Function ScaleJ8

lib/PBA/SparseBundleCPU.cpp:177–180

  _mm_store_ps(jcy, _mm_mul_ps(_mm_load_ps(jcy), ps));
}
inline void ScaleJ8(float* jcx, float* jcy, const float* sj) {
  ScaleJ4(jcx, jcy, sj);
  ScaleJ4(jcx + 4, jcy + 4, sj + 4);
}
inline void ScaleJ4(double* jcx, double* jcy, const double* sj) {
  __m128d ps1 = _mm_load_pd(sj), ps2 = _mm_load_pd(sj + 2);
  _mm_store_pd(jcx, _mm_mul_pd(_mm_load_pd(jcx), ps1));