Merge remote-tracking branch 'origin/resource_mapping_optim' into integration_2025_w07 (!3127)

Optimizations of PDSCH Resource Mapping in nr_dlsch.c/nr_modulation.c

These changes add SIMD optimizations for Neon/AVX2/AVX512 in the PDSCH
transmit path. The timing improvements are listed here based on the

    nr_dlsim -e25 -R273 -b273 -s30 -x "layers" -y 4 -z 4 -P

benchmark with "layers" 2,3,4 and comparing "PHY proc tx":

273 PRBS, mcs25, 64QAM

peafowl (gcc11,AMD EPYC 9374F)

    2-layer, 4 TX : 431 us (develop 565 us)
    3-layer, 4 TX : 692 us (develop 849 us)
    4-layer, 4 TX : 963 us (develop 1172 us)

stupix (gcc10, Xeon Gold 6354)

    2-layer, 4 TX : 568 us (develop 652 us)
    3-layer, 4 TX : 901 us (develop 1030 us)
    4-layer, 4 TX : 1250 us (develop 1396 us)

matix (gcc14, Ryzen 9 PRO 7945)

    2-layer, 4 TX : 317 us (develop 505 us)
    3-layer, 4 TX : 538 us (develop 779 us)
    4-layer, 4 TX : 767 us (develop 1233 us)

openair1/PHY/MODULATION/nr_modulation.c

@@ -22,9 +22,14 @@
 #include "nr_modulation.h"
 #include "PHY/NR_REFSIG/nr_mod_table.h"
 #include "executables/softmodem-common.h"
+#include <simde/x86/avx512.h>
+// Lacking declaration in present simde external package, will be detected as compilation error when they will add it
+#define simde_mm512_extracti64x2_epi64(a...) _mm512_extracti64x2_epi64(a)
 
 // #define DEBUG_DLSCH_PRECODING_PRINT_WITH_TRIVIAL // TODO: For debug, to be removed if want to merge to develop
-
+// #define DEBUG_LAYER_MAPPING
+#define USE_NEON
+// #define USE_GATHER
 //  Table 6.3.1.5-1 Precoding Matrix W 1 layer 2 antenna ports 'n' = -1 and 'o' = -j
 const char nr_W_1l_2p[6][2][1] = {
     {{'1'}, {'0'}}, // pmi 0
@@ -112,10 +117,7 @@ const char nr_W_4l_4p[5][4][4] = {
     {{'1', '1', '1', '1'}, {'1', 'n', '1', 'n'}, {'j', 'j', 'o', 'o'}, {'j', 'o', 'o', 'j'}} // pmi 4
 };
 
-void nr_modulation(const uint32_t *in,
+void nr_modulation(const uint32_t *in, uint32_t length, uint16_t mod_order, int16_t *out)
-                   uint32_t length,
-                   uint16_t mod_order,
-                   int16_t *out)
 {
   uint16_t mask = ((1 << mod_order) - 1);
   int32_t *nr_mod_table32;
@@ -133,7 +135,6 @@ void nr_modulation(const uint32_t *in,
   LOG_D(PHY, "nr_modulation: length %d, mod_order %d\n", length, mod_order);
 
   switch (mod_order) {
-
 #if defined(__SSE2__)
     case 2:
       nr_mod_table128 = (simde__m128i *)nr_qpsk_byte_mod_table;
@@ -251,36 +252,272 @@ void nr_layer_mapping(int nbCodes,
                       uint8_t n_layers,
                       int layerSz,
                       uint32_t n_symbs,
-                      c16_t tx_layer[layerSz],
+                      c16_t tx_layers[][layerSz])
-                      int layer)
 {
   LOG_D(PHY, "Doing layer mapping for %d layers, %d symbols\n", n_layers, n_symbs);
-
+  c16_t *mod = mod_symbs[0];
   switch (n_layers) {
-
     case 1:
-      memcpy(tx_layer, mod_symbs[0], n_symbs * sizeof(**mod_symbs));
+      memcpy(tx_layers[0], mod, n_symbs * sizeof(**mod_symbs));
       break;
 
-    case 2:
+    case 2: {
-    case 3:
+      int i = 0;
-    case 4:
+      c16_t *tx0 = tx_layers[0];
-    for (int i = 0; i < n_symbs / n_layers; i++) {
+      c16_t *tx1 = tx_layers[1];
-      const c16_t *base = mod_symbs[0] + n_layers * i;
+#if defined(__AVX512BW__)
-      tx_layer[i] = base[layer];
+      simde__m512i perm2a = simde_mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
+      simde__m512i perm2b = simde_mm512_set_epi32(31, 29, 27, 25, 23, 21, 19, 17, 15, 13, 11, 9, 7, 5, 3, 1);
+      for (; i < (n_symbs & ~31); i += 32) {
+        simde__m512i a = *(simde__m512i *)(mod + i);
+        simde__m512i b = *(simde__m512i *)(mod + i + 16);
+        *(simde__m512i *)tx0 = simde_mm512_permutex2var_epi32(a, perm2a, b);
+        *(simde__m512i *)tx1 = simde_mm512_permutex2var_epi32(a, perm2b, b);
+        tx0 += 16;
+        tx1 += 16;
       }
-      break;
+#endif
+#ifdef __AVX2__
+      simde__m256i perm2 = simde_mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
+      for (; i < (n_symbs & ~7); i += 8) {
+        simde__m256i d = simde_mm256_permutevar8x32_epi32(*(simde__m256i *)(mod + i), perm2);
+        *(simde__m128i *)tx0 = simde_mm256_extractf128_si256(d, 0);
+        *(simde__m128i *)tx1 = simde_mm256_extractf128_si256(d, 1);
+        tx0 += 4;
+        tx1 += 4;
+      }
+#endif
+#if defined(__aarch64__) && defined(USE_NEON)
+      // SIMDe doesn't handle this properly, gcc up to 14.2 neither
+      uint8_t const perm0[16] = {0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15};
+      uint8x16_t perm = vld1q_u8(perm0);
+      uint8x16_t d;
+      for (; i < (n_symbs & (~3)); i += 4) {
+        d = vqtbl1q_u8(*(uint8x16_t *)(mod + i), perm);
+        *(int64_t *)tx0 = vgetq_lane_u64((uint64x2_t)d, 0);
+        *(int64_t *)tx1 = vgetq_lane_u64((uint64x2_t)d, 1);
+        tx0 += 2;
+        tx1 += 2;
+      }
+#endif
+      for (; i < n_symbs; i += 2) {
+        *tx0++ = mod[i];
+        *tx1++ = mod[i + 1];
+      }
+    } break;
+    case 3: {
+      int i = 0;
+      c16_t *tx0 = tx_layers[0];
+      c16_t *tx1 = tx_layers[1];
+      c16_t *tx2 = tx_layers[2];
+#if defined(__AVX512BW__)
+      simde__m512i perm3_0 = simde_mm512_set_epi32(13 + 16,
+                                                   10 + 16,
+                                                   7 + 16,
+                                                   4 + 16,
+                                                   1 + 16,
+                                                   14 + 16,
+                                                   11 + 16,
+                                                   8 + 16,
+                                                   5 + 16,
+                                                   2 + 16,
+                                                   15,
+                                                   12,
+                                                   9,
+                                                   6,
+                                                   3,
+                                                   0);
+      simde__m512i perm3_0b = simde_mm512_set_epi32(13 + 16, 10 + 16, 7 + 16, 4 + 16, 1 + 16, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
+      simde__m512i perm3_1 = simde_mm512_set_epi32(14 + 16,
+                                                   11 + 16,
+                                                   8 + 16,
+                                                   5 + 16,
+                                                   2 + 16,
+                                                   15 + 16,
+                                                   12 + 16,
+                                                   9 + 16,
+                                                   6 + 16,
+                                                   3 + 16,
+                                                   0 + 16,
+                                                   13,
+                                                   10,
+                                                   7,
+                                                   4,
+                                                   1);
+      simde__m512i perm3_1b = simde_mm512_set_epi32(14 + 16, 11 + 16, 8 + 16, 5 + 16, 2 + 16, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
+      simde__m512i perm3_2 = simde_mm512_set_epi32(15 + 16,
+                                                   12 + 16,
+                                                   9 + 16,
+                                                   6 + 16,
+                                                   3 + 16,
+                                                   0 + 16,
+                                                   13 + 16,
+                                                   10 + 16,
+                                                   7 + 16,
+                                                   4 + 16,
+                                                   1 + 16,
+                                                   14,
+                                                   11,
+                                                   8,
+                                                   5,
+                                                   2);
+      simde__m512i perm3_2b = simde_mm512_set_epi32(15 + 16, 12 + 16, 9 + 16, 6 + 16, 3 + 16, 0 + 16, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
+      for (; i < (n_symbs & ~63); i += 48) {
+        simde__m512i i0 = *(simde__m512i *)(mod + i);
+        simde__m512i i1 = *(simde__m512i *)(mod + i + 16);
+        simde__m512i i2 = *(simde__m512i *)(mod + i + 32);
+        simde__m512i d0 = simde_mm512_permutex2var_epi32(i0, perm3_0, i1);
+        *(simde__m512i *)tx0 = simde_mm512_permutex2var_epi32(d0, perm3_0b, i2); // 11000000
+        tx0 += 16;
+        d0 = simde_mm512_permutex2var_epi32(i0, perm3_1, i1);
+        *(simde__m512i *)tx1 = simde_mm512_permutex2var_epi32(d0, perm3_1b, i2); // 11000000
+        tx1 += 16;
+        d0 = simde_mm512_permutex2var_epi32(i0, perm3_2, i1);
+        *(simde__m512i *)tx2 = simde_mm512_permutex2var_epi32(d0, perm3_2b, i2); // 11000000
+        tx2 += 16;
+      }
+#endif
+#ifdef __AVX2__
+      {
+        simde__m256i perm3_0 = simde_mm256_set_epi32(5, 2, 7, 4, 1, 6, 3, 0);
+        simde__m256i perm3_1 = simde_mm256_set_epi32(6, 3, 0, 5, 2, 7, 4, 1);
+        simde__m256i perm3_2 = simde_mm256_set_epi32(7, 4, 1, 6, 3, 0, 5, 2);
+        for (; i < (n_symbs & ~31); i += 24) {
+          simde__m256i i0 = *(simde__m256i *)(mod + i);
+          simde__m256i i1 = *(simde__m256i *)(mod + i + 8);
+          simde__m256i i2 = *(simde__m256i *)(mod + i + 16);
+          simde__m256i d0 = simde_mm256_permutevar8x32_epi32(i0, perm3_0);
+          simde__m256i d1 = simde_mm256_permutevar8x32_epi32(i1, perm3_0);
+          simde__m256i d2 = simde_mm256_permutevar8x32_epi32(i2, perm3_0);
+          simde__m256i d3 = simde_mm256_blend_epi32(d0, d1, 0x38); // 00111000
+          *(simde__m256i *)tx0 = simde_mm256_blend_epi32(d3, d2, 0xc0); // 11000000
+          tx0 += 8;
+          d0 = simde_mm256_permutevar8x32_epi32(i0, perm3_1);
+          d1 = simde_mm256_permutevar8x32_epi32(i1, perm3_1);
+          d2 = simde_mm256_permutevar8x32_epi32(i2, perm3_1);
+          d3 = simde_mm256_blend_epi32(d0, d1, 0x18); // 00011000
+          *(simde__m256i *)tx1 = simde_mm256_blend_epi32(d3, d2, 0xe0); // 11100000
+          tx1 += 8;
+          d0 = simde_mm256_permutevar8x32_epi32(i0, perm3_2);
+          d1 = simde_mm256_permutevar8x32_epi32(i1, perm3_2);
+          d2 = simde_mm256_permutevar8x32_epi32(i2, perm3_2);
+          d3 = simde_mm256_blend_epi32(d0, d1, 0x1c); // 00011100
+          *(simde__m256i *)tx2 = simde_mm256_blend_epi32(d3, d2, 0xe0); // 11100000
+          tx2 += 8;
+        }
+      }
+#endif
+      for (; i < n_symbs; i += 3) {
+        *tx0++ = mod[i];
+        *tx1++ = mod[i + 1];
+        *tx2++ = mod[i + 2];
+      }
+
+#ifdef DEBUG_LAYER_MAPPING
+      printf("\nsymb %d/%d\n", i << 3, n_symbs);
+      printf(" layer 0:\t");
+      for (int j = 0; j < 8 * 6; j += 6) {
+        printf("%d %d ", ((int16_t *)&mod[i << 3])[j], ((int16_t *)&mod[i << 3])[j + 1]);
+      }
+      printf("\n layer 1:\t");
+      for (int j = 2; j < 8 * 6; j += 6) {
+        printf("%d %d ", ((int16_t *)&mod[i << 3])[j], ((int16_t *)&mod[i << 3])[j + 1]);
+      }
+      printf("\n layer 2:\t");
+      for (int j = 4; j < 8 * 6; j += 6) {
+        printf("%d %d ", ((int16_t *)&mod[i << 3])[j], ((int16_t *)&mod[i << 3])[j + 1]);
+      }
+      printf("\n Mapping layer 0:\t");
+      for (int j = 0; j < 16; j++) {
+        printf("%d ", ((int16_t *)&tx_layers[0][n << 3])[j]);
+      }
+      printf("\n Mapping layer 1:\t");
+      for (int j = 0; j < 16; j++) {
+        printf("%d ", ((int16_t *)&tx_layers[1][n << 3])[j]);
+      }
+      printf("\n Mapping layer 2:\t");
+      for (int j = 0; j < 16; j++) {
+        printf("%d ", ((int16_t *)&tx_layers[2][n << 3])[j]);
+      }
+#endif
+    } break;
+
+    case 4: {
+      int i = 0;
+      c16_t *tx0 = tx_layers[0];
+      c16_t *tx1 = tx_layers[1];
+      c16_t *tx2 = tx_layers[2];
+      c16_t *tx3 = tx_layers[3];
+#if defined(__AVX512BW__)
+      simde__m512i perm4 = simde_mm512_set_epi32(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0);
+      for (; i < (n_symbs & ~15); i += 16) {
+        simde__m512i e = simde_mm512_permutexvar_epi32(perm4, *(simde__m512i *)(mod + i));
+        *(simde__m128i *)tx0 = simde_mm512_extracti64x2_epi64(e, 0);
+        tx0 += 4;
+        *(simde__m128i *)tx1 = simde_mm512_extracti64x2_epi64(e, 1);
+        tx1 += 4;
+        *(simde__m128i *)tx2 = simde_mm512_extracti64x2_epi64(e, 2);
+        tx2 += 4;
+        *(simde__m128i *)tx3 = simde_mm512_extracti64x2_epi64(e, 3);
+        tx3 += 4;
+      }
+#endif
+#ifdef __AVX2__
+      {
+        simde__m256i perm4 = simde_mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
+        for (; i < (n_symbs & ~7); i += 8) {
+          simde__m256i e = simde_mm256_permutevar8x32_epi32(*(simde__m256i *)(mod + i), perm4);
+          *(uint64_t *)tx0 = simde_mm256_extract_epi64(e, 0);
+          tx0 += 2;
+          *(uint64_t *)tx1 = simde_mm256_extract_epi64(e, 1);
+          tx1 += 2;
+          *(uint64_t *)tx2 = simde_mm256_extract_epi64(e, 2);
+          tx2 += 2;
+          *(uint64_t *)tx3 = simde_mm256_extract_epi64(e, 3);
+          tx3 += 2;
+        }
+      }
+#endif
+#if defined(__aarch64__) && defined(USE_NEON)
+      // SIMDe doesn't handle this properly, gcc up to 14.2 neither
+      for (; i < (n_symbs & ~3); i += 4) {
+        uint32x4_t d4 = *(uint32x4_t *)(mod + i);
+        *(uint32_t *)tx0 = vgetq_lane_u32(d4, 0); 
+	tx0++;
+        *(uint32_t *)tx1 = vgetq_lane_u32(d4, 1); 
+	tx1++;
+        *(uint32_t *)tx2 = vgetq_lane_u32(d4, 0); 
+	tx2++;
+        *(uint32_t *)tx3 = vgetq_lane_u32(d4, 1); 
+	tx3++;
+      }
+#endif
+      for (; i < n_symbs; i += 4) {
+        *tx0++ = mod[i];
+        *tx1++ = mod[i + 1];
+        *tx2++ = mod[i + 2];
+        *tx3++ = mod[i + 3];
+      }
+    } break;
 
     case 5:
-      if (layer < 2)
+    case 6:
+    case 7:
+    case 8:
+      /*
+      // Layer 0,1
       for (int i = 0; i < n_symbs; i += 2) {
 const int txIdx = i / 2;
-          tx_layer[txIdx] = mod_symbs[0][i + layer];
+tx_layer[0][txIdx] = mod_symbs[0][i];
+tx_layer[1][txIdx] = mod_symbs[0][i + 1];
       }
+      // layers 2,3,4
       else
 for (int i = 0; i < n_symbs; i += 3) {
 const int txIdx = i / 3;
-          tx_layer[txIdx] = mod_symbs[1][i + layer];
+tx_layer[2][txIdx] = mod_symbs[1][i + 2];
+tx_layer[3][txIdx] = mod_symbs[1][i + 3];
+tx_layer[4][txIdx] = mod_symbs[1][i + 4];
 }
       break;
 
@@ -288,7 +525,12 @@ void nr_layer_mapping(int nbCodes,
       for (int q=0; q<2; q++)
 for (int i = 0; i < n_symbs; i += 3) {
 const int txIdx = i / 3;
-          tx_layer[txIdx] = mod_symbs[q][i + layer];
+tx_layer[0][txIdx] = mod_symbs[q][i + layer];
+tx_layer[1][txIdx] = mod_symbs[q][i + layer];
+tx_layer[2][txIdx] = mod_symbs[q][i + layer];
+tx_layer[3][txIdx] = mod_symbs[q][i + layer];
+tx_layer[4][txIdx] = mod_symbs[q][i + layer];
+tx_layer[5][txIdx] = mod_symbs[q][i + layer];
 }
       break;
 
@@ -312,7 +554,7 @@ void nr_layer_mapping(int nbCodes,
 tx_layer[txIdx] = mod_symbs[q][i + layer];
       }
       break;
-
+*/
     default:
       AssertFatal(0, "Invalid number of layers %d\n", n_layers);
   }
@@ -360,7 +602,6 @@ void nr_dft(c16_t *z, c16_t *d, uint32_t Msc_PUSCH)
     for (i = 0, ip = 0; i < Msc_PUSCH; i++, ip += 4)
       z[i] = dft_out0[ip];
   }
-
 }
 
 void perform_symbol_rotation(NR_DL_FRAME_PARMS *fp, double f0, c16_t *symbol_rotation)
@@ -430,7 +671,9 @@ void init_timeshift_rotation(NR_DL_FRAME_PARMS *fp)
     fp->timeshift_symbol_rotation[i].i = (int16_t)round(exp_im * 32767);
 
     if (i < 10)
-      LOG_D(PHY,"Timeshift symbol rotation %d => (%d,%d) %f\n",i,
+      LOG_D(PHY,
+            "Timeshift symbol rotation %d => (%d,%d) %f\n",
+            i,
             fp->timeshift_symbol_rotation[i].r,
             fp->timeshift_symbol_rotation[i].i,
             poff);
@@ -474,30 +717,26 @@ c16_t nr_layer_precoder(int sz, c16_t datatx_F_precoding[][sz], const char *prec
 }
 
 c16_t nr_layer_precoder_cm(int n_layers,
-                           int n_symbols,
                            int symSz,
-                           c16_t datatx_F_precoding[n_layers][n_symbols][symSz],
+                           c16_t datatx_F_precoding[n_layers][symSz],
                            int ap,
                            nfapi_nr_pm_pdu_t *pmi_pdu,
-                           int symbol,
                            int offset)
 {
   c16_t precodatatx_F = {0};
   for (int al = 0; al < n_layers; al++) {
     nfapi_nr_pm_weights_t *w = &pmi_pdu->weights[al][ap];
     c16_t prec_weight = {.r = w->precoder_weight_Re, .i = w->precoder_weight_Im};
-    precodatatx_F = c16maddShift(datatx_F_precoding[al][symbol][offset], prec_weight, precodatatx_F, 15);
+    precodatatx_F = c16maddShift(datatx_F_precoding[al][offset], prec_weight, precodatatx_F, 15);
   }
   return precodatatx_F;
 }
 
 void nr_layer_precoder_simd(const int n_layers,
-                           const int n_symbols,
                             const int symSz,
-                           const c16_t txdataF_res_mapped[n_layers][n_symbols][symSz],
+                            const c16_t txdataF_res_mapped[n_layers][symSz],
                             const int ant,
                             const nfapi_nr_pm_pdu_t *pmi_pdu,
-                           const int symbol,
                             const int sc_offset,
                             const int re_cnt,
                             c16_t *txdataF_precoded)
@@ -517,15 +756,18 @@ void nr_layer_precoder_simd(const int n_layers,
       prec_weight.r = pmi_pdu->weights[nl][ant].precoder_weight_Re;
       prec_weight.i = pmi_pdu->weights[nl][ant].precoder_weight_Im;
 
-      const simde__m256i x = simde_mm256_loadu_si256(&txdataF_res_mapped[nl][symbol][sc]);
+      const simde__m256i x = simde_mm256_loadu_si256(&txdataF_res_mapped[nl][sc]);
 
       // Rearrange precoding matrix weight to match complex multiplication and broadcast it to match SIMD size
       const simde__m256i w_c = simde_mm256_set1_epi32(c16toI32(c16conj(prec_weight))); // broadcast conjugate of w
       const simde__m256i w_s = simde_mm256_set1_epi32(c16toI32(c16swap(prec_weight))); // broadcast swapped real and img of w
 
       // Multiplication and shift
-      const simde__m256i reals = simde_mm256_srai_epi32(simde_mm256_madd_epi16(x, w_c), 15); // (int32_t) .r = (x.r * w.r - x.i * w.i) >> 15
+      const simde__m256i reals =
-      const simde__m256i imags = simde_mm256_slli_epi32(simde_mm256_madd_epi16(x, w_s), 1);  // (int32_t) .i = (x.r * w.i + x.i * w.r) << 1, since higher 16 bit of each 32 bit is taken by blend_epi16
+          simde_mm256_srai_epi32(simde_mm256_madd_epi16(x, w_c), 15); // (int32_t) .r = (x.r * w.r - x.i * w.i) >> 15
+      const simde__m256i imags = simde_mm256_slli_epi32(
+          simde_mm256_madd_epi16(x, w_s),
+          1); // (int32_t) .i = (x.r * w.i + x.i * w.r) << 1, since higher 16 bit of each 32 bit is taken by blend_epi16
 
       // Re-arrange to match c16_t format
       const simde__m256i produ = simde_mm256_blend_epi16(reals, imags, 0xAA);
@@ -544,14 +786,7 @@ void nr_layer_precoder_simd(const int n_layers,
 #ifdef DEBUG_DLSCH_PRECODING_PRINT_WITH_TRIVIAL // Get result with trivial solution, TODO: To be removed
     c16_t y_triv[4];
     for (int i = 0; i < 4; i++)
-        y_triv[i] = nr_layer_precoder_cm(n_layers,
+      y_triv[i] = nr_layer_precoder_cm(n_layers, symSz, txdataF_res_mapped, ant, pmi_pdu, sc + i);
-                                         NR_SYMBOLS_PER_SLOT,
-                                         symSz,
-                                         txdataF_res_mapped,
-                                         ant,
-                                         pmi_pdu,
-                                         symbol,
-                                         sc + i);
     memcpy(&txdataF_precoded[sc], y_triv, sizeof(y_triv));
 #endif
 
@@ -561,22 +796,26 @@ void nr_layer_precoder_simd(const int n_layers,
       prec_weight.r = pmi_pdu->weights[nl][ant].precoder_weight_Re;
       prec_weight.i = pmi_pdu->weights[nl][ant].precoder_weight_Im;
 
-      const simde__m128i x = simde_mm_loadu_si128(&txdataF_res_mapped[nl][symbol][sc]);
+      const simde__m128i x = simde_mm_loadu_si128(&txdataF_res_mapped[nl][sc]);
 
       // Rearrange precoding matrix weight to match complex multiplication and broadcast it to match SIMD size
       const simde__m128i w_c = simde_mm_set1_epi32(c16toI32(c16conj(prec_weight))); // broadcast conjugate of w
       const simde__m128i w_s = simde_mm_set1_epi32(c16toI32(c16swap(prec_weight))); // broadcast swapped real and img of w
 
       // Multiplication and shift
-      const simde__m128i reals = simde_mm_srai_epi32(simde_mm_madd_epi16(x, w_c), 15); // (int32_t) .r = (x.r * w.r - x.i * w.i) >> 15
+      const simde__m128i reals =
-      const simde__m128i imags = simde_mm_slli_epi32(simde_mm_madd_epi16(x, w_s), 1);  // (int32_t) .i = (x.r * w.i + x.i * w.r) << 1, since higher 16 bit of each 32 bit is taken by blend_epi16
+          simde_mm_srai_epi32(simde_mm_madd_epi16(x, w_c), 15); // (int32_t) .r = (x.r * w.r - x.i * w.i) >> 15
+      const simde__m128i imags = simde_mm_slli_epi32(
+          simde_mm_madd_epi16(x, w_s),
+          1); // (int32_t) .i = (x.r * w.i + x.i * w.r) << 1, since higher 16 bit of each 32 bit is taken by blend_epi16
 
       /* Re-arrange to match c16_t format
          bit index: 0            | 16              | 32           | 48              | 64           | 80              | 96 | 112
-        reals =   {R0.r[15..30] | R0.r[31] (0)*15 | R1.r[15..30] | R1.r[31] (0)*15 | R2.r[15..30] | R2.r[31] (0)*15 | R3.r[15..30] | R3.r[31] (0)*15}
+         reals =   {R0.r[15..30] | R0.r[31] (0)*15 | R1.r[15..30] | R1.r[31] (0)*15 | R2.r[15..30] | R2.r[31] (0)*15 | R3.r[15..30]
-        imags =   {0 R0.i[0..14]| R0.i[15..30]    | 0 R1.i[0..14]| R1.i[15..30]    | 0 R2.i[0..14]| R2.i[15..30]    | 0 R3.i[0..14]| R3.i[15..30]   }
+         | R3.r[31] (0)*15} imags =   {0 R0.i[0..14]| R0.i[15..30]    | 0 R1.i[0..14]| R1.i[15..30]    | 0 R2.i[0..14]| R2.i[15..30]
-        16b from  {reals        | imags           | reals        | imags           | reals        | imags           | reals        | imags          }
+         | 0 R3.i[0..14]| R3.i[15..30]   } 16b from  {reals        | imags           | reals        | imags | reals | imags | reals
-        produ =   {R0.r[15..30] | R0.i[15..30]    | R1.r[15..30] | R1.i[15..30]    | R2.r[15..30] | R2.i[15..30]    | R3.r[15..30] | R3.i[15..30]   }
+         | imags          } produ =   {R0.r[15..30] | R0.i[15..30]    | R1.r[15..30] | R1.i[15..30] | R2.r[15..30] | R2.i[15..30] |
+         R3.r[15..30] | R3.i[15..30]   }
       */
       const simde__m128i produ = simde_mm_blend_epi16(reals, imags, 0xAA);
 

openair1/PHY/MODULATION/nr_modulation.h

@@ -60,8 +60,7 @@ void nr_layer_mapping(int nbCodes,
                       uint8_t n_layers,
                       int layerSz,
                       uint32_t n_symbs,
-                      c16_t tx_layers[layerSz],
+                      c16_t tx_layers[][layerSz]);
-                      int l);
 
 /*! \brief Perform NR layer mapping. TS 38.211 V15.4.0 subclause 7.3.1.3
   @param[in] ulsch_ue, double Pointer to NR_UE_ULSCH_t struct
@@ -136,12 +135,10 @@ void apply_nr_rotation_RX(const NR_DL_FRAME_PARMS *frame_parms,
 c16_t nr_layer_precoder(int sz, c16_t datatx_F_precoding[][sz], const char *prec_matrix, uint8_t n_layers, int32_t re_offset);
 
 c16_t nr_layer_precoder_cm(int n_layers,
-                           int n_symbols,
                            int symSz,
-                           c16_t datatx_F_precoding[n_layers][n_symbols][symSz],
+                           c16_t datatx_F_precoding[n_layers][symSz],
                            int ap,
                            nfapi_nr_pm_pdu_t *pmi_pdu,
-                           int symbol,
                            int offset);
 
 /*! \brief Precoding with SIMDe, txdataF_precoded[] = prec_matrix[] * txdataF_res_mapped[]
@@ -151,12 +148,10 @@ c16_t nr_layer_precoder_cm(int n_layers,
   @param[out] txdataF_precoded   Precoded antenna data
 */
 void nr_layer_precoder_simd(const int n_layers,
-                           const int n_symbols,
                             const int symSz,
-                           const c16_t txdataF_res_mapped[n_layers][n_symbols][symSz],
+                            const c16_t txdataF_res_mapped[n_layers][symSz],
                             const int ant,
                             const nfapi_nr_pm_pdu_t *pmi_pdu,
-                           const int symbol,
                             const int sc_offset,
                             const int re_cnt,
                             c16_t *txdataF_precoded);

openair1/PHY/NR_TRANSPORT/nr_dci.c

@@ -249,9 +249,11 @@ void nr_generate_dci_top(processingData_L1tx_t *msgTx, int slot, int txdataF_off
 {
   PHY_VARS_gNB *gNB = msgTx->gNB;
   NR_DL_FRAME_PARMS *frame_parms = &gNB->frame_parms;
+  start_meas(&gNB->dci_generation_stats);
   for (int i = 0; i < msgTx->num_ul_pdcch; i++)
     nr_generate_dci(msgTx->gNB, &msgTx->ul_pdcch_pdu[i].pdcch_pdu.pdcch_pdu_rel15, txdataF_offset, frame_parms, slot);
   for (int i = 0; i < msgTx->num_dl_pdcch; i++)
     nr_generate_dci(msgTx->gNB, &msgTx->pdcch_pdu[i].pdcch_pdu_rel15, txdataF_offset, frame_parms, slot);
+  stop_meas(&gNB->dci_generation_stats);
 }
 

openair1/PHY/defs_gNB.h

@@ -520,6 +520,8 @@ typedef struct PHY_VARS_gNB_s {
   time_stats_t dlsch_interleaving_stats;
   time_stats_t dlsch_segmentation_stats;
 
+  time_stats_t dci_generation_stats;
+  time_stats_t phase_comp_stats;
   time_stats_t rx_pusch_stats;
   time_stats_t rx_pusch_init_stats;
   time_stats_t rx_pusch_symbol_processing_stats;

openair1/SCHED_NR/phy_procedures_nr_gNB.c

@@ -290,6 +290,7 @@ void phy_procedures_gNB_TX(processingData_L1tx_t *msgTx,
 
   //apply the OFDM symbol rotation here
   if (gNB->phase_comp) {
+    start_meas(&gNB->phase_comp_stats);
     for(int i = 0; i < gNB->common_vars.num_beams_period; ++i) {
       for (int aa = 0; aa < cfg->carrier_config.num_tx_ant.value; aa++) {
         apply_nr_rotation_TX(fp,
@@ -304,6 +305,7 @@ void phy_procedures_gNB_TX(processingData_L1tx_t *msgTx,
           T_INT(aa), T_BUFFER(&gNB->common_vars.txdataF[aa][txdataF_offset], fp->samples_per_slot_wCP*sizeof(int32_t)));
       }
     }
+    stop_meas(&gNB->phase_comp_stats);
   }
 
   VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_PHY_PROCEDURES_gNB_TX + gNB->CC_id, 0);

openair1/SIMULATION/NR_PHY/dlsim.c

@@ -965,10 +965,12 @@ printf("%d\n", slot);
     reset_meas(&gNB->dlsch_segmentation_stats);
     reset_meas(&gNB->dlsch_modulation_stats);
     reset_meas(&gNB->dlsch_encoding_stats);
+    reset_meas(&gNB->dci_generation_stats);
     reset_meas(&gNB->tinput);
     reset_meas(&gNB->tprep);
     reset_meas(&gNB->tparity);
     reset_meas(&gNB->toutput);
+    reset_meas(&gNB->phase_comp_stats);
 
     uint32_t errors_scrambling[16] = {0};
     int n_errors[16] = {0};
@@ -1264,6 +1266,7 @@ printf("%d\n", slot);
              UE->dl_harq_processes[0][slot].C,
              msgDataTx->dlsch[0][0].harq_process.pdsch_pdu.pdsch_pdu_rel15.TBSize[0] << 3);
       printDistribution(&gNB->phy_proc_tx,table_tx,"PHY proc tx");
+      printStatIndent2(&gNB->dci_generation_stats, "DCI encoding time");
       printStatIndent2(&gNB->dlsch_encoding_stats,"DLSCH encoding time");
       printStatIndent3(&gNB->dlsch_segmentation_stats,"DLSCH segmentation time");
       printStatIndent3(&gNB->tinput,"DLSCH LDPC input processing time");
@@ -1274,8 +1277,9 @@ printf("%d\n", slot);
       printStatIndent3(&gNB->dlsch_interleaving_stats,  "DLSCH Interleaving time");
       printStatIndent2(&gNB->dlsch_modulation_stats,"DLSCH modulation time");
       printStatIndent2(&gNB->dlsch_scrambling_stats, "DLSCH scrambling time");
-      printStatIndent2(&gNB->dlsch_resource_mapping_stats, "DLSCH Resource Mapping time");
+      printStatIndent2(&gNB->dlsch_precoding_stats,"DLSCH Mapping/Precoding time");
-      printStatIndent2(&gNB->dlsch_precoding_stats,"DLSCH Layer Precoding time");
+      if (gNB->phase_comp)
+        printStatIndent2(&gNB->phase_comp_stats, "Phase Compensation");
 
       printf("\nUE function statistics (per %d us slot)\n", 1000 >> *scc->ssbSubcarrierSpacing);
       for (int i = RX_PDSCH_STATS; i <= DLSCH_PROCEDURES_STATS; i++) {