thrown away) (4×) which adds up to 4× the image allocated in this function. Then it calls `calculate_rms`, which has: - a difference between them as int16 (2×) - the difference cast to 64-bit float (8×) - the square of the difference as 64-bit float (though possibly the original difference was thrown away) (8×) which at its peak has 16× the image allocated in parallel. If the RMS is over the desired tolerance, then `save_diff_image` is called, which: - loads the actual and expected images _again_ as uint8 (2× image) - converts both to 64-bit float (throwing away the original) (16×) - calculates the difference (8×) - calculates the absolute value (8×) - multiples that by 10 (in-place, so no allocation) - clips to 0-255 (8×) - casts to uint8 (1×) which at peak uses 32× the image. So at their peak, `compare_images`→`calculate_rms` will have 20× the image allocated, and then `compare_images`→`save_diff_image` will have 36× the image allocated. This is generally not a problem, but on resource-constrained places like WASM, it can sometimes run out of memory just in `calculate_rms`. This implementation in C++ always allocates the diff image, even when not needed, but doesn't have all the temporaries, so it's a maximum of 3× the image size (plus a few scalar temporaries).
|
So I no longer have any memory-based skips on the PR adding WASM, but maybe we still want to do this to save memory in general? |
|
This seems to make sense! Should we also use this in |
| PNG via the `.converter` dictionary. The underlying RMS is calculated | ||
| with the `.calculate_rms` function. | ||
| in a similar way to the `.calculate_rms` function. |
There was a problem hiding this comment.
I think what's important here is how these methods differ? (what's the takeaway supposed to be here?)
There was a problem hiding this comment.
The only reason I wrote it vaguely is that I didn't want people to think they could monkeypatch calculate_rms and expect the image_comparison decorator / compare_images to use it. The algorithm is otherwise the same, I think.
| if (expected_image.ndim() != 3) { | ||
| auto exceptions = py::module_::import("matplotlib.testing.exceptions"); | ||
| auto ImageComparisonFailure = exceptions.attr("ImageComparisonFailure"); | ||
| py::set_error( | ||
| ImageComparisonFailure, | ||
| "Expected image must be 3-dimensional, but is {ndim}-dimensional"_s.format( | ||
| "ndim"_a=expected_image.ndim())); | ||
| throw py::error_already_set(); | ||
| } | ||
|
|
||
| if (actual_image.ndim() != 3) { | ||
| auto exceptions = py::module_::import("matplotlib.testing.exceptions"); | ||
| auto ImageComparisonFailure = exceptions.attr("ImageComparisonFailure"); | ||
| py::set_error( | ||
| ImageComparisonFailure, | ||
| "Actual image must be 3-dimensional, but is {ndim}-dimensional"_s.format( | ||
| "ndim"_a=actual_image.ndim())); | ||
| throw py::error_already_set(); | ||
| } |
There was a problem hiding this comment.
Can this be done in a loop since it's the same test/error message?
| for (auto i = 0; i < height; i++) { | ||
| for (auto j = 0; j < width; j++) { | ||
| for (auto k = 0; k < depth; k++) { | ||
| auto pixel_diff = static_cast<double>(expected(i, j, k)) - |
There was a problem hiding this comment.
is it more efficient to case the individual values than the whole array?
There was a problem hiding this comment.
I'm not sure what you mean here? There's no condition.
|
|
||
| if (k != 3) { // Hard-code a fully solid alpha channel by omitting it. | ||
| diff(i, j, k) = static_cast<unsigned char>(std::clamp( | ||
| abs(pixel_diff) * 10, // Expand differences in luminance domain. |
There was a problem hiding this comment.
why are you only doing this is for rgba?
There was a problem hiding this comment.
The alpha channel is ignored, just as with compare_images.
PR summary
The current implementation is not slow, but uses a lot of memory per image.
In
compare_images, we have:which adds up to 4× the image allocated in this function.
Then it calls
calculate_rms, which has:which at its peak has 16× the image allocated in parallel.
If the RMS is over the desired tolerance, then
save_diff_imageis called, which:which at peak uses 32× the image.
So at their peak,
compare_images→calculate_rmswill have 20× the image allocated, and thencompare_images→save_diff_imagewill have 36× the image allocated. This is generally not a problem, but on resource-constrained places like WASM, it can sometimes run out of memory just incalculate_rms.This implementation in C++ always allocates the diff image, even when not needed, but doesn't have all the temporaries, so it's a maximum of 3× the image size (plus a few scalar temporaries).
PR checklist