Use SSE/MMX/Altivec for faster CTMF #432
Conversation
| * Modified from code | ||
| * Copyright (C) 2006 Simon Perreault | ||
| * | ||
| * Reference: S. Perreault and P. Hébert, "Median Filtering in Constant Time", |
|
Is anyone able to test this on Windows? /cc @blink1073 |
|
Will do, tonight |
| @@ -17,6 +17,11 @@ cimport cython | |||
| from libc.stdlib cimport malloc, free | |||
| from libc.string cimport memset | |||
|
|
|||
| cdef extern from "_histogram.h": | |||
| ctypedef unsigned short int uint16_t | |||
There was a problem hiding this comment.
Use np.uint16_t here and below?
|
Good to go, but failed the following two unrelated tests: |
|
ping @thouis |
|
I think I got everything above (plus cleaned up compilation a bit with some casts). |
| @@ -323,33 +327,17 @@ cdef void set_stride(Histograms *ph, SCoord *psc): | |||
| # | |||
| ############################################################################ | |||
| cdef inline np.int32_t tl_br_colidx(Histograms *ph, np.int32_t colidx): | |||
| return (colidx + 3*ph.radius + ph.current_row) % ph.stripe_length | |||
| return <np.int32_t> (colidx + 3*ph.radius + ph.current_row) % ph.stripe_length | |||
|
This looks good to me! Thanks, @thouis . If you can, just do a pass through the PEP8 checker, otherwise I can do that when I merge. |
|
Bah. I accidentally deleted the branch instead of pushing a new version to it. Anyway, fixed the pep8. Some long lines remain, but I thought they were more readable than breaking them up. |
|
Merged. |
Use SSE/MMX/Altivec for faster CTMF
This speeds up CTMF by about a factor of 2 on my machine.
I don't have access to a non SSE machine, so some testing is in order.