Back to home page

Project CMSSW displayed by LXR

 
 

    


Warning, /RecoTracker/MkFitCore/src/MultHelixProp.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 a_0 = LD(a, 0);
0005   IntrVec_t b_0 = LD(b, 0);
0006   IntrVec_t c_0 = MUL(a_0, b_0);
0007   IntrVec_t b_1 = LD(b, 1);
0008   IntrVec_t c_1 = MUL(a_0, b_1);
0009   IntrVec_t b_3 = LD(b, 3);
0010   IntrVec_t c_2 = MUL(a_0, b_3);
0011   IntrVec_t b_6 = LD(b, 6);
0012   IntrVec_t c_3 = MUL(a_0, b_6);
0013   IntrVec_t b_10 = LD(b, 10);
0014   IntrVec_t c_4 = MUL(a_0, b_10);
0015   IntrVec_t b_15 = LD(b, 15);
0016   IntrVec_t c_5 = MUL(a_0, b_15);
0017 
0018   IntrVec_t a_1 = LD(a, 1);
0019   c_0 = FMA(a_1, b_1, c_0);
0020   IntrVec_t b_2 = LD(b, 2);
0021   c_1 = FMA(a_1, b_2, c_1);
0022   IntrVec_t b_4 = LD(b, 4);
0023   c_2 = FMA(a_1, b_4, c_2);
0024   IntrVec_t b_7 = LD(b, 7);
0025   c_3 = FMA(a_1, b_7, c_3);
0026   IntrVec_t b_11 = LD(b, 11);
0027   c_4 = FMA(a_1, b_11, c_4);
0028   IntrVec_t b_16 = LD(b, 16);
0029   c_5 = FMA(a_1, b_16, c_5);
0030 
0031   IntrVec_t a_3 = LD(a, 3);
0032   c_0 = FMA(a_3, b_6, c_0);
0033   c_1 = FMA(a_3, b_7, c_1);
0034   IntrVec_t b_8 = LD(b, 8);
0035   c_2 = FMA(a_3, b_8, c_2);
0036   IntrVec_t b_9 = LD(b, 9);
0037   c_3 = FMA(a_3, b_9, c_3);
0038   IntrVec_t b_13 = LD(b, 13);
0039   c_4 = FMA(a_3, b_13, c_4);
0040   IntrVec_t b_18 = LD(b, 18);
0041   c_5 = FMA(a_3, b_18, c_5);
0042 
0043   IntrVec_t a_4 = LD(a, 4);
0044   c_0 = FMA(a_4, b_10, c_0);
0045   c_1 = FMA(a_4, b_11, c_1);
0046   IntrVec_t b_12 = LD(b, 12);
0047   c_2 = FMA(a_4, b_12, c_2);
0048   c_3 = FMA(a_4, b_13, c_3);
0049   IntrVec_t b_14 = LD(b, 14);
0050   c_4 = FMA(a_4, b_14, c_4);
0051   IntrVec_t b_19 = LD(b, 19);
0052   c_5 = FMA(a_4, b_19, c_5);
0053 
0054   IntrVec_t a_6 = LD(a, 6);
0055   IntrVec_t c_6 = MUL(a_6, b_0);
0056   IntrVec_t c_7 = MUL(a_6, b_1);
0057   IntrVec_t c_8 = MUL(a_6, b_3);
0058   ST(c, 0, c_0);
0059   ST(c, 1, c_1);
0060   ST(c, 2, c_2);
0061   ST(c, 3, c_3);
0062   ST(c, 4, c_4);
0063   ST(c, 5, c_5);
0064   IntrVec_t c_9 = MUL(a_6, b_6);
0065   IntrVec_t c_10 = MUL(a_6, b_10);
0066   IntrVec_t c_11 = MUL(a_6, b_15);
0067 
0068   IntrVec_t a_7 = LD(a, 7);
0069   c_6 = FMA(a_7, b_1, c_6);
0070   c_7 = FMA(a_7, b_2, c_7);
0071   c_8 = FMA(a_7, b_4, c_8);
0072   c_9 = FMA(a_7, b_7, c_9);
0073   c_10 = FMA(a_7, b_11, c_10);
0074   c_11 = FMA(a_7, b_16, c_11);
0075 
0076   IntrVec_t a_9 = LD(a, 9);
0077   c_6 = FMA(a_9, b_6, c_6);
0078   c_7 = FMA(a_9, b_7, c_7);
0079   c_8 = FMA(a_9, b_8, c_8);
0080   c_9 = FMA(a_9, b_9, c_9);
0081   c_10 = FMA(a_9, b_13, c_10);
0082   c_11 = FMA(a_9, b_18, c_11);
0083 
0084   IntrVec_t a_10 = LD(a, 10);
0085   c_6 = FMA(a_10, b_10, c_6);
0086   c_7 = FMA(a_10, b_11, c_7);
0087   c_8 = FMA(a_10, b_12, c_8);
0088   c_9 = FMA(a_10, b_13, c_9);
0089   c_10 = FMA(a_10, b_14, c_10);
0090   c_11 = FMA(a_10, b_19, c_11);
0091 
0092   IntrVec_t a_12 = LD(a, 12);
0093   IntrVec_t c_12 = MUL(a_12, b_0);
0094   IntrVec_t c_13 = MUL(a_12, b_1);
0095   IntrVec_t c_14 = MUL(a_12, b_3);
0096   ST(c, 6, c_6);
0097   ST(c, 7, c_7);
0098   ST(c, 8, c_8);
0099   ST(c, 9, c_9);
0100   ST(c, 10, c_10);
0101   ST(c, 11, c_11);
0102   IntrVec_t c_15 = MUL(a_12, b_6);
0103   IntrVec_t c_16 = MUL(a_12, b_10);
0104   IntrVec_t c_17 = MUL(a_12, b_15);
0105 
0106   IntrVec_t a_13 = LD(a, 13);
0107   c_12 = FMA(a_13, b_1, c_12);
0108   c_13 = FMA(a_13, b_2, c_13);
0109   c_14 = FMA(a_13, b_4, c_14);
0110   c_15 = FMA(a_13, b_7, c_15);
0111   c_16 = FMA(a_13, b_11, c_16);
0112   c_17 = FMA(a_13, b_16, c_17);
0113 
0114   c_12 = ADD(b_3, c_12);
0115   c_13 = ADD(b_4, c_13);
0116   IntrVec_t b_5 = LD(b, 5);
0117   c_14 = ADD(b_5, c_14);
0118   c_15 = ADD(b_8, c_15);
0119   c_16 = ADD(b_12, c_16);
0120   IntrVec_t b_17 = LD(b, 17);
0121   c_17 = ADD(b_17, c_17);
0122 
0123   IntrVec_t a_15 = LD(a, 15);
0124   c_12 = FMA(a_15, b_6, c_12);
0125   c_13 = FMA(a_15, b_7, c_13);
0126   c_14 = FMA(a_15, b_8, c_14);
0127   c_15 = FMA(a_15, b_9, c_15);
0128   c_16 = FMA(a_15, b_13, c_16);
0129   c_17 = FMA(a_15, b_18, c_17);
0130 
0131   IntrVec_t a_16 = LD(a, 16);
0132   c_12 = FMA(a_16, b_10, c_12);
0133   c_13 = FMA(a_16, b_11, c_13);
0134   c_14 = FMA(a_16, b_12, c_14);
0135   c_15 = FMA(a_16, b_13, c_15);
0136   c_16 = FMA(a_16, b_14, c_16);
0137   c_17 = FMA(a_16, b_19, c_17);
0138 
0139   IntrVec_t a_17 = LD(a, 17);
0140   c_12 = FMA(a_17, b_15, c_12);
0141   c_13 = FMA(a_17, b_16, c_13);
0142   c_14 = FMA(a_17, b_17, c_14);
0143   c_15 = FMA(a_17, b_18, c_15);
0144   c_16 = FMA(a_17, b_19, c_16);
0145   ST(c, 12, c_12);
0146   ST(c, 13, c_13);
0147   ST(c, 14, c_14);
0148   ST(c, 15, c_15);
0149   ST(c, 16, c_16);
0150   IntrVec_t b_20 = LD(b, 20);
0151   c_17 = FMA(a_17, b_20, c_17);
0152 
0153   IntrVec_t a_18 = LD(a, 18);
0154   IntrVec_t c_18 = MUL(a_18, b_0);
0155   IntrVec_t c_19 = MUL(a_18, b_1);
0156   IntrVec_t c_20 = MUL(a_18, b_3);
0157   ST(c, 17, c_17);
0158   IntrVec_t c_21 = MUL(a_18, b_6);
0159   IntrVec_t c_22 = MUL(a_18, b_10);
0160   IntrVec_t c_23 = MUL(a_18, b_15);
0161 
0162   IntrVec_t a_19 = LD(a, 19);
0163   c_18 = FMA(a_19, b_1, c_18);
0164   c_19 = FMA(a_19, b_2, c_19);
0165   c_20 = FMA(a_19, b_4, c_20);
0166   c_21 = FMA(a_19, b_7, c_21);
0167   c_22 = FMA(a_19, b_11, c_22);
0168   c_23 = FMA(a_19, b_16, c_23);
0169 
0170   IntrVec_t a_21 = LD(a, 21);
0171   c_18 = FMA(a_21, b_6, c_18);
0172   c_19 = FMA(a_21, b_7, c_19);
0173   c_20 = FMA(a_21, b_8, c_20);
0174   c_21 = FMA(a_21, b_9, c_21);
0175   c_22 = FMA(a_21, b_13, c_22);
0176   c_23 = FMA(a_21, b_18, c_23);
0177 
0178   IntrVec_t a_22 = LD(a, 22);
0179   c_18 = FMA(a_22, b_10, c_18);
0180   c_19 = FMA(a_22, b_11, c_19);
0181   c_20 = FMA(a_22, b_12, c_20);
0182   c_21 = FMA(a_22, b_13, c_21);
0183   c_22 = FMA(a_22, b_14, c_22);
0184   c_23 = FMA(a_22, b_19, c_23);
0185 
0186   IntrVec_t a_24 = LD(a, 24);
0187   IntrVec_t c_24 = MUL(a_24, b_0);
0188   IntrVec_t c_25 = MUL(a_24, b_1);
0189   IntrVec_t c_26 = MUL(a_24, b_3);
0190   ST(c, 18, c_18);
0191   ST(c, 19, c_19);
0192   ST(c, 20, c_20);
0193   ST(c, 21, c_21);
0194   ST(c, 22, c_22);
0195   ST(c, 23, c_23);
0196   IntrVec_t c_27 = MUL(a_24, b_6);
0197   IntrVec_t c_28 = MUL(a_24, b_10);
0198   IntrVec_t c_29 = MUL(a_24, b_15);
0199 
0200   IntrVec_t a_25 = LD(a, 25);
0201   c_24 = FMA(a_25, b_1, c_24);
0202   c_25 = FMA(a_25, b_2, c_25);
0203   c_26 = FMA(a_25, b_4, c_26);
0204   c_27 = FMA(a_25, b_7, c_27);
0205   c_28 = FMA(a_25, b_11, c_28);
0206   c_29 = FMA(a_25, b_16, c_29);
0207 
0208   IntrVec_t a_27 = LD(a, 27);
0209   c_24 = FMA(a_27, b_6, c_24);
0210   c_25 = FMA(a_27, b_7, c_25);
0211   c_26 = FMA(a_27, b_8, c_26);
0212   c_27 = FMA(a_27, b_9, c_27);
0213   c_28 = FMA(a_27, b_13, c_28);
0214   c_29 = FMA(a_27, b_18, c_29);
0215 
0216   IntrVec_t a_28 = LD(a, 28);
0217   c_24 = FMA(a_28, b_10, c_24);
0218   c_25 = FMA(a_28, b_11, c_25);
0219   c_26 = FMA(a_28, b_12, c_26);
0220   c_27 = FMA(a_28, b_13, c_27);
0221   c_28 = FMA(a_28, b_14, c_28);
0222   c_29 = FMA(a_28, b_19, c_29);
0223 
0224   IntrVec_t c_30 = b_15;
0225   IntrVec_t c_31 = b_16;
0226   IntrVec_t c_32 = b_17;
0227   IntrVec_t c_33 = b_18;
0228   ST(c, 24, c_24);
0229   ST(c, 25, c_25);
0230   ST(c, 26, c_26);
0231   ST(c, 27, c_27);
0232   ST(c, 28, c_28);
0233   ST(c, 29, c_29);
0234   ST(c, 30, c_30);
0235   ST(c, 31, c_31);
0236   ST(c, 32, c_32);
0237   ST(c, 33, c_33);
0238   IntrVec_t c_34 = b_19;
0239   IntrVec_t c_35 = b_20;
0240   ST(c, 34, c_34);
0241   ST(c, 35, c_35);
0242 }
0243 
0244 #else
0245 
0246 #pragma omp simd
0247 for (int n = 0; n < N; ++n) {
0248   c[0 * N + n] = a[0 * N + n] * b[0 * N + n] + a[1 * N + n] * b[1 * N + n] + a[3 * N + n] * b[6 * N + n] +
0249                  a[4 * N + n] * b[10 * N + n];
0250   c[1 * N + n] = a[0 * N + n] * b[1 * N + n] + a[1 * N + n] * b[2 * N + n] + a[3 * N + n] * b[7 * N + n] +
0251                  a[4 * N + n] * b[11 * N + n];
0252   c[2 * N + n] = a[0 * N + n] * b[3 * N + n] + a[1 * N + n] * b[4 * N + n] + a[3 * N + n] * b[8 * N + n] +
0253                  a[4 * N + n] * b[12 * N + n];
0254   c[3 * N + n] = a[0 * N + n] * b[6 * N + n] + a[1 * N + n] * b[7 * N + n] + a[3 * N + n] * b[9 * N + n] +
0255                  a[4 * N + n] * b[13 * N + n];
0256   c[4 * N + n] = a[0 * N + n] * b[10 * N + n] + a[1 * N + n] * b[11 * N + n] + a[3 * N + n] * b[13 * N + n] +
0257                  a[4 * N + n] * b[14 * N + n];
0258   c[5 * N + n] = a[0 * N + n] * b[15 * N + n] + a[1 * N + n] * b[16 * N + n] + a[3 * N + n] * b[18 * N + n] +
0259                  a[4 * N + n] * b[19 * N + n];
0260   c[6 * N + n] = a[6 * N + n] * b[0 * N + n] + a[7 * N + n] * b[1 * N + n] + a[9 * N + n] * b[6 * N + n] +
0261                  a[10 * N + n] * b[10 * N + n];
0262   c[7 * N + n] = a[6 * N + n] * b[1 * N + n] + a[7 * N + n] * b[2 * N + n] + a[9 * N + n] * b[7 * N + n] +
0263                  a[10 * N + n] * b[11 * N + n];
0264   c[8 * N + n] = a[6 * N + n] * b[3 * N + n] + a[7 * N + n] * b[4 * N + n] + a[9 * N + n] * b[8 * N + n] +
0265                  a[10 * N + n] * b[12 * N + n];
0266   c[9 * N + n] = a[6 * N + n] * b[6 * N + n] + a[7 * N + n] * b[7 * N + n] + a[9 * N + n] * b[9 * N + n] +
0267                  a[10 * N + n] * b[13 * N + n];
0268   c[10 * N + n] = a[6 * N + n] * b[10 * N + n] + a[7 * N + n] * b[11 * N + n] + a[9 * N + n] * b[13 * N + n] +
0269                   a[10 * N + n] * b[14 * N + n];
0270   c[11 * N + n] = a[6 * N + n] * b[15 * N + n] + a[7 * N + n] * b[16 * N + n] + a[9 * N + n] * b[18 * N + n] +
0271                   a[10 * N + n] * b[19 * N + n];
0272   c[12 * N + n] = a[12 * N + n] * b[0 * N + n] + a[13 * N + n] * b[1 * N + n] + b[3 * N + n] +
0273                   a[15 * N + n] * b[6 * N + n] + a[16 * N + n] * b[10 * N + n] + a[17 * N + n] * b[15 * N + n];
0274   c[13 * N + n] = a[12 * N + n] * b[1 * N + n] + a[13 * N + n] * b[2 * N + n] + b[4 * N + n] +
0275                   a[15 * N + n] * b[7 * N + n] + a[16 * N + n] * b[11 * N + n] + a[17 * N + n] * b[16 * N + n];
0276   c[14 * N + n] = a[12 * N + n] * b[3 * N + n] + a[13 * N + n] * b[4 * N + n] + b[5 * N + n] +
0277                   a[15 * N + n] * b[8 * N + n] + a[16 * N + n] * b[12 * N + n] + a[17 * N + n] * b[17 * N + n];
0278   c[15 * N + n] = a[12 * N + n] * b[6 * N + n] + a[13 * N + n] * b[7 * N + n] + b[8 * N + n] +
0279                   a[15 * N + n] * b[9 * N + n] + a[16 * N + n] * b[13 * N + n] + a[17 * N + n] * b[18 * N + n];
0280   c[16 * N + n] = a[12 * N + n] * b[10 * N + n] + a[13 * N + n] * b[11 * N + n] + b[12 * N + n] +
0281                   a[15 * N + n] * b[13 * N + n] + a[16 * N + n] * b[14 * N + n] + a[17 * N + n] * b[19 * N + n];
0282   c[17 * N + n] = a[12 * N + n] * b[15 * N + n] + a[13 * N + n] * b[16 * N + n] + b[17 * N + n] +
0283                   a[15 * N + n] * b[18 * N + n] + a[16 * N + n] * b[19 * N + n] + a[17 * N + n] * b[20 * N + n];
0284   c[18 * N + n] = a[18 * N + n] * b[0 * N + n] + a[19 * N + n] * b[1 * N + n] + a[21 * N + n] * b[6 * N + n] +
0285                   a[22 * N + n] * b[10 * N + n];
0286   c[19 * N + n] = a[18 * N + n] * b[1 * N + n] + a[19 * N + n] * b[2 * N + n] + a[21 * N + n] * b[7 * N + n] +
0287                   a[22 * N + n] * b[11 * N + n];
0288   c[20 * N + n] = a[18 * N + n] * b[3 * N + n] + a[19 * N + n] * b[4 * N + n] + a[21 * N + n] * b[8 * N + n] +
0289                   a[22 * N + n] * b[12 * N + n];
0290   c[21 * N + n] = a[18 * N + n] * b[6 * N + n] + a[19 * N + n] * b[7 * N + n] + a[21 * N + n] * b[9 * N + n] +
0291                   a[22 * N + n] * b[13 * N + n];
0292   c[22 * N + n] = a[18 * N + n] * b[10 * N + n] + a[19 * N + n] * b[11 * N + n] + a[21 * N + n] * b[13 * N + n] +
0293                   a[22 * N + n] * b[14 * N + n];
0294   c[23 * N + n] = a[18 * N + n] * b[15 * N + n] + a[19 * N + n] * b[16 * N + n] + a[21 * N + n] * b[18 * N + n] +
0295                   a[22 * N + n] * b[19 * N + n];
0296   c[24 * N + n] = a[24 * N + n] * b[0 * N + n] + a[25 * N + n] * b[1 * N + n] + a[27 * N + n] * b[6 * N + n] +
0297                   a[28 * N + n] * b[10 * N + n];
0298   c[25 * N + n] = a[24 * N + n] * b[1 * N + n] + a[25 * N + n] * b[2 * N + n] + a[27 * N + n] * b[7 * N + n] +
0299                   a[28 * N + n] * b[11 * N + n];
0300   c[26 * N + n] = a[24 * N + n] * b[3 * N + n] + a[25 * N + n] * b[4 * N + n] + a[27 * N + n] * b[8 * N + n] +
0301                   a[28 * N + n] * b[12 * N + n];
0302   c[27 * N + n] = a[24 * N + n] * b[6 * N + n] + a[25 * N + n] * b[7 * N + n] + a[27 * N + n] * b[9 * N + n] +
0303                   a[28 * N + n] * b[13 * N + n];
0304   c[28 * N + n] = a[24 * N + n] * b[10 * N + n] + a[25 * N + n] * b[11 * N + n] + a[27 * N + n] * b[13 * N + n] +
0305                   a[28 * N + n] * b[14 * N + n];
0306   c[29 * N + n] = a[24 * N + n] * b[15 * N + n] + a[25 * N + n] * b[16 * N + n] + a[27 * N + n] * b[18 * N + n] +
0307                   a[28 * N + n] * b[19 * N + n];
0308   c[30 * N + n] = b[15 * N + n];
0309   c[31 * N + n] = b[16 * N + n];
0310   c[32 * N + n] = b[17 * N + n];
0311   c[33 * N + n] = b[18 * N + n];
0312   c[34 * N + n] = b[19 * N + n];
0313   c[35 * N + n] = b[20 * N + n];
0314 }
0315 #endif