added AVX2 optimization for 64QAM LLR computation in NR ULSCH

openair1/PHY/NR_TRANSPORT/nr_ulsch_llr_computation.c

@@ -239,10 +239,22 @@ void nr_ulsch_64qam_llr(int32_t *rxdataF_comp,
                         uint8_t  symbol)
 {
 
+#ifdef __AVX2__
+  int off = ((nb_rb&1) == 1)? 4:0;
+#else
+  int off = 0;
+#endif
+
 #if defined(__x86_64__) || defined(__i386__)
+#ifdef __AVX2__
+  __m256i *rxF = (__m256i*)rxdataF_comp;
+  __m256i *ch_mag,*ch_magb;
+  register __m256i xmm0,xmm1,xmm2;
+#else
   __m128i *rxF = (__m128i*)rxdataF_comp;
   __m128i *ch_mag,*ch_magb;
   register __m128i xmm0,xmm1,xmm2;
+#endif
 #elif defined(__arm__)
   int16x8_t *rxF = (int16x8_t*)&rxdataF_comp;
   int16x8_t *ch_mag,*ch_magb; // [hna] This should be uncommented once channel estimation is implemented
@@ -250,28 +262,44 @@ void nr_ulsch_64qam_llr(int32_t *rxdataF_comp,
 #endif
 
   int i;
-  unsigned char len_mod4;
 
 #if defined(__x86_64__) || defined(__i386__)
+#ifdef __AVX2__
+  ch_mag = (__m256i*)&ul_ch_mag[0][(symbol*(off+(nb_rb*12)))];
+  ch_magb = (__m256i*)&ul_ch_magb[0][(symbol*(off+(nb_rb*12)))];
+#else
   ch_mag = (__m128i*)&ul_ch_mag[0][(symbol*nb_rb*12)];
   ch_magb = (__m128i*)&ul_ch_magb[0][(symbol*nb_rb*12)];
+#endif
 #elif defined(__arm__)
   ch_mag = (int16x8_t*)&ul_ch_mag[0][(symbol*nb_rb*12)];
   ch_magb = (int16x8_t*)&ul_ch_magb[0][(symbol*nb_rb*12)];
 #endif
 
-  len_mod4 = nb_re&3;
+#ifdef __AVX2__
+  int len_mod8 = nb_re&7;
+  nb_re    = nb_re>>3;  // length in quad words (4 REs)
+  nb_re   += ((len_mod8 == 0) ? 0 : 1);
+#else
+  int len_mod4 = nb_re&3;
   nb_re    = nb_re>>2;  // length in quad words (4 REs)
   nb_re   += ((len_mod4 == 0) ? 0 : 1);
-
+#endif
 
   for (i=0; i<nb_re; i++) {
     xmm0 = rxF[i];
 #if defined(__x86_64__) || defined(__i386__)
+#ifdef __AVX2__
+    xmm1 = _mm256_abs_epi16(xmm0);
+    xmm1 = _mm256_subs_epi16(ch_mag[i],xmm1);
+    xmm2 = _mm256_abs_epi16(xmm1);
+    xmm2 = _mm256_subs_epi16(ch_magb[i],xmm2);
+#else
     xmm1 = _mm_abs_epi16(xmm0);
     xmm1 = _mm_subs_epi16(ch_mag[i],xmm1);
     xmm2 = _mm_abs_epi16(xmm1);
     xmm2 = _mm_subs_epi16(ch_magb[i],xmm2);
+#endif
 #elif defined(__arm__)
     xmm1 = vabsq_s16(xmm0);
     xmm1 = vsubq_s16(ch_mag[i],xmm1);
@@ -283,12 +311,21 @@ void nr_ulsch_64qam_llr(int32_t *rxdataF_comp,
     // 1st RE
     // ---------------------------------------
 #if defined(__x86_64__) || defined(__i386__)
+#ifdef __AVX2__
+    ulsch_llr[0] = _mm256_extract_epi16(xmm0,0);
+    ulsch_llr[1] = _mm256_extract_epi16(xmm0,1);
+    ulsch_llr[2] = _mm256_extract_epi16(xmm1,0);
+    ulsch_llr[3] = _mm256_extract_epi16(xmm1,1);
+    ulsch_llr[4] = _mm256_extract_epi16(xmm2,0);
+    ulsch_llr[5] = _mm256_extract_epi16(xmm2,1);
+#else
     ulsch_llr[0] = _mm_extract_epi16(xmm0,0);
     ulsch_llr[1] = _mm_extract_epi16(xmm0,1);
     ulsch_llr[2] = _mm_extract_epi16(xmm1,0);
     ulsch_llr[3] = _mm_extract_epi16(xmm1,1);
     ulsch_llr[4] = _mm_extract_epi16(xmm2,0);
     ulsch_llr[5] = _mm_extract_epi16(xmm2,1);
+#endif
 #elif defined(__arm__)
     ulsch_llr[0] = vgetq_lane_s16(xmm0,0);
     ulsch_llr[1] = vgetq_lane_s16(xmm0,1);
@@ -305,12 +342,21 @@ void nr_ulsch_64qam_llr(int32_t *rxdataF_comp,
     // 2nd RE
     // ---------------------------------------
 #if defined(__x86_64__) || defined(__i386__)
+#ifdef __AVX2__
+    ulsch_llr[0] = _mm256_extract_epi16(xmm0,2);
+    ulsch_llr[1] = _mm256_extract_epi16(xmm0,3);
+    ulsch_llr[2] = _mm256_extract_epi16(xmm1,2);
+    ulsch_llr[3] = _mm256_extract_epi16(xmm1,3);
+    ulsch_llr[4] = _mm256_extract_epi16(xmm2,2);
+    ulsch_llr[5] = _mm256_extract_epi16(xmm2,3);
+#else
     ulsch_llr[0] = _mm_extract_epi16(xmm0,2);
     ulsch_llr[1] = _mm_extract_epi16(xmm0,3);
     ulsch_llr[2] = _mm_extract_epi16(xmm1,2);
     ulsch_llr[3] = _mm_extract_epi16(xmm1,3);
     ulsch_llr[4] = _mm_extract_epi16(xmm2,2);
     ulsch_llr[5] = _mm_extract_epi16(xmm2,3);
+#endif
 #elif defined(__arm__)
     ulsch_llr[2] = vgetq_lane_s16(xmm0,2);
     ulsch_llr[3] = vgetq_lane_s16(xmm0,3);
@@ -327,12 +373,21 @@ void nr_ulsch_64qam_llr(int32_t *rxdataF_comp,
     // 3rd RE
     // ---------------------------------------
 #if defined(__x86_64__) || defined(__i386__)
+#ifdef __AVX2__
+    ulsch_llr[0] = _mm256_extract_epi16(xmm0,4);
+    ulsch_llr[1] = _mm256_extract_epi16(xmm0,5);
+    ulsch_llr[2] = _mm256_extract_epi16(xmm1,4);
+    ulsch_llr[3] = _mm256_extract_epi16(xmm1,5);
+    ulsch_llr[4] = _mm256_extract_epi16(xmm2,4);
+    ulsch_llr[5] = _mm256_extract_epi16(xmm2,5);
+#else
     ulsch_llr[0] = _mm_extract_epi16(xmm0,4);
     ulsch_llr[1] = _mm_extract_epi16(xmm0,5);
     ulsch_llr[2] = _mm_extract_epi16(xmm1,4);
     ulsch_llr[3] = _mm_extract_epi16(xmm1,5);
     ulsch_llr[4] = _mm_extract_epi16(xmm2,4);
     ulsch_llr[5] = _mm_extract_epi16(xmm2,5);
+#endif
 #elif defined(__arm__)
     ulsch_llr[0] = vgetq_lane_s16(xmm0,4);
     ulsch_llr[1] = vgetq_lane_s16(xmm0,5);
@@ -349,12 +404,21 @@ void nr_ulsch_64qam_llr(int32_t *rxdataF_comp,
     // 4th RE
     // ---------------------------------------
 #if defined(__x86_64__) || defined(__i386__)
+#ifdef __AVX2__
+    ulsch_llr[0] = _mm256_extract_epi16(xmm0,6);
+    ulsch_llr[1] = _mm256_extract_epi16(xmm0,7);
+    ulsch_llr[2] = _mm256_extract_epi16(xmm1,6);
+    ulsch_llr[3] = _mm256_extract_epi16(xmm1,7);
+    ulsch_llr[4] = _mm256_extract_epi16(xmm2,6);
+    ulsch_llr[5] = _mm256_extract_epi16(xmm2,7);
+#else
     ulsch_llr[0] = _mm_extract_epi16(xmm0,6);
     ulsch_llr[1] = _mm_extract_epi16(xmm0,7);
     ulsch_llr[2] = _mm_extract_epi16(xmm1,6);
     ulsch_llr[3] = _mm_extract_epi16(xmm1,7);
     ulsch_llr[4] = _mm_extract_epi16(xmm2,6);
     ulsch_llr[5] = _mm_extract_epi16(xmm2,7);
+#endif
 #elif defined(__arm__)
     ulsch_llr[0] = vgetq_lane_s16(xmm0,6);
     ulsch_llr[1] = vgetq_lane_s16(xmm0,7);
@@ -366,6 +430,37 @@ void nr_ulsch_64qam_llr(int32_t *rxdataF_comp,
     // ---------------------------------------
 
     ulsch_llr+=6;
+#ifdef __AVX2__
+    ulsch_llr[0] = _mm256_extract_epi16(xmm0,8);
+    ulsch_llr[1] = _mm256_extract_epi16(xmm0,9);
+    ulsch_llr[2] = _mm256_extract_epi16(xmm1,8);
+    ulsch_llr[3] = _mm256_extract_epi16(xmm1,9);
+    ulsch_llr[4] = _mm256_extract_epi16(xmm2,8);
+    ulsch_llr[5] = _mm256_extract_epi16(xmm2,9);
+
+    ulsch_llr[6] = _mm256_extract_epi16(xmm0,10);
+    ulsch_llr[7] = _mm256_extract_epi16(xmm0,11);
+    ulsch_llr[8] = _mm256_extract_epi16(xmm1,10);
+    ulsch_llr[9] = _mm256_extract_epi16(xmm1,11);
+    ulsch_llr[10] = _mm256_extract_epi16(xmm2,10);
+    ulsch_llr[11] = _mm256_extract_epi16(xmm2,11);
+
+    ulsch_llr[12] = _mm256_extract_epi16(xmm0,12);
+    ulsch_llr[13] = _mm256_extract_epi16(xmm0,13);
+    ulsch_llr[14] = _mm256_extract_epi16(xmm1,12);
+    ulsch_llr[15] = _mm256_extract_epi16(xmm1,13);
+    ulsch_llr[16] = _mm256_extract_epi16(xmm2,12);
+    ulsch_llr[17] = _mm256_extract_epi16(xmm2,13);
+
+    ulsch_llr[18] = _mm256_extract_epi16(xmm0,14);
+    ulsch_llr[19] = _mm256_extract_epi16(xmm0,15);
+    ulsch_llr[20] = _mm256_extract_epi16(xmm1,14);
+    ulsch_llr[21] = _mm256_extract_epi16(xmm1,15);
+    ulsch_llr[22] = _mm256_extract_epi16(xmm2,14);
+    ulsch_llr[23] = _mm256_extract_epi16(xmm2,15);
+
+    ulsch_llr+=24;
+#endif
   }
 
 #if defined(__x86_64__) || defined(__i386__)