Original commit message from CVS:
* gst/deinterlace2/tvtime/greedy.c:
(deinterlace_greedy_packed422_scanline_mmx),
(deinterlace_greedy_packed422_scanline_mmxext):
Optimize MMX/MMXEXT implementations a bit by requiring two less
memory accesses and fix the workaround for the missing right shift
on bytes to unset the highest bit of every byte.
Original commit message from CVS:
* gst/deinterlace2/tvtime/greedy.c:
(deinterlace_greedy_packed422_scanline_mmxext):
Remove sfence instruction as it's not needed and actually is an SSE
instruction.
Original commit message from CVS:
* gst/deinterlace2/tvtime/greedy.c:
(deinterlace_greedy_packed422_scanline_mmx),
(deinterlace_greedy_packed422_scanline):
Add plain MMX implementation for the greedyl method.
Original commit message from CVS:
* gst/deinterlace2/tvtime/greedy.c:
(deinterlace_greedy_packed422_scanline_c),
(deinterlace_greedy_packed422_scanline_mmxext),
(deinterlace_greedy_packed422_scanline):
Fix the C implementation to produce correct results and optimize the
MMXEXT implementation.
Handle odd widths and don't read over array boundaries in the MMXEXT
implementation.
* gst/deinterlace2/tvtime/vfir.c: (deinterlace_line_c),
(deinterlace_line_mmx), (deinterlace_scanline_vfir):
Fix a small rounding bug in the MMX implementation, the MMX
implementation doesn't actually need MMXEXT instructions so don't mark
it as such.
Handle odd widths in both implementations.
Original commit message from CVS:
* gst/deinterlace2/tvtime/greedy.c:
(deinterlace_greedy_packed422_scanline_sse),
(deinterlace_greedy_packed422_scanline_c),
(deinterlace_greedy_packed422_scanline):
Implement a C version of the greedy low motion algorithm and mark the
assembly optimized version as SSE as it uses SSE instructions
additional to MMX instructions.