Warning, /RecoTracker/MkFitCore/src/MultHelixPropTranspEndcap.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 #ifdef AVX512_INTRINSICS
0005 IntrVec_t all_zeros = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
0006 #else
0007 IntrVec_t all_zeros = {0, 0, 0, 0, 0, 0, 0, 0};
0008 #endif
0009
0010 IntrVec_t b_0 = LD(b, 0);
0011 IntrVec_t c_0 = b_0;
0012
0013 IntrVec_t b_2 = LD(b, 2);
0014 IntrVec_t a_2 = LD(a, 2);
0015 c_0 = FMA(b_2, a_2, c_0);
0016
0017 IntrVec_t b_3 = LD(b, 3);
0018 IntrVec_t a_3 = LD(a, 3);
0019 c_0 = FMA(b_3, a_3, c_0);
0020
0021 IntrVec_t b_4 = LD(b, 4);
0022 IntrVec_t a_4 = LD(a, 4);
0023 c_0 = FMA(b_4, a_4, c_0);
0024
0025 IntrVec_t b_5 = LD(b, 5);
0026 IntrVec_t a_5 = LD(a, 5);
0027 c_0 = FMA(b_5, a_5, c_0);
0028
0029 IntrVec_t b_6 = LD(b, 6);
0030 IntrVec_t c_1 = b_6;
0031
0032 IntrVec_t b_7 = LD(b, 7);
0033 IntrVec_t c_2 = b_7;
0034 ST(c, 0, c_0);
0035
0036 IntrVec_t b_8 = LD(b, 8);
0037 c_1 = FMA(b_8, a_2, c_1);
0038 IntrVec_t a_8 = LD(a, 8);
0039 c_2 = FMA(b_8, a_8, c_2);
0040
0041 IntrVec_t b_9 = LD(b, 9);
0042 c_1 = FMA(b_9, a_3, c_1);
0043 IntrVec_t a_9 = LD(a, 9);
0044 c_2 = FMA(b_9, a_9, c_2);
0045
0046 IntrVec_t b_10 = LD(b, 10);
0047 c_1 = FMA(b_10, a_4, c_1);
0048 IntrVec_t a_10 = LD(a, 10);
0049 c_2 = FMA(b_10, a_10, c_2);
0050
0051 IntrVec_t b_11 = LD(b, 11);
0052 c_1 = FMA(b_11, a_5, c_1);
0053 IntrVec_t a_11 = LD(a, 11);
0054 c_2 = FMA(b_11, a_11, c_2);
0055
0056 IntrVec_t b_12 = LD(b, 12);
0057 IntrVec_t c_3 = b_12;
0058 ST(c, 1, c_1);
0059
0060 IntrVec_t b_13 = LD(b, 13);
0061 IntrVec_t c_4 = b_13;
0062 ST(c, 2, c_2);
0063
0064 IntrVec_t b_14 = LD(b, 14);
0065 c_3 = FMA(b_14, a_2, c_3);
0066 c_4 = FMA(b_14, a_8, c_4);
0067
0068 IntrVec_t b_15 = LD(b, 15);
0069 c_3 = FMA(b_15, a_3, c_3);
0070 c_4 = FMA(b_15, a_9, c_4);
0071
0072 IntrVec_t b_16 = LD(b, 16);
0073 c_3 = FMA(b_16, a_4, c_3);
0074 c_4 = FMA(b_16, a_10, c_4);
0075
0076 IntrVec_t b_17 = LD(b, 17);
0077 c_3 = FMA(b_17, a_5, c_3);
0078 c_4 = FMA(b_17, a_11, c_4);
0079 ST(c, 5, all_zeros);
0080
0081 IntrVec_t b_18 = LD(b, 18);
0082 IntrVec_t c_6 = b_18;
0083
0084 IntrVec_t b_19 = LD(b, 19);
0085 IntrVec_t c_7 = b_19;
0086 ST(c, 3, c_3);
0087 ST(c, 4, c_4);
0088
0089 IntrVec_t b_20 = LD(b, 20);
0090 c_6 = FMA(b_20, a_2, c_6);
0091 c_7 = FMA(b_20, a_8, c_7);
0092
0093 IntrVec_t b_21 = LD(b, 21);
0094 c_6 = FMA(b_21, a_3, c_6);
0095 c_7 = FMA(b_21, a_9, c_7);
0096 IntrVec_t c_9 = b_21;
0097
0098 IntrVec_t b_22 = LD(b, 22);
0099 c_6 = FMA(b_22, a_4, c_6);
0100 c_7 = FMA(b_22, a_10, c_7);
0101
0102 IntrVec_t b_23 = LD(b, 23);
0103 c_6 = FMA(b_23, a_5, c_6);
0104 c_7 = FMA(b_23, a_11, c_7);
0105 ST(c, 8, all_zeros);
0106
0107 IntrVec_t b_24 = LD(b, 24);
0108 IntrVec_t c_10 = b_24;
0109
0110 IntrVec_t b_25 = LD(b, 25);
0111 IntrVec_t c_11 = b_25;
0112 ST(c, 6, c_6);
0113 ST(c, 7, c_7);
0114 ST(c, 9, c_9);
0115
0116 IntrVec_t b_26 = LD(b, 26);
0117 c_10 = FMA(b_26, a_2, c_10);
0118 c_11 = FMA(b_26, a_8, c_11);
0119 IntrVec_t a_26 = LD(a, 26);
0120 IntrVec_t c_14 = MUL(b_26, a_26);
0121
0122 IntrVec_t b_27 = LD(b, 27);
0123 c_10 = FMA(b_27, a_3, c_10);
0124 c_11 = FMA(b_27, a_9, c_11);
0125 IntrVec_t c_13 = b_27;
0126 IntrVec_t a_27 = LD(a, 27);
0127 c_14 = FMA(b_27, a_27, c_14);
0128
0129 IntrVec_t b_28 = LD(b, 28);
0130 c_10 = FMA(b_28, a_4, c_10);
0131 c_11 = FMA(b_28, a_10, c_11);
0132 c_14 = ADD(b_28, c_14);
0133
0134 IntrVec_t b_29 = LD(b, 29);
0135 c_10 = FMA(b_29, a_5, c_10);
0136 c_11 = FMA(b_29, a_11, c_11);
0137 ST(c, 12, all_zeros);
0138 IntrVec_t a_29 = LD(a, 29);
0139 c_14 = FMA(b_29, a_29, c_14);
0140
0141 IntrVec_t b_30 = LD(b, 30);
0142 IntrVec_t c_15 = b_30;
0143 ST(c, 10, c_10);
0144 ST(c, 11, c_11);
0145 ST(c, 13, c_13);
0146 ST(c, 14, c_14);
0147
0148 IntrVec_t b_31 = LD(b, 31);
0149 IntrVec_t c_16 = b_31;
0150
0151 IntrVec_t b_32 = LD(b, 32);
0152 c_15 = FMA(b_32, a_2, c_15);
0153 c_16 = FMA(b_32, a_8, c_16);
0154 IntrVec_t c_19 = MUL(b_32, a_26);
0155
0156 IntrVec_t b_33 = LD(b, 33);
0157 c_15 = FMA(b_33, a_3, c_15);
0158 c_16 = FMA(b_33, a_9, c_16);
0159 IntrVec_t c_18 = b_33;
0160 c_19 = FMA(b_33, a_27, c_19);
0161
0162 IntrVec_t b_34 = LD(b, 34);
0163 c_15 = FMA(b_34, a_4, c_15);
0164 c_16 = FMA(b_34, a_10, c_16);
0165 c_19 = ADD(b_34, c_19);
0166
0167 IntrVec_t b_35 = LD(b, 35);
0168 c_15 = FMA(b_35, a_5, c_15);
0169 c_16 = FMA(b_35, a_11, c_16);
0170 ST(c, 17, all_zeros);
0171 c_19 = FMA(b_35, a_29, c_19);
0172 IntrVec_t c_20 = b_35;
0173 ST(c, 15, c_15);
0174 ST(c, 16, c_16);
0175 ST(c, 18, c_18);
0176 ST(c, 19, c_19);
0177 ST(c, 20, c_20);
0178 }
0179
0180 #else
0181
0182 #pragma omp simd
0183 for (int n = 0; n < N; ++n) {
0184 c[0 * N + n] = b[0 * N + n] + b[2 * N + n] * a[2 * N + n] + b[3 * N + n] * a[3 * N + n] +
0185 b[4 * N + n] * a[4 * N + n] + b[5 * N + n] * a[5 * N + n];
0186 c[1 * N + n] = b[6 * N + n] + b[8 * N + n] * a[2 * N + n] + b[9 * N + n] * a[3 * N + n] +
0187 b[10 * N + n] * a[4 * N + n] + b[11 * N + n] * a[5 * N + n];
0188 c[2 * N + n] = b[7 * N + n] + b[8 * N + n] * a[8 * N + n] + b[9 * N + n] * a[9 * N + n] +
0189 b[10 * N + n] * a[10 * N + n] + b[11 * N + n] * a[11 * N + n];
0190 c[3 * N + n] = b[12 * N + n] + b[14 * N + n] * a[2 * N + n] + b[15 * N + n] * a[3 * N + n] +
0191 b[16 * N + n] * a[4 * N + n] + b[17 * N + n] * a[5 * N + n];
0192 c[4 * N + n] = b[13 * N + n] + b[14 * N + n] * a[8 * N + n] + b[15 * N + n] * a[9 * N + n] +
0193 b[16 * N + n] * a[10 * N + n] + b[17 * N + n] * a[11 * N + n];
0194 c[5 * N + n] = 0;
0195 c[6 * N + n] = b[18 * N + n] + b[20 * N + n] * a[2 * N + n] + b[21 * N + n] * a[3 * N + n] +
0196 b[22 * N + n] * a[4 * N + n] + b[23 * N + n] * a[5 * N + n];
0197 c[7 * N + n] = b[19 * N + n] + b[20 * N + n] * a[8 * N + n] + b[21 * N + n] * a[9 * N + n] +
0198 b[22 * N + n] * a[10 * N + n] + b[23 * N + n] * a[11 * N + n];
0199 c[8 * N + n] = 0;
0200 c[9 * N + n] = b[21 * N + n];
0201 c[10 * N + n] = b[24 * N + n] + b[26 * N + n] * a[2 * N + n] + b[27 * N + n] * a[3 * N + n] +
0202 b[28 * N + n] * a[4 * N + n] + b[29 * N + n] * a[5 * N + n];
0203 c[11 * N + n] = b[25 * N + n] + b[26 * N + n] * a[8 * N + n] + b[27 * N + n] * a[9 * N + n] +
0204 b[28 * N + n] * a[10 * N + n] + b[29 * N + n] * a[11 * N + n];
0205 c[12 * N + n] = 0;
0206 c[13 * N + n] = b[27 * N + n];
0207 c[14 * N + n] =
0208 b[26 * N + n] * a[26 * N + n] + b[27 * N + n] * a[27 * N + n] + b[28 * N + n] + b[29 * N + n] * a[29 * N + n];
0209 c[15 * N + n] = b[30 * N + n] + b[32 * N + n] * a[2 * N + n] + b[33 * N + n] * a[3 * N + n] +
0210 b[34 * N + n] * a[4 * N + n] + b[35 * N + n] * a[5 * N + n];
0211 c[16 * N + n] = b[31 * N + n] + b[32 * N + n] * a[8 * N + n] + b[33 * N + n] * a[9 * N + n] +
0212 b[34 * N + n] * a[10 * N + n] + b[35 * N + n] * a[11 * N + n];
0213 c[17 * N + n] = 0;
0214 c[18 * N + n] = b[33 * N + n];
0215 c[19 * N + n] =
0216 b[32 * N + n] * a[26 * N + n] + b[33 * N + n] * a[27 * N + n] + b[34 * N + n] + b[35 * N + n] * a[29 * N + n];
0217 c[20 * N + n] = b[35 * N + n];
0218 }
0219 #endif
