Merge from Chromium at DEPS revision r198571

This commit was generated by merge_to_master.py.

Change-Id: I951118a03836157090561764dd2627f0add8118f

1 files changed, 76 insertions, 3 deletions

diff --git a/media/base/vector_math.cc b/media/base/vector_math.cc
index 96f94d95e4..603ae0b04f 100644
--- a/media/base/vector_math.cc
+++ b/media/base/vector_math.cc
@@ -7,6 +7,11 @@
 
 #include "base/cpu.h"
 #include "base/logging.h"
+#include "build/build_config.h"
+
+#if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+#include <arm_neon.h>
+#endif
 
 namespace media {
 namespace vector_math {
@@ -16,17 +21,23 @@ void FMAC(const float src[], float scale, int len, float dest[]) {
   DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(src) & (kRequiredAlignment - 1));
   DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(dest) & (kRequiredAlignment - 1));
 
-  // Rely on function level static initialization to keep VectorFMACProc
-  // selection thread safe.
   typedef void (*VectorFMACProc)(const float src[], float scale, int len,
                                  float dest[]);
-#if defined(ARCH_CPU_X86_FAMILY)
+
+  // No NaCl code uses the SSE functionality of AudioBus and plumbing the -msse
+  // built library is non-trivial, so simply disable for now.  iOS lies about
+  // its architecture, so we need to exclude it here.
+#if defined(ARCH_CPU_X86_FAMILY) && !defined(OS_NACL) && !defined(OS_IOS)
 #if defined(__SSE__)
   static const VectorFMACProc kVectorFMACProc = FMAC_SSE;
 #else
+  // TODO(dalecurtis): Remove function level static initialization, it's not
+  // thread safe: http://crbug.com/224662.
   static const VectorFMACProc kVectorFMACProc =
       base::CPU().has_sse() ? FMAC_SSE : FMAC_C;
 #endif
+#elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+  static const VectorFMACProc kVectorFMACProc = FMAC_NEON;
 #else
   static const VectorFMACProc kVectorFMACProc = FMAC_C;
 #endif
@@ -39,5 +50,67 @@ void FMAC_C(const float src[], float scale, int len, float dest[]) {
     dest[i] += src[i] * scale;
 }
 
+void FMUL(const float src[], float scale, int len, float dest[]) {
+  // Ensure |src| and |dest| are 16-byte aligned.
+  DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(src) & (kRequiredAlignment - 1));
+  DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(dest) & (kRequiredAlignment - 1));
+
+  typedef void (*VectorFMULProc)(const float src[], float scale, int len,
+                                 float dest[]);
+
+  // No NaCl code uses the SSE functionality of AudioBus and plumbing the -msse
+  // built library is non-trivial, so simply disable for now.  iOS lies about
+  // its architecture, so we need to exclude it here.
+#if defined(ARCH_CPU_X86_FAMILY) && !defined(OS_NACL) && !defined(OS_IOS)
+#if defined(__SSE__)
+  static const VectorFMULProc kVectorFMULProc = FMUL_SSE;
+#else
+  // TODO(dalecurtis): Remove function level static initialization, it's not
+  // thread safe: http://crbug.com/224662.
+  static const VectorFMULProc kVectorFMULProc =
+      base::CPU().has_sse() ? FMUL_SSE : FMUL_C;
+#endif
+#elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+  static const VectorFMULProc kVectorFMULProc = FMUL_NEON;
+#else
+  static const VectorFMULProc kVectorFMULProc = FMUL_C;
+#endif
+
+  return kVectorFMULProc(src, scale, len, dest);
+}
+
+void FMUL_C(const float src[], float scale, int len, float dest[]) {
+  for (int i = 0; i < len; ++i)
+    dest[i] = src[i] * scale;
+}
+
+#if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+void FMAC_NEON(const float src[], float scale, int len, float dest[]) {
+  const int rem = len % 4;
+  const int last_index = len - rem;
+  float32x4_t m_scale = vmovq_n_f32(scale);
+  for (int i = 0; i < last_index; i += 4) {
+    vst1q_f32(dest + i, vmlaq_f32(
+        vld1q_f32(dest + i), vld1q_f32(src + i), m_scale));
+  }
+
+  // Handle any remaining values that wouldn't fit in an NEON pass.
+  for (int i = last_index; i < len; ++i)
+    dest[i] += src[i] * scale;
+}
+
+void FMUL_NEON(const float src[], float scale, int len, float dest[]) {
+  const int rem = len % 4;
+  const int last_index = len - rem;
+  float32x4_t m_scale = vmovq_n_f32(scale);
+  for (int i = 0; i < last_index; i += 4)
+    vst1q_f32(dest + i, vmulq_f32(vld1q_f32(src + i), m_scale));
+
+  // Handle any remaining values that wouldn't fit in an NEON pass.
+  for (int i = last_index; i < len; ++i)
+    dest[i] = src[i] * scale;
+}
+#endif
+
 }  // namespace vector_math
 }  // namespace media