优化2D旋转

Question

Given the classic formula for rotating a point in 2D space:

cv::Point pt[NPOINTS];
cv::Point rotated[NPOINTS];
float angle = WHATEVER;
float cosine = cos(angle);
float sine = sin(angle);

for (int i = 0; i < NPOINTS; i++)
{
    rotated[i].x = pt[i].x * cosine - pt[i].y * sine;
    rotated[i].y = pt[i].x * sine   + pt[i].y * cosine;
}

Given NPOINTS is 32 and the arrays are aligned, how would one go about optimising the code for SSE or AVX? Searching around here and elsewhere didn't turn up anything useful, and I got lost about here:

__m128i onePoint = _mm_set_epi32(pt[i].x, pt[i].y, pt[i].x, pt[i].y);
__m128 onefPoint = _m128_cvtepi32_ps(onePoint);
__m128 sinCos = _mm_set_ps(cosine, -sine, sine, cosine);
__m128 rotated = _mm_mul_ps(onefPoint, sinCos);

But how to go from [y*cosine, -x*sine, x*sine, y*cosine] to [y*cosine + -x*sine, x*sine + y*cosine] ? Is this the best approach? Does it easily scale to __m512 ?

UPDATE : I did a little more research and I now have approximately:

__m128i onePoint = _mm_set_epi32(pt[i].x, pt[i].y, pt[i].x, pt[i].y);
__m128 onefPoint = _m128_cvtepi32_ps(onePoint);
__m128i twoPoint = _mm_set_epi32(pt[i+1].x, pt[i+1].y, pt[i+1].x, pt[i+1].y);
__m128 twofPoint = _m128_cvtepi32_ps(twoPoint);
__m128 sinCos = _mm_set_ps(cosine, -sine, sine, cosine);
__m128 rotated1 = _mm_mul_ps(onefPoint, sinCos);
__m128 rotated2 = _mm_mul_ps(twofPoint, sinCos);
__m128 added = _mm_hadd_ps(rotated1, rotated2);
__m128i intResult = _mm_cvtps_epi32(added);
int results[4];
_mm_storeu_si128((__m128i*)results, intResult);

This gives a 50% speed-up from 11% of the processor time to about 6%. Expanding to __m256 and doing four points at a time gives another speed-up. This looks quite awful code, but am I heading the right direction?

Answer 1

Use an array of struct of arrays (AoSoA) and process eight points at a time. In the code below point8 is struct of arrays containing eight points. The function rotate_point8 rotates eight points and has the same algebraic structure as the the function rotate_point which rotates a single point. The function rotate_all8 rotates 32 points using the AoSoA point8* .

The single point rotation code does 4 multiplications, one addition, and one subtraction.

If we look at the assembly for rotate_point8 we see that GCC unrolled the loop and does 4 SIMD multiplications, one SIMD addition, and one SIMD subtraction per unroll. That's the best you can do: eight for the price of one.

#include <x86intrin.h>
#include <stdio.h>
#include <math.h>

struct point8 {
  __m256 x;
  __m256 y;
};

struct point {
  float x;
  float y;
};

static point rotate_point(point p, float a, float b) {
  point r;
  r.x = p.x*a - p.y*b;
  r.y = p.x*b + p.y*a;
  return r;
}

static point8 rotate_point8(point8 p, float a, float b) {
  __m256 va = _mm256_set1_ps(a), vb = _mm256_set1_ps(b);
  point8 r;
  r.x = _mm256_sub_ps(_mm256_mul_ps(p.x,va), _mm256_mul_ps(p.y,vb));
  r.y = _mm256_add_ps(_mm256_mul_ps(p.x,vb), _mm256_mul_ps(p.y,va));
  return r;
}

void rotate_all(point* points, point* r, float angle) {
  float a = cos(angle), b = sin(angle);
  for(int i=0; i<32; i++) r[i] = rotate_point(points[i], a, b);
}

void rotate_all8(point8* points, point8* r8, float angle) {
  float a = cos(angle), b = sin(angle);
  for(int i=0; i<4; i++) r8[i] = rotate_point8(points[i], a, b);
}

int main(void) {
  float x[32], y[32];
  point p[32], r[32];
  point8 p8[4], r8[4];
  float angle = 3.14159f/4;

  for(int i=0; i<32; i++) y[i] = 1.0*i/31, x[i] = sqrt(1-y[i]*y[i]);
  for(int i=0; i<32; i++) p[i].x = x[i], p[i].y = y[i];
  for(int i=0; i<4; i++) p8[i].x = _mm256_load_ps(&x[8*i]), p8[i].y = _mm256_load_ps(&y[8*i]); 

  for(int i=0; i<32; i++) printf("%f %f\n", p[i].x, p[i].y); puts("");

  rotate_all(p, r, angle);
  for(int i=0; i<32; i++) printf("%f %f\n", r[i].x, r[i].y); puts("");

  rotate_all8(p8, r8, angle);
  for(int i=0; i<4; i++) {
    _mm256_storeu_ps(x, r8[i].x), _mm256_storeu_ps(y, r8[i].y);
    for(int j=0; j<8; j++) printf("%f %f\n", x[j], y[j]);
  }
}