Experiment with AVX optimizations for FMAC, FMUL operations.

media/filters/wsola_internals.cc

Index: media/filters/wsola_internals.cc
diff --git a/media/filters/wsola_internals.cc b/media/filters/wsola_internals.cc
index 9e12f032f3cc021a151c933aec88a2fe82e31bf7..353232679af01ce70dbe291bff8f2d4e02420f47 100644
--- a/media/filters/wsola_internals.cc
+++ b/media/filters/wsola_internals.cc
@@ -14,14 +14,7 @@
 
 #include "base/logging.h"
 #include "media/base/audio_bus.h"
-
-#if defined(ARCH_CPU_X86_FAMILY)
-#define USE_SIMD 1
-#include <xmmintrin.h>
-#elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
-#define USE_SIMD 1
-#include <arm_neon.h>
-#endif
+#include "media/base/vector_math.h"
 
 namespace media {
 
@@ -56,55 +49,11 @@ void MultiChannelDotProduct(const AudioBus* a,
   DCHECK_LE(frame_offset_a + num_frames, a->frames());
   DCHECK_LE(frame_offset_b + num_frames, b->frames());
 
-// SIMD optimized variants can provide a massive speedup to this operation.
-#if defined(USE_SIMD)
-  const int rem = num_frames % 4;
-  const int last_index = num_frames - rem;
   const int channels = a->channels();
   for (int ch = 0; ch < channels; ++ch) {
-    const float* a_src = a->channel(ch) + frame_offset_a;
-    const float* b_src = b->channel(ch) + frame_offset_b;
-
-#if defined(ARCH_CPU_X86_FAMILY)
-    // First sum all components.
-    __m128 m_sum = _mm_setzero_ps();
-    for (int s = 0; s < last_index; s += 4) {
-      m_sum = _mm_add_ps(
-          m_sum, _mm_mul_ps(_mm_loadu_ps(a_src + s), _mm_loadu_ps(b_src + s)));
-    }
-
-    // Reduce to a single float for this channel. Sadly, SSE1,2 doesn't have a
-    // horizontal sum function, so we have to condense manually.
-    m_sum = _mm_add_ps(_mm_movehl_ps(m_sum, m_sum), m_sum);
-    _mm_store_ss(dot_product + ch,
-                 _mm_add_ss(m_sum, _mm_shuffle_ps(m_sum, m_sum, 1)));
-#elif defined(ARCH_CPU_ARM_FAMILY)
-    // First sum all components.
-    float32x4_t m_sum = vmovq_n_f32(0);
-    for (int s = 0; s < last_index; s += 4)
-      m_sum = vmlaq_f32(m_sum, vld1q_f32(a_src + s), vld1q_f32(b_src + s));
-
-    // Reduce to a single float for this channel.
-    float32x2_t m_half = vadd_f32(vget_high_f32(m_sum), vget_low_f32(m_sum));
-    dot_product[ch] = vget_lane_f32(vpadd_f32(m_half, m_half), 0);
-#endif
-  }
-
-  if (!rem)
-    return;
-  num_frames = rem;
-  frame_offset_a += last_index;
-  frame_offset_b += last_index;
-#else
-  memset(dot_product, 0, sizeof(*dot_product) * a->channels());
-#endif  // defined(USE_SIMD)
-
-  // C version is required to handle remainder of frames (% 4 != 0)
-  for (int k = 0; k < a->channels(); ++k) {
-    const float* ch_a = a->channel(k) + frame_offset_a;
-    const float* ch_b = b->channel(k) + frame_offset_b;
-    for (int n = 0; n < num_frames; ++n)
-      dot_product[k] += *ch_a++ * *ch_b++;
+    dot_product[ch] =
+        vector_math::DotProduct(a->channel(ch) + frame_offset_a,
+                                b->channel(ch) + frame_offset_b, num_frames);
   }
 }