AVX2 optimization activated for gamma computation in 16-bit turbo decoder (single-codeword)

cmake_targets/CMakeLists.txt

@@ -134,7 +134,7 @@ else (CMAKE_SYSTEM_PROCESSOR STREQUAL "armv7l")
     set(C_FLAGS_PROCESSOR "${C_FLAGS_PROCESSOR} -mavx2")
   endif()
   if (CPUINFO MATCHES "sse4_2")
-    set(C_FLAGS_PROCESSOR "${C_FLAGS_PROCESSOR} -mavx2 -msse4.2")
+    set(C_FLAGS_PROCESSOR "${C_FLAGS_PROCESSOR} -mavx2 -msse4.2 -fno-tree-vectorize")
   endif()
   if (CPUINFO MATCHES "sse4_1")
     set(C_FLAGS_PROCESSOR "${C_FLAGS_PROCESSOR} -msse4.1")

openair1/PHY/CODING/3gpplte_sse.c

@@ -223,6 +223,7 @@ char interleave_compact_byte(short * base_interleaver,unsigned char * input, uns
                                     0b00000001};
 #endif
  
+
 #ifndef __AVX2__
   if ((n&15) > 0)
     loop++;

openair1/PHY/CODING/3gpplte_turbo_decoder_avx2_16bit.c

@@ -186,12 +186,16 @@ void compute_alpha16avx2(llr_t* alpha,llr_t* beta,llr_t* m_11,llr_t* m_10,uint16
   __m256i new0,new1,new2,new3,new4,new5,new6,new7;
   __m256i alpha_max;
 
+  unsigned long long timein,timeout;
+
   l2 = L>>3;
   K1 = (frame_length>>3);
 #ifdef DEBUG_LOGMAP
   fprintf(fdavx2,"Compute alpha (avx2_16bit)\n");
   fprintf(fdavx2b,"Compute alpha (avx2_16bit)\n");
 #endif
+  timein = rdtsc_oai();
+
   for (l=K1;; l=l2,rerun_flag=1) {
     alpha128 = (__m256i *)alpha;
 
@@ -378,6 +382,9 @@ void compute_alpha16avx2(llr_t* alpha,llr_t* beta,llr_t* m_11,llr_t* m_10,uint16
     if (rerun_flag==1)
       break;
   }
+  timeout = rdtsc_oai();
+  printf("alpha: inner loop time %llu\n",timeout-timein);
+
 }
 
 
@@ -386,9 +393,10 @@ void compute_beta16avx2(llr_t* alpha,llr_t* beta,llr_t *m_11,llr_t* m_10,uint16_
 
   int k,rerun_flag=0;
 
-  __m256i m11_128,m10_128;
-  __m256i m_b0,m_b1,m_b2,m_b3,m_b4,m_b5,m_b6,m_b7;
-  __m256i new0,new1,new2,new3,new4,new5,new6,new7;
+  __m256i *m11p,*m10p;
+  register __m256i b0,b1,b2,b3,b4,b5,b6,b7;
+  register __m256i m_b0,m_b1,m_b2,m_b3,m_b4,m_b5,m_b6,m_b7;
+  register __m256i new0,new1,new2,new3,new4,new5,new6,new7;
 
   __m256i *beta128,*alpha128,*beta_ptr;
   __m256i beta_max;
@@ -398,6 +406,8 @@ void compute_beta16avx2(llr_t* alpha,llr_t* beta,llr_t *m_11,llr_t* m_10,uint16_
   llr_t beta0,beta1;
   llr_t beta0_cw2,beta1_cw2;
 
+  unsigned long long timein,timeout;
+
 #ifdef DEBUG_LOGMAP
   fprintf(fdavx2,"compute_beta (avx2_16bit), %p,%p,%p,%p,framelength %d,F %d\n",
       beta,m_11,m_10,alpha,frame_length,F);
@@ -590,56 +600,74 @@ void compute_beta16avx2(llr_t* alpha,llr_t* beta,llr_t *m_11,llr_t* m_10,uint16_
 #endif
     int loopval=((rerun_flag==0)?0:((frame_length-L)>>3));
 
+    printf("beta: rerun %d => loopval %d\n",rerun_flag,loopval);
+
+    timein = rdtsc_oai();
+
+    m11p = (frame_length>>3)-1+(__m256i*)m_11;
+    m10p = (frame_length>>3)-1+(__m256i*)m_10;
+
     for (k=(frame_length>>3)-1; k>=loopval; k--) {
 
-      m11_128=((__m256i*)m_11)[k];
-      m10_128=((__m256i*)m_10)[k];
-
-      m_b0 = _mm256_adds_epi16(beta_ptr[4],m11_128);  //m11
-      m_b1 = _mm256_subs_epi16(beta_ptr[4],m11_128);  //m00
-      m_b2 = _mm256_subs_epi16(beta_ptr[5],m10_128);  //m01
-      m_b3 = _mm256_adds_epi16(beta_ptr[5],m10_128);  //m10
-      m_b4 = _mm256_adds_epi16(beta_ptr[6],m10_128);  //m10
-      m_b5 = _mm256_subs_epi16(beta_ptr[6],m10_128);  //m01
-      m_b6 = _mm256_subs_epi16(beta_ptr[7],m11_128);  //m00
-      m_b7 = _mm256_adds_epi16(beta_ptr[7],m11_128);  //m11
-
-      new0 = _mm256_subs_epi16(beta_ptr[0],m11_128);  //m00
-      new1 = _mm256_adds_epi16(beta_ptr[0],m11_128);  //m11
-      new2 = _mm256_adds_epi16(beta_ptr[1],m10_128);  //m10
-      new3 = _mm256_subs_epi16(beta_ptr[1],m10_128);  //m01
-      new4 = _mm256_subs_epi16(beta_ptr[2],m10_128);  //m01
-      new5 = _mm256_adds_epi16(beta_ptr[2],m10_128);  //m10
-      new6 = _mm256_adds_epi16(beta_ptr[3],m11_128);  //m11
-      new7 = _mm256_subs_epi16(beta_ptr[3],m11_128);  //m00
+      
+      b4 = _mm256_load_si256(&beta_ptr[4]);
+      b5 = _mm256_load_si256(&beta_ptr[5]);
+      b6 = _mm256_load_si256(&beta_ptr[6]);
+      b7 = _mm256_load_si256(&beta_ptr[7]);
+
+      m_b0 = _mm256_adds_epi16(b4,*m11p);  //m11
+      m_b1 = _mm256_subs_epi16(b4,*m11p);  //m00
+      m_b2 = _mm256_subs_epi16(b5,*m10p);  //m01
+      m_b3 = _mm256_adds_epi16(b5,*m10p);  //m10
+      m_b4 = _mm256_adds_epi16(b6,*m10p);  //m10
+      m_b5 = _mm256_subs_epi16(b6,*m10p);  //m01
+      m_b6 = _mm256_subs_epi16(b7,*m11p);  //m00
+      m_b7 = _mm256_adds_epi16(b7,*m11p);  //m11
+
+      b0 = _mm256_load_si256(&beta_ptr[0]);
+      b1 = _mm256_load_si256(&beta_ptr[1]);
+      b2 = _mm256_load_si256(&beta_ptr[2]);
+      b3 = _mm256_load_si256(&beta_ptr[3]);
+
+      new0 = _mm256_subs_epi16(b0,*m11p);  //m00
+      new1 = _mm256_adds_epi16(b0,*m11p);  //m11
+      new2 = _mm256_adds_epi16(b1,*m10p);  //m10
+      new3 = _mm256_subs_epi16(b1,*m10p);  //m01
+      new4 = _mm256_subs_epi16(b2,*m10p);  //m01
+      new5 = _mm256_adds_epi16(b2,*m10p);  //m10
+      new6 = _mm256_adds_epi16(b3,*m11p);  //m11
+      new7 = _mm256_subs_epi16(b3,*m11p);  //m00
+
+
+      b0 = _mm256_max_epi16(m_b0,new0);
+      b1 = _mm256_max_epi16(m_b1,new1);
+      b2 = _mm256_max_epi16(m_b2,new2);
+      b3 = _mm256_max_epi16(m_b3,new3);
+      b4 = _mm256_max_epi16(m_b4,new4);
+      b5 = _mm256_max_epi16(m_b5,new5);
+      b6 = _mm256_max_epi16(m_b6,new6);
+      b7 = _mm256_max_epi16(m_b7,new7);
+
+      beta_max = _mm256_max_epi16(b0,b1);
+      beta_max = _mm256_max_epi16(beta_max   ,b2);
+      beta_max = _mm256_max_epi16(beta_max   ,b3);
+      beta_max = _mm256_max_epi16(beta_max   ,b4);
+      beta_max = _mm256_max_epi16(beta_max   ,b5);
+      beta_max = _mm256_max_epi16(beta_max   ,b6);
+      beta_max = _mm256_max_epi16(beta_max   ,b7);
 
       beta_ptr-=8;
-
-      beta_ptr[0] = _mm256_max_epi16(m_b0,new0);
-      beta_ptr[1] = _mm256_max_epi16(m_b1,new1);
-      beta_ptr[2] = _mm256_max_epi16(m_b2,new2);
-      beta_ptr[3] = _mm256_max_epi16(m_b3,new3);
-      beta_ptr[4] = _mm256_max_epi16(m_b4,new4);
-      beta_ptr[5] = _mm256_max_epi16(m_b5,new5);
-      beta_ptr[6] = _mm256_max_epi16(m_b6,new6);
-      beta_ptr[7] = _mm256_max_epi16(m_b7,new7);
-
-      beta_max = _mm256_max_epi16(beta_ptr[0],beta_ptr[1]);
-      beta_max = _mm256_max_epi16(beta_max   ,beta_ptr[2]);
-      beta_max = _mm256_max_epi16(beta_max   ,beta_ptr[3]);
-      beta_max = _mm256_max_epi16(beta_max   ,beta_ptr[4]);
-      beta_max = _mm256_max_epi16(beta_max   ,beta_ptr[5]);
-      beta_max = _mm256_max_epi16(beta_max   ,beta_ptr[6]);
-      beta_max = _mm256_max_epi16(beta_max   ,beta_ptr[7]);
-
-      beta_ptr[0] = _mm256_subs_epi16(beta_ptr[0],beta_max);
-      beta_ptr[1] = _mm256_subs_epi16(beta_ptr[1],beta_max);
-      beta_ptr[2] = _mm256_subs_epi16(beta_ptr[2],beta_max);
-      beta_ptr[3] = _mm256_subs_epi16(beta_ptr[3],beta_max);
-      beta_ptr[4] = _mm256_subs_epi16(beta_ptr[4],beta_max);
-      beta_ptr[5] = _mm256_subs_epi16(beta_ptr[5],beta_max);
-      beta_ptr[6] = _mm256_subs_epi16(beta_ptr[6],beta_max);
-      beta_ptr[7] = _mm256_subs_epi16(beta_ptr[7],beta_max);
+      m11p--;
+      m10p--;
+
+      beta_ptr[0] = _mm256_subs_epi16(b0,beta_max);
+      beta_ptr[1] = _mm256_subs_epi16(b1,beta_max);
+      beta_ptr[2] = _mm256_subs_epi16(b2,beta_max);
+      beta_ptr[3] = _mm256_subs_epi16(b3,beta_max);
+      beta_ptr[4] = _mm256_subs_epi16(b4,beta_max);
+      beta_ptr[5] = _mm256_subs_epi16(b5,beta_max);
+      beta_ptr[6] = _mm256_subs_epi16(b6,beta_max);
+      beta_ptr[7] = _mm256_subs_epi16(b7,beta_max);
 
 #ifdef DEBUG_LOGMAP
       fprintf(fdavx2,"Loop index %d, mb\n",k);
@@ -658,6 +686,8 @@ void compute_beta16avx2(llr_t* alpha,llr_t* beta,llr_t *m_11,llr_t* m_10,uint16_
 
 #endif
     }
+    timeout = rdtsc_oai();
+    printf("beta: inner loop time %llu\n",timeout-timein);
 
     if (rerun_flag==1)
       break;
@@ -968,7 +998,7 @@ unsigned char phy_threegpplte_turbo_decoder16avx2(int16_t *y,
   yp2 = yparity2;
 
 
-
+#if 0
   for (i=0; i<n; i+=8) {
     pi2_p = &pi2tab16avx2[iind][i];
 
@@ -1084,9 +1114,23 @@ unsigned char phy_threegpplte_turbo_decoder16avx2(int16_t *y,
     yp128_cw2+=3;
 
   }
-
   yp=(llr_t*)yp128;
   yp_cw2=(llr_t*)yp128_cw2;
+#else
+  
+  pi2_p    = &pi2tab16avx2[iind][0];
+  for (i=0,j=0; i<n; i++) {
+    s[*pi2_p]     = y[j];
+    s[*pi2_p+8]   = y2[j++];
+    yp1[*pi2_p]   = y[j];
+    yp1[*pi2_p+8] = y2[j++];
+    yp2[*pi2_p]   = y[j];
+    yp2[(*pi2_p++)+8] = y2[j++];
+  }    
+  yp=(llr_t*)&y[j];
+  yp_cw2=(llr_t*)&y2[j];
+#endif
+
 
   // Termination
   for (i=0; i<3; i++) {

openair1/PHY/LTE_REFSIG/lte_gold.c

@@ -61,21 +61,18 @@ void lte_gold(LTE_DL_FRAME_PARMS *frame_parms,uint32_t lte_gold_table[20][2][14]
 
       x2 = Ncp +
            (Nid_cell<<1) +
-           (((1+(Nid_cell<<1))*(1 + (((frame_parms->Ncp==0)?4:3)*l) + (7*(1+ns))))<<10); //cinit
+	(((1+(Nid_cell<<1))*(1 + (((frame_parms->Ncp==0)?4:3)*l) + (7*(1+ns))))<<10); //cinit
       //x2 = frame_parms->Ncp + (Nid_cell<<1) + (1+(Nid_cell<<1))*(1 + (3*l) + (7*(1+ns))); //cinit
       //n = 0
-      //      printf("cinit (ns %d, l %d) => %d\n",ns,l,x2);
       x1 = 1+ (1<<31);
       x2=x2 ^ ((x2 ^ (x2>>1) ^ (x2>>2) ^ (x2>>3))<<31);
 
       // skip first 50 double words (1600 bits)
-      //printf("n=0 : x1 %x, x2 %x\n",x1,x2);
       for (n=1; n<50; n++) {
         x1 = (x1>>1) ^ (x1>>4);
         x1 = x1 ^ (x1<<31) ^ (x1<<28);
         x2 = (x2>>1) ^ (x2>>2) ^ (x2>>3) ^ (x2>>4);
         x2 = x2 ^ (x2<<31) ^ (x2<<30) ^ (x2<<29) ^ (x2<<28);
-        //  printf("x1 : %x, x2 : %x\n",x1,x2);
       }
 
       for (n=0; n<14; n++) {
@@ -84,7 +81,6 @@ void lte_gold(LTE_DL_FRAME_PARMS *frame_parms,uint32_t lte_gold_table[20][2][14]
         x2 = (x2>>1) ^ (x2>>2) ^ (x2>>3) ^ (x2>>4);
         x2 = x2 ^ (x2<<31) ^ (x2<<30) ^ (x2<<29) ^ (x2<<28);
         lte_gold_table[ns][l][n] = x1^x2;
-        //  printf("n=%d : c %x\n",n,x1^x2);
       }
 
     }

openair1/PHY/LTE_TRANSPORT/dlsch_decoding.c

@@ -446,7 +446,8 @@ uint32_t  dlsch_decoding(PHY_VARS_UE *phy_vars_ue,
     printf("\n");
     */
 
-#ifndef __AVX2__
+    //#ifndef __AVX2__
+#if 1
     if (err_flag == 0) {
 
       start_meas(dlsch_turbo_decoding_stats);

openair1/PHY/LTE_TRANSPORT/dlsch_demodulation.c

@@ -1898,17 +1898,17 @@ void dlsch_channel_compensation_TM3(LTE_DL_FRAME_PARMS *frame_parms,
 
   for (aarx=0; aarx<frame_parms->nb_antennas_rx; aarx++) {
 
-    dl_ch0_128          = (__m128i *)&dl_ch_estimates_ext[aarx][symbol*frame_parms->N_RB_DL*12];
-    dl_ch1_128          = (__m128i *)&dl_ch_estimates_ext[2+aarx][symbol*frame_parms->N_RB_DL*12];
+    dl_ch0_128          = (__m128i *)&dl_ch_estimates_ext[aarx][symbol*frame_parms->N_RB_DL*12];    // hr,0
+    dl_ch1_128          = (__m128i *)&dl_ch_estimates_ext[2+aarx][symbol*frame_parms->N_RB_DL*12];  // hr,1
 
 
     dl_ch_mag0_128      = (__m128i *)&dl_ch_mag0[aarx][symbol*frame_parms->N_RB_DL*12];
     dl_ch_mag0_128b     = (__m128i *)&dl_ch_magb0[aarx][symbol*frame_parms->N_RB_DL*12];
     dl_ch_mag1_128      = (__m128i *)&dl_ch_mag1[aarx][symbol*frame_parms->N_RB_DL*12];
     dl_ch_mag1_128b     = (__m128i *)&dl_ch_magb1[aarx][symbol*frame_parms->N_RB_DL*12];
-    rxdataF128          = (__m128i *)&rxdataF_ext[aarx][symbol*frame_parms->N_RB_DL*12];
-    rxdataF_comp0_128   = (__m128i *)&rxdataF_comp0[aarx][symbol*frame_parms->N_RB_DL*12];
-    rxdataF_comp1_128   = (__m128i *)&rxdataF_comp1[aarx][symbol*frame_parms->N_RB_DL*12];
+    rxdataF128          = (__m128i *)&rxdataF_ext[aarx][symbol*frame_parms->N_RB_DL*12];           // yr
+    rxdataF_comp0_128   = (__m128i *)&rxdataF_comp0[aarx][symbol*frame_parms->N_RB_DL*12];         // yr,0 = yr * conj(hr,0)
+    rxdataF_comp1_128   = (__m128i *)&rxdataF_comp1[aarx][symbol*frame_parms->N_RB_DL*12];         // yr,1 = yr * conj(hr,1)
 
 
     for (rb=0; rb<nb_rb; rb++) {