Turbo decoder working with hard decision

srslte/include/srslte/utils/vector_simd.h

@@ -43,6 +43,10 @@ SRSLTE_API void srslte_vec_sub_sss_simd(short *x, short *y, short *z, uint32_t l
 
 SRSLTE_API void srslte_vec_sc_div2_sss_simd(short *x, int n_rightshift, short *z, uint32_t len); 
 
+SRSLTE_API void srslte_vec_lut_sss_simd(short *x, unsigned short *lut, short *y, uint32_t len); 
+
+SRSLTE_API void srslte_vec_convert_fi_simd(float *x, int16_t *z, float scale, uint32_t len); 
+
 #ifdef __cplusplus
 }
 #endif

srslte/lib/fec/src/turbodecoder.c

@@ -207,7 +207,7 @@ void srslte_map_gen_alpha(srslte_map_gen_t * s, uint32_t long_cb)
   __m128i gv; 
   __m128i *gPtr = (__m128i*) s->branch;
   __m128i g, ap, an; 
-  
+    
   __m128i alpha_k = _mm_set_epi16(-INF, -INF, -INF, -INF, -INF, -INF, -INF, 0);
   
 #define ALPHA_STEP(c)  g = _mm_shuffle_epi8(gv, shuf_g[c]); \
@@ -564,13 +564,13 @@ void srslte_tdec_iteration(srslte_tdec_t * h, float * input, uint32_t long_cb)
     }
     
     // Interleave extrinsic output of DEC1 to form apriori info for decoder 2
-    srslte_vec_lut_sss(h->ext1, inter, h->app2, long_cb);
+    srslte_vec_lut_sss(h->ext1, deinter, h->app2, long_cb);
 
     // Run MAP DEC #2. 2nd decoder uses apriori information as systematic bits
     srslte_map_gen_dec(&h->dec, h->app2, NULL, h->parity1, h->ext2, long_cb);
 
     // Deinterleaved extrinsic bits become apriori info for decoder 1 
-    srslte_vec_lut_sss(h->ext2, deinter, h->app1, long_cb);
+    srslte_vec_lut_sss(h->ext2, inter, h->app1, long_cb);
     
     h->n_iter++;
   } else {
@@ -626,14 +626,14 @@ void srslte_tdec_decision_byte(srslte_tdec_t * h, uint8_t *output, uint32_t long
   
   // long_cb is always byte aligned
   for (uint32_t i = 0; i < long_cb/8; i++) {
-    uint8_t out0 = h->app1[i+0]>0?mask[0]:0;
-    uint8_t out1 = h->app1[i+1]>0?mask[1]:0;
-    uint8_t out2 = h->app1[i+2]>0?mask[2]:0;
-    uint8_t out3 = h->app1[i+3]>0?mask[3]:0;
-    uint8_t out4 = h->app1[i+4]>0?mask[4]:0;
-    uint8_t out5 = h->app1[i+5]>0?mask[5]:0;
-    uint8_t out6 = h->app1[i+6]>0?mask[6]:0;
-    uint8_t out7 = h->app1[i+7]>0?mask[7]:0;
+    uint8_t out0 = h->app1[8*i+0]>0?mask[0]:0;
+    uint8_t out1 = h->app1[8*i+1]>0?mask[1]:0;
+    uint8_t out2 = h->app1[8*i+2]>0?mask[2]:0;
+    uint8_t out3 = h->app1[8*i+3]>0?mask[3]:0;
+    uint8_t out4 = h->app1[8*i+4]>0?mask[4]:0;
+    uint8_t out5 = h->app1[8*i+5]>0?mask[5]:0;
+    uint8_t out6 = h->app1[8*i+6]>0?mask[6]:0;
+    uint8_t out7 = h->app1[8*i+7]>0?mask[7]:0;
     
     output[i] = out0 | out1 | out2 | out3 | out4 | out5 | out6 | out7; 
   }
@@ -650,7 +650,7 @@ int srslte_tdec_run_all(srslte_tdec_t * h, float * input, uint8_t *output,
     srslte_tdec_iteration(h, input, long_cb);
   } while (h->n_iter < nof_iterations);
 
-  srslte_tdec_decision(h, output, long_cb);
+  srslte_tdec_decision_byte(h, output, long_cb);
   
   return SRSLTE_SUCCESS;
 }

srslte/lib/fec/test/turbodecoder_test.c

@@ -113,7 +113,7 @@ int main(int argc, char **argv) {
   uint32_t frame_cnt;
   float *llr;
   uint8_t *llr_c;
-  uint8_t *data_tx, *data_rx, *symbols;
+  uint8_t *data_tx, *data_rx, *data_rx_bytes, *symbols;
   uint32_t i, j;
   float var[SNR_POINTS];
   uint32_t snr_points;
@@ -157,6 +157,11 @@ int main(int argc, char **argv) {
     perror("malloc");
     exit(-1);
   }
+  data_rx_bytes = srslte_vec_malloc(frame_length * sizeof(uint8_t));
+  if (!data_rx_bytes) {
+    perror("malloc");
+    exit(-1);
+  }
 
   symbols = srslte_vec_malloc(coded_length * sizeof(uint8_t));
   if (!symbols) {
@@ -248,12 +253,14 @@ int main(int argc, char **argv) {
 
       gettimeofday(&tdata[1], NULL); 
       for (int k=0;k<nof_repetitions;k++) {     
-        srslte_tdec_run_all(&tdec, llr, data_rx, t, frame_length);        
+        srslte_tdec_run_all(&tdec, llr, data_rx_bytes, t, frame_length);        
       }
       gettimeofday(&tdata[2], NULL);
       get_time_interval(tdata);
       mean_usec = (float) mean_usec * 0.9 + (float) (tdata[0].tv_usec/nof_repetitions) * 0.1;
       
+      srslte_bit_unpack_vector(data_rx_bytes, data_rx, frame_length);
+
       errors += srslte_bit_diff(data_tx, data_rx, frame_length);
       
       gettimeofday(&tdata[1], NULL); 

srslte/lib/phch/test/pdsch_test.c

@@ -228,7 +228,7 @@ int main(int argc, char **argv) {
     int r = srslte_pdsch_decode(&pdsch, &pdsch_cfg, &softbuffer_rx, slot_symbols[0], ce, 0, data);
     gettimeofday(&t[2], NULL);
     get_time_interval(t);
-    printf("DECODED %d in %d:%d (%.2f Mbps)\n", r?"Error":"OK",
+    printf("DECODED %s in %d:%d (%.2f Mbps)\n", r?"Error":"OK",
              (int) t[0].tv_sec, (int) t[0].tv_usec, (float) grant.mcs.tbs/t[0].tv_usec);                
     if (r) {
       ret = -1;

srslte/lib/utils/src/vector.c

@@ -241,26 +241,30 @@ void srslte_vec_sc_prod_ccc(cf_t *x, cf_t h, cf_t *z, uint32_t len) {
 }
 
 void srslte_vec_convert_fi(float *x, int16_t *z, float scale, uint32_t len) {
-#ifdef HAVE_VOLK_CONVERT_FI_FUNCTION
-  volk_32f_s32f_convert_16i(z, x, scale, len);
-#else 
+#ifndef HAVE_VECTOR_SIMD
   int i;
   for (i=0;i<len;i++) {
     z[i] = (int16_t) (x[i]*scale);
   }
+#else 
+  srslte_vec_convert_fi_simd(x, z, scale, len);
 #endif
 }
 
 void srslte_vec_lut_fuf(float *x, uint32_t *lut, float *y, uint32_t len) {
   for (int i=0;i<len;i++) {
-    y[i] = x[lut[i]];
+    y[lut[i]] = x[i];
   }
 }
 
 void srslte_vec_lut_sss(short *x, unsigned short *lut, short *y, uint32_t len) {
+#ifndef HAVE_VECTOR_SIMD
   for (int i=0;i<len;i++) {
-    y[i] = x[lut[i]];
+    y[lut[i]] = x[i];
   }
+#else
+  srslte_vec_lut_sss_simd(x, lut, y, len);
+#endif
 }
 
 void srslte_vec_interleave_cf(float *real, float *imag, cf_t *x, uint32_t len) {

srslte/lib/utils/src/vector_simd.c

@@ -135,3 +135,82 @@ void srslte_vec_sc_div2_sss_simd(short *x, int k, short *z, uint32_t len)
   }
 }
 
+void srslte_vec_lut_sss_simd(short *x, unsigned short *lut, short *y, uint32_t len)
+{
+  unsigned int number = 0;
+  const unsigned int points = len / 8;
+
+  const __m128i* xPtr = (const __m128i*) x;
+  const __m128i* lutPtr = (__m128i*) lut;
+
+  __m128i xVal, lutVal;
+  for(;number < points; number++){
+
+    xVal   = _mm_load_si128(xPtr);
+    lutVal = _mm_load_si128(lutPtr);
+    
+    for (int i=0;i<8;i++) {
+      uint16_t x = (uint16_t) _mm_extract_epi16(xVal, i); 
+      uint16_t l = (uint16_t) _mm_extract_epi16(lutVal, i);
+      y[l] = x;
+    }
+    xPtr ++;
+    lutPtr ++;
+  }
+
+  number = points * 8;
+  for(;number < len; number++){
+    y[lut[number]] = x[number];
+  }
+  
+}
+
+/* Modified from volk_32f_s32f_convert_16i_a_sse2. Removed clipping */
+void srslte_vec_convert_fi_simd(float *x, int16_t *z, float scale, uint32_t len)
+{
+    unsigned int number = 0;
+
+  const unsigned int eighthPoints = len / 8;
+
+  const float* inputVectorPtr = (const float*)x;
+  int16_t* outputVectorPtr = z;
+
+  float min_val = -32768;
+  float max_val = 32767;
+  float r;
+
+  __m128 vScalar = _mm_set_ps1(scale);
+  __m128 inputVal1, inputVal2;
+  __m128i intInputVal1, intInputVal2;
+  __m128 ret1, ret2;
+  __m128 vmin_val = _mm_set_ps1(min_val);
+  __m128 vmax_val = _mm_set_ps1(max_val);
+
+  for(;number < eighthPoints; number++){
+    inputVal1 = _mm_load_ps(inputVectorPtr); inputVectorPtr += 4;
+    inputVal2 = _mm_load_ps(inputVectorPtr); inputVectorPtr += 4;
+
+    // Scale and clip
+    ret1 = _mm_mul_ps(inputVal1, vScalar);
+    ret2 = _mm_mul_ps(inputVal2, vScalar);
+
+    intInputVal1 = _mm_cvtps_epi32(ret1);
+    intInputVal2 = _mm_cvtps_epi32(ret2);
+
+    intInputVal1 = _mm_packs_epi32(intInputVal1, intInputVal2);
+
+    _mm_store_si128((__m128i*)outputVectorPtr, intInputVal1);
+    outputVectorPtr += 8;
+  }
+
+  number = eighthPoints * 8;
+  for(; number < num_points; number++){
+    r = inputVector[number] * scalar;
+    if(r > max_val)
+      r = max_val;
+    else if(r < min_val)
+      r = min_val;
+    outputVector[number] = (int16_t)rintf(r);
+  }
+
+}