Experiment with AVX optimizations for FMAC, FMUL operations.

media/base/vector_math.cc

Index: media/base/vector_math.cc
diff --git a/media/base/vector_math.cc b/media/base/vector_math.cc
index 578290538ac26cee8080b144a39124ae344f8646..dd963d785b155e6d7bb2f229890ff56e728138f1 100644
--- a/media/base/vector_math.cc
+++ b/media/base/vector_math.cc
@@ -3,80 +3,158 @@
 // found in the LICENSE file.
 
 #include "media/base/vector_math.h"
-#include "media/base/vector_math_testing.h"
 
 #include <algorithm>
 
+#include "base/cpu.h"
 #include "base/logging.h"
 #include "build/build_config.h"
+#include "media/base/vector_math_testing.h"
 
 // NaCl does not allow intrinsics.
 #if defined(ARCH_CPU_X86_FAMILY) && !defined(OS_NACL)
 #include <xmmintrin.h>
+
+using FmacProc = void (*)(const float*, float, int, float*);
+static FmacProc g_fmac_proc_ = nullptr;
+using DotProductProc = float (*)(const float*, const float*, int);
+static DotProductProc g_dotproduct_proc_ = nullptr;
+using ConvolveProc = float (*)(const float*,
+                               const float*,
+                               const float*,
+                               double);
+static ConvolveProc g_convolve_proc_ = nullptr;
+
+// AVX FMAC only performs well on AVX2+ machines due to those machines actually
+// having 256-bit processing units vs wrappers around 128 bit units.
+#define INITIALIZE()                                                           \
+  do {                                                                         \
+    CHECK(!g_fmac_proc_);                                                      \
+    CHECK(!g_convolve_proc_);                                                  \
+    CHECK(!g_dotproduct_proc_);                                                \
+    base::CPU cpu_info;                                                        \
+    g_fmac_proc_ = cpu_info.has_avx2() ? FMAC_AVX : FMAC_SSE;                  \
+    g_convolve_proc_ = cpu_info.has_avx() ? Convolve_AVX : Convolve_SSE;       \
+    g_dotproduct_proc_ = cpu_info.has_avx() ? DotProduct_AVX : DotProduct_SSE; \
+  } while (0)
+
+#define CONVOLVE_FUNC g_convolve_proc_
+#define DOTPRODUCT_FUNC g_dotproduct_proc_
+#define FMAC_FUNC g_fmac_proc_
+
 // Don't use custom SSE versions where the auto-vectorized C version performs
 // better, which is anywhere clang is used.
 #if !defined(__clang__)
-#define FMAC_FUNC FMAC_SSE
 #define FMUL_FUNC FMUL_SSE
 #else
-#define FMAC_FUNC FMAC_C
 #define FMUL_FUNC FMUL_C
 #endif
 #define EWMAAndMaxPower_FUNC EWMAAndMaxPower_SSE
+
 #elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+
+// NEON optimized versions.
 #include <arm_neon.h>
+#define INITIALIZE()
+#define CONVOLVE_FUNC Convolve_NEON
+#define DOTPRODUCT_FUNC DotProduct_NEON
 #define FMAC_FUNC FMAC_NEON
 #define FMUL_FUNC FMUL_NEON
 #define EWMAAndMaxPower_FUNC EWMAAndMaxPower_NEON
+
 #else
+
+// No SIMD optimization versions.
+#define INITIALIZE()
+#define CONVOLVE_FUNC Convolve_C
+#define DOTPRODUCT_FUNC DotProduct_C
 #define FMAC_FUNC FMAC_C
 #define FMUL_FUNC FMUL_C
 #define EWMAAndMaxPower_FUNC EWMAAndMaxPower_C
+
 #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.
+void Initialize() {
+  INITIALIZE();
+}
+
+float Convolve(const float* src,
+               const float* k1,
+               const float* k2,
+               double kernel_interpolation_factor) {
+  return CONVOLVE_FUNC(src, k1, k2, kernel_interpolation_factor);
+}
+
+float Convolve_C(const float* src,
+                 const float* k1,
+                 const float* k2,
+                 double kernel_interpolation_factor) {
+  float sum1 = 0;
+  float sum2 = 0;
+
+  // Generate a single output sample.
+  int n = kKernelSize;
+  while (n--) {
+    sum1 += *src * *k1++;
+    sum2 += *src++ * *k2++;
+  }
+
+  // Linearly interpolate the two "convolutions".
+  return static_cast<float>((1.0 - kernel_interpolation_factor) * sum1 +
+                            kernel_interpolation_factor * sum2);
+}
+
+float DotProduct(const float* a, const float* b, int len) {
+  return DOTPRODUCT_FUNC(a, b, len);
+}
+
+float DotProduct_C(const float* a, const float* b, int len) {
+  float sum = 0;
+  for (int i = 0; i < len; ++i)
+    sum += a[i] * b[i];
+  return sum;
+}
+
+void FMAC(const float* src, float scale, int len, float* dest) {
+  // Ensure |src| and |dest| are aligned.
   DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(src) & (kRequiredAlignment - 1));
   DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(dest) & (kRequiredAlignment - 1));
   return FMAC_FUNC(src, scale, len, dest);
 }
 
-void FMAC_C(const float src[], float scale, int len, float dest[]) {
+void FMAC_C(const float* src, float scale, int len, float* dest) {
   for (int i = 0; i < len; ++i)
     dest[i] += src[i] * scale;
 }
 
-void FMUL(const float src[], float scale, int len, float dest[]) {
-  // Ensure |src| and |dest| are 16-byte aligned.
+void FMUL(const float* src, float scale, int len, float* dest) {
+  // Ensure |src| and |dest| are aligned.
   DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(src) & (kRequiredAlignment - 1));
   DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(dest) & (kRequiredAlignment - 1));
   return FMUL_FUNC(src, scale, len, dest);
 }
 
-void FMUL_C(const float src[], float scale, int len, float 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;
 }
 
-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) {
+std::pair<float, float> EWMAAndMaxPower(float initial_value,
+                                        const float* src,
+                                        int len,
+                                        float smoothing_factor) {
   // Ensure |src| is 16-byte aligned.
   DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(src) & (kRequiredAlignment - 1));
   return EWMAAndMaxPower_FUNC(initial_value, src, len, smoothing_factor);
 }
 
-std::pair<float, float> EWMAAndMaxPower_C(
-    float initial_value, const float src[], int len, float smoothing_factor) {
+std::pair<float, float> EWMAAndMaxPower_C(float initial_value,
+                                          const float* src,
+                                          int len,
+                                          float smoothing_factor) {
   std::pair<float, float> result(initial_value, 0.0f);
   const float weight_prev = 1.0f - smoothing_factor;
   for (int i = 0; i < len; ++i) {
@@ -90,10 +168,75 @@ std::pair<float, float> EWMAAndMaxPower_C(
 }
 
 #if defined(ARCH_CPU_X86_FAMILY) && !defined(OS_NACL)
-void FMUL_SSE(const float src[], float scale, int len, float dest[]) {
+float Convolve_SSE(const float* src,
+                   const float* k1,
+                   const float* k2,
+                   double kernel_interpolation_factor) {
+  __m128 m_input;
+  __m128 m_sums1 = _mm_setzero_ps();
+  __m128 m_sums2 = _mm_setzero_ps();
+
+  // Based on |input_ptr| alignment, we need to use loadu or load.  Unrolling
+  // these loops hurt performance in local testing.
+  if (reinterpret_cast<uintptr_t>(src) & 0x0F) {
+    for (int i = 0; i < kKernelSize; i += 4) {
+      m_input = _mm_loadu_ps(src + i);
+      m_sums1 = _mm_add_ps(m_sums1, _mm_mul_ps(m_input, _mm_load_ps(k1 + i)));
+      m_sums2 = _mm_add_ps(m_sums2, _mm_mul_ps(m_input, _mm_load_ps(k2 + i)));
+    }
+  } else {
+    for (int i = 0; i < kKernelSize; i += 4) {
+      m_input = _mm_load_ps(src + i);
+      m_sums1 = _mm_add_ps(m_sums1, _mm_mul_ps(m_input, _mm_load_ps(k1 + i)));
+      m_sums2 = _mm_add_ps(m_sums2, _mm_mul_ps(m_input, _mm_load_ps(k2 + i)));
+    }
+  }
+
+  // Linearly interpolate the two "convolutions".
+  m_sums1 = _mm_mul_ps(
+      m_sums1,
+      _mm_set_ps1(static_cast<float>(1.0 - kernel_interpolation_factor)));
+  m_sums2 = _mm_mul_ps(
+      m_sums2, _mm_set_ps1(static_cast<float>(kernel_interpolation_factor)));
+  m_sums1 = _mm_add_ps(m_sums1, m_sums2);
+
+  // Sum components together.
+  float result;
+  m_sums2 = _mm_add_ps(_mm_movehl_ps(m_sums1, m_sums1), m_sums1);
+  _mm_store_ss(&result,
+               _mm_add_ss(m_sums2, _mm_shuffle_ps(m_sums2, m_sums2, 1)));
+
+  return result;
+}
+
+float DotProduct_SSE(const float* a, const float* b, int len) {
   const int rem = len % 4;
   const int last_index = len - rem;
-  __m128 m_scale = _mm_set_ps1(scale);
+
+  // 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 + s), _mm_loadu_ps(b + 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.
+  float sum;
+  m_sum = _mm_add_ps(_mm_movehl_ps(m_sum, m_sum), m_sum);
+  _mm_store_ss(&sum, _mm_add_ss(m_sum, _mm_shuffle_ps(m_sum, m_sum, 1)));
+
+  // Handle any remaining values that wouldn't fit in an SSE pass.
+  for (int i = last_index; i < len; ++i)
+    sum += a[i] * b[i];
+
+  return sum;
+}
+
+void FMUL_SSE(const float* src, float scale, int len, float* dest) {
+  const int rem = len % 4;
+  const int last_index = len - rem;
+  const __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));
 
@@ -102,13 +245,14 @@ void FMUL_SSE(const float src[], float scale, int len, float dest[]) {
     dest[i] = src[i] * scale;
 }
 
-void FMAC_SSE(const float src[], float scale, int len, float dest[]) {
+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);
+  const __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)));
+    _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.
@@ -120,12 +264,12 @@ void FMAC_SSE(const float src[], float scale, int len, float dest[]) {
 // 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)))
+  (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) {
+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:
   //
@@ -159,10 +303,8 @@ std::pair<float, float> EWMAAndMaxPower_SSE(
     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));
+    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])
@@ -195,13 +337,61 @@ std::pair<float, float> EWMAAndMaxPower_SSE(
 #endif
 
 #if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
-void FMAC_NEON(const float src[], float scale, int len, float dest[]) {
+float Convolve_NEON(const float* src,
+                    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);
+
+  const float* upper = src + kKernelSize;
+  for (; input_ptr < upper;) {
+    m_input = vld1q_f32(src);
+    src += 4;
+    m_sums1 = vmlaq_f32(m_sums1, m_input, vld1q_f32(k1));
+    k1 += 4;
+    m_sums2 = vmlaq_f32(m_sums2, m_input, vld1q_f32(k2));
+    k2 += 4;
+  }
+
+  // Linearly interpolate the two "convolutions".
+  m_sums1 = vmlaq_f32(
+      vmulq_f32(m_sums1, vmovq_n_f32(1.0 - kernel_interpolation_factor)),
+      m_sums2, vmovq_n_f32(kernel_interpolation_factor));
+
+  // 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);
+}
+
+float DotProduct_NEON(const float* a, const float* b, int len) {
+  const int rem = len % 4;
+  const int last_index = len - rem;
+
+  // 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 + s), vld1q_f32(b + 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));
+  float sum = vget_lane_f32(vpadd_f32(m_half, m_half), 0);
+
+  // Handle any remaining values that wouldn't fit in an NEON pass.
+  for (int i = last_index; i < len; ++i)
+    sum += a[i] * b[i];
+
+  return sum;
+}
+
+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);
+  const 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));
+    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.
@@ -209,10 +399,10 @@ void FMAC_NEON(const float src[], float scale, int len, float dest[]) {
     dest[i] += src[i] * scale;
 }
 
-void FMUL_NEON(const float src[], float scale, int len, float dest[]) {
+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);
+  const 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));
 
@@ -221,8 +411,10 @@ void FMUL_NEON(const float src[], float scale, int len, float dest[]) {
     dest[i] = src[i] * scale;
 }
 
-std::pair<float, float> EWMAAndMaxPower_NEON(
-    float initial_value, const float src[], int len, float smoothing_factor) {
+std::pair<float, float> EWMAAndMaxPower_NEON(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:
   //