Add ARM NEON intrinsic optimizations for SincResampler.

media/base/sinc_resampler.cc

Index: media/base/sinc_resampler.cc
diff --git a/media/base/sinc_resampler.cc b/media/base/sinc_resampler.cc
index 869af1b3738dd07f207fca840d079688e90280ba..be3e71d336ea100f61d06fea7f97b415b43a9e15 100644
--- a/media/base/sinc_resampler.cc
+++ b/media/base/sinc_resampler.cc
@@ -36,13 +36,19 @@
 
 #include "media/base/sinc_resampler.h"
 
-#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
-#include <xmmintrin.h>
-#endif
 #include <cmath>
 
 #include "base/cpu.h"
 #include "base/logging.h"
+#include "build/build_config.h"
+
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+#include <xmmintrin.h>
+#endif
+
+#if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+#include <arm_neon.h>
+#endif
 
 namespace media {
 
@@ -231,6 +237,8 @@ float SincResampler::Convolve(const float* input_ptr, const float* k1,
 #if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
   static const ConvolveProc kConvolveProc =
       base::CPU().has_sse() ? Convolve_SSE : Convolve_C;
+#elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+  static const ConvolveProc kConvolveProc = Convolve_NEON;
 #else
   static const ConvolveProc kConvolveProc = Convolve_C;
 #endif
@@ -301,4 +309,29 @@ float SincResampler::Convolve_SSE(const float* input_ptr, const float* k1,
 }
 #endif
 
+#if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+float SincResampler::Convolve_NEON(const float* input_ptr, const float* k1,
+                                   const float* k2,
+                                   double kernel_interpolation_factor) {
+  float32x4_t m_input;
+  float32x4_t m_sums1 = vmovq_n_f32(0);
+  float32x4_t m_sums2 = vmovq_n_f32(0);
+
+  for (int i = 0; i < kKernelSize; i += 4) {
+    m_input = vld1q_f32(input_ptr + i);
+    m_sums1 = vaddq_f32(m_sums1, vmulq_f32(m_input, vld1q_f32(k1 + i)));
+    m_sums2 = vaddq_f32(m_sums2, vmulq_f32(m_input, vld1q_f32(k2 + i)));
+  }
+
+  // Linearly interpolate the two "convolutions".
+  m_sums1 = vmulq_f32(m_sums1, vmovq_n_f32(1.0 - kernel_interpolation_factor));
+  m_sums2 = vmulq_f32(m_sums2, vmovq_n_f32(kernel_interpolation_factor));
+  m_sums1 = vaddq_f32(m_sums1, m_sums2);
+
+  // Sum components together.
+  float32x2_t m_half = vadd_f32(vget_high_f32(m_sums1), vget_low_f32(m_sums1));
+  return vget_lane_f32(vpadd_f32(m_half, m_half), 0);
+}
+#endif
+
 }  // namespace media