[Rapp-dev] This is an initial SIMD implementation of integral image

[Willie Betschart]

Hi!

This is an initial SIMD implementation of integral image for 8 bits to 16 and
32 bit integrals.

The current solution is a mixture of hardcoded SWAR and SIMD. 

I'm will continue to experiment on a full SIMD implementation. At the moment 
this implementation is not faster than generic.

Note:
It was required to add the rows below in file /rapp/tune/rapptune.h 

#define rc_integral_sum_u8_u16_IMPL                          RC_IMPL_SIMD
#define rc_integral_sum_u8_u16_UNROLL                        1
#define rc_integral_sum_u8_u16_SCORE                         1.69e+09

#define rc_integral_sum_u8_u32_IMPL                          RC_IMPL_SIMD
#define rc_integral_sum_u8_u32_UNROLL                        1
#define rc_integral_sum_u8_u32_SCORE                         1.57e+09

The score is just copied but this was necessary to make it build when adding a 
vector implementation. This wasn't trivial. I had to build once first.

Best regards
Willie

simd_integral.patch
Description: simd_integral.patch