RecoTracker/MkFitCore/src/MultHelixPropTransp.ah - master - githubmirror/cmssw

RecoTracker/MkFitCore/src/MultHelixPropTransp.ah
246 строк8 KB
Shahzad Malik Muzaffar
[RECONSTRUCTION] [LLVM14] Apply clang-tidy and clang-format changes
26 апр 2023, 17:50
26 апр 2023, 17:50cef8b06
#ifdef MPLEX_INTRINSICS

for (int n = 0; n < N; n += MPLEX_INTRINSICS_WIDTH_BYTES / sizeof(T)) {
  IntrVec_t b_0 = LD(b, 0);
  IntrVec_t a_0 = LD(a, 0);
  IntrVec_t c_0 = MUL(b_0, a_0);

  IntrVec_t b_1 = LD(b, 1);
  IntrVec_t a_1 = LD(a, 1);
  c_0 = FMA(b_1, a_1, c_0);

  IntrVec_t b_3 = LD(b, 3);
  IntrVec_t a_3 = LD(a, 3);
  c_0 = FMA(b_3, a_3, c_0);

  IntrVec_t b_4 = LD(b, 4);
  IntrVec_t a_4 = LD(a, 4);
  c_0 = FMA(b_4, a_4, c_0);

  IntrVec_t b_6 = LD(b, 6);
  IntrVec_t c_1 = MUL(b_6, a_0);
  IntrVec_t a_6 = LD(a, 6);
  IntrVec_t c_2 = MUL(b_6, a_6);
  ST(c, 0, c_0);

  IntrVec_t b_7 = LD(b, 7);
  c_1 = FMA(b_7, a_1, c_1);
  IntrVec_t a_7 = LD(a, 7);
  c_2 = FMA(b_7, a_7, c_2);

  IntrVec_t b_9 = LD(b, 9);
  c_1 = FMA(b_9, a_3, c_1);
  IntrVec_t a_9 = LD(a, 9);
  c_2 = FMA(b_9, a_9, c_2);

  IntrVec_t b_10 = LD(b, 10);
  c_1 = FMA(b_10, a_4, c_1);
  IntrVec_t a_10 = LD(a, 10);
  c_2 = FMA(b_10, a_10, c_2);

  IntrVec_t b_12 = LD(b, 12);
  IntrVec_t c_3 = MUL(b_12, a_0);
  IntrVec_t c_4 = MUL(b_12, a_6);
  IntrVec_t a_12 = LD(a, 12);
  IntrVec_t c_5 = MUL(b_12, a_12);
  ST(c, 1, c_1);
  ST(c, 2, c_2);

  IntrVec_t b_13 = LD(b, 13);
  c_3 = FMA(b_13, a_1, c_3);
  c_4 = FMA(b_13, a_7, c_4);
  IntrVec_t a_13 = LD(a, 13);
  c_5 = FMA(b_13, a_13, c_5);

  IntrVec_t b_14 = LD(b, 14);
  c_5 = ADD(b_14, c_5);

  IntrVec_t b_15 = LD(b, 15);
  c_3 = FMA(b_15, a_3, c_3);
  c_4 = FMA(b_15, a_9, c_4);
  IntrVec_t a_15 = LD(a, 15);
  c_5 = FMA(b_15, a_15, c_5);

  IntrVec_t b_16 = LD(b, 16);
  c_3 = FMA(b_16, a_4, c_3);
  c_4 = FMA(b_16, a_10, c_4);
  IntrVec_t a_16 = LD(a, 16);
  c_5 = FMA(b_16, a_16, c_5);

  IntrVec_t b_17 = LD(b, 17);
  IntrVec_t a_17 = LD(a, 17);
  c_5 = FMA(b_17, a_17, c_5);

  IntrVec_t b_18 = LD(b, 18);
  IntrVec_t c_6 = MUL(b_18, a_0);
  ST(c, 3, c_3);
  ST(c, 4, c_4);
  ST(c, 5, c_5);
  IntrVec_t c_7 = MUL(b_18, a_6);
  IntrVec_t c_8 = MUL(b_18, a_12);
  IntrVec_t a_18 = LD(a, 18);
  IntrVec_t c_9 = MUL(b_18, a_18);

  IntrVec_t b_19 = LD(b, 19);
  c_6 = FMA(b_19, a_1, c_6);
  c_7 = FMA(b_19, a_7, c_7);
  c_8 = FMA(b_19, a_13, c_8);
  IntrVec_t a_19 = LD(a, 19);
  c_9 = FMA(b_19, a_19, c_9);

  IntrVec_t b_20 = LD(b, 20);
  c_8 = ADD(b_20, c_8);

  IntrVec_t b_21 = LD(b, 21);
  c_6 = FMA(b_21, a_3, c_6);
  c_7 = FMA(b_21, a_9, c_7);
  c_8 = FMA(b_21, a_15, c_8);
  IntrVec_t a_21 = LD(a, 21);
  c_9 = FMA(b_21, a_21, c_9);

  IntrVec_t b_22 = LD(b, 22);
  c_6 = FMA(b_22, a_4, c_6);
  c_7 = FMA(b_22, a_10, c_7);
  c_8 = FMA(b_22, a_16, c_8);
  IntrVec_t a_22 = LD(a, 22);
  c_9 = FMA(b_22, a_22, c_9);

  IntrVec_t b_23 = LD(b, 23);
  c_8 = FMA(b_23, a_17, c_8);

  IntrVec_t b_24 = LD(b, 24);
  IntrVec_t c_10 = MUL(b_24, a_0);
  ST(c, 6, c_6);
  ST(c, 7, c_7);
  ST(c, 8, c_8);
  ST(c, 9, c_9);
  IntrVec_t c_11 = MUL(b_24, a_6);
  IntrVec_t c_12 = MUL(b_24, a_12);
  IntrVec_t c_13 = MUL(b_24, a_18);
  IntrVec_t a_24 = LD(a, 24);
  IntrVec_t c_14 = MUL(b_24, a_24);

  IntrVec_t b_25 = LD(b, 25);
  c_10 = FMA(b_25, a_1, c_10);
  c_11 = FMA(b_25, a_7, c_11);
  c_12 = FMA(b_25, a_13, c_12);
  c_13 = FMA(b_25, a_19, c_13);
  IntrVec_t a_25 = LD(a, 25);
  c_14 = FMA(b_25, a_25, c_14);

  IntrVec_t b_26 = LD(b, 26);
  c_12 = ADD(b_26, c_12);

  IntrVec_t b_27 = LD(b, 27);
  c_10 = FMA(b_27, a_3, c_10);
  c_11 = FMA(b_27, a_9, c_11);
  c_12 = FMA(b_27, a_15, c_12);
  c_13 = FMA(b_27, a_21, c_13);
  IntrVec_t a_27 = LD(a, 27);
  c_14 = FMA(b_27, a_27, c_14);

  IntrVec_t b_28 = LD(b, 28);
  c_10 = FMA(b_28, a_4, c_10);
  c_11 = FMA(b_28, a_10, c_11);
  c_12 = FMA(b_28, a_16, c_12);
  c_13 = FMA(b_28, a_22, c_13);
  IntrVec_t a_28 = LD(a, 28);
  c_14 = FMA(b_28, a_28, c_14);

  IntrVec_t b_29 = LD(b, 29);
  c_12 = FMA(b_29, a_17, c_12);

  IntrVec_t b_30 = LD(b, 30);
  IntrVec_t c_15 = MUL(b_30, a_0);
  ST(c, 10, c_10);
  ST(c, 11, c_11);
  ST(c, 12, c_12);
  ST(c, 13, c_13);
  ST(c, 14, c_14);
  IntrVec_t c_16 = MUL(b_30, a_6);
  IntrVec_t c_17 = MUL(b_30, a_12);
  IntrVec_t c_18 = MUL(b_30, a_18);
  IntrVec_t c_19 = MUL(b_30, a_24);

  IntrVec_t b_31 = LD(b, 31);
  c_15 = FMA(b_31, a_1, c_15);
  c_16 = FMA(b_31, a_7, c_16);
  c_17 = FMA(b_31, a_13, c_17);
  c_18 = FMA(b_31, a_19, c_18);
  c_19 = FMA(b_31, a_25, c_19);

  IntrVec_t b_32 = LD(b, 32);
  c_17 = ADD(b_32, c_17);

  IntrVec_t b_33 = LD(b, 33);
  c_15 = FMA(b_33, a_3, c_15);
  c_16 = FMA(b_33, a_9, c_16);
  c_17 = FMA(b_33, a_15, c_17);
  c_18 = FMA(b_33, a_21, c_18);
  c_19 = FMA(b_33, a_27, c_19);

  IntrVec_t b_34 = LD(b, 34);
  c_15 = FMA(b_34, a_4, c_15);
  c_16 = FMA(b_34, a_10, c_16);
  c_17 = FMA(b_34, a_16, c_17);
  c_18 = FMA(b_34, a_22, c_18);
  c_19 = FMA(b_34, a_28, c_19);

  IntrVec_t b_35 = LD(b, 35);
  c_17 = FMA(b_35, a_17, c_17);
  IntrVec_t c_20 = b_35;
  ST(c, 15, c_15);
  ST(c, 16, c_16);
  ST(c, 17, c_17);
  ST(c, 18, c_18);
  ST(c, 19, c_19);
  ST(c, 20, c_20);
}

#else

#pragma omp simd
for (int n = 0; n < N; ++n) {
  c[0 * N + n] = b[0 * N + n] * a[0 * N + n] + b[1 * N + n] * a[1 * N + n] + b[3 * N + n] * a[3 * N + n] +
                 b[4 * N + n] * a[4 * N + n];
  c[1 * N + n] = b[6 * N + n] * a[0 * N + n] + b[7 * N + n] * a[1 * N + n] + b[9 * N + n] * a[3 * N + n] +
                 b[10 * N + n] * a[4 * N + n];
  c[2 * N + n] = b[6 * N + n] * a[6 * N + n] + b[7 * N + n] * a[7 * N + n] + b[9 * N + n] * a[9 * N + n] +
                 b[10 * N + n] * a[10 * N + n];
  c[3 * N + n] = b[12 * N + n] * a[0 * N + n] + b[13 * N + n] * a[1 * N + n] + b[15 * N + n] * a[3 * N + n] +
                 b[16 * N + n] * a[4 * N + n];
  c[4 * N + n] = b[12 * N + n] * a[6 * N + n] + b[13 * N + n] * a[7 * N + n] + b[15 * N + n] * a[9 * N + n] +
                 b[16 * N + n] * a[10 * N + n];
  c[5 * N + n] = b[12 * N + n] * a[12 * N + n] + b[13 * N + n] * a[13 * N + n] + b[14 * 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];
  c[6 * N + n] = b[18 * N + n] * a[0 * N + n] + b[19 * N + n] * a[1 * N + n] + b[21 * N + n] * a[3 * N + n] +
                 b[22 * N + n] * a[4 * N + n];
  c[7 * N + n] = b[18 * N + n] * a[6 * N + n] + b[19 * N + n] * a[7 * N + n] + b[21 * N + n] * a[9 * N + n] +
                 b[22 * N + n] * a[10 * N + n];
  c[8 * N + n] = b[18 * N + n] * a[12 * N + n] + b[19 * N + n] * a[13 * N + n] + b[20 * N + n] +
                 b[21 * N + n] * a[15 * N + n] + b[22 * N + n] * a[16 * N + n] + b[23 * N + n] * a[17 * N + n];
  c[9 * N + n] = b[18 * N + n] * a[18 * N + n] + b[19 * N + n] * a[19 * N + n] + b[21 * N + n] * a[21 * N + n] +
                 b[22 * N + n] * a[22 * N + n];
  c[10 * N + n] = b[24 * N + n] * a[0 * N + n] + b[25 * N + n] * a[1 * N + n] + b[27 * N + n] * a[3 * N + n] +
                  b[28 * N + n] * a[4 * N + n];
  c[11 * N + n] = b[24 * N + n] * a[6 * N + n] + b[25 * N + n] * a[7 * N + n] + b[27 * N + n] * a[9 * N + n] +
                  b[28 * N + n] * a[10 * N + n];
  c[12 * N + n] = b[24 * N + n] * a[12 * N + n] + b[25 * N + n] * a[13 * N + n] + b[26 * N + n] +
                  b[27 * N + n] * a[15 * N + n] + b[28 * N + n] * a[16 * N + n] + b[29 * N + n] * a[17 * N + n];
  c[13 * N + n] = b[24 * N + n] * a[18 * N + n] + b[25 * N + n] * a[19 * N + n] + b[27 * N + n] * a[21 * N + n] +
                  b[28 * N + n] * a[22 * N + n];
  c[14 * N + n] = b[24 * N + n] * a[24 * N + n] + b[25 * N + n] * a[25 * N + n] + b[27 * N + n] * a[27 * N + n] +
                  b[28 * N + n] * a[28 * N + n];
  c[15 * N + n] = b[30 * N + n] * a[0 * N + n] + b[31 * N + n] * a[1 * N + n] + b[33 * N + n] * a[3 * N + n] +
                  b[34 * N + n] * a[4 * N + n];
  c[16 * N + n] = b[30 * N + n] * a[6 * N + n] + b[31 * N + n] * a[7 * N + n] + b[33 * N + n] * a[9 * N + n] +
                  b[34 * N + n] * a[10 * N + n];
  c[17 * N + n] = b[30 * N + n] * a[12 * N + n] + b[31 * N + n] * a[13 * N + n] + b[32 * N + n] +
                  b[33 * N + n] * a[15 * N + n] + b[34 * N + n] * a[16 * N + n] + b[35 * N + n] * a[17 * N + n];
  c[18 * N + n] = b[30 * N + n] * a[18 * N + n] + b[31 * N + n] * a[19 * N + n] + b[33 * N + n] * a[21 * N + n] +
                  b[34 * N + n] * a[22 * N + n];
  c[19 * N + n] = b[30 * N + n] * a[24 * N + n] + b[31 * N + n] * a[25 * N + n] + b[33 * N + n] * a[27 * N + n] +
                  b[34 * N + n] * a[28 * N + n];
  c[20 * N + n] = b[35 * N + n];
}
#endif