Warning, /RecoTracker/MkFitCore/src/OutErrCCSTransp.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_3 = LD(b, 3);
0005 IntrVec_t a_3 = LD(a, 3);
0006 IntrVec_t c_0 = MUL(b_3, a_3);
0007
0008 IntrVec_t b_4 = LD(b, 4);
0009 IntrVec_t a_4 = LD(a, 4);
0010 c_0 = FMA(b_4, a_4, c_0);
0011
0012 IntrVec_t b_8 = LD(b, 8);
0013 IntrVec_t c_1 = MUL(b_8, a_3);
0014 IntrVec_t a_8 = LD(a, 8);
0015 IntrVec_t c_2 = MUL(b_8, a_8);
0016 ST(c, 0, c_0);
0017
0018 IntrVec_t b_9 = LD(b, 9);
0019 c_1 = FMA(b_9, a_4, c_1);
0020 IntrVec_t a_9 = LD(a, 9);
0021 c_2 = FMA(b_9, a_9, c_2);
0022
0023 IntrVec_t b_13 = LD(b, 13);
0024 IntrVec_t c_3 = MUL(b_13, a_3);
0025 ST(c, 1, c_1);
0026 IntrVec_t c_4 = MUL(b_13, a_8);
0027 ST(c, 2, c_2);
0028 IntrVec_t a_13 = LD(a, 13);
0029 IntrVec_t c_5 = MUL(b_13, a_13);
0030
0031 IntrVec_t b_14 = LD(b, 14);
0032 c_3 = FMA(b_14, a_4, c_3);
0033 c_4 = FMA(b_14, a_9, c_4);
0034 IntrVec_t a_14 = LD(a, 14);
0035 c_5 = FMA(b_14, a_14, c_5);
0036
0037 IntrVec_t b_15 = LD(b, 15);
0038 IntrVec_t a_15 = LD(a, 15);
0039 IntrVec_t c_9 = MUL(b_15, a_15);
0040 ST(c, 3, c_3);
0041 ST(c, 4, c_4);
0042 ST(c, 5, c_5);
0043
0044 IntrVec_t b_16 = LD(b, 16);
0045 IntrVec_t a_16 = LD(a, 16);
0046 c_9 = FMA(b_16, a_16, c_9);
0047
0048 IntrVec_t b_17 = LD(b, 17);
0049 IntrVec_t a_17 = LD(a, 17);
0050 c_9 = FMA(b_17, a_17, c_9);
0051
0052 IntrVec_t b_18 = LD(b, 18);
0053 IntrVec_t c_6 = MUL(b_18, a_3);
0054 IntrVec_t c_7 = MUL(b_18, a_8);
0055 IntrVec_t c_8 = MUL(b_18, a_13);
0056
0057 IntrVec_t b_19 = LD(b, 19);
0058 c_6 = FMA(b_19, a_4, c_6);
0059 c_7 = FMA(b_19, a_9, c_7);
0060 c_8 = FMA(b_19, a_14, c_8);
0061
0062 IntrVec_t b_20 = LD(b, 20);
0063 IntrVec_t c_13 = MUL(b_20, a_15);
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_21 = LD(b, 21);
0070 c_13 = FMA(b_21, a_16, c_13);
0071 IntrVec_t a_21 = LD(a, 21);
0072 IntrVec_t c_14 = MUL(b_21, a_21);
0073
0074 IntrVec_t b_22 = LD(b, 22);
0075 c_13 = FMA(b_22, a_17, c_13);
0076 IntrVec_t a_22 = LD(a, 22);
0077 c_14 = FMA(b_22, a_22, c_14);
0078
0079 IntrVec_t b_23 = LD(b, 23);
0080 IntrVec_t c_10 = MUL(b_23, a_3);
0081 IntrVec_t c_11 = MUL(b_23, a_8);
0082 IntrVec_t c_12 = MUL(b_23, a_13);
0083 IntrVec_t a_23 = LD(a, 23);
0084 c_14 = FMA(b_23, a_23, c_14);
0085
0086 IntrVec_t b_24 = LD(b, 24);
0087 c_10 = FMA(b_24, a_4, c_10);
0088 c_11 = FMA(b_24, a_9, c_11);
0089 c_12 = FMA(b_24, a_14, c_12);
0090 IntrVec_t a_24 = LD(a, 24);
0091 c_14 = FMA(b_24, a_24, c_14);
0092 ST(c, 10, c_10);
0093 ST(c, 11, c_11);
0094 ST(c, 12, c_12);
0095 ST(c, 13, c_13);
0096 ST(c, 14, c_14);
0097
0098 IntrVec_t b_25 = LD(b, 25);
0099 IntrVec_t c_18 = MUL(b_25, a_15);
0100
0101 IntrVec_t b_26 = LD(b, 26);
0102 c_18 = FMA(b_26, a_16, c_18);
0103 IntrVec_t c_19 = MUL(b_26, a_21);
0104 IntrVec_t a_26 = LD(a, 26);
0105 IntrVec_t c_20 = MUL(b_26, a_26);
0106
0107 IntrVec_t b_27 = LD(b, 27);
0108 c_18 = FMA(b_27, a_17, c_18);
0109 c_19 = FMA(b_27, a_22, c_19);
0110 IntrVec_t a_27 = LD(a, 27);
0111 c_20 = FMA(b_27, a_27, c_20);
0112
0113 IntrVec_t b_28 = LD(b, 28);
0114 IntrVec_t c_15 = MUL(b_28, a_3);
0115 IntrVec_t c_16 = MUL(b_28, a_8);
0116 IntrVec_t c_17 = MUL(b_28, a_13);
0117 c_19 = FMA(b_28, a_23, c_19);
0118
0119 IntrVec_t b_29 = LD(b, 29);
0120 c_15 = FMA(b_29, a_4, c_15);
0121 c_16 = FMA(b_29, a_9, c_16);
0122 c_17 = FMA(b_29, a_14, c_17);
0123 c_19 = FMA(b_29, a_24, c_19);
0124 ST(c, 15, c_15);
0125 ST(c, 16, c_16);
0126 ST(c, 17, c_17);
0127 ST(c, 18, c_18);
0128 ST(c, 19, c_19);
0129 ST(c, 20, c_20);
0130 }
0131
0132 #else
0133
0134 #pragma omp simd
0135 for (int n = 0; n < N; ++n) {
0136 c[0 * N + n] = b[3 * N + n] * a[3 * N + n] + b[4 * N + n] * a[4 * N + n];
0137 c[1 * N + n] = b[8 * N + n] * a[3 * N + n] + b[9 * N + n] * a[4 * N + n];
0138 c[2 * N + n] = b[8 * N + n] * a[8 * N + n] + b[9 * N + n] * a[9 * N + n];
0139 c[3 * N + n] = b[13 * N + n] * a[3 * N + n] + b[14 * N + n] * a[4 * N + n];
0140 c[4 * N + n] = b[13 * N + n] * a[8 * N + n] + b[14 * N + n] * a[9 * N + n];
0141 c[5 * N + n] = b[13 * N + n] * a[13 * N + n] + b[14 * N + n] * a[14 * N + n];
0142 c[6 * N + n] = b[18 * N + n] * a[3 * N + n] + b[19 * N + n] * a[4 * N + n];
0143 c[7 * N + n] = b[18 * N + n] * a[8 * N + n] + b[19 * N + n] * a[9 * N + n];
0144 c[8 * N + n] = b[18 * N + n] * a[13 * N + n] + b[19 * N + n] * a[14 * N + n];
0145 c[9 * N + n] = b[15 * N + n] * a[15 * N + n] + b[16 * N + n] * a[16 * N + n] + b[17 * N + n] * a[17 * N + n];
0146 c[10 * N + n] = b[23 * N + n] * a[3 * N + n] + b[24 * N + n] * a[4 * N + n];
0147 c[11 * N + n] = b[23 * N + n] * a[8 * N + n] + b[24 * N + n] * a[9 * N + n];
0148 c[12 * N + n] = b[23 * N + n] * a[13 * N + n] + b[24 * N + n] * a[14 * N + n];
0149 c[13 * N + n] = b[20 * N + n] * a[15 * N + n] + b[21 * N + n] * a[16 * N + n] + b[22 * N + n] * a[17 * N + n];
0150 c[14 * N + n] = b[21 * N + n] * a[21 * N + n] + b[22 * N + n] * a[22 * N + n] + b[23 * N + n] * a[23 * N + n] +
0151 b[24 * N + n] * a[24 * N + n];
0152 c[15 * N + n] = b[28 * N + n] * a[3 * N + n] + b[29 * N + n] * a[4 * N + n];
0153 c[16 * N + n] = b[28 * N + n] * a[8 * N + n] + b[29 * N + n] * a[9 * N + n];
0154 c[17 * N + n] = b[28 * N + n] * a[13 * N + n] + b[29 * N + n] * a[14 * N + n];
0155 c[18 * N + n] = b[25 * N + n] * a[15 * N + n] + b[26 * N + n] * a[16 * N + n] + b[27 * N + n] * a[17 * N + n];
0156 c[19 * N + n] = b[26 * N + n] * a[21 * N + n] + b[27 * N + n] * a[22 * N + n] + b[28 * N + n] * a[23 * N + n] +
0157 b[29 * N + n] * a[24 * N + n];
0158 c[20 * N + n] = b[26 * N + n] * a[26 * N + n] + b[27 * N + n] * a[27 * N + n];
0159 }
0160 #endif
