[Commit-gnuradio] [gnuradio] 11/22: volk: add neon version of 32u_byteswap

[git]

This is an automated email from the git hooks/post-receive script.

jcorgan pushed a commit to branch master
in repository gnuradio.

commit f2bc76cc65ffba51a141950f98e75364e49df874
Author: Nathan West <address@hidden>
Date:   Mon Oct 20 19:05:11 2014 -0500

    volk: add neon version of 32u_byteswap
    
    The QA needed a puppet because qa_utils only checks output buffers,
    so this also introduces a puppet
---
 volk/kernels/volk/volk_32u_byteswap.h           | 55 ++++++++++++++++++++++++-
 volk/kernels/volk/volk_32u_byteswappuppet_32u.h | 45 ++++++++++++++++++++
 2 files changed, 98 insertions(+), 2 deletions(-)

diff --git a/volk/kernels/volk/volk_32u_byteswap.h 
b/volk/kernels/volk/volk_32u_byteswap.h
index 74d9a0b..0194efc 100644
--- a/volk/kernels/volk/volk_32u_byteswap.h
+++ b/volk/kernels/volk/volk_32u_byteswap.h
@@ -73,6 +73,59 @@ static inline void volk_32u_byteswap_u_sse2(uint32_t* 
intsToSwap, unsigned int n
 }
 #endif /* LV_HAVE_SSE2 */
 
+#ifdef LV_HAVE_NEON
+#include <arm_neon.h>
+/*!
+  \brief Byteswaps (in-place) an aligned vector of int32_t's.
+  \param intsToSwap The vector of data to byte swap
+  \param numDataPoints The number of data points
+*/
+static inline void volk_32u_byteswap_neon(uint32_t* intsToSwap, unsigned int 
num_points){
+  uint32_t* inputPtr = intsToSwap;
+  unsigned int number = 0;
+  unsigned int n8points = num_points / 8;
+
+  uint8x8x4_t input_table;
+  uint8x8_t int_lookup01, int_lookup23, int_lookup45, int_lookup67;
+  uint8x8_t swapped_int01, swapped_int23, swapped_int45, swapped_int67;
+
+  /* these magic numbers are used as byte-indeces in the LUT.
+     they are pre-computed to save time. A simple C program
+     can calculate them; for example for lookup01:
+    uint8_t chars[8] = {24, 16, 8, 0, 25, 17, 9, 1};
+    for(ii=0; ii < 8; ++ii) {
+        index += ((uint64_t)(*(chars+ii))) << (ii*8);
+    }
+  */
+  int_lookup01 = vcreate_u8(74609667900706840);
+  int_lookup23 = vcreate_u8(219290013576860186);
+  int_lookup45 = vcreate_u8(363970359253013532);
+  int_lookup67 = vcreate_u8(508650704929166878);
+  
+  for(number = 0; number < n8points; ++number){
+    input_table = vld4_u8((uint8_t*) inputPtr);
+    swapped_int01 = vtbl4_u8(input_table, int_lookup01);
+    swapped_int23 = vtbl4_u8(input_table, int_lookup23);
+    swapped_int45 = vtbl4_u8(input_table, int_lookup45);
+    swapped_int67 = vtbl4_u8(input_table, int_lookup67);
+    vst1_u8((uint8_t*) inputPtr, swapped_int01);
+    vst1_u8((uint8_t*) (inputPtr+2), swapped_int23);
+    vst1_u8((uint8_t*) (inputPtr+4), swapped_int45);
+    vst1_u8((uint8_t*) (inputPtr+6), swapped_int67);
+
+    inputPtr += 8;
+  }
+
+  for(number = n8points * 8; number < num_points; ++number){
+    uint32_t output = *inputPtr;
+    output = (((output >> 24) & 0xff) | ((output >> 8) & 0x0000ff00) | 
((output << 8) & 0x00ff0000) | ((output << 24) & 0xff000000));
+
+    *inputPtr = output;
+    inputPtr++;
+  }
+}
+#endif /* LV_HAVE_NEON */
+
 #ifdef LV_HAVE_GENERIC
 /*!
   \brief Byteswaps (in-place) an aligned vector of int32_t's.
@@ -94,8 +147,6 @@ static inline void volk_32u_byteswap_generic(uint32_t* 
intsToSwap, unsigned int
 #endif /* LV_HAVE_GENERIC */
 
 
-
-
 #endif /* INCLUDED_volk_32u_byteswap_u_H */
 #ifndef INCLUDED_volk_32u_byteswap_a_H
 #define INCLUDED_volk_32u_byteswap_a_H
diff --git a/volk/kernels/volk/volk_32u_byteswappuppet_32u.h 
b/volk/kernels/volk/volk_32u_byteswappuppet_32u.h
new file mode 100644
index 0000000..e50ad1d
--- /dev/null
+++ b/volk/kernels/volk/volk_32u_byteswappuppet_32u.h
@@ -0,0 +1,45 @@
+#ifndef INCLUDED_volk_32u_byteswappuppet_32u_H
+#define INCLUDED_volk_32u_byteswappuppet_32u_H
+
+
+#include <stdint.h>
+#include <volk/volk_32u_byteswap.h>
+
+#ifdef LV_HAVE_GENERIC
+static inline void volk_32u_byteswappuppet_32u_generic(uint32_t*output, 
uint32_t* intsToSwap, unsigned int num_points){
+
+    volk_32u_byteswap_generic((uint32_t*)intsToSwap, num_points);
+    memcpy((void*)output, (void*)intsToSwap, num_points * sizeof(uint32_t));
+
+}
+#endif
+
+#ifdef LV_HAVE_NEON
+static inline void volk_32u_byteswappuppet_32u_neon(uint32_t*output, uint32_t* 
intsToSwap, unsigned int num_points){
+
+    volk_32u_byteswap_neon((uint32_t*)intsToSwap, num_points);
+    memcpy((void*)output, (void*)intsToSwap, num_points * sizeof(uint32_t));
+
+}
+#endif
+
+#ifdef LV_HAVE_SSE2
+static inline void volk_32u_byteswappuppet_u_sse2(uint32_t* intsToSwap, 
unsigned int num_points){
+
+    volk_32u_byteswap_u_sse2((uint32_t*)intsToSwap, num_points);
+    memcpy((void*)output, (void*)intsToSwap, num_points * sizeof(uint32_t));
+
+}
+#endif
+
+#ifdef LV_HAVE_SSE2
+static inline void volk_32u_byteswappuppet_a_sse2(uint32_t* intsToSwap, 
unsigned int num_points){
+
+    volk_32u_byteswap_a_sse2((uint32_t*)intsToSwap, num_points);
+    memcpy((void*)output, (void*)intsToSwap, num_points * sizeof(uint32_t));
+
+}
+#endif
+
+
+#endif