8

我试图使用 ARM Neon 指令移植一些 SSE2 代码（快速角检测器分数计算）。代码乍一看很简单，但由于某种原因结果不同。问题是，有时差异可能非常显着，有时相差 2 或 3 个值。如果有人能解释为什么会发生，那就太好了。

这是代码

原始 SSE2：

__m128i q0 = _mm_set1_epi16(-1000), q1 = _mm_set1_epi16(1000);
for( k = 0; k < 16; k += 8 )
{
    __m128i v0 = _mm_loadu_si128((__m128i*)(d+k+1));
    __m128i v1 = _mm_loadu_si128((__m128i*)(d+k+2));
    __m128i a = _mm_min_epi16(v0, v1);
    __m128i b = _mm_max_epi16(v0, v1);
    v0 = _mm_loadu_si128((__m128i*)(d+k+3));
    a = _mm_min_epi16(a, v0);
    b = _mm_max_epi16(b, v0);
    v0 = _mm_loadu_si128((__m128i*)(d+k+4));
    a = _mm_min_epi16(a, v0);
    b = _mm_max_epi16(b, v0);
    v0 = _mm_loadu_si128((__m128i*)(d+k+5));
    a = _mm_min_epi16(a, v0);
    b = _mm_max_epi16(b, v0);
    v0 = _mm_loadu_si128((__m128i*)(d+k+6));
    a = _mm_min_epi16(a, v0);
    b = _mm_max_epi16(b, v0);
    v0 = _mm_loadu_si128((__m128i*)(d+k+7));
    a = _mm_min_epi16(a, v0);
    b = _mm_max_epi16(b, v0);
    v0 = _mm_loadu_si128((__m128i*)(d+k+8));
    a = _mm_min_epi16(a, v0);
    b = _mm_max_epi16(b, v0);
    v0 = _mm_loadu_si128((__m128i*)(d+k));
    q0 = _mm_max_epi16(q0, _mm_min_epi16(a, v0));
    q1 = _mm_min_epi16(q1, _mm_max_epi16(b, v0));
    v0 = _mm_loadu_si128((__m128i*)(d+k+9));
    q0 = _mm_max_epi16(q0, _mm_min_epi16(a, v0));
    q1 = _mm_min_epi16(q1, _mm_max_epi16(b, v0));
}
q0 = _mm_max_epi16(q0, _mm_sub_epi16(_mm_setzero_si128(), q1));
q0 = _mm_max_epi16(q0, _mm_unpackhi_epi64(q0, q0));
q0 = _mm_max_epi16(q0, _mm_srli_si128(q0, 4));
q0 = _mm_max_epi16(q0, _mm_srli_si128(q0, 2));
threshold = (short)_mm_cvtsi128_si32(q0) - 1;

这是工作臂霓虹灯代码：

int16x8_t q0 = vdupq_n_s16(-1000), q1 = vdupq_n_s16(1000);
int16x8_t zero = vdupq_n_s16(0);
for( k = 0; k < 16; k += 8 )
{
    int16x8_t v0 = vld1q_s16((const int16_t*)(d+k+1));
    int16x8_t v1 = vld1q_s16((const int16_t*)(d+k+2));
    int16x8_t a = vminq_s16(v0, v1);
    int16x8_t b = vmaxq_s16(v0, v1);
    v0 = vld1q_s16((const int16_t*)(d+k+3));
    a = vminq_s16(a, v0);
    b = vmaxq_s16(b, v0);
    v0 = vld1q_s16((const int16_t*)(d+k+4));
    a = vminq_s16(a, v0);
    b = vmaxq_s16(b, v0);
    v0 = vld1q_s16((const int16_t*)(d+k+5));
    a = vminq_s16(a, v0);
    b = vmaxq_s16(b, v0);
    v0 = vld1q_s16((const int16_t*)(d+k+6));
    a = vminq_s16(a, v0);
    b = vmaxq_s16(b, v0);
    v0 = vld1q_s16((const int16_t*)(d+k+7));
    a = vminq_s16(a, v0);
    b = vmaxq_s16(b, v0);
    v0 = vld1q_s16((const int16_t*)(d+k+8));
    a = vminq_s16(a, v0);
    b = vmaxq_s16(b, v0);
    v0 = vld1q_s16((const int16_t*)(d+k));
    q0 = vmaxq_s16(q0, vminq_s16(a, v0));
    q1 = vminq_s16(q1, vmaxq_s16(b, v0));
    v0 = vld1q_s16((const int16_t*)(d+k+9));
    q0 = vmaxq_s16(q0, vminq_s16(a, v0));
    q1 = vminq_s16(q1, vmaxq_s16(b, v0));
}
q0 = vmaxq_s16(q0, vsubq_s16(zero, q1));
// first mistake it produce wrong result
//q0 = vmaxq_s16(q0, vzipq_s16(q0, q0).val[1]);
// may be someone knows faster/better way?
int16x4_t a_hi = vget_high_s16(q0);
q1 = vcombine_s16(a_hi, a_hi);
q0 = vmaxq_s16(q0, q1);

// this is _mm_srli_si128(q0, 4)
q1 = vextq_s16(q0, zero, 2);

q0 = vmaxq_s16(q0, q1);

// this is _mm_srli_si128(q0, 2)
q1 = vextq_s16(q0, zero, 1);
q0 = vmaxq_s16(q0, q1);

// read the result
int16_t __attribute__ ((aligned (16))) x[8];
vst1q_s16(x, q0);

threshold = x[0] - 1;

这段代码似乎是正确的，所以如果有人觉得它有用的话……快速角检测器是目前最快的。通过如此小的优化，它可能在手机上运行得更快。

4

0 回答 0