4 files changed, 85 insertions, 10 deletions

diff --git a/makefile b/makefile
index 41b32d433..af097c5dd 100644
--- a/makefile
+++ b/makefile
@@ -10,7 +10,7 @@
 # Macros
 #
 
-CC = /usr/bin/g++
+CC = /usr/bin/g++ -msse4
 CC_OPTIONS = -O2 -s
 CCE_OPTIONS = -O2 -s -x c
 LNK_OPTIONS = -lstdc++ -pthread
diff --git a/source/cChunk.cpp b/source/cChunk.cpp
index a8e8938a0..8e03ff6c7 100644
--- a/source/cChunk.cpp
+++ b/source/cChunk.cpp
@@ -678,13 +678,13 @@ void cChunk::SpreadLight(char* a_LightBuffer)
 

 float GetNoise( float x, float y, cNoise & a_Noise )

 {

-	float oct1 = a_Noise.CubicNoise2D( x*cGenSettings::HeightFreq1, y*cGenSettings::HeightFreq1 )*cGenSettings::HeightAmp1;

-	float oct2 = a_Noise.CubicNoise2D( x*cGenSettings::HeightFreq2, y*cGenSettings::HeightFreq2 )*cGenSettings::HeightAmp2;

-	float oct3 = a_Noise.CubicNoise2D( x*cGenSettings::HeightFreq3, y*cGenSettings::HeightFreq3 )*cGenSettings::HeightAmp3;

+	float oct1 = a_Noise.SSE_CubicNoise2D( x*cGenSettings::HeightFreq1, y*cGenSettings::HeightFreq1 )*cGenSettings::HeightAmp1;

+	float oct2 = a_Noise.SSE_CubicNoise2D( x*cGenSettings::HeightFreq2, y*cGenSettings::HeightFreq2 )*cGenSettings::HeightAmp2;

+	float oct3 = a_Noise.SSE_CubicNoise2D( x*cGenSettings::HeightFreq3, y*cGenSettings::HeightFreq3 )*cGenSettings::HeightAmp3;

 

-	float height = a_Noise.CubicNoise2D( x*0.1f, y*0.1f )*2;

+	float height = a_Noise.SSE_CubicNoise2D( x*0.1f, y*0.1f )*2;

 

-	float flatness = ((a_Noise.CubicNoise2D( x*0.5f, y*0.5f ) + 1.f ) * 0.5f) * 1.1f; // 0 ... 1.5

+	float flatness = ((a_Noise.SSE_CubicNoise2D( x*0.5f, y*0.5f ) + 1.f ) * 0.5f) * 1.1f; // 0 ... 1.5

 	flatness *= flatness * flatness;

 

 	return (oct1 + oct2 + oct3) * flatness + height;

@@ -836,8 +836,8 @@ void cChunk::GenerateTerrain()
 //				int yy = TopY;

 				int zz = z + m_PosZ*16;

 				

-				float val1 = m_Noise.CubicNoise2D( xx*0.1f, zz*0.1f );

-				float val2 = m_Noise.CubicNoise2D( xx*0.01f, zz*0.01f );

+				float val1 = m_Noise.SSE_CubicNoise2D( xx*0.1f, zz*0.1f );

+				float val2 = m_Noise.SSE_CubicNoise2D( xx*0.01f, zz*0.01f );

 				if( m_BlockType[index] == SandID )

 				{

 					if( (val1 + val2 > 0.f) && (rand()%128) > 124 && m_BlockType[index] == E_BLOCK_SAND )

@@ -852,8 +852,8 @@ void cChunk::GenerateTerrain()
 				}

 				else if( m_BlockType[index] == GrassID )

 				{

-					float val3 = m_Noise.CubicNoise2D( xx*0.01f+10, zz*0.01f+10 );

-					float val4 = m_Noise.CubicNoise2D( xx*0.05f+20, zz*0.05f+20 );

+					float val3 = m_Noise.SSE_CubicNoise2D( xx*0.01f+10, zz*0.01f+10 );

+					float val4 = m_Noise.SSE_CubicNoise2D( xx*0.05f+20, zz*0.05f+20 );

 					if( val1 + val2 > 0.2f && (rand()%128) > 124 )

 						cRoot::Get()->GetWorld()->GrowTree( xx, TopY, zz );

 					else if( val3 > 0.2f && (rand()%128) > 124 )

diff --git a/source/cNoise.cpp b/source/cNoise.cpp
index 1d321c089..1bdc52698 100644
--- a/source/cNoise.cpp
+++ b/source/cNoise.cpp
@@ -1,6 +1,8 @@
 #include "cNoise.h"

 #include <math.h>

 

+#include <smmintrin.h> //_mm_mul_epi32

+

 #define FAST_FLOOR( x ) ( (x) < 0 ? ((int)x)-1 : ((int)x) )

 

 cNoise::cNoise( unsigned int a_Seed )

@@ -36,6 +38,41 @@ float cNoise::IntNoise3D( int a_X, int a_Y, int a_Z ) const
     return ( 1.0f - ( (n * (n * n * 15731 + 789221) + 1376312589) & 0x7fffffff) / 1073741824.0f);  

 }

 

+/****************

+ * SSE Random value generator

+ **/

+__m128 cNoise::SSE_IntNoise2D( int a_X1, int a_Y1, int a_X2, int a_Y2, int a_X3, int a_Y3, int a_X4, int a_Y4 ) const

+{

+	const __m128i X4 = _mm_set_epi32(a_X4, a_X3, a_X2, a_X1);

+	const __m128i Y4 = _mm_set_epi32(a_Y4, a_Y3, a_Y2, a_Y1);

+

+	const __m128 One4 = _mm_set_ps1( 1.f );

+	const __m128i YScale4 = _mm_set1_epi32( 57 );

+

+	const __m128i i15731 = _mm_set1_epi32( 15731 );

+	const __m128i i789221 = _mm_set1_epi32( 789221 );

+	const __m128i i1376312589 = _mm_set1_epi32(1376312589);

+	const __m128i MaskValue4 = _mm_set1_epi32(0x7fffffff);

+	const __m128 f1073741824 = _mm_set_ps1( 1073741824.0f );

+

+	const __m128i Seed4 = _mm_mullo_epi32( _mm_mullo_epi32( _mm_set1_epi32( m_Seed ), YScale4 ), YScale4 );

+

+	const __m128i ScaledY4 = _mm_mullo_epi32( Y4, YScale4 );

+	const __m128i n4 = _mm_add_epi32( _mm_add_epi32( X4, ScaledY4 ), Seed4 );

+	const __m128i nn4 = _mm_slli_epi32( n4, 13 );

+	const __m128i nnn4 = _mm_xor_si128( nn4, n4 );

+

+	const __m128i StepA4 = _mm_mullo_epi32( nnn4, nnn4 );

+	const __m128i StepAA4 = _mm_add_epi32( _mm_mullo_epi32( StepA4, i15731 ), i789221 );

+	const __m128i StepB4 = _mm_add_epi32( _mm_mullo_epi32( nnn4, StepAA4 ), i1376312589 );

+	const __m128i StepC4 = _mm_and_si128( StepB4, MaskValue4 );

+	const __m128 StepD4 = _mm_div_ps( _mm_cvtepi32_ps( StepC4 ), f1073741824 );

+	const __m128 Result4 = _mm_sub_ps( One4, StepD4 );

+

+	return Result4;

+}

+

+

 /***************

  * Interpolated (and 1 smoothed) noise in 1-dimension

  **/

@@ -129,6 +166,26 @@ float cNoise::CubicNoise2D( float a_X, float a_Y ) const
 	return CubicInterpolate( interp1, interp2, interp3, interp4, FracY );

 }

 

+float cNoise::SSE_CubicNoise2D( float a_X, float a_Y ) const

+{

+	const int	BaseX = FAST_FLOOR( a_X );

+	const int	BaseY = FAST_FLOOR( a_Y );

+

+	__m128 points4[4] = {

+		SSE_IntNoise2D( BaseX-1, BaseY-1,	BaseX-1, BaseY, BaseX-1, BaseY+1,	BaseX-1, BaseY+2 ),

+		SSE_IntNoise2D( BaseX,	 BaseY-1,	BaseX, BaseY,	BaseX, BaseY+1,		BaseX, BaseY+2 ),

+		SSE_IntNoise2D( BaseX+1, BaseY-1,	BaseX+1, BaseY, BaseX+1, BaseY+1,	BaseX+1, BaseY+2 ),

+		SSE_IntNoise2D( BaseX+2, BaseY-1,	BaseX+2, BaseY,	BaseX+2, BaseY+1,	BaseX+2, BaseY+2 ),

+	};

+

+	const float	FracX = (a_X) - BaseX;

+	union { __m128 p4; float p[4]; } 

+	AllInterp = { CubicInterpolate4( points4[0], points4[1], points4[2], points4[3], FracX ) };

+

+	const float	FracY = (a_Y) - BaseY;

+	return CubicInterpolate( AllInterp.p[0], AllInterp.p[1], AllInterp.p[2], AllInterp.p[3], FracY );

+}

+

 /******************

  * Interpolated (and 1 smoothed) noise in 3-dimensions

  **/

@@ -240,6 +297,19 @@ float cNoise::CubicInterpolate( float a_A, float a_B, float a_C, float a_D, floa
 	return P*(a_Pct*a_Pct*a_Pct) + Q*(a_Pct*a_Pct) + R*a_Pct + S;

 }

 

+__m128 cNoise::CubicInterpolate4( const __m128 & a_A, const __m128 & a_B, const __m128 & a_C, const __m128 & a_D, float a_Pct ) const

+{

+	const __m128 P = _mm_sub_ps( _mm_sub_ps( a_D, a_C ), _mm_sub_ps( a_A, a_B ) );

+	const __m128 Q = _mm_sub_ps( _mm_sub_ps( a_A, a_B ), P );

+	const __m128 R = _mm_sub_ps( a_C, a_A );

+

+	const __m128 Pct = _mm_set_ps1( a_Pct );

+	const __m128 Pct2 = _mm_mul_ps( Pct, Pct );

+	const __m128 Pct3 = _mm_mul_ps( Pct2, Pct );

+

+	return _mm_add_ps( _mm_add_ps( _mm_add_ps( _mm_mul_ps(P, Pct3), _mm_mul_ps( Q, Pct2 ) ), _mm_mul_ps( R, Pct ) ), a_B );

+}

+

 float cNoise::CosineInterpolate( float a_A, float a_B, float a_Pct ) const

 {

 	const float ft = a_Pct * 3.1415927f;

diff --git a/source/cNoise.h b/source/cNoise.h
index 538970da3..9511ab6e4 100644
--- a/source/cNoise.h
+++ b/source/cNoise.h
@@ -1,5 +1,7 @@
 #pragma once

 

+#include <emmintrin.h>

+

 class cNoise

 {

 public:

@@ -8,6 +10,7 @@ public:
 

 	float IntNoise( int a_X ) const;

 	float IntNoise2D( int a_X, int a_Y ) const;

+	__m128 SSE_IntNoise2D( int a_X1, int a_Y1, int a_X2, int a_Y2, int a_X3, int a_Y3, int a_X4, int a_Y4 ) const;

 	float IntNoise3D( int a_X, int a_Y, int a_Z ) const;

 

 	float LinearNoise1D( float a_X ) const;

@@ -18,6 +21,7 @@ public:
 	float LinearNoise2D( float a_X, float a_Y ) const;

 	float CosineNoise2D( float a_X, float a_Y ) const;

 	float CubicNoise2D( float a_X, float a_Y ) const;

+	float SSE_CubicNoise2D( float a_X, float a_Y ) const;

 

 	float CosineNoise3D( float a_X, float a_Y, float a_Z ) const;

 	float CubicNoise3D( float a_X, float a_Y, float a_Z ) const;

@@ -25,6 +29,7 @@ public:
 	void SetSeed( unsigned int a_Seed ) { m_Seed = a_Seed; }

 private:

 	float CubicInterpolate( float a_A, float a_B, float a_C, float a_D, float a_Pct ) const;

+	__m128 CubicInterpolate4( const __m128 & a_A, const __m128 & a_B, const __m128 & a_C, const __m128 & a_D, float a_Pct ) const;

 	float CosineInterpolate( float a_A, float a_B, float a_Pct ) const;

 	float LinearInterpolate( float a_A, float a_B, float a_Pct ) const;