Back to home page

Project CMSSW displayed by LXR

 
 

    


Warning, /RecoTracker/MkFitCore/src/MultHelixPropTransp.ah is written in an unsupported language. File is not indexed.

0001 #ifdef MPLEX_INTRINSICS
0002 
0003 for (int n = 0; n < N; n += MPLEX_INTRINSICS_WIDTH_BYTES / sizeof(T)) {
0004   IntrVec_t b_0 = LD(b, 0);
0005   IntrVec_t a_0 = LD(a, 0);
0006   IntrVec_t c_0 = MUL(b_0, a_0);
0007 
0008   IntrVec_t b_1 = LD(b, 1);
0009   IntrVec_t a_1 = LD(a, 1);
0010   c_0 = FMA(b_1, a_1, c_0);
0011 
0012   IntrVec_t b_3 = LD(b, 3);
0013   IntrVec_t a_3 = LD(a, 3);
0014   c_0 = FMA(b_3, a_3, c_0);
0015 
0016   IntrVec_t b_4 = LD(b, 4);
0017   IntrVec_t a_4 = LD(a, 4);
0018   c_0 = FMA(b_4, a_4, c_0);
0019 
0020   IntrVec_t b_6 = LD(b, 6);
0021   IntrVec_t c_1 = MUL(b_6, a_0);
0022   IntrVec_t a_6 = LD(a, 6);
0023   IntrVec_t c_2 = MUL(b_6, a_6);
0024   ST(c, 0, c_0);
0025 
0026   IntrVec_t b_7 = LD(b, 7);
0027   c_1 = FMA(b_7, a_1, c_1);
0028   IntrVec_t a_7 = LD(a, 7);
0029   c_2 = FMA(b_7, a_7, c_2);
0030 
0031   IntrVec_t b_9 = LD(b, 9);
0032   c_1 = FMA(b_9, a_3, c_1);
0033   IntrVec_t a_9 = LD(a, 9);
0034   c_2 = FMA(b_9, a_9, c_2);
0035 
0036   IntrVec_t b_10 = LD(b, 10);
0037   c_1 = FMA(b_10, a_4, c_1);
0038   IntrVec_t a_10 = LD(a, 10);
0039   c_2 = FMA(b_10, a_10, c_2);
0040 
0041   IntrVec_t b_12 = LD(b, 12);
0042   IntrVec_t c_3 = MUL(b_12, a_0);
0043   IntrVec_t c_4 = MUL(b_12, a_6);
0044   IntrVec_t a_12 = LD(a, 12);
0045   IntrVec_t c_5 = MUL(b_12, a_12);
0046   ST(c, 1, c_1);
0047   ST(c, 2, c_2);
0048 
0049   IntrVec_t b_13 = LD(b, 13);
0050   c_3 = FMA(b_13, a_1, c_3);
0051   c_4 = FMA(b_13, a_7, c_4);
0052   IntrVec_t a_13 = LD(a, 13);
0053   c_5 = FMA(b_13, a_13, c_5);
0054 
0055   IntrVec_t b_14 = LD(b, 14);
0056   c_5 = ADD(b_14, c_5);
0057 
0058   IntrVec_t b_15 = LD(b, 15);
0059   c_3 = FMA(b_15, a_3, c_3);
0060   c_4 = FMA(b_15, a_9, c_4);
0061   IntrVec_t a_15 = LD(a, 15);
0062   c_5 = FMA(b_15, a_15, c_5);
0063 
0064   IntrVec_t b_16 = LD(b, 16);
0065   c_3 = FMA(b_16, a_4, c_3);
0066   c_4 = FMA(b_16, a_10, c_4);
0067   IntrVec_t a_16 = LD(a, 16);
0068   c_5 = FMA(b_16, a_16, c_5);
0069 
0070   IntrVec_t b_17 = LD(b, 17);
0071   IntrVec_t a_17 = LD(a, 17);
0072   c_5 = FMA(b_17, a_17, c_5);
0073 
0074   IntrVec_t b_18 = LD(b, 18);
0075   IntrVec_t c_6 = MUL(b_18, a_0);
0076   ST(c, 3, c_3);
0077   ST(c, 4, c_4);
0078   ST(c, 5, c_5);
0079   IntrVec_t c_7 = MUL(b_18, a_6);
0080   IntrVec_t c_8 = MUL(b_18, a_12);
0081   IntrVec_t a_18 = LD(a, 18);
0082   IntrVec_t c_9 = MUL(b_18, a_18);
0083 
0084   IntrVec_t b_19 = LD(b, 19);
0085   c_6 = FMA(b_19, a_1, c_6);
0086   c_7 = FMA(b_19, a_7, c_7);
0087   c_8 = FMA(b_19, a_13, c_8);
0088   IntrVec_t a_19 = LD(a, 19);
0089   c_9 = FMA(b_19, a_19, c_9);
0090 
0091   IntrVec_t b_20 = LD(b, 20);
0092   c_8 = ADD(b_20, c_8);
0093 
0094   IntrVec_t b_21 = LD(b, 21);
0095   c_6 = FMA(b_21, a_3, c_6);
0096   c_7 = FMA(b_21, a_9, c_7);
0097   c_8 = FMA(b_21, a_15, c_8);
0098   IntrVec_t a_21 = LD(a, 21);
0099   c_9 = FMA(b_21, a_21, c_9);
0100 
0101   IntrVec_t b_22 = LD(b, 22);
0102   c_6 = FMA(b_22, a_4, c_6);
0103   c_7 = FMA(b_22, a_10, c_7);
0104   c_8 = FMA(b_22, a_16, c_8);
0105   IntrVec_t a_22 = LD(a, 22);
0106   c_9 = FMA(b_22, a_22, c_9);
0107 
0108   IntrVec_t b_23 = LD(b, 23);
0109   c_8 = FMA(b_23, a_17, c_8);
0110 
0111   IntrVec_t b_24 = LD(b, 24);
0112   IntrVec_t c_10 = MUL(b_24, a_0);
0113   ST(c, 6, c_6);
0114   ST(c, 7, c_7);
0115   ST(c, 8, c_8);
0116   ST(c, 9, c_9);
0117   IntrVec_t c_11 = MUL(b_24, a_6);
0118   IntrVec_t c_12 = MUL(b_24, a_12);
0119   IntrVec_t c_13 = MUL(b_24, a_18);
0120   IntrVec_t a_24 = LD(a, 24);
0121   IntrVec_t c_14 = MUL(b_24, a_24);
0122 
0123   IntrVec_t b_25 = LD(b, 25);
0124   c_10 = FMA(b_25, a_1, c_10);
0125   c_11 = FMA(b_25, a_7, c_11);
0126   c_12 = FMA(b_25, a_13, c_12);
0127   c_13 = FMA(b_25, a_19, c_13);
0128   IntrVec_t a_25 = LD(a, 25);
0129   c_14 = FMA(b_25, a_25, c_14);
0130 
0131   IntrVec_t b_26 = LD(b, 26);
0132   c_12 = ADD(b_26, c_12);
0133 
0134   IntrVec_t b_27 = LD(b, 27);
0135   c_10 = FMA(b_27, a_3, c_10);
0136   c_11 = FMA(b_27, a_9, c_11);
0137   c_12 = FMA(b_27, a_15, c_12);
0138   c_13 = FMA(b_27, a_21, c_13);
0139   IntrVec_t a_27 = LD(a, 27);
0140   c_14 = FMA(b_27, a_27, c_14);
0141 
0142   IntrVec_t b_28 = LD(b, 28);
0143   c_10 = FMA(b_28, a_4, c_10);
0144   c_11 = FMA(b_28, a_10, c_11);
0145   c_12 = FMA(b_28, a_16, c_12);
0146   c_13 = FMA(b_28, a_22, c_13);
0147   IntrVec_t a_28 = LD(a, 28);
0148   c_14 = FMA(b_28, a_28, c_14);
0149 
0150   IntrVec_t b_29 = LD(b, 29);
0151   c_12 = FMA(b_29, a_17, c_12);
0152 
0153   IntrVec_t b_30 = LD(b, 30);
0154   IntrVec_t c_15 = MUL(b_30, a_0);
0155   ST(c, 10, c_10);
0156   ST(c, 11, c_11);
0157   ST(c, 12, c_12);
0158   ST(c, 13, c_13);
0159   ST(c, 14, c_14);
0160   IntrVec_t c_16 = MUL(b_30, a_6);
0161   IntrVec_t c_17 = MUL(b_30, a_12);
0162   IntrVec_t c_18 = MUL(b_30, a_18);
0163   IntrVec_t c_19 = MUL(b_30, a_24);
0164 
0165   IntrVec_t b_31 = LD(b, 31);
0166   c_15 = FMA(b_31, a_1, c_15);
0167   c_16 = FMA(b_31, a_7, c_16);
0168   c_17 = FMA(b_31, a_13, c_17);
0169   c_18 = FMA(b_31, a_19, c_18);
0170   c_19 = FMA(b_31, a_25, c_19);
0171 
0172   IntrVec_t b_32 = LD(b, 32);
0173   c_17 = ADD(b_32, c_17);
0174 
0175   IntrVec_t b_33 = LD(b, 33);
0176   c_15 = FMA(b_33, a_3, c_15);
0177   c_16 = FMA(b_33, a_9, c_16);
0178   c_17 = FMA(b_33, a_15, c_17);
0179   c_18 = FMA(b_33, a_21, c_18);
0180   c_19 = FMA(b_33, a_27, c_19);
0181 
0182   IntrVec_t b_34 = LD(b, 34);
0183   c_15 = FMA(b_34, a_4, c_15);
0184   c_16 = FMA(b_34, a_10, c_16);
0185   c_17 = FMA(b_34, a_16, c_17);
0186   c_18 = FMA(b_34, a_22, c_18);
0187   c_19 = FMA(b_34, a_28, c_19);
0188 
0189   IntrVec_t b_35 = LD(b, 35);
0190   c_17 = FMA(b_35, a_17, c_17);
0191   IntrVec_t c_20 = b_35;
0192   ST(c, 15, c_15);
0193   ST(c, 16, c_16);
0194   ST(c, 17, c_17);
0195   ST(c, 18, c_18);
0196   ST(c, 19, c_19);
0197   ST(c, 20, c_20);
0198 }
0199 
0200 #else
0201 
0202 #pragma omp simd
0203 for (int n = 0; n < N; ++n) {
0204   c[0 * N + n] = b[0 * N + n] * a[0 * N + n] + b[1 * N + n] * a[1 * N + n] + b[3 * N + n] * a[3 * N + n] +
0205                  b[4 * N + n] * a[4 * N + n];
0206   c[1 * N + n] = b[6 * N + n] * a[0 * N + n] + b[7 * N + n] * a[1 * N + n] + b[9 * N + n] * a[3 * N + n] +
0207                  b[10 * N + n] * a[4 * N + n];
0208   c[2 * N + n] = b[6 * N + n] * a[6 * N + n] + b[7 * N + n] * a[7 * N + n] + b[9 * N + n] * a[9 * N + n] +
0209                  b[10 * N + n] * a[10 * N + n];
0210   c[3 * N + n] = b[12 * N + n] * a[0 * N + n] + b[13 * N + n] * a[1 * N + n] + b[15 * N + n] * a[3 * N + n] +
0211                  b[16 * N + n] * a[4 * N + n];
0212   c[4 * N + n] = b[12 * N + n] * a[6 * N + n] + b[13 * N + n] * a[7 * N + n] + b[15 * N + n] * a[9 * N + n] +
0213                  b[16 * N + n] * a[10 * N + n];
0214   c[5 * N + n] = b[12 * N + n] * a[12 * N + n] + b[13 * N + n] * a[13 * N + n] + b[14 * N + n] +
0215                  b[15 * N + n] * a[15 * N + n] + b[16 * N + n] * a[16 * N + n] + b[17 * N + n] * a[17 * N + n];
0216   c[6 * N + n] = b[18 * N + n] * a[0 * N + n] + b[19 * N + n] * a[1 * N + n] + b[21 * N + n] * a[3 * N + n] +
0217                  b[22 * N + n] * a[4 * N + n];
0218   c[7 * N + n] = b[18 * N + n] * a[6 * N + n] + b[19 * N + n] * a[7 * N + n] + b[21 * N + n] * a[9 * N + n] +
0219                  b[22 * N + n] * a[10 * N + n];
0220   c[8 * N + n] = b[18 * N + n] * a[12 * N + n] + b[19 * N + n] * a[13 * N + n] + b[20 * N + n] +
0221                  b[21 * N + n] * a[15 * N + n] + b[22 * N + n] * a[16 * N + n] + b[23 * N + n] * a[17 * N + n];
0222   c[9 * N + n] = b[18 * N + n] * a[18 * N + n] + b[19 * N + n] * a[19 * N + n] + b[21 * N + n] * a[21 * N + n] +
0223                  b[22 * N + n] * a[22 * N + n];
0224   c[10 * N + n] = b[24 * N + n] * a[0 * N + n] + b[25 * N + n] * a[1 * N + n] + b[27 * N + n] * a[3 * N + n] +
0225                   b[28 * N + n] * a[4 * N + n];
0226   c[11 * N + n] = b[24 * N + n] * a[6 * N + n] + b[25 * N + n] * a[7 * N + n] + b[27 * N + n] * a[9 * N + n] +
0227                   b[28 * N + n] * a[10 * N + n];
0228   c[12 * N + n] = b[24 * N + n] * a[12 * N + n] + b[25 * N + n] * a[13 * N + n] + b[26 * N + n] +
0229                   b[27 * N + n] * a[15 * N + n] + b[28 * N + n] * a[16 * N + n] + b[29 * N + n] * a[17 * N + n];
0230   c[13 * N + n] = b[24 * N + n] * a[18 * N + n] + b[25 * N + n] * a[19 * N + n] + b[27 * N + n] * a[21 * N + n] +
0231                   b[28 * N + n] * a[22 * N + n];
0232   c[14 * N + n] = b[24 * N + n] * a[24 * N + n] + b[25 * N + n] * a[25 * N + n] + b[27 * N + n] * a[27 * N + n] +
0233                   b[28 * N + n] * a[28 * N + n];
0234   c[15 * N + n] = b[30 * N + n] * a[0 * N + n] + b[31 * N + n] * a[1 * N + n] + b[33 * N + n] * a[3 * N + n] +
0235                   b[34 * N + n] * a[4 * N + n];
0236   c[16 * N + n] = b[30 * N + n] * a[6 * N + n] + b[31 * N + n] * a[7 * N + n] + b[33 * N + n] * a[9 * N + n] +
0237                   b[34 * N + n] * a[10 * N + n];
0238   c[17 * N + n] = b[30 * N + n] * a[12 * N + n] + b[31 * N + n] * a[13 * N + n] + b[32 * N + n] +
0239                   b[33 * N + n] * a[15 * N + n] + b[34 * N + n] * a[16 * N + n] + b[35 * N + n] * a[17 * N + n];
0240   c[18 * N + n] = b[30 * N + n] * a[18 * N + n] + b[31 * N + n] * a[19 * N + n] + b[33 * N + n] * a[21 * N + n] +
0241                   b[34 * N + n] * a[22 * N + n];
0242   c[19 * N + n] = b[30 * N + n] * a[24 * N + n] + b[31 * N + n] * a[25 * N + n] + b[33 * N + n] * a[27 * N + n] +
0243                   b[34 * N + n] * a[28 * N + n];
0244   c[20 * N + n] = b[35 * N + n];
0245 }
0246 #endif