call_inverse_func incorrectly shortens width

[kornelski]

If the input is a 2d array, then width.min(32) is inappropriate - it will cause a skew instead of narrowing the width. It needs to be something like input.chunks_exact(width).map(|row| &row[..width.min(32)]) to preserve the stride.

rav1e/src/asm/shared/transform/inverse.rs

Line 27 in 21bb4cb

let input: &[T::Coeff] = &input[..width.min(32) * height.min(32)];

[barrbrain]

This is an example of a misleading comment. It is a precondition of call_inverse_func() that the stride of input is width.min(32).

[kornelski]

Stride can't be shorter than the width? I assume you mean the width can't be > 32, but if I put debug_assert!(width <= 32), then these tests fail:

asm::shared::transform::inverse::test::inv_txfm2d_add_dct_dct_64x16
asm::shared::transform::inverse::test::inv_txfm2d_add_dct_dct_64x32
asm::shared::transform::inverse::test::inv_txfm2d_add_dct_dct_64x64