From 99b87744ba22df78e3c476c92945b75acd267b87 Mon Sep 17 00:00:00 2001
From: Konstantin <const@const.me>
Date: Sun, 22 Jan 2023 13:28:24 +0100
Subject: VAD CPU performance, slightly better code generation

---
 Whisper/Whisper/melSpectrogram.cpp         |  2 --
 Whisper/Whisper/voiceActivityDetection.cpp | 10 ++++++++--
 Whisper/stdafx.h                           |  4 ++++
 3 files changed, 12 insertions(+), 4 deletions(-)

diff --git a/Whisper/Whisper/melSpectrogram.cpp b/Whisper/Whisper/melSpectrogram.cpp
index a178d16..8056e23 100644
--- a/Whisper/Whisper/melSpectrogram.cpp
+++ b/Whisper/Whisper/melSpectrogram.cpp
@@ -1,8 +1,6 @@
 #include "stdafx.h"
 #include <cmath>
 #include "melSpectrogram.h"
-#define _XM_SSE4_INTRINSICS_
-#include <DirectXMath.h>
 
 namespace Whisper
 {
diff --git a/Whisper/Whisper/voiceActivityDetection.cpp b/Whisper/Whisper/voiceActivityDetection.cpp
index 31763ec..c0eb7ef 100644
--- a/Whisper/Whisper/voiceActivityDetection.cpp
+++ b/Whisper/Whisper/voiceActivityDetection.cpp
@@ -1,6 +1,5 @@
 #include "stdafx.h"
 #include "voiceActivityDetection.h"
-#include <DirectXMath.h>
 using namespace Whisper;
 
 // Initially ported (poorly) from there https://github.com/panmasuo/voice-activity-detection MIT license
@@ -56,6 +55,13 @@ void VAD::fft() const
 
 constexpr float mulInt16FromFloat = 32768.0;
 
+inline float squareRoot( float x )
+{
+	__m128 v = _mm_set_ss( x );
+	v = _mm_sqrt_ss( v );
+	return _mm_cvtss_f32( v );
+}
+
 float VAD::computeEnergy( const float* rsi )
 {
 	// calculate_energy
@@ -67,7 +73,7 @@ float VAD::computeEnergy( const float* rsi )
 		f *= f;
 		sum += f;
 	}
-	return std::sqrtf( (float)( sum * ( 1.0 / FFT_POINTS ) ) );
+	return squareRoot( (float)( sum * ( 1.0 / FFT_POINTS ) ) );
 }
 
 float VAD::computeDominant( const cplx* spectrum )
diff --git a/Whisper/stdafx.h b/Whisper/stdafx.h
index 5df6e1b..907980e 100644
--- a/Whisper/stdafx.h
+++ b/Whisper/stdafx.h
@@ -17,7 +17,11 @@
 #include <sdkddkver.h>
 
 #include <windows.h>
+
+#define _XM_SSE4_INTRINSICS_
 #include <d3d11.h>
+#include <DirectXMath.h>
+
 #include <atlcomcli.h>
 #include "Utils/Logger.h"
 #include "Utils/miscUtils.h"
-- 
cgit v1.2.3