Merge branch 'convert_types' into usrp1

1 files changed, 88 insertions, 7 deletions

diff --git a/host/lib/transport/convert_types_impl.hpp b/host/lib/transport/convert_types_impl.hpp
index 5958b08cb..fdc859883 100644
--- a/host/lib/transport/convert_types_impl.hpp
+++ b/host/lib/transport/convert_types_impl.hpp
@@ -28,6 +28,10 @@
     #define USE_EMMINTRIN_H //use sse2 intrinsics
 #endif
 
+#if defined(USE_EMMINTRIN_H)
+    #include <emmintrin.h>
+#endif
+
 /***********************************************************************
  * Typedefs
  **********************************************************************/
@@ -72,7 +76,7 @@ static UHD_INLINE void item32_to_sc16_bswap(
 }
 
 /***********************************************************************
- * Convert complex float buffer to items32
+ * Convert complex float buffer to items32 (no swap)
  **********************************************************************/
 static const float shorts_per_float = float(32767);
 
@@ -82,6 +86,41 @@ static UHD_INLINE item32_t fc32_to_item32(fc32_t num){
     return (item32_t(real) << 16) | (item32_t(imag) << 0);
 }
 
+////////////////////////////////////
+// none-swap
+////////////////////////////////////
+#if defined(USE_EMMINTRIN_H)
+static UHD_INLINE void fc32_to_item32_nswap(
+    const fc32_t *input, item32_t *output, size_t nsamps
+){
+    __m128 scalar = _mm_set_ps1(shorts_per_float);
+
+    //convert blocks of samples with intrinsics
+    size_t i = 0; for (; i < (nsamps & ~0x3); i+=4){
+        //load from input
+        __m128 tmplo = _mm_loadu_ps(reinterpret_cast<const float *>(input+i+0));
+        __m128 tmphi = _mm_loadu_ps(reinterpret_cast<const float *>(input+i+2));
+
+        //convert and scale
+        __m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar));
+        __m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar));
+
+        //pack + swap 16-bit pairs
+        __m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi);
+        tmpi = _mm_shufflelo_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1));
+        tmpi = _mm_shufflehi_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1));
+
+        //store to output
+        _mm_storeu_si128(reinterpret_cast<__m128i *>(output+i), tmpi);
+    }
+
+    //convert remainder
+    for (; i < nsamps; i++){
+        output[i] = fc32_to_item32(input[i]);
+    }
+}
+
+#else
 static UHD_INLINE void fc32_to_item32_nswap(
     const fc32_t *input, item32_t *output, size_t nsamps
 ){
@@ -90,9 +129,12 @@ static UHD_INLINE void fc32_to_item32_nswap(
     }
 }
 
-#if defined(USE_EMMINTRIN_H)
-#include <emmintrin.h>
+#endif
 
+////////////////////////////////////
+// byte-swap
+////////////////////////////////////
+#if defined(USE_EMMINTRIN_H)
 static UHD_INLINE void fc32_to_item32_bswap(
     const fc32_t *input, item32_t *output, size_t nsamps
 ){
@@ -108,7 +150,7 @@ static UHD_INLINE void fc32_to_item32_bswap(
         __m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar));
         __m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar));
 
-        //pack + byteswap -> byteswap 32 bit words
+        //pack + byteswap -> byteswap 16 bit words
         __m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi);
         tmpi = _mm_or_si128(_mm_srli_epi16(tmpi, 8), _mm_slli_epi16(tmpi, 8));
 
@@ -145,6 +187,43 @@ static UHD_INLINE fc32_t item32_to_fc32(item32_t item){
     );
 }
 
+////////////////////////////////////
+// none-swap
+////////////////////////////////////
+#if defined(USE_EMMINTRIN_H)
+static UHD_INLINE void item32_to_fc32_nswap(
+    const item32_t *input, fc32_t *output, size_t nsamps
+){
+    __m128 scalar = _mm_set_ps1(floats_per_short/(1 << 16));
+    __m128i zeroi = _mm_setzero_si128();
+
+    //convert blocks of samples with intrinsics
+    size_t i = 0; for (; i < (nsamps & ~0x3); i+=4){
+        //load from input
+        __m128i tmpi = _mm_loadu_si128(reinterpret_cast<const __m128i *>(input+i));
+
+        //unpack + swap 16-bit pairs
+        tmpi = _mm_shufflelo_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1));
+        tmpi = _mm_shufflehi_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1));
+        __m128i tmpilo = _mm_unpacklo_epi16(zeroi, tmpi); //value in upper 16 bits
+        __m128i tmpihi = _mm_unpackhi_epi16(zeroi, tmpi);
+
+        //convert and scale
+        __m128 tmplo = _mm_mul_ps(_mm_cvtepi32_ps(tmpilo), scalar);
+        __m128 tmphi = _mm_mul_ps(_mm_cvtepi32_ps(tmpihi), scalar);
+
+        //store to output
+        _mm_storeu_ps(reinterpret_cast<float *>(output+i+0), tmplo);
+        _mm_storeu_ps(reinterpret_cast<float *>(output+i+2), tmphi);
+    }
+
+    //convert remainder
+    for (; i < nsamps; i++){
+        output[i] = item32_to_fc32(input[i]);
+    }
+}
+
+#else
 static UHD_INLINE void item32_to_fc32_nswap(
     const item32_t *input, fc32_t *output, size_t nsamps
 ){
@@ -152,10 +231,12 @@ static UHD_INLINE void item32_to_fc32_nswap(
         output[i] = item32_to_fc32(input[i]);
     }
 }
+#endif
 
+////////////////////////////////////
+// byte-swap
+////////////////////////////////////
 #if defined(USE_EMMINTRIN_H)
-#include <emmintrin.h>
-
 static UHD_INLINE void item32_to_fc32_bswap(
     const item32_t *input, fc32_t *output, size_t nsamps
 ){
@@ -167,7 +248,7 @@ static UHD_INLINE void item32_to_fc32_bswap(
         //load from input
         __m128i tmpi = _mm_loadu_si128(reinterpret_cast<const __m128i *>(input+i));
 
-        //byteswap + unpack -> byteswap 32 bit words
+        //byteswap + unpack -> byteswap 16 bit words
         tmpi = _mm_or_si128(_mm_srli_epi16(tmpi, 8), _mm_slli_epi16(tmpi, 8));
         __m128i tmpilo = _mm_unpacklo_epi16(zeroi, tmpi); //value in upper 16 bits
         __m128i tmpihi = _mm_unpackhi_epi16(zeroi, tmpi);