diff options
| author | Carl Hetherington <cth@carlh.net> | 2020-11-16 23:56:11 +0100 |
|---|---|---|
| committer | Carl Hetherington <cth@carlh.net> | 2020-11-17 00:16:58 +0100 |
| commit | fcc0515087a0b2bf9af5043af1c324367f1c152c (patch) | |
| tree | 2b4651e72eba63da753c6463fcc22319b0d242bb /src/lib/image.cc | |
| parent | d676cd5115d8c79a547742364f2037435c14a54a (diff) | |
Use 64-byte alignment for aligned image memory.
FFmpeg uses this values if AVX512 is available, and with only
32-byte alignment I am seeing strange scaling effects whereby
crop_scale_window_test7 gives black bars down the right side of
cropped images (when run on an i7 7700).
Diffstat (limited to 'src/lib/image.cc')
| -rw-r--r-- | src/lib/image.cc | 15 |
1 files changed, 10 insertions, 5 deletions
diff --git a/src/lib/image.cc b/src/lib/image.cc index 9dae94f7c..2d3a4a1d5 100644 --- a/src/lib/image.cc +++ b/src/lib/image.cc @@ -55,6 +55,11 @@ using std::runtime_error; using boost::shared_ptr; using dcp::Size; + +/** The memory alignment, in bytes, used for each row of an image if aligment is requested */ +#define ALIGNMENT 64 + + int Image::vertical_factor (int n) const { @@ -838,7 +843,7 @@ Image::bytes_per_pixel (int c) const * * @param p Pixel format. * @param s Size in pixels. - * @param aligned true to make each row of this image aligned to a 32-byte boundary. + * @param aligned true to make each row of this image aligned to a ALIGNMENT-byte boundary. */ Image::Image (AVPixelFormat p, dcp::Size s, bool aligned) : _size (s) @@ -862,7 +867,7 @@ Image::allocate () for (int i = 0; i < planes(); ++i) { _line_size[i] = ceil (_size.width * bytes_per_pixel(i)); - _stride[i] = stride_round_up (i, _line_size, _aligned ? 32 : 1); + _stride[i] = stride_round_up (i, _line_size, _aligned ? ALIGNMENT : 1); /* The assembler function ff_rgb24ToY_avx (in libswscale/x86/input.asm) uses a 16-byte fetch to read three bytes (R/G/B) of image data. @@ -875,7 +880,7 @@ Image::allocate () Further to the above, valgrind is now telling me that ff_rgb24ToY_ssse3 over-reads by more then _avx. I can't follow the code to work out how much, - so I'll just over-allocate by 32 bytes and have done with it. Empirical + so I'll just over-allocate by ALIGNMENT bytes and have done with it. Empirical testing suggests that it works. In addition to these concerns, we may read/write as much as a whole extra line @@ -901,12 +906,12 @@ Image::allocate () |XXXwrittenXXX|<------line-size------------->|XXXwrittenXXXXXXwrittenXXX ^^^^ out of bounds */ - _data[i] = (uint8_t *) wrapped_av_malloc (_stride[i] * (sample_size(i).height + 1) + 32); + _data[i] = (uint8_t *) wrapped_av_malloc (_stride[i] * (sample_size(i).height + 1) + ALIGNMENT); #if HAVE_VALGRIND_MEMCHECK_H /* The data between the end of the line size and the stride is undefined but processed by libswscale, causing lots of valgrind errors. Mark it all defined to quell these errors. */ - VALGRIND_MAKE_MEM_DEFINED (_data[i], _stride[i] * (sample_size(i).height + 1) + 32); + VALGRIND_MAKE_MEM_DEFINED (_data[i], _stride[i] * (sample_size(i).height + 1) + ALIGNMENT); #endif } } |
