SSE.h

Go to the documentation of this file.

//**************************************************************************/
// Copyright (c) 2008 Autodesk, Inc.
// All rights reserved.
//
// Use of this software is subject to the terms of the Autodesk license
// agreement provided at the time of installation or download, or which
// otherwise accompanies this software in either electronic or hard copy form.
//
//**************************************************************************/
// DESCRIPTION:
// CREATED: October 2008
//**************************************************************************/

#ifndef __MUDBOXSDK_SSE_H__
#define __MUDBOXSDK_SSE_H__

#if defined(JAMBUILD)
#include <Mudbox/mudbox.h>
#else
#include "mudbox.h"
#endif

#if defined(__GNUC__)

#ifndef __SSE3__
#error This file was intended to compiled with SSE3 instruction set enabled.
#endif

//
// If you are using GCC instead of the Intel C Compiler, don't forget 
// to specify -I/usr/lib/gcc/i686-apple-darwin9/4.0.1/include when compiling
// a file that use this header.
//
#include <xmmintrin.h>
#include <pmmintrin.h>
#endif

#if defined(WIN32) || defined(WIN64)
#include <xmmintrin.h>
#include <intrin.h>
#endif

#ifndef MB_SSE_ALIGN16_VAR

#if defined(_MSC_VER)
#define MB_SSE_ALIGN16_VAR(v) __declspec(align(16)) v
#define MB_SSE_ALIGN16_CLASS  __declspec(align(16))
#ifndef __attribute__
#define __attribute__(_a)
#endif
#elif defined(__GNUC__)
#define MB_SSE_ALIGN16_VAR(v) v __attribute__ ((aligned(16)))
#define MB_SSE_ALIGN16_CLASS  __attribute__ ((aligned(16)))
#ifndef __forceinline
#define __forceinline inline
#endif
#else
#error The MB_SSE_ALIGN16_VAR needs to be ported for this compiler.
#endif

#endif

struct MBDLL_DECL HWVector
{
    inline HWVector( void ) __attribute__((always_inline)){};

    inline void setZero( void ) __attribute__((always_inline))
    {
        v = _mm_setzero_ps();
    }

    inline HWVector( float f ) __attribute__((always_inline))
    {
        v = _mm_set_ps1( f );
    }

    inline HWVector( const mudbox::Vector &h ) __attribute__((always_inline))
    { 
        v = _mm_set_ps( h.x, h.y, h.z, 0); 
    };
    inline HWVector( float x, float y, float z, float w = 0 ) __attribute__((always_inline))
    { 
        v = _mm_set_ps( x,y,z,w ); 
    };
    void Fill( const mudbox::Vector &vVector );
    inline HWVector( const float a[4] ) __attribute__((always_inline))
    {
        v = _mm_loadu_ps( a );
    };
    inline HWVector ShiftLeft( void ) const    __attribute__((always_inline)){ HWVector r; r.v = _mm_shuffle_ps(v,v,_MM_SHUFFLE(2,1,3,0)); return r; };
    inline void ShiftLeft(HWVector &result)    __attribute__((always_inline)){ result.v = _mm_shuffle_ps(v,v,_MM_SHUFFLE(2,1,3,0)); }
    inline void ShiftLeftInPlace()             __attribute__((always_inline)){ v = _mm_shuffle_ps(v,v,_MM_SHUFFLE(2,1,3,0)); }
    inline HWVector ShiftRight( void ) const   __attribute__((always_inline)){ HWVector r; r.v = _mm_shuffle_ps(v,v,_MM_SHUFFLE(1,3,2,0)); return r; };
    inline void ShiftRight( HWVector &result ) __attribute__((always_inline)){ result.v = _mm_shuffle_ps(v,v,_MM_SHUFFLE(1,3,2,0)); };
    inline HWVector operator &(  HWVector &o ) __attribute__((always_inline))
    {
        //HWVector al = (*this).ShiftLeft(), bl = o.ShiftLeft(), ar = (*this).ShiftRight(), br = o.ShiftRight();
        HWVector al, bl, ar, br;
        ShiftLeft(al);
        o.ShiftLeft(bl);
        ShiftRight(ar);
        o.ShiftRight(br);

        return al*br-ar*bl;
    };

    inline HWVector operator |( const HWVector &o ) const __attribute__((always_inline))
    {
        HWVector r;
        r.v = _mm_mul_ps( v, o.v );
        // Requires SSE3
        r.v = _mm_hadd_ps( r.v, r.v );
        r.v = _mm_hadd_ps( r.v, r.v );
        return r;
    };

    inline HWVector Length( void ) const __attribute__((always_inline))
    {
        HWVector r = operator |( *this );
        r.v = _mm_sqrt_ss( r.v );
        return r;
    };

    inline HWVector LengthSquared( void ) const __attribute__((always_inline))
    {
        HWVector r = operator |( *this );
        return r;
    };

    inline float DistanceFromLine( const HWVector &vStart,const HWVector &vEnd ) const __attribute__((always_inline))
    {
        HWVector n = vEnd - vStart;
        HWVector m = (*this)-vStart;
        HWVector h = n&m;
        HWVector d = n&h;
        d.Normalize();
        float fDistance = (d|(*this))-(d|vStart);
        return fDistance>0?fDistance:(-fDistance);
    }

    inline void Normalize( void ) __attribute__((always_inline))
    {
        HWVector f = operator |( *this );
        f.v = _mm_rsqrt_ps( f.v );
        v = _mm_mul_ps( f.v, v );
    };

    inline HWVector Floor( void ) __attribute__((always_inline))
    {
        static const unsigned int a        = (1 << 23);
        static const float  twoTo23AsFloat = (float)a;
        static const __m128 twoTo23 = _mm_set_ps( twoTo23AsFloat, twoTo23AsFloat, twoTo23AsFloat, twoTo23AsFloat );
        // b = fabs(v)
        __m128 b = _mm_castsi128_ps(_mm_srli_epi32( _mm_slli_epi32( _mm_castps_si128(v),1 ), 1 ));
        // The essence of the floor routine
        __m128 d = _mm_sub_ps( _mm_add_ps( _mm_add_ps( _mm_sub_ps( v, twoTo23 ), twoTo23 ), twoTo23 ), twoTo23 );
        // set mask to all 1s if v >= 2**23
        __m128 largeMaskE = _mm_cmpgt_ps( b, twoTo23 );
        // Check for possible off by one error
        __m128 g = _mm_cmplt_ps( v, d );
        // Convert positive check result to -1.0, negative to 0.0
        __m128 h = _mm_cvtepi32_ps( _mm_castps_si128(g) );
        // Add in the error if there is one
        __m128 t = _mm_add_ps( d, h );
        // Select between output result and input value based on v >= 2**23
        __m128 w = _mm_and_ps( v, largeMaskE );
        t = _mm_andnot_ps( largeMaskE, t );
        HWVector vResult;
        vResult.v = _mm_or_ps( t, w );
        return vResult;
    };

    inline HWVector Minimum( const HWVector &o ) const    __attribute__((always_inline)){ HWVector r; r.v = _mm_min_ps( v, o.v ); return r; };
    inline HWVector Maximum( const HWVector &o ) const    __attribute__((always_inline)){ HWVector r; r.v = _mm_max_ps( v, o.v ); return r; };
    inline HWVector operator +( const HWVector &o ) const __attribute__((always_inline)){ HWVector r; r.v = _mm_add_ps( v, o.v ); return r; };
    inline HWVector operator -( const HWVector &o ) const __attribute__((always_inline)){ HWVector r; r.v = _mm_sub_ps( v, o.v ); return r; };
    inline void operator +=( const HWVector &o ) __attribute__((always_inline)){ v = _mm_add_ps( v, o.v ); };
    inline void operator -=( const HWVector &o ) __attribute__((always_inline)){ v = _mm_sub_ps( v, o.v ); };
    inline void operator *=( const HWVector &o ) __attribute__((always_inline)){ v = _mm_mul_ps( v, o.v ); };
    inline void operator /=( const HWVector &o ) __attribute__((always_inline)){ v = _mm_div_ps( v, o.v ); };
    inline HWVector operator *( const HWVector &o ) const __attribute__((always_inline)){ HWVector r; r.v = _mm_mul_ps( v, o.v ); return r; };
    inline HWVector operator *( float f ) const __attribute__((always_inline)){ HWVector r; r.v = _mm_mul_ps( v, _mm_set1_ps( f ) ); return r; };
    inline HWVector operator /( const HWVector &o ) const __attribute__((always_inline)){ HWVector r; r.v = _mm_div_ps( v, o.v ); return r; };
    inline void operator *=( float f )  __attribute__((always_inline)){ v = _mm_mul_ps( v, _mm_set1_ps(f ) ); };
    inline void Store( float *p ) __attribute__((always_inline)){ _mm_storeu_ps( p, v ); };
    inline void Load( float f ) __attribute__((always_inline))  { v = _mm_set_ps1( f ); };
    inline void StoreNormalAsInt( int *pBuffer ) const __attribute__((always_inline))
    {
        //static __declspec(align(16)) float c[4] = { 32766.0f, 32766.0f, 32766.0f, 32766.0f };
        // instead of the correct value, we use a little bit smaller number, because after 
        // normalization a component can be a littlebit bigger than 1.0. in that case storing it in
        // a 16 bit integer would overflow, and artifacts on the surface would appear.
        static const MB_SSE_ALIGN16_VAR(float c[4]) = { 32740.0f, 32740.0f, 32740.0f, 32740.0f };

        __m128 f = _mm_load_ps( c );
        f = _mm_mul_ps( f, v );
        __m128i i = _mm_cvtps_epi32( f );
        _mm_storeu_si128( (__m128i *)pBuffer, i );
    };
    inline void StoreAsInt( int *pBuffer ) const __attribute__((always_inline))
    {
        __m128i i = _mm_cvtps_epi32( v );
        _mm_storeu_si128( (__m128i *)pBuffer, i );
    };
    inline void StoreAsIntTruncate( int *pBuffer ) const __attribute__((always_inline))
    {
        __m128i i = _mm_cvttps_epi32( v );
        _mm_storeu_si128( (__m128i *)pBuffer, i );
    };
    inline operator float( void ) const __attribute__((always_inline))
    {
        MB_SSE_ALIGN16_VAR(float f);
        _mm_store_ss( &f, v );
        return f;
    };
    inline operator mudbox::Vector( void ) const __attribute__((always_inline))
    {
        mudbox::Vector r;
        __m128 t = _mm_shuffle_ps( v, v, _MM_SHUFFLE(2,1,0,3) );
        _mm_store_ss( &r.x, t );
        t = _mm_shuffle_ps( t, t, _MM_SHUFFLE(2,1,0,3) );
        _mm_store_ss( &r.y, t );
        t = _mm_shuffle_ps( t, t, _MM_SHUFFLE(2,1,0,3) );
        _mm_store_ss( &r.z, t );
        return r;
    };
    inline operator mudbox::Vector4( void ) const __attribute__((always_inline))
    {
        mudbox::Vector4 r;
        __m128 t = _mm_shuffle_ps( v, v, _MM_SHUFFLE(2,1,0,3) );
        _mm_store_ss( &r.x, t );
        t = _mm_shuffle_ps( t, t, _MM_SHUFFLE(2,1,0,3) );
        _mm_store_ss( &r.y, t );
        t = _mm_shuffle_ps( t, t, _MM_SHUFFLE(2,1,0,3) );
        _mm_store_ss( &r.z, t );
        t = _mm_shuffle_ps( t, t, _MM_SHUFFLE(2,1,0,3) );
        _mm_store_ss( &r.w, t );
        return r;
    };

    MB_SSE_ALIGN16_VAR(__m128 v);
};

inline mudbox::Vector &operator <<( mudbox::Vector &v, const HWVector &r ) 
{
    __m128 t = _mm_shuffle_ps( r.v, r.v, _MM_SHUFFLE(2,1,0,3) );
    _mm_store_ss( &v.x, t );
    t = _mm_shuffle_ps( t, t, _MM_SHUFFLE(2,1,0,3) );
    _mm_store_ss( &v.y, t );
    t = _mm_shuffle_ps( t, t, _MM_SHUFFLE(2,1,0,3) );
    _mm_store_ss( &v.z, t );
    return v;

    //__m128 t = _mm_shuffle_ps( r.v, r.v, _MM_SHUFFLE(0,1,2,3) );
    //int a = ((int *)&v)[3];
    //_mm_storeu_ps( &v.x, t );
    //((int *)&v)[3] = a;
    //return v;
};


struct HWMatrix
{
    inline HWMatrix( void )  __attribute__((always_inline)){};
    inline HWMatrix( const mudbox::Matrix &m ) __attribute__((always_inline))
    {
        r0 = _mm_loadu_ps( &m._11 );
        r1 = _mm_loadu_ps( &m._21 );
        r2 = _mm_loadu_ps( &m._31 );
        r3 = _mm_loadu_ps( &m._41 );
    };
    void MirrorX( void ) __attribute__((always_inline))
    {
        r0 = _mm_shuffle_ps( r0, r0, _MM_SHUFFLE(0, 1, 2, 3) );
        r1 = _mm_shuffle_ps( r1, r1, _MM_SHUFFLE(0, 1, 2, 3) );
        r2 = _mm_shuffle_ps( r2, r2, _MM_SHUFFLE(0, 1, 2, 3) );
        r3 = _mm_shuffle_ps( r3, r3, _MM_SHUFFLE(0, 1, 2, 3) );
    };
    inline HWVector Transform( const HWVector &v, float w = 1.0f ) const __attribute__((always_inline))
    {
        // Represents v * M
        // v is a row vector.

        MB_SSE_ALIGN16_VAR(float c[4]) = { w, w, w, w };

        __m128 v0 = _mm_shuffle_ps( v.v, v.v, _MM_SHUFFLE(3,3,3,3) );
        __m128 v1 = _mm_shuffle_ps( v.v, v.v, _MM_SHUFFLE(2,2,2,2) );
        __m128 v2 = _mm_shuffle_ps( v.v, v.v, _MM_SHUFFLE(1,1,1,1) );
        __m128 v3 = _mm_load_ps( c );

        __m128 a0 = _mm_mul_ps( v0, r0 );
        __m128 a1 = _mm_mul_ps( v1, r1 );
        __m128 a2 = _mm_mul_ps( v2, r2 );
        __m128 a3 = _mm_mul_ps( v3, r3 );

        __m128 r = _mm_add_ps( a0, _mm_add_ps( a1, _mm_add_ps( a2, a3 ) ) );

        HWVector z;
        z.v = r;
        return z;
    };
    inline HWVector ProjectedTransform( const HWVector &v ) const __attribute__((always_inline))
    {
        static MB_SSE_ALIGN16_VAR(float c[4]) = { 1, 1, 1, 1 };

        __m128 v0 = _mm_shuffle_ps( v.v, v.v, _MM_SHUFFLE(3,3,3,3) );
        __m128 v1 = _mm_shuffle_ps( v.v, v.v, _MM_SHUFFLE(2,2,2,2) );
        __m128 v2 = _mm_shuffle_ps( v.v, v.v, _MM_SHUFFLE(1,1,1,1) );
        __m128 v3 = _mm_load_ps( c );

        __m128 a0 = _mm_mul_ps( v0, r0 );
        __m128 a1 = _mm_mul_ps( v1, r1 );
        __m128 a2 = _mm_mul_ps( v2, r2 );
        __m128 a3 = _mm_mul_ps( v3, r3 );

        __m128 r = _mm_add_ps( a0, _mm_add_ps( a1, _mm_add_ps( a2, a3 ) ) );
        __m128 d = _mm_shuffle_ps( r, r, _MM_SHUFFLE(3,3,3,3) );

        HWVector z;
        z.v = _mm_div_ps( r, d );
        return z;
    };
    
    MB_SSE_ALIGN16_VAR(__m128 r0);
    MB_SSE_ALIGN16_VAR(__m128 r1);
    MB_SSE_ALIGN16_VAR(__m128 r2);
    MB_SSE_ALIGN16_VAR(__m128 r3);
};


//-----------------------------------------------------------------------------
bool MBDLL_DECL hasSSE3();

bool MBDLL_DECL hasSSE41();

bool MBDLL_DECL hasSSE42();

bool MBDLL_DECL hasAVX256();

//-----------------------------------------------------------------------------

#endif