Warning, /RecoTracker/MkFitCore/src/CartesianErrTransp.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 {
0005 IntrVec_t b_0 = LD(b, 0);
0006 IntrVec_t c_0 = b_0;
0007
0008
0009
0010
0011
0012
0013 IntrVec_t b_6 = LD(b, 6);
0014 IntrVec_t c_1 = b_6;
0015
0016 IntrVec_t b_7 = LD(b, 7);
0017 IntrVec_t c_2 = b_7;
0018 ST(c, 0, c_0);
0019
0020
0021
0022
0023
0024 IntrVec_t b_12 = LD(b, 12);
0025 IntrVec_t c_3 = b_12;
0026
0027 IntrVec_t b_13 = LD(b, 13);
0028 IntrVec_t c_4 = b_13;
0029 ST(c, 1, c_1);
0030 ST(c, 2, c_2);
0031
0032 IntrVec_t b_14 = LD(b, 14);
0033 IntrVec_t c_5 = b_14;
0034
0035
0036
0037
0038 IntrVec_t b_18 = LD(b, 18);
0039 IntrVec_t c_6 = b_18;
0040
0041 IntrVec_t b_19 = LD(b, 19);
0042 IntrVec_t c_7 = b_19;
0043 ST(c, 3, c_3);
0044 ST(c, 4, c_4);
0045 ST(c, 5, c_5);
0046
0047 IntrVec_t b_20 = LD(b, 20);
0048 IntrVec_t c_8 = b_20;
0049
0050 IntrVec_t b_21 = LD(b, 21);
0051 IntrVec_t a_21 = LD(a, 21);
0052 IntrVec_t c_9 = MUL(b_21, a_21);
0053
0054 IntrVec_t b_22 = LD(b, 22);
0055 IntrVec_t a_22 = LD(a, 22);
0056 c_9 = FMA(b_22, a_22, c_9);
0057
0058
0059 IntrVec_t b_24 = LD(b, 24);
0060 IntrVec_t c_10 = b_24;
0061
0062 IntrVec_t b_25 = LD(b, 25);
0063 IntrVec_t c_11 = b_25;
0064 ST(c, 6, c_6);
0065 ST(c, 7, c_7);
0066 ST(c, 8, c_8);
0067 ST(c, 9, c_9);
0068
0069 IntrVec_t b_26 = LD(b, 26);
0070 IntrVec_t c_12 = b_26;
0071
0072 IntrVec_t b_27 = LD(b, 27);
0073 IntrVec_t c_13 = MUL(b_27, a_21);
0074 IntrVec_t a_27 = LD(a, 27);
0075 IntrVec_t c_14 = MUL(b_27, a_27);
0076
0077 IntrVec_t b_28 = LD(b, 28);
0078 c_13 = FMA(b_28, a_22, c_13);
0079 IntrVec_t a_28 = LD(a, 28);
0080 c_14 = FMA(b_28, a_28, c_14);
0081
0082
0083 IntrVec_t b_30 = LD(b, 30);
0084 IntrVec_t c_15 = b_30;
0085
0086 IntrVec_t b_31 = LD(b, 31);
0087 IntrVec_t c_16 = b_31;
0088 ST(c, 10, c_10);
0089 ST(c, 11, c_11);
0090 ST(c, 12, c_12);
0091 ST(c, 13, c_13);
0092 ST(c, 14, c_14);
0093
0094 IntrVec_t b_32 = LD(b, 32);
0095 IntrVec_t c_17 = b_32;
0096
0097 IntrVec_t b_33 = LD(b, 33);
0098 IntrVec_t c_18 = MUL(b_33, a_21);
0099 IntrVec_t c_19 = MUL(b_33, a_27);
0100 IntrVec_t a_33 = LD(a, 33);
0101 IntrVec_t c_20 = MUL(b_33, a_33);
0102
0103 IntrVec_t b_34 = LD(b, 34);
0104 c_18 = FMA(b_34, a_22, c_18);
0105 c_19 = FMA(b_34, a_28, c_19);
0106
0107 IntrVec_t b_35 = LD(b, 35);
0108 IntrVec_t a_35 = LD(a, 35);
0109 c_20 = FMA(b_35, a_35, c_20);
0110 ST(c, 15, c_15);
0111 ST(c, 16, c_16);
0112 ST(c, 17, c_17);
0113 ST(c, 18, c_18);
0114 ST(c, 19, c_19);
0115 ST(c, 20, c_20);
0116 }
0117
0118 #else
0119
0120 #pragma omp simd
0121 for (int n = 0; n < N; ++n)
0122 {
0123 c[ 0*N+n] = b[ 0*N+n];
0124 c[ 1*N+n] = b[ 6*N+n];
0125 c[ 2*N+n] = b[ 7*N+n];
0126 c[ 3*N+n] = b[12*N+n];
0127 c[ 4*N+n] = b[13*N+n];
0128 c[ 5*N+n] = b[14*N+n];
0129 c[ 6*N+n] = b[18*N+n];
0130 c[ 7*N+n] = b[19*N+n];
0131 c[ 8*N+n] = b[20*N+n];
0132 c[ 9*N+n] = b[21*N+n]*a[21*N+n] + b[22*N+n]*a[22*N+n];
0133 c[10*N+n] = b[24*N+n];
0134 c[11*N+n] = b[25*N+n];
0135 c[12*N+n] = b[26*N+n];
0136 c[13*N+n] = b[27*N+n]*a[21*N+n] + b[28*N+n]*a[22*N+n];
0137 c[14*N+n] = b[27*N+n]*a[27*N+n] + b[28*N+n]*a[28*N+n];
0138 c[15*N+n] = b[30*N+n];
0139 c[16*N+n] = b[31*N+n];
0140 c[17*N+n] = b[32*N+n];
0141 c[18*N+n] = b[33*N+n]*a[21*N+n] + b[34*N+n]*a[22*N+n];
0142 c[19*N+n] = b[33*N+n]*a[27*N+n] + b[34*N+n]*a[28*N+n];
0143 c[20*N+n] = b[33*N+n]*a[33*N+n] + b[35*N+n]*a[35*N+n];
0144 }
0145 #endif
