Back to home page

Project CMSSW displayed by LXR

 
 

    


Warning, /RecoTracker/MkFitCore/src/Matriplex/intr_sym_6x6.ah is written in an unsupported language. File is not indexed.

0001       IntrVec_t a_0 = LD(a, 0);
0002       IntrVec_t b_0 = LD(b, 0);
0003       IntrVec_t c_0 = MUL(a_0, b_0);
0004       IntrVec_t b_1 = LD(b, 1);
0005       IntrVec_t c_1 = MUL(a_0, b_1);
0006       IntrVec_t b_3 = LD(b, 3);
0007       IntrVec_t c_2 = MUL(a_0, b_3);
0008       IntrVec_t b_6 = LD(b, 6);
0009       IntrVec_t c_3 = MUL(a_0, b_6);
0010       IntrVec_t b_10 = LD(b, 10);
0011       IntrVec_t c_4 = MUL(a_0, b_10);
0012       IntrVec_t b_15 = LD(b, 15);
0013       IntrVec_t c_5 = MUL(a_0, b_15);
0014 
0015       IntrVec_t a_1 = LD(a, 1);
0016       c_0 = FMA(a_1, b_1, c_0);
0017       IntrVec_t b_2 = LD(b, 2);
0018       c_1 = FMA(a_1, b_2, c_1);
0019       IntrVec_t b_4 = LD(b, 4);
0020       c_2 = FMA(a_1, b_4, c_2);
0021       IntrVec_t b_7 = LD(b, 7);
0022       c_3 = FMA(a_1, b_7, c_3);
0023       IntrVec_t b_11 = LD(b, 11);
0024       c_4 = FMA(a_1, b_11, c_4);
0025       IntrVec_t b_16 = LD(b, 16);
0026       c_5 = FMA(a_1, b_16, c_5);
0027 
0028       IntrVec_t a_3 = LD(a, 3);
0029       c_0 = FMA(a_3, b_3, c_0);
0030       c_1 = FMA(a_3, b_4, c_1);
0031       IntrVec_t b_5 = LD(b, 5);
0032       c_2 = FMA(a_3, b_5, c_2);
0033       IntrVec_t b_8 = LD(b, 8);
0034       c_3 = FMA(a_3, b_8, c_3);
0035       IntrVec_t b_12 = LD(b, 12);
0036       c_4 = FMA(a_3, b_12, c_4);
0037       IntrVec_t b_17 = LD(b, 17);
0038       c_5 = FMA(a_3, b_17, c_5);
0039 
0040       IntrVec_t a_6 = LD(a, 6);
0041       c_0 = FMA(a_6, b_6, c_0);
0042       c_1 = FMA(a_6, b_7, c_1);
0043       c_2 = FMA(a_6, b_8, c_2);
0044       IntrVec_t b_9 = LD(b, 9);
0045       c_3 = FMA(a_6, b_9, c_3);
0046       IntrVec_t b_13 = LD(b, 13);
0047       c_4 = FMA(a_6, b_13, c_4);
0048       IntrVec_t b_18 = LD(b, 18);
0049       c_5 = FMA(a_6, b_18, c_5);
0050 
0051       IntrVec_t a_10 = LD(a, 10);
0052       c_0 = FMA(a_10, b_10, c_0);
0053       c_1 = FMA(a_10, b_11, c_1);
0054       c_2 = FMA(a_10, b_12, c_2);
0055       c_3 = FMA(a_10, b_13, c_3);
0056       IntrVec_t b_14 = LD(b, 14);
0057       c_4 = FMA(a_10, b_14, c_4);
0058       IntrVec_t b_19 = LD(b, 19);
0059       c_5 = FMA(a_10, b_19, c_5);
0060 
0061       IntrVec_t a_15 = LD(a, 15);
0062       c_0 = FMA(a_15, b_15, c_0);
0063       c_1 = FMA(a_15, b_16, c_1);
0064       c_2 = FMA(a_15, b_17, c_2);
0065       c_3 = FMA(a_15, b_18, c_3);
0066       c_4 = FMA(a_15, b_19, c_4);
0067       ST(c, 0, c_0);
0068       ST(c, 1, c_1);
0069       ST(c, 2, c_2);
0070       ST(c, 3, c_3);
0071       ST(c, 4, c_4);
0072       IntrVec_t b_20 = LD(b, 20);
0073       c_5 = FMA(a_15, b_20, c_5);
0074 
0075       IntrVec_t c_6 = MUL(a_1, b_0);
0076       IntrVec_t c_7 = MUL(a_1, b_1);
0077       IntrVec_t c_8 = MUL(a_1, b_3);
0078       IntrVec_t c_9 = MUL(a_1, b_6);
0079       ST(c, 5, c_5);
0080       IntrVec_t c_10 = MUL(a_1, b_10);
0081       IntrVec_t c_11 = MUL(a_1, b_15);
0082 
0083       IntrVec_t a_2 = LD(a, 2);
0084       c_6 = FMA(a_2, b_1, c_6);
0085       c_7 = FMA(a_2, b_2, c_7);
0086       c_8 = FMA(a_2, b_4, c_8);
0087       c_9 = FMA(a_2, b_7, c_9);
0088       c_10 = FMA(a_2, b_11, c_10);
0089       c_11 = FMA(a_2, b_16, c_11);
0090 
0091       IntrVec_t a_4 = LD(a, 4);
0092       c_6 = FMA(a_4, b_3, c_6);
0093       c_7 = FMA(a_4, b_4, c_7);
0094       c_8 = FMA(a_4, b_5, c_8);
0095       c_9 = FMA(a_4, b_8, c_9);
0096       c_10 = FMA(a_4, b_12, c_10);
0097       c_11 = FMA(a_4, b_17, c_11);
0098 
0099       IntrVec_t a_7 = LD(a, 7);
0100       c_6 = FMA(a_7, b_6, c_6);
0101       c_7 = FMA(a_7, b_7, c_7);
0102       c_8 = FMA(a_7, b_8, c_8);
0103       c_9 = FMA(a_7, b_9, c_9);
0104       c_10 = FMA(a_7, b_13, c_10);
0105       c_11 = FMA(a_7, b_18, c_11);
0106 
0107       IntrVec_t a_11 = LD(a, 11);
0108       c_6 = FMA(a_11, b_10, c_6);
0109       c_7 = FMA(a_11, b_11, c_7);
0110       c_8 = FMA(a_11, b_12, c_8);
0111       c_9 = FMA(a_11, b_13, c_9);
0112       c_10 = FMA(a_11, b_14, c_10);
0113       c_11 = FMA(a_11, b_19, c_11);
0114 
0115       IntrVec_t a_16 = LD(a, 16);
0116       c_6 = FMA(a_16, b_15, c_6);
0117       c_7 = FMA(a_16, b_16, c_7);
0118       c_8 = FMA(a_16, b_17, c_8);
0119       c_9 = FMA(a_16, b_18, c_9);
0120       c_10 = FMA(a_16, b_19, c_10);
0121       ST(c, 6, c_6);
0122       ST(c, 7, c_7);
0123       ST(c, 8, c_8);
0124       ST(c, 9, c_9);
0125       ST(c, 10, c_10);
0126       c_11 = FMA(a_16, b_20, c_11);
0127 
0128       IntrVec_t c_12 = MUL(a_3, b_0);
0129       IntrVec_t c_13 = MUL(a_3, b_1);
0130       IntrVec_t c_14 = MUL(a_3, b_3);
0131       IntrVec_t c_15 = MUL(a_3, b_6);
0132       ST(c, 11, c_11);
0133       IntrVec_t c_16 = MUL(a_3, b_10);
0134       IntrVec_t c_17 = MUL(a_3, b_15);
0135 
0136       c_12 = FMA(a_4, b_1, c_12);
0137       c_13 = FMA(a_4, b_2, c_13);
0138       c_14 = FMA(a_4, b_4, c_14);
0139       c_15 = FMA(a_4, b_7, c_15);
0140       c_16 = FMA(a_4, b_11, c_16);
0141       c_17 = FMA(a_4, b_16, c_17);
0142 
0143       IntrVec_t a_5 = LD(a, 5);
0144       c_12 = FMA(a_5, b_3, c_12);
0145       c_13 = FMA(a_5, b_4, c_13);
0146       c_14 = FMA(a_5, b_5, c_14);
0147       c_15 = FMA(a_5, b_8, c_15);
0148       c_16 = FMA(a_5, b_12, c_16);
0149       c_17 = FMA(a_5, b_17, c_17);
0150 
0151       IntrVec_t a_8 = LD(a, 8);
0152       c_12 = FMA(a_8, b_6, c_12);
0153       c_13 = FMA(a_8, b_7, c_13);
0154       c_14 = FMA(a_8, b_8, c_14);
0155       c_15 = FMA(a_8, b_9, c_15);
0156       c_16 = FMA(a_8, b_13, c_16);
0157       c_17 = FMA(a_8, b_18, c_17);
0158 
0159       IntrVec_t a_12 = LD(a, 12);
0160       c_12 = FMA(a_12, b_10, c_12);
0161       c_13 = FMA(a_12, b_11, c_13);
0162       c_14 = FMA(a_12, b_12, c_14);
0163       c_15 = FMA(a_12, b_13, c_15);
0164       c_16 = FMA(a_12, b_14, c_16);
0165       c_17 = FMA(a_12, b_19, c_17);
0166 
0167       IntrVec_t a_17 = LD(a, 17);
0168       c_12 = FMA(a_17, b_15, c_12);
0169       c_13 = FMA(a_17, b_16, c_13);
0170       c_14 = FMA(a_17, b_17, c_14);
0171       c_15 = FMA(a_17, b_18, c_15);
0172       c_16 = FMA(a_17, b_19, c_16);
0173       ST(c, 12, c_12);
0174       ST(c, 13, c_13);
0175       ST(c, 14, c_14);
0176       ST(c, 15, c_15);
0177       ST(c, 16, c_16);
0178       c_17 = FMA(a_17, b_20, c_17);
0179 
0180       IntrVec_t c_18 = MUL(a_6, b_0);
0181       IntrVec_t c_19 = MUL(a_6, b_1);
0182       IntrVec_t c_20 = MUL(a_6, b_3);
0183       IntrVec_t c_21 = MUL(a_6, b_6);
0184       ST(c, 17, c_17);
0185       IntrVec_t c_22 = MUL(a_6, b_10);
0186       IntrVec_t c_23 = MUL(a_6, b_15);
0187 
0188       c_18 = FMA(a_7, b_1, c_18);
0189       c_19 = FMA(a_7, b_2, c_19);
0190       c_20 = FMA(a_7, b_4, c_20);
0191       c_21 = FMA(a_7, b_7, c_21);
0192       c_22 = FMA(a_7, b_11, c_22);
0193       c_23 = FMA(a_7, b_16, c_23);
0194 
0195       c_18 = FMA(a_8, b_3, c_18);
0196       c_19 = FMA(a_8, b_4, c_19);
0197       c_20 = FMA(a_8, b_5, c_20);
0198       c_21 = FMA(a_8, b_8, c_21);
0199       c_22 = FMA(a_8, b_12, c_22);
0200       c_23 = FMA(a_8, b_17, c_23);
0201 
0202       IntrVec_t a_9 = LD(a, 9);
0203       c_18 = FMA(a_9, b_6, c_18);
0204       c_19 = FMA(a_9, b_7, c_19);
0205       c_20 = FMA(a_9, b_8, c_20);
0206       c_21 = FMA(a_9, b_9, c_21);
0207       c_22 = FMA(a_9, b_13, c_22);
0208       c_23 = FMA(a_9, b_18, c_23);
0209 
0210       IntrVec_t a_13 = LD(a, 13);
0211       c_18 = FMA(a_13, b_10, c_18);
0212       c_19 = FMA(a_13, b_11, c_19);
0213       c_20 = FMA(a_13, b_12, c_20);
0214       c_21 = FMA(a_13, b_13, c_21);
0215       c_22 = FMA(a_13, b_14, c_22);
0216       c_23 = FMA(a_13, b_19, c_23);
0217 
0218       IntrVec_t a_18 = LD(a, 18);
0219       c_18 = FMA(a_18, b_15, c_18);
0220       c_19 = FMA(a_18, b_16, c_19);
0221       c_20 = FMA(a_18, b_17, c_20);
0222       c_21 = FMA(a_18, b_18, c_21);
0223       c_22 = FMA(a_18, b_19, c_22);
0224       ST(c, 18, c_18);
0225       ST(c, 19, c_19);
0226       ST(c, 20, c_20);
0227       ST(c, 21, c_21);
0228       ST(c, 22, c_22);
0229       c_23 = FMA(a_18, b_20, c_23);
0230 
0231       IntrVec_t c_24 = MUL(a_10, b_0);
0232       IntrVec_t c_25 = MUL(a_10, b_1);
0233       IntrVec_t c_26 = MUL(a_10, b_3);
0234       IntrVec_t c_27 = MUL(a_10, b_6);
0235       ST(c, 23, c_23);
0236       IntrVec_t c_28 = MUL(a_10, b_10);
0237       IntrVec_t c_29 = MUL(a_10, b_15);
0238 
0239       c_24 = FMA(a_11, b_1, c_24);
0240       c_25 = FMA(a_11, b_2, c_25);
0241       c_26 = FMA(a_11, b_4, c_26);
0242       c_27 = FMA(a_11, b_7, c_27);
0243       c_28 = FMA(a_11, b_11, c_28);
0244       c_29 = FMA(a_11, b_16, c_29);
0245 
0246       c_24 = FMA(a_12, b_3, c_24);
0247       c_25 = FMA(a_12, b_4, c_25);
0248       c_26 = FMA(a_12, b_5, c_26);
0249       c_27 = FMA(a_12, b_8, c_27);
0250       c_28 = FMA(a_12, b_12, c_28);
0251       c_29 = FMA(a_12, b_17, c_29);
0252 
0253       c_24 = FMA(a_13, b_6, c_24);
0254       c_25 = FMA(a_13, b_7, c_25);
0255       c_26 = FMA(a_13, b_8, c_26);
0256       c_27 = FMA(a_13, b_9, c_27);
0257       c_28 = FMA(a_13, b_13, c_28);
0258       c_29 = FMA(a_13, b_18, c_29);
0259 
0260       IntrVec_t a_14 = LD(a, 14);
0261       c_24 = FMA(a_14, b_10, c_24);
0262       c_25 = FMA(a_14, b_11, c_25);
0263       c_26 = FMA(a_14, b_12, c_26);
0264       c_27 = FMA(a_14, b_13, c_27);
0265       c_28 = FMA(a_14, b_14, c_28);
0266       c_29 = FMA(a_14, b_19, c_29);
0267 
0268       IntrVec_t a_19 = LD(a, 19);
0269       c_24 = FMA(a_19, b_15, c_24);
0270       c_25 = FMA(a_19, b_16, c_25);
0271       c_26 = FMA(a_19, b_17, c_26);
0272       c_27 = FMA(a_19, b_18, c_27);
0273       c_28 = FMA(a_19, b_19, c_28);
0274       ST(c, 24, c_24);
0275       ST(c, 25, c_25);
0276       ST(c, 26, c_26);
0277       ST(c, 27, c_27);
0278       ST(c, 28, c_28);
0279       c_29 = FMA(a_19, b_20, c_29);
0280 
0281       IntrVec_t c_30 = MUL(a_15, b_0);
0282       IntrVec_t c_31 = MUL(a_15, b_1);
0283       IntrVec_t c_32 = MUL(a_15, b_3);
0284       IntrVec_t c_33 = MUL(a_15, b_6);
0285       ST(c, 29, c_29);
0286       IntrVec_t c_34 = MUL(a_15, b_10);
0287       IntrVec_t c_35 = MUL(a_15, b_15);
0288 
0289       c_30 = FMA(a_16, b_1, c_30);
0290       c_31 = FMA(a_16, b_2, c_31);
0291       c_32 = FMA(a_16, b_4, c_32);
0292       c_33 = FMA(a_16, b_7, c_33);
0293       c_34 = FMA(a_16, b_11, c_34);
0294       c_35 = FMA(a_16, b_16, c_35);
0295 
0296       c_30 = FMA(a_17, b_3, c_30);
0297       c_31 = FMA(a_17, b_4, c_31);
0298       c_32 = FMA(a_17, b_5, c_32);
0299       c_33 = FMA(a_17, b_8, c_33);
0300       c_34 = FMA(a_17, b_12, c_34);
0301       c_35 = FMA(a_17, b_17, c_35);
0302 
0303       c_30 = FMA(a_18, b_6, c_30);
0304       c_31 = FMA(a_18, b_7, c_31);
0305       c_32 = FMA(a_18, b_8, c_32);
0306       c_33 = FMA(a_18, b_9, c_33);
0307       c_34 = FMA(a_18, b_13, c_34);
0308       c_35 = FMA(a_18, b_18, c_35);
0309 
0310       c_30 = FMA(a_19, b_10, c_30);
0311       c_31 = FMA(a_19, b_11, c_31);
0312       c_32 = FMA(a_19, b_12, c_32);
0313       c_33 = FMA(a_19, b_13, c_33);
0314       c_34 = FMA(a_19, b_14, c_34);
0315       c_35 = FMA(a_19, b_19, c_35);
0316 
0317       IntrVec_t a_20 = LD(a, 20);
0318       c_30 = FMA(a_20, b_15, c_30);
0319       c_31 = FMA(a_20, b_16, c_31);
0320       c_32 = FMA(a_20, b_17, c_32);
0321       c_33 = FMA(a_20, b_18, c_33);
0322       c_34 = FMA(a_20, b_19, c_34);
0323       ST(c, 30, c_30);
0324       ST(c, 31, c_31);
0325       ST(c, 32, c_32);
0326       ST(c, 33, c_33);
0327       ST(c, 34, c_34);
0328       c_35 = FMA(a_20, b_20, c_35);
0329 
0330       ST(c, 35, c_35);