1 files changed, 118 insertions, 40 deletions
diff --git a/chromium/media/base/vector_math.cc b/chromium/media/base/vector_math.cc
index 32584f5cf64..71721b69858 100644
--- a/chromium/media/base/vector_math.cc
+++ b/chromium/media/base/vector_math.cc
@@ -7,63 +7,29 @@
 
 #include <algorithm>
 
-#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 {
-
-// If we know the minimum architecture at compile time, avoid CPU detection.
-// Force NaCl code to use C routines since (at present) nothing there uses these
-// methods and plumbing the -msse built library is non-trivial.
+// NaCl does not allow intrinsics.
 #if defined(ARCH_CPU_X86_FAMILY) && !defined(OS_NACL)
-#if defined(__SSE__)
+#include <xmmintrin.h>
 #define FMAC_FUNC FMAC_SSE
 #define FMUL_FUNC FMUL_SSE
 #define EWMAAndMaxPower_FUNC EWMAAndMaxPower_SSE
-void Initialize() {}
-#else
-// X86 CPU detection required.  Functions will be set by Initialize().
-// TODO(dalecurtis): Once Chrome moves to an SSE baseline this can be removed.
-#define FMAC_FUNC g_fmac_proc_
-#define FMUL_FUNC g_fmul_proc_
-#define EWMAAndMaxPower_FUNC g_ewma_power_proc_
-
-typedef void (*MathProc)(const float src[], float scale, int len, float dest[]);
-static MathProc g_fmac_proc_ = NULL;
-static MathProc g_fmul_proc_ = NULL;
-typedef std::pair<float, float> (*EWMAAndMaxPowerProc)(
-    float initial_value, const float src[], int len, float smoothing_factor);
-static EWMAAndMaxPowerProc g_ewma_power_proc_ = NULL;
-
-void Initialize() {
-  CHECK(!g_fmac_proc_);
-  CHECK(!g_fmul_proc_);
-  CHECK(!g_ewma_power_proc_);
-  const bool kUseSSE = base::CPU().has_sse();
-  g_fmac_proc_ = kUseSSE ? FMAC_SSE : FMAC_C;
-  g_fmul_proc_ = kUseSSE ? FMUL_SSE : FMUL_C;
-  g_ewma_power_proc_ = kUseSSE ? EWMAAndMaxPower_SSE : EWMAAndMaxPower_C;
-}
-#endif
 #elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+#include <arm_neon.h>
 #define FMAC_FUNC FMAC_NEON
 #define FMUL_FUNC FMUL_NEON
 #define EWMAAndMaxPower_FUNC EWMAAndMaxPower_NEON
-void Initialize() {}
 #else
-// Unknown architecture.
 #define FMAC_FUNC FMAC_C
 #define FMUL_FUNC FMUL_C
 #define EWMAAndMaxPower_FUNC EWMAAndMaxPower_C
-void Initialize() {}
 #endif
 
+namespace media {
+namespace vector_math {
+
 void FMAC(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));
@@ -88,6 +54,13 @@ void FMUL_C(const float src[], float scale, int len, float dest[]) {
     dest[i] = src[i] * scale;
 }
 
+void Crossfade(const float src[], int len, float dest[]) {
+  float cf_ratio = 0;
+  const float cf_increment = 1.0f / len;
+  for (int i = 0; i < len; ++i, cf_ratio += cf_increment)
+    dest[i] = (1.0f - cf_ratio) * src[i] + cf_ratio * dest[i];
+}
+
 std::pair<float, float> EWMAAndMaxPower(
     float initial_value, const float src[], int len, float smoothing_factor) {
   // Ensure |src| is 16-byte aligned.
@@ -109,6 +82,111 @@ std::pair<float, float> EWMAAndMaxPower_C(
   return result;
 }
 
+#if defined(ARCH_CPU_X86_FAMILY) && !defined(OS_NACL)
+void FMUL_SSE(const float src[], float scale, int len, float dest[]) {
+  const int rem = len % 4;
+  const int last_index = len - rem;
+  __m128 m_scale = _mm_set_ps1(scale);
+  for (int i = 0; i < last_index; i += 4)
+    _mm_store_ps(dest + i, _mm_mul_ps(_mm_load_ps(src + i), m_scale));
+
+  // Handle any remaining values that wouldn't fit in an SSE pass.
+  for (int i = last_index; i < len; ++i)
+    dest[i] = src[i] * scale;
+}
+
+void FMAC_SSE(const float src[], float scale, int len, float dest[]) {
+  const int rem = len % 4;
+  const int last_index = len - rem;
+  __m128 m_scale = _mm_set_ps1(scale);
+  for (int i = 0; i < last_index; i += 4) {
+    _mm_store_ps(dest + i, _mm_add_ps(_mm_load_ps(dest + i),
+                 _mm_mul_ps(_mm_load_ps(src + i), m_scale)));
+  }
+
+  // Handle any remaining values that wouldn't fit in an SSE pass.
+  for (int i = last_index; i < len; ++i)
+    dest[i] += src[i] * scale;
+}
+
+// Convenience macro to extract float 0 through 3 from the vector |a|.  This is
+// needed because compilers other than clang don't support access via
+// operator[]().
+#define EXTRACT_FLOAT(a, i) \
+    (i == 0 ? \
+         _mm_cvtss_f32(a) : \
+         _mm_cvtss_f32(_mm_shuffle_ps(a, a, i)))
+
+std::pair<float, float> EWMAAndMaxPower_SSE(
+    float initial_value, const float src[], int len, float smoothing_factor) {
+  // When the recurrence is unrolled, we see that we can split it into 4
+  // separate lanes of evaluation:
+  //
+  // y[n] = a(S[n]^2) + (1-a)(y[n-1])
+  //      = a(S[n]^2) + (1-a)^1(aS[n-1]^2) + (1-a)^2(aS[n-2]^2) + ...
+  //      = z[n] + (1-a)^1(z[n-1]) + (1-a)^2(z[n-2]) + (1-a)^3(z[n-3])
+  //
+  // where z[n] = a(S[n]^2) + (1-a)^4(z[n-4]) + (1-a)^8(z[n-8]) + ...
+  //
+  // Thus, the strategy here is to compute z[n], z[n-1], z[n-2], and z[n-3] in
+  // each of the 4 lanes, and then combine them to give y[n].
+
+  const int rem = len % 4;
+  const int last_index = len - rem;
+
+  const __m128 smoothing_factor_x4 = _mm_set_ps1(smoothing_factor);
+  const float weight_prev = 1.0f - smoothing_factor;
+  const __m128 weight_prev_x4 = _mm_set_ps1(weight_prev);
+  const __m128 weight_prev_squared_x4 =
+      _mm_mul_ps(weight_prev_x4, weight_prev_x4);
+  const __m128 weight_prev_4th_x4 =
+      _mm_mul_ps(weight_prev_squared_x4, weight_prev_squared_x4);
+
+  // Compute z[n], z[n-1], z[n-2], and z[n-3] in parallel in lanes 3, 2, 1 and
+  // 0, respectively.
+  __m128 max_x4 = _mm_setzero_ps();
+  __m128 ewma_x4 = _mm_setr_ps(0.0f, 0.0f, 0.0f, initial_value);
+  int i;
+  for (i = 0; i < last_index; i += 4) {
+    ewma_x4 = _mm_mul_ps(ewma_x4, weight_prev_4th_x4);
+    const __m128 sample_x4 = _mm_load_ps(src + i);
+    const __m128 sample_squared_x4 = _mm_mul_ps(sample_x4, sample_x4);
+    max_x4 = _mm_max_ps(max_x4, sample_squared_x4);
+    // Note: The compiler optimizes this to a single multiply-and-accumulate
+    // instruction:
+    ewma_x4 = _mm_add_ps(ewma_x4,
+                         _mm_mul_ps(sample_squared_x4, smoothing_factor_x4));
+  }
+
+  // y[n] = z[n] + (1-a)^1(z[n-1]) + (1-a)^2(z[n-2]) + (1-a)^3(z[n-3])
+  float ewma = EXTRACT_FLOAT(ewma_x4, 3);
+  ewma_x4 = _mm_mul_ps(ewma_x4, weight_prev_x4);
+  ewma += EXTRACT_FLOAT(ewma_x4, 2);
+  ewma_x4 = _mm_mul_ps(ewma_x4, weight_prev_x4);
+  ewma += EXTRACT_FLOAT(ewma_x4, 1);
+  ewma_x4 = _mm_mul_ss(ewma_x4, weight_prev_x4);
+  ewma += EXTRACT_FLOAT(ewma_x4, 0);
+
+  // Fold the maximums together to get the overall maximum.
+  max_x4 = _mm_max_ps(max_x4,
+                      _mm_shuffle_ps(max_x4, max_x4, _MM_SHUFFLE(3, 3, 1, 1)));
+  max_x4 = _mm_max_ss(max_x4, _mm_shuffle_ps(max_x4, max_x4, 2));
+
+  std::pair<float, float> result(ewma, EXTRACT_FLOAT(max_x4, 0));
+
+  // Handle remaining values at the end of |src|.
+  for (; i < len; ++i) {
+    result.first *= weight_prev;
+    const float sample = src[i];
+    const float sample_squared = sample * sample;
+    result.first += sample_squared * smoothing_factor;
+    result.second = std::max(result.second, sample_squared);
+  }
+
+  return result;
+}
+#endif
+
 #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;