From 767d69622770a4bb886f527eaa2e1e2a15a71309 Mon Sep 17 00:00:00 2001
From: "Matthias P. Braendli" <matthias.braendli@mpb.li>
Date: Mon, 11 Nov 2024 09:42:47 +0100
Subject: Use ARM NEON in FormatConverter

---
 src/Buffer.h            |  2 +-
 src/CicEqualizer.h      |  3 ---
 src/FormatConverter.cpp | 43 ++++++++++++++++++++++++++++++++++---------
 src/Utils.cpp           |  3 +++
 4 files changed, 38 insertions(+), 13 deletions(-)

diff --git a/src/Buffer.h b/src/Buffer.h
index d5aa802..2c2a65e 100644
--- a/src/Buffer.h
+++ b/src/Buffer.h
@@ -43,7 +43,7 @@ typedef std::complex<fixed_16> complexfix;
 typedef std::complex<fpm::fixed_16_16> complexfix_wide;
 
 /* Buffer is a container for a byte array, which is memory-aligned
- * to 32 bytes for SSE performance.
+ * to 32 bytes for SIMD performance.
  *
  * The allocation/freeing of the data is handled internally.
  */
diff --git a/src/CicEqualizer.h b/src/CicEqualizer.h
index 70c3ae9..4510d0c 100644
--- a/src/CicEqualizer.h
+++ b/src/CicEqualizer.h
@@ -29,9 +29,6 @@
 
 #include <vector>
 #include <sys/types.h>
-#ifdef __SSE__
-#   include <xmmintrin.h>
-#endif
 
 class CicEqualizer : public ModCodec
 {
diff --git a/src/FormatConverter.cpp b/src/FormatConverter.cpp
index 0821191..94dfa2c 100644
--- a/src/FormatConverter.cpp
+++ b/src/FormatConverter.cpp
@@ -28,11 +28,15 @@
 
 #include "FormatConverter.h"
 #include "PcDebug.h"
+#include "Log.h"
 
-#include <sys/types.h>
-#include <string.h>
 #include <stdexcept>
+#include <cstring>
 #include <assert.h>
+#include <sys/types.h>
+#if defined(__ARM_NEON)
+#include <arm_neon.h>
+#endif
 
 FormatConverter::FormatConverter(bool input_is_complexfix_wide, const std::string& format_out) :
     ModCodec(),
@@ -42,9 +46,16 @@ FormatConverter::FormatConverter(bool input_is_complexfix_wide, const std::strin
 
 FormatConverter::~FormatConverter()
 {
-    etiLog.level(debug) << "FormatConverter: "
-        << m_num_clipped_samples.load() <<
-        " clipped samples";
+    if (
+#if defined(__ARM_NEON)
+    not m_input_complexfix_wide
+#else
+    true
+#endif
+    ) {
+        etiLog.level(debug) << "FormatConverter: " <<
+            m_num_clipped_samples.load() << " clipped";
+    }
 }
 
 
@@ -56,16 +67,29 @@ int FormatConverter::process(Buffer* const dataIn, Buffer* dataOut)
 
     size_t num_clipped_samples = 0;
 
-
     if (m_input_complexfix_wide) {
         size_t sizeIn = dataIn->getLength() / sizeof(int32_t);
-        int32_t* in = reinterpret_cast<int32_t*>(dataIn->getData());
         if (m_format_out == "s16") {
             dataOut->setLength(sizeIn * sizeof(int16_t));
+            const int32_t *in = reinterpret_cast<int32_t*>(dataIn->getData());
             int16_t* out = reinterpret_cast<int16_t*>(dataOut->getData());
 
+            constexpr int shift = 7;
+
+#if defined(__ARM_NEON)
+            if (sizeIn % 4 != 0) {
+                throw std::logic_error("Unexpected length not multiple of 4");
+            }
+
+            for (size_t i = 0; i < sizeIn; i += 4) {
+                int32x4_t input_vec = vld1q_s32(&in[i]);
+                // Apply shift right, saturate on conversion to int16_t
+                int16x4_t output_vec = vqshrn_n_s32(input_vec, shift);
+                vst1_s16(&out[i], output_vec);
+            }
+#else
             for (size_t i = 0; i < sizeIn; i++) {
-                const int32_t val = in[i] >> 7;
+                const int32_t val = in[i] >> shift;
                 if (val < INT16_MIN) {
                     out[i] = INT16_MIN;
                     num_clipped_samples++;
@@ -78,6 +102,7 @@ int FormatConverter::process(Buffer* const dataIn, Buffer* dataOut)
                     out[i] = val;
                 }
             }
+#endif
         }
         else {
             throw std::runtime_error("FormatConverter: Invalid fix format " + m_format_out);
@@ -85,7 +110,7 @@ int FormatConverter::process(Buffer* const dataIn, Buffer* dataOut)
     }
     else {
         size_t sizeIn = dataIn->getLength() / sizeof(float);
-        float* in = reinterpret_cast<float*>(dataIn->getData());
+        const float* in = reinterpret_cast<float*>(dataIn->getData());
 
         if (m_format_out == "s16") {
             dataOut->setLength(sizeIn * sizeof(int16_t));
diff --git a/src/Utils.cpp b/src/Utils.cpp
index 0065bc1..f54122c 100644
--- a/src/Utils.cpp
+++ b/src/Utils.cpp
@@ -62,6 +62,9 @@ static void printHeader()
 #endif
 #if defined(__SSE__)
         "SSE " <<
+#endif
+#if defined(__ARM_NEON)
+        "NEON " <<
 #endif
         "\n";
 }
-- 
cgit v1.2.3