Impala Documentation

Mat32 * Impala::Core::Matrix::SIMD_MatSquareAndSumAxis1 ( Mat * m ) [inline]

Definition at line 26 of file MatNorm2DistSSE.h.

References Impala::IntAlignDown(), MatE(), MatNrCol(), and MatNrRow().

00027 {
00028     int n = MatNrRow(m);
00029     Mat* res = MatCreate<Mat>(n, 1);
00030     for (int i=0 ; i<n ; i++)
00031     {
00032         __m128d tmp = _mm_setzero_pd();
00033         __m128d* base = (__m128d*) MatE(m, i, 0);
00034         for (int j=0 ; j<IntAlignDown(MatNrCol(m), 2) / 2 ; j++)
00035             tmp = _mm_add_pd(tmp, _mm_mul_pd(base[j], base[j]));
00036 
00037         __m128d shuffle = _mm_shuffle_pd(tmp, tmp, _MM_SHUFFLE2(0, 1));
00038         Real64 result = _mm_cvtsd_f64(_mm_add_pd(tmp, shuffle));
00039         for (int j=IntAlignDown(MatNrCol(m), 2); j<MatNrCol(m) ; j++)
00040             result += result + *MatE(m, i, j);
00041         *MatE(res, i, 0) = result;
00042     }
00043     return res;
00044 }

Here is the call graph for this function: