Optimised AVX512 LDPC decoder hard decision

lib/src/phy/fec/ldpc/ldpc_dec_c_avx512.c

@@ -232,9 +232,7 @@ int extract_ldpc_message_c_avx512(void* p, uint8_t* message, uint16_t liftK)
 
   int ini = 0;
   for (int i = 0; i < liftK; i = i + vp->ls) {
-    for (int k = 0; k < vp->ls; k++) {
-      message[i + k] = (vp->soft_bits.c[ini + k] < 0);
-    }
+    fec_avx512_hard_decision_c(&vp->soft_bits.c[ini], &message[i], vp->ls);
     ini = ini + SRSRAN_AVX512_B_SIZE;
   }
 

lib/src/phy/fec/ldpc/ldpc_dec_c_avx512long.c

@@ -292,9 +292,7 @@ int extract_ldpc_message_c_avx512long(void* p, uint8_t* message, uint16_t liftK)
 
   int ini = 0;
   for (int i = 0; i < liftK; i = i + vp->ls) {
-    for (int k = 0; k < vp->ls; k++) {
-      message[i + k] = (vp->soft_bits->c[ini + k] < 0);
-    }
+    fec_avx512_hard_decision_c(&vp->soft_bits->c[ini], &message[i], vp->ls);
     ini = ini + vp->node_size;
   }
 

lib/src/phy/fec/utils_avx512.h

@@ -26,4 +26,21 @@
 #define SRSRAN_AVX512_B_SIZE 64    /*!< \brief Number of packed bytes in an AVX512 instruction. */
 #define SRSRAN_AVX512_B_SIZE_LOG 6 /*!< \brief \f$\log_2\f$ of \ref SRSRAN_AVX512_B_SIZE. */
 
+#ifdef LV_HAVE_AVX512
+
+#include <immintrin.h>
+
+static inline void fec_avx512_hard_decision_c(const int8_t* llr, uint8_t* message, int nof_llr)
+{
+  int k = 0;
+  for (; k < nof_llr - (SRSRAN_AVX512_B_SIZE - 1); k += SRSRAN_AVX512_B_SIZE) {
+    __mmask64 mask = _mm512_cmpge_epi8_mask(_mm512_load_si512((__m512i*)&llr[k]), _mm512_set1_epi8(0));
+    _mm512_storeu_si512((__m512i*)&message[k], _mm512_mask_blend_epi8(mask, _mm512_set1_epi8(1), _mm512_set1_epi8(0)));
+  }
+  for (; k < nof_llr; k++) {
+    message[k] = (llr[k] < 0);
+  }
+}
+#endif // LV_HAVE_AVX512
+
 #endif // SRSRAN_UTILS_AVX512_H