I was inspecting cblas_matrixproduct in cblasfuncs.c (numpy/numpy/blob/main/numpy/core/src/common/cblasfuncs.c), which handles matrix multiplication.

Currently, the function checks for negative strides, broadcasts, and possible misalignment via _bad_strides(). If any of these conditions are met, it performs a full copy of either a or b at the beginning of the function to ensure compatibility with BLAS.

An enhancement here could involve extending _bad_strides() to return more descriptive statuses (e.g., STRIDE_OK, STRIDE_NEGATIVE, etc.) rather than just a boolean. Then to add logic to avoid copies in specific cases — such as vector-vector, matrix-vector, or vector-matrix operations — by adjusting parameters lda, incX, and incY instead of copying the array.

As @ricardoV94 mentioned, it’s unclear whether the gain is worth the added complexity. Still, I wanted to get a second opinion — if this seems useful, I’d be happy to take it on.

cc: @eendebakpt

Making a copy of the array with negative strides is also quite expensive(%timeit A_flipped.copy()), so making a copy first and then performing the np.dot does not help.

@niranjanorkat I cannot judge whether this is useful or not. It depends on how much complexity the additional code adds (probably hard to judge without actually implementing), and how often this case of negative strides occurs in actual code.

and how often this case of negative strides occurs in actual code.

Negative strides are common for us when computing gradient of Cumsum/Convolution functions