Image thresholding using NEON instructions

I'm working on some image processing app for an iOS and thresholding is really a huge bottleneck. So I'm trying to optimize it using NEON. Here is the C version of function. Is there any way to rewrite this using NEON (unfortunately I have absolutely no experience in this) ?

static void thresh_8u( const Image& _src, Image& _dst, uchar thresh, uchar maxval, int type ) {
    int i, j;
    uchar tab[256];
    Size roi = _src.size();
    roi.width *= _src.channels();

    memset(&tab[0], 0, thresh);
    memset(&tab[thresh], maxval, 256-thresh);

    for( i = 0; i < roi.height; i++ ) {
        const uchar* src = (const uchar*)(_src.data + _src.step*i);
        uchar* dst = (uchar*)(_dst.data + _dst.step*i);
        j = 0;

        for(; j <= roi.width; ++j) {
            dst[j] = tab[src[j]];
        }
    }
}

Solution

It's actually pretty easy, if you can make sure your rows are always a multiple of 16 bytes wide, because the compiler (clang) has special types representing the NEON vector registers, and knows how to apply the normal C operators to them. Here's my little test function:

#ifdef __ARM_NEON__

#include <arm_neon.h>

void computeThreshold(void *input, void *output, int count, uint8_t threshold, uint8_t highValue) {
    uint8x16_t thresholdVector = vdupq_n_u8(threshold);
    uint8x16_t highValueVector = vdupq_n_u8(highValue);
    uint8x16_t *__restrict inputVector = (uint8x16_t *)input;
    uint8x16_t *__restrict outputVector = (uint8x16_t *)output;
    for ( ; count > 0; count -= 16, ++inputVector, ++outputVector) {
        *outputVector = (*inputVector > thresholdVector) & highValueVector;
    }
}

#endif

This operates on 16 bytes at a time. A uint8x16_t is a vector register containing 16 8-bit unsigned ints. The vdupq_n_u8 returns a vector uint8x16_t filled with 16 copies of its argument.

The > operator, applied to two uint8x16_t values, does 16 comparisons between pairs of 8-bit unsigned ints. Where the left input is greater than the right input, it returns 0xff (which is different from a normal C >, which just returns 0x01). Where the left input is less than or equal to the right input, it returns 0. (It compiles into the VCGT.U8 instruction.)

The & operator, applied to two uint8x16_t values, computes the boolean AND of 128 pairs of bits.

The loop compiles to this in a release build:

0x6e668:  vldmia r2, {d4, d5}
0x6e66c:  subs   r0, #16
0x6e66e:  vcgt.u8 q10, q10, q8
0x6e672:  adds   r2, #16
0x6e674:  cmp    r0, #0
0x6e676:  vand   q10, q10, q9
0x6e67a:  vstmia r1, {d4, d5}
0x6e67e:  add.w  r1, r1, #16
0x6e682:  bgt    0x6e668