Running C-RAN emulations

openair1/SIMULATION/RF/dac.c

@@ -384,7 +384,7 @@ float dac_fixed_gain_prach_SSE_float(float *s_re[2],
       _mm_storeu_ps(&s_re[aa][4*i],input_re128);
       _mm_storeu_ps(&s_im[aa][4*i],input_im128);
 
-      if (2*i+input_offset==12*2*ofdm_symbol_size)
+      if (4*i+input_offset==12*2*ofdm_symbol_size)
  	i=0;
     }
   }
@@ -449,7 +449,7 @@ float dac_fixed_gain_prach_AVX_float(float *s_re[2],
       _mm256_storeu_ps(&s_re[aa][8*i],input_re256);
       _mm256_storeu_ps(&s_im[aa][8*i],input_im256);
 
-      if (2*i+input_offset==12*2*ofdm_symbol_size)
+      if (8*i+input_offset==12*2*ofdm_symbol_size)
  	i=0;
     }
   }

openair1/SIMULATION/RF/rf.c

@@ -432,9 +432,9 @@ clock_t start=clock();*/
 		      //gauss_1_128_sqrt_NOW = _mm_set_ps(gaussdouble(0.0,1.0),gaussdouble(0.0,1.0),gaussdouble(0.0,1.0),gaussdouble(0.0,1.0));
 		      //gauss_0_128_sqrt_NOW = _mm_set_ps(ziggurat(0.0,1.0),ziggurat(0.0,1.0),ziggurat(0.0,1.0),ziggurat(0.0,1.0));
 		      //gauss_1_128_sqrt_NOW = _mm_set_ps(ziggurat(0.0,1.0),ziggurat(0.0,1.0),ziggurat(0.0,1.0),ziggurat(0.0,1.0));
-		      //boxmuller_SSE_float(&gauss_0_128_sqrt_NOW, &gauss_1_128_sqrt_NOW);
-		      gauss_0_128_sqrt_NOW = ziggurat_SSE_float();
-		      gauss_1_128_sqrt_NOW = ziggurat_SSE_float();
+		      boxmuller_SSE_float(&gauss_0_128_sqrt_NOW, &gauss_1_128_sqrt_NOW);
+		      //gauss_0_128_sqrt_NOW = ziggurat_SSE_float();
+		      //gauss_1_128_sqrt_NOW = ziggurat_SSE_float();
 		      //stop_meas(&desc->ziggurat);
 		      gauss_0_128_sqrt_NOW = _mm_mul_ps(gauss_0_128_sqrt_NOW,_mm_set1_ps(sqrt_NOW));
 		      gauss_1_128_sqrt_NOW = _mm_mul_ps(gauss_1_128_sqrt_NOW,_mm_set1_ps(sqrt_NOW));

openair1/SIMULATION/TOOLS/abstraction.c

@@ -232,7 +232,7 @@ int init_freq_channel_SSE_float(channel_desc_t *desc,uint16_t nb_rb,int16_t n_sa
       //printf("f %d,l %d (cos,sin) (%e,%e):\n",4*f,l,cos_lut_f[(n_samples>>1)][l],sin_lut_f[(n_samples>>1)][l]);
   }
 
-  for (f=1; f<=(n_samples>>3)+1; f++) {
+  for (f=1; f<=(n_samples>>3); f++) {
     //count++;
     //freq=delta_f*(double)f*1e-6;// due to the fact that delays is in mus
     for (l=0; l<(int)desc->nb_taps; l++) {
@@ -314,7 +314,7 @@ int init_freq_channel_AVX_float(channel_desc_t *desc,uint16_t nb_rb,int16_t n_sa
         delay = desc->delays[l];
       else
         delay = desc->delays[l]+NB_SAMPLES_CHANNEL_OFFSET/desc->sampling_rate;
-        sincos256_ps(_mm256_set_ps(twopi*(4*f)*delay,twopi*(4*f-1)*delay,twopi*(4*f-2)*delay,twopi*(4*f-3)*delay,twopi*(4*f-4)*delay,twopi*(4*f-5)*delay,twopi*(4*f-6)*delay,twopi*(4*f-7)*delay), &sin_lut256, &cos_lut256);
+        sincos256_ps(_mm256_set_ps(twopi*(8*f)*delay,twopi*(8*f-1)*delay,twopi*(8*f-2)*delay,twopi*(8*f-3)*delay,twopi*(8*f-4)*delay,twopi*(8*f-5)*delay,twopi*(8*f-6)*delay,twopi*(8*f-7)*delay), &sin_lut256, &cos_lut256);
       //cos_lut256=_mm256_set_ps(cos(twopi*(4*f)*delay),cos(twopi*(4*f-1)*delay),cos(twopi*(4*f-2)*delay),cos(twopi*(4*f-3)*delay),cos(twopi*(4*f-4)*delay),cos(twopi*(4*f-5)*delay),cos(twopi*(4*f-6)*delay),cos(twopi*(4*f-7)*delay));
       //sin_lut256=_mm256_set_ps(sin(twopi*(4*f)*delay),sin(twopi*(4*f-1)*delay),sin(twopi*(4*f-2)*delay),sin(twopi*(4*f-3)*delay),sin(twopi*(4*f-4)*delay),sin(twopi*(4*f-5)*delay),sin(twopi*(4*f-6)*delay),sin(twopi*(4*f-7)*delay));
       _mm256_storeu_ps(&cos_lut_f[l][8*f-7+(n_samples>>1)],cos_lut256);

targets/SIMU/USER/channel_sim.c

@@ -1520,14 +1520,34 @@ void do_UL_sig_freq_prach(channel_desc_t *UE2eNB[NUMBER_OF_UE_MAX][NUMBER_OF_eNB
 	if (UE2eNB[UE_id][eNB_id][CC_id]->first_run == 1)
 	  UE2eNB[UE_id][eNB_id][CC_id]->first_run = 0;
 	
-	
+	__m256 r_re0_f_prach_256,r_im0_f_prach_256,r_re_UL_f_prach_256,r_im_UL_f_prach_256;
 	pthread_mutex_lock(&UE_output_mutex[eNB_id]);
+#ifdef    SSE_float
+	for (aa=0; aa<nb_antennas_rx; aa++) {
+		for (i=0; i<(frame_parms->ofdm_symbol_size*frame_parms->symbols_per_tti)>>3; i++) {
+		    //r_re_UL_f_prach[eNB_id][aa][i]+=r_re0_f_prach[aa][i];
+		    //r_im_UL_f_prach[eNB_id][aa][i]+=r_im0_f_prach[aa][i];
+
+		    r_re0_f_prach_256 = _mm256_loadu_ps(&r_re0_f_prach[aa][8*i]);
+		    r_im0_f_prach_256 = _mm256_loadu_ps(&r_im0_f_prach[aa][8*i]);
+		    r_re_UL_f_prach_256 = _mm256_loadu_ps(&r_re_UL_f_prach[eNB_id][aa][8*i]);
+		    r_im_UL_f_prach_256 = _mm256_loadu_ps(&r_im_UL_f_prach[eNB_id][aa][8*i]);
+
+		    r_re_UL_f_prach_256 = _mm256_add_ps(r_re_UL_f_prach_256,r_re0_f_prach_256);
+		    r_im_UL_f_prach_256 = _mm256_add_ps(r_im_UL_f_prach_256,r_im0_f_prach_256);
+
+		    _mm256_storeu_ps(&r_re_UL_f_prach[eNB_id][aa][8*i],r_re_UL_f_prach_256);
+		    _mm256_storeu_ps(&r_im_UL_f_prach[eNB_id][aa][8*i],r_im_UL_f_prach_256);
+		}
+	}
+#else
 	for (aa=0; aa<nb_antennas_rx; aa++) {
 		for (i=0; i<frame_parms->ofdm_symbol_size*frame_parms->symbols_per_tti; i++) {
 		    r_re_UL_f_prach[eNB_id][aa][i]+=r_re0_f_prach[aa][i];
 		    r_im_UL_f_prach[eNB_id][aa][i]+=r_im0_f_prach[aa][i];
 		}
 	}
+#endif
 	pthread_mutex_unlock(&UE_output_mutex[eNB_id]);
       }
     } //UE_id