Adding new functionality for SIMD optimizations in AEC3

webrtc/modules/audio_processing/aec3/vector_math.h

Index: webrtc/modules/audio_processing/aec3/vector_math.h
diff --git a/webrtc/modules/audio_processing/aec3/vector_math.h b/webrtc/modules/audio_processing/aec3/vector_math.h
new file mode 100644
index 0000000000000000000000000000000000000000..44bde257e2bca7151d368b0f5ce3ab9c91858f03
--- /dev/null
+++ b/webrtc/modules/audio_processing/aec3/vector_math.h
@@ -0,0 +1,128 @@
+/*
+ *  Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AEC3_VECTOR_MATH_H_
+#define WEBRTC_MODULES_AUDIO_PROCESSING_AEC3_VECTOR_MATH_H_
+
+#include "webrtc/typedefs.h"

[aleloi, 2017/04/11 14:27:30]

Project headers should be after system headers.

+#if defined(WEBRTC_ARCH_X86_FAMILY)
+#include <emmintrin.h>
+#endif
+#include <math.h>
+#include <algorithm>
+#include <array>
+#include <functional>
+
+#include "webrtc/base/array_view.h"
+#include "webrtc/base/checks.h"
+#include "webrtc/modules/audio_processing/aec3/aec3_common.h"
+
+namespace webrtc {
+namespace aec3 {
+
+// Provides optimizations for mathematical operations based on vectors.
+class VectorMath {
+ public:
+  explicit VectorMath(Aec3Optimization optimization)
+      : optimization_(optimization) {}
+
+  // Elementwise square root.
+  void Sqrt(rtc::ArrayView<float> x) {
+    switch (optimization_) {
+#if defined(WEBRTC_ARCH_X86_FAMILY)
+      case Aec3Optimization::kSse2: {
+        const size_t kVectorLimit = x.size() >> 2;
+        const size_t kRemaining = x.size() - kVectorLimit * 4;

[aleloi, 2017/04/11 14:27:30]

I think automatic variables are usually named with lower_case in webrtc. The
style guide is ok with kConstantNaming though.

+
+        int j = 0;
+        for (size_t k = 0; k < kVectorLimit; ++k, j += 4) {

[ivoc, 2017/04/11 12:03:01]

I think the 2 different loop variables are a bit confusing. An alternative is to
do it like this:

int j = 0;
for (; j<kVectorLimit*4; j+=4) {
  ...
}
for (; j<x.size(); j++) {
  ...
}

[peah-webrtc, 2017/04/11 13:09:26]

Awesome! Thanks!

Done.

+          __m128 g = _mm_loadu_ps(&x[j]);
+          g = _mm_sqrt_ps(g);
+          _mm_storeu_ps(&x[j], g);
+        }
+
+        for (size_t k = 0; k < kRemaining; ++k, ++j) {
+          x[j] = sqrtf(x[j]);
+        }
+      } break;
+#endif
+      default:
+        std::for_each(x.begin(), x.end(), [](float& a) { a = sqrtf(a); });
+    }
+  }
+
+  // Elementwise vector multiplication z = x * y.
+  void Multiply(rtc::ArrayView<const float> x,
+                rtc::ArrayView<const float> y,
+                rtc::ArrayView<float> z) {
+    RTC_DCHECK_EQ(z.size(), x.size());
+    RTC_DCHECK_EQ(z.size(), y.size());
+    switch (optimization_) {
+#if defined(WEBRTC_ARCH_X86_FAMILY)
+      case Aec3Optimization::kSse2: {
+        const size_t kVectorLimit = x.size() >> 2;
+        const size_t kRemaining = x.size() - kVectorLimit * 4;
+
+        int j = 0;
+        for (size_t k = 0; k < kVectorLimit; ++k, j += 4) {

[ivoc, 2017/04/11 12:03:01]

Same here.

[peah-webrtc, 2017/04/11 13:09:26]

Done.

+          const __m128 x_k = _mm_loadu_ps(&x[j]);
+          const __m128 y_k = _mm_loadu_ps(&y[j]);
+          const __m128 z_k = _mm_mul_ps(x_k, y_k);
+          _mm_storeu_ps(&z[j], z_k);
+        }
+
+        for (size_t k = 0; k < kRemaining; ++k, ++j) {
+          z[j] = x[j] * y[j];
+        }
+      } break;
+#endif
+      default:
+        std::transform(x.begin(), x.end(), y.begin(), z.begin(),
+                       std::multiplies<float>());

[aleloi, 2017/04/11 14:27:30]

Are we sure std::transform doesn't do this kind of optimization already? I don't
know what flags to pass to the compiler to enable SSE2 in godbolt.

+    }
+  }
+
+  // Elementwise vector accumulation z += x.
+  void Accumulate(rtc::ArrayView<const float> x, rtc::ArrayView<float> z) {
+    RTC_DCHECK_EQ(z.size(), x.size());
+    switch (optimization_) {
+#if defined(WEBRTC_ARCH_X86_FAMILY)
+      case Aec3Optimization::kSse2: {
+        const size_t kVectorLimit = x.size() >> 2;
+        const size_t kRemaining = x.size() - kVectorLimit * 4;
+
+        int j = 0;
+        for (size_t k = 0; k < kVectorLimit; ++k, j += 4) {

[ivoc, 2017/04/11 12:03:01]

And here.

[peah-webrtc, 2017/04/11 13:09:26]

Done.

+          const __m128 x_k = _mm_loadu_ps(&x[j]);
+          __m128 z_k = _mm_loadu_ps(&z[j]);
+          z_k = _mm_add_ps(x_k, z_k);
+          _mm_storeu_ps(&z[j], z_k);
+        }
+
+        for (size_t k = 0; k < kRemaining; ++k, ++j) {
+          z[j] += x[j];
+        }
+      } break;
+#endif
+      default:
+        std::transform(x.begin(), x.end(), z.begin(), z.begin(),
+                       std::plus<float>());
+    }
+  }
+
+ private:
+  Aec3Optimization optimization_;
+};
+
+}  // namespace aec3
+
+}  // namespace webrtc
+
+#endif  // WEBRTC_MODULES_AUDIO_PROCESSING_AEC3_VECTOR_MATH_H_