87     int fullRank = 1, diag,k;  in arm_mat_ldlt_f32()  local
97     for(int k=0;k < n; k+=8)  in arm_mat_ldlt_f32()  local
102       vecP = vidupq_u16((uint16_t)k, 1);  in arm_mat_ldlt_f32()
104       vstrhq_p(&pp[k], vecP, p0);  in arm_mat_ldlt_f32()
110     for(k=0;k < n; k++)  in arm_mat_ldlt_f32()
114         int j=k;  in arm_mat_ldlt_f32()
117         for(int r=k;r<n;r++)  in arm_mat_ldlt_f32()
126         if(j != k)  in arm_mat_ldlt_f32()
128           SWAP_ROWS_F32(pl,0,k,j);  in arm_mat_ldlt_f32()
129           SWAP_COLS_F32(pl,0,k,j);  in arm_mat_ldlt_f32()
133         pp[k] = j;  in arm_mat_ldlt_f32()
135         a = pA[k*n+k];  in arm_mat_ldlt_f32()
153         for(w=k+1; w<n; w+=4)  in arm_mat_ldlt_f32()
155           int cnt = n - k - 1;  in arm_mat_ldlt_f32()
164           vecW0 = vdupq_n_f32(pA[(w + 0)*n+k]);  in arm_mat_ldlt_f32()
165           vecW1 = vdupq_n_f32(pA[(w + 1)*n+k]);  in arm_mat_ldlt_f32()
166           vecW2 = vdupq_n_f32(pA[(w + 2)*n+k]);  in arm_mat_ldlt_f32()
167           vecW3 = vdupq_n_f32(pA[(w + 3)*n+k]);  in arm_mat_ldlt_f32()
169           for(int x=k+1;x<n;x += 4)  in arm_mat_ldlt_f32()
176              vecX = vldrwq_gather_shifted_offset_z_f32(&pA[x*n+k], (uint32x4_t)vecOffs, p0);  in arm_mat_ldlt_f32()
202           int cnt = n - k - 1;  in arm_mat_ldlt_f32()
209           vecW = vdupq_n_f32(pA[w*n+k]);  in arm_mat_ldlt_f32()
211           for(int x=k+1;x<n;x += 4)  in arm_mat_ldlt_f32()
219              vecX = vldrwq_gather_shifted_offset_z_f32(&pA[x*n+k], (uint32x4_t)vecOffs, p0);  in arm_mat_ldlt_f32()
230         for(int w=k+1;w<n;w++)  in arm_mat_ldlt_f32()
232                pA[w*n+k] = pA[w*n+k] * invA;  in arm_mat_ldlt_f32()
241     diag=k;  in arm_mat_ldlt_f32()
248         int cnt= n-k;  in arm_mat_ldlt_f32()
251         for(int col=k; col < n;col += 4)  in arm_mat_ldlt_f32()
346     int fullRank = 1, diag,k;  in arm_mat_ldlt_f32()  local
354     for(k=0;k < n; k++)  in arm_mat_ldlt_f32()
356       pp[k] = k;  in arm_mat_ldlt_f32()
360     for(k=0;k < n; k++)  in arm_mat_ldlt_f32()
364         int j=k;  in arm_mat_ldlt_f32()
369         for(r=k;r<n;r++)  in arm_mat_ldlt_f32()
378         if(j != k)  in arm_mat_ldlt_f32()
380           SWAP_ROWS_F32(pl,0,k,j);  in arm_mat_ldlt_f32()
381           SWAP_COLS_F32(pl,0,k,j);  in arm_mat_ldlt_f32()
385         pp[k] = j;  in arm_mat_ldlt_f32()
387         a = pA[k*n+k];  in arm_mat_ldlt_f32()
396         for(int w=k+1;w<n;w++)  in arm_mat_ldlt_f32()
399           for(x=k+1;x<n;x++)  in arm_mat_ldlt_f32()
401              pA[w*n+x] = pA[w*n+x] - pA[w*n+k] * pA[x*n+k] / a;  in arm_mat_ldlt_f32()
405         for(int w=k+1;w<n;w++)  in arm_mat_ldlt_f32()
407                pA[w*n+k] = pA[w*n+k] / a;  in arm_mat_ldlt_f32()
416     diag=k;  in arm_mat_ldlt_f32()
423         for(col=k; col < n;col++)  in arm_mat_ldlt_f32()