Added zero-copy option to dft module

srslte/examples/pdsch_enodeb.c

@@ -511,9 +511,9 @@ int main(int argc, char **argv) {
   sigaddset(&sigset, SIGINT);
   sigprocmask(SIG_UNBLOCK, &sigset, NULL);
   signal(SIGINT, sig_int_handler);
-  cuhd_set_master_clock_rate(uhd, 30.72e6);        
 
   if (!output_file_name) {
+    cuhd_set_master_clock_rate(uhd, 30.72e6);        
     printf("Set TX rate: %.2f MHz\n",
         cuhd_set_tx_srate(uhd, srslte_sampling_freq_hz(cell.nof_prb)) / 1000000);
     printf("Set TX gain: %.1f dB\n", cuhd_set_tx_gain(uhd, uhd_gain));

srslte/include/srslte/dft/dft.h

@@ -109,6 +109,10 @@ SRSLTE_API void srslte_dft_run(srslte_dft_plan_t *plan,
                                void *in, 
                                void *out);
 
+SRSLTE_API void srslte_dft_run_c_zerocopy(srslte_dft_plan_t *plan, 
+                                          cf_t *in, 
+                                          cf_t *out); 
+
 SRSLTE_API void srslte_dft_run_c(srslte_dft_plan_t *plan, 
                                  cf_t *in, 
                                  cf_t *out);

srslte/lib/common/src/phy_common.c

@@ -36,7 +36,7 @@
 #include "srslte/common/sequence.h"
 
 
-#define USE_REDUCED_SAMPLING_RATES
+//#define USE_REDUCED_SAMPLING_RATES
 
 
 

srslte/lib/dft/src/dft.c

@@ -30,6 +30,7 @@
 #include <complex.h>
 #include <fftw3.h>
 #include <string.h>
+#include <volk/volk.h>
 
 #include "srslte/dft/dft.h"
 #include "srslte/utils/vector.h"
@@ -136,6 +137,10 @@ void srslte_dft_run(srslte_dft_plan_t *plan, void *in, void *out) {
   }
 }
 
+void srslte_dft_run_c_zerocopy(srslte_dft_plan_t *plan, cf_t *in, cf_t *out) {
+  fftwf_execute_dft(plan->p, in, out);  
+}
+
 void srslte_dft_run_c(srslte_dft_plan_t *plan, cf_t *in, cf_t *out) {
   float norm;
   int i;

srslte/lib/dft/src/ofdm.c

@@ -163,13 +163,25 @@ void srslte_ofdm_rx_slot(srslte_ofdm_t *q, cf_t *input, cf_t *output) {
   }
 }
 
+void srslte_ofdm_rx_slot_zerocopy(srslte_ofdm_t *q, cf_t *input, cf_t *output) {
+  uint32_t i;
+  for (i=0;i<q->nof_symbols;i++) {
+    input += SRSLTE_CP_ISNORM(q->cp)?SRSLTE_CP_LEN_NORM(i, q->symbol_sz):SRSLTE_CP_LEN_EXT(q->symbol_sz);
+    srslte_dft_run_c_zerocopy(&q->fft_plan, input, q->tmp);
+    memcpy(output, &q->tmp[q->symbol_sz/2+q->nof_guards], sizeof(cf_t)*q->nof_re/2);
+    memcpy(&output[q->nof_re/2], &q->tmp[1], sizeof(cf_t)*q->nof_re/2);
+    input += q->symbol_sz;
+    output += q->nof_re;
+  }  
+}
+
 void srslte_ofdm_rx_sf(srslte_ofdm_t *q, cf_t *input, cf_t *output) {
   uint32_t n; 
   if (q->freq_shift) {
     srslte_vec_prod_ccc(input, q->shift_buffer, input, 2*q->slot_sz);
   }
   for (n=0;n<2;n++) {
-    srslte_ofdm_rx_slot(q, &input[n*q->slot_sz], &output[n*q->nof_re*q->nof_symbols]);
+    srslte_ofdm_rx_slot_zerocopy(q, &input[n*q->slot_sz], &output[n*q->nof_re*q->nof_symbols]);
   }
 }
 

srslte/lib/ue/src/ue_dl.c

@@ -183,7 +183,7 @@ int srslte_ue_dl_decode_fft_estimate(srslte_ue_dl_t *q, cf_t *input, uint32_t sf
     
     /* Run FFT for all subframe data */
     srslte_ofdm_rx_sf(&q->fft, input, q->sf_symbols);
-
+    
     /* Get channel estimates for each port */
     srslte_chest_dl_estimate(&q->chest, q->sf_symbols, q->ce, sf_idx);