3 files changed, 79 insertions, 1 deletions
diff --git a/volk/apps/volk_profile.cc b/volk/apps/volk_profile.cc
index 6244abb35..648f4d878 100644
--- a/volk/apps/volk_profile.cc
+++ b/volk/apps/volk_profile.cc
@@ -33,6 +33,7 @@ int main(int argc, char *argv[]) {
     //VOLK_PROFILE(volk_16i_permute_and_scalar_add_a, 1e-4, 0, 2046, 10000, &results);
     //VOLK_PROFILE(volk_16i_x4_quad_max_star_16i_a, 1e-4, 0, 2046, 10000, &results);
     VOLK_PROFILE(volk_16u_byteswap_a, 0, 0, 204600, 10000, &results);
+    VOLK_PROFILE(volk_16u_byteswap_u, 0, 0, 204600, 10000, &results);
     VOLK_PROFILE(volk_16i_32fc_dot_prod_32fc_a, 1e-4, 0, 204600, 10000, &results);
     VOLK_PROFILE(volk_32f_accumulator_s32f_a, 1e-4, 0, 204600, 10000, &results);
     VOLK_PROFILE(volk_32f_x2_add_32f_a, 1e-4, 0, 204600, 10000, &results);
diff --git a/volk/include/volk/volk_16u_byteswap_u.h b/volk/include/volk/volk_16u_byteswap_u.h
new file mode 100644
index 000000000..37bed7fce
--- /dev/null
+++ b/volk/include/volk/volk_16u_byteswap_u.h
@@ -0,0 +1,76 @@
+#ifndef INCLUDED_volk_16u_byteswap_u_H
+#define INCLUDED_volk_16u_byteswap_u_H
+
+#include <inttypes.h>
+#include <stdio.h>
+
+#ifdef LV_HAVE_SSE2
+#include <emmintrin.h>
+
+/*!
+  \brief Byteswaps (in-place) an unaligned vector of int16_t's.
+  \param intsToSwap The vector of data to byte swap
+  \param numDataPoints The number of data points
+*/
+static inline void volk_16u_byteswap_u_sse2(uint16_t* intsToSwap, unsigned int num_points){
+  unsigned int number = 0;
+  uint16_t* inputPtr = intsToSwap;
+  __m128i input, left, right, output;
+
+  const unsigned int eighthPoints = num_points / 8;
+  for(;number < eighthPoints; number++){
+    // Load the 16t values, increment inputPtr later since we're doing it in-place.
+    input = _mm_loadu_si128((__m128i*)inputPtr);
+    // Do the two shifts
+    left = _mm_slli_epi16(input, 8);
+    right = _mm_srli_epi16(input, 8);
+    // Or the left and right halves together
+    output = _mm_or_si128(left, right);
+    // Store the results
+    _mm_storeu_si128((__m128i*)inputPtr, output);
+    inputPtr += 8;
+  }
+
+  // Byteswap any remaining points:
+  number = eighthPoints*8;
+  for(; number < num_points; number++){
+    uint16_t outputVal = *inputPtr;
+    outputVal = (((outputVal >> 8) & 0xff) | ((outputVal << 8) & 0xff00));
+    *inputPtr = outputVal;
+    inputPtr++;
+  }
+}
+#endif /* LV_HAVE_SSE2 */
+
+#ifdef LV_HAVE_GENERIC
+/*!
+  \brief Byteswaps (in-place) an unaligned vector of int16_t's.
+  \param intsToSwap The vector of data to byte swap
+  \param numDataPoints The number of data points
+*/
+static inline void volk_16u_byteswap_u_generic(uint16_t* intsToSwap, unsigned int num_points){
+  unsigned int point;
+  uint16_t* inputPtr = intsToSwap;
+  for(point = 0; point < num_points; point++){
+    uint16_t output = *inputPtr;
+    output = (((output >> 8) & 0xff) | ((output << 8) & 0xff00));
+    *inputPtr = output;
+    inputPtr++;
+  }
+}
+#endif /* LV_HAVE_GENERIC */
+
+#ifdef LV_HAVE_ORC
+/*!
+  \brief Byteswaps (in-place) an unaligned vector of int16_t's.
+  \param intsToSwap The vector of data to byte swap
+  \param numDataPoints The number of data points
+*/
+extern void volk_16u_byteswap_u_orc_impl(uint16_t* intsToSwap, unsigned int num_points);
+static inline void volk_16u_byteswap_u_orc(uint16_t* intsToSwap, unsigned int num_points){
+    volk_16u_byteswap_u_orc_impl(intsToSwap, num_points);
+}
+#endif /* LV_HAVE_ORC */
+
+
+#endif /* INCLUDED_volk_16u_byteswap_u_H */
diff --git a/volk/lib/testqa.cc b/volk/lib/testqa.cc
index 507c78777..2e41c25da 100644
--- a/volk/lib/testqa.cc
+++ b/volk/lib/testqa.cc
@@ -20,6 +20,7 @@ VOLK_RUN_TESTS(volk_16i_convert_8i_u, 0, 0, 20460, 1);
 //VOLK_RUN_TESTS(volk_16i_permute_and_scalar_add_a, 1e-4, 0, 2046, 1000);
 //VOLK_RUN_TESTS(volk_16i_x4_quad_max_star_16i_a, 1e-4, 0, 2046, 1000);
 VOLK_RUN_TESTS(volk_16u_byteswap_a, 0, 0, 20460, 1);
+VOLK_RUN_TESTS(volk_16u_byteswap_u, 0, 0, 20460, 1);
 //VOLK_RUN_TESTS(volk_16i_32fc_dot_prod_32fc_a, 1e-4, 0, 204600, 1);
 VOLK_RUN_TESTS(volk_32f_accumulator_s32f_a, 1e-4, 0, 20460, 1);
 VOLK_RUN_TESTS(volk_32f_x2_add_32f_a, 1e-4, 0, 20460, 1);
@@ -36,7 +37,7 @@ VOLK_RUN_TESTS(volk_32fc_s32f_deinterleave_real_16i_a, 0, 32768, 20460, 1);
 VOLK_RUN_TESTS(volk_32fc_deinterleave_real_32f_a, 1e-4, 0, 20460, 1);
 VOLK_RUN_TESTS(volk_32fc_deinterleave_real_64f_a, 1e-4, 0, 20460, 1);
 VOLK_RUN_TESTS(volk_32fc_x2_dot_prod_32fc_a, 1e-4, 0, 2046000, 1);
-VOLK_RUN_TESTS(volk_32fc_32f_dot_prod_32fc_a, 1e-4, 0, 2046000, 1);
+VOLK_RUN_TESTS(volk_32fc_32f_dot_prod_32fc_a, 1e-4, 0, 204600, 1);
 VOLK_RUN_TESTS(volk_32fc_index_max_16u_a, 3, 0, 20460, 1);
 VOLK_RUN_TESTS(volk_32fc_s32f_magnitude_16i_a, 1, 32768, 20460, 1);
 VOLK_RUN_TESTS(volk_32fc_magnitude_32f_a, 1e-4, 0, 20460, 1);