diff options
Diffstat (limited to 'indra/llmath')
42 files changed, 4663 insertions, 1054 deletions
diff --git a/indra/llmath/CMakeLists.txt b/indra/llmath/CMakeLists.txt index e93fe90650..9dadad7dd3 100644 --- a/indra/llmath/CMakeLists.txt +++ b/indra/llmath/CMakeLists.txt @@ -15,13 +15,16 @@ set(llmath_SOURCE_FILES llcamera.cpp llcoordframe.cpp llline.cpp + llmatrix3a.cpp llmodularmath.cpp llperlin.cpp llquaternion.cpp llrect.cpp llsphere.cpp + llvector4a.cpp llvolume.cpp llvolumemgr.cpp + llvolumeoctree.cpp llsdutil_math.cpp m3math.cpp m4math.cpp @@ -49,21 +52,32 @@ set(llmath_HEADER_FILES llinterp.h llline.h llmath.h + llmatrix3a.h + llmatrix3a.inl llmodularmath.h lloctree.h llperlin.h llplane.h llquantize.h llquaternion.h + llquaternion2.h + llquaternion2.inl llrect.h + llsimdmath.h + llsimdtypes.h + llsimdtypes.inl llsphere.h lltreenode.h + llvector4a.h + llvector4a.inl + llvector4logical.h llv4math.h llv4matrix3.h llv4matrix4.h llv4vector3.h llvolume.h llvolumemgr.h + llvolumeoctree.h llsdutil_math.h m3math.h m4math.h diff --git a/indra/llmath/llcamera.cpp b/indra/llmath/llcamera.cpp index 487ed6451f..beb5c48624 100644 --- a/indra/llmath/llcamera.cpp +++ b/indra/llmath/llcamera.cpp @@ -48,10 +48,10 @@ LLCamera::LLCamera() : mPlaneCount(6), mFrustumCornerDist(0.f) { + alignPlanes(); calculateFrustumPlanes(); } - LLCamera::LLCamera(F32 vertical_fov_rads, F32 aspect_ratio, S32 view_height_in_pixels, F32 near_plane, F32 far_plane) : LLCoordFrame(), mViewHeightInPixels(view_height_in_pixels), @@ -59,6 +59,7 @@ LLCamera::LLCamera(F32 vertical_fov_rads, F32 aspect_ratio, S32 view_height_in_p mPlaneCount(6), mFrustumCornerDist(0.f) { + alignPlanes(); mAspect = llclamp(aspect_ratio, MIN_ASPECT_RATIO, MAX_ASPECT_RATIO); mNearPlane = llclamp(near_plane, MIN_NEAR_PLANE, MAX_NEAR_PLANE); if(far_plane < 0) far_plane = DEFAULT_FAR_PLANE; @@ -67,6 +68,23 @@ LLCamera::LLCamera(F32 vertical_fov_rads, F32 aspect_ratio, S32 view_height_in_p setView(vertical_fov_rads); } +LLCamera::~LLCamera() +{ + +} + +const LLCamera& LLCamera::operator=(const LLCamera& rhs) +{ + memcpy(this, &rhs, sizeof(LLCamera)); + alignPlanes(); + LLVector4a::memcpyNonAliased16((F32*) mAgentPlanes, (F32*) rhs.mAgentPlanes, 4*7*sizeof(F32)); + return *this; +} + +void LLCamera::alignPlanes() +{ + mAgentPlanes = (LLPlane*) LL_NEXT_ALIGNED_ADDRESS<U8>(mAgentPlaneBuffer); +} // ---------------- LLCamera::getFoo() member functions ---------------- @@ -91,8 +109,8 @@ F32 LLCamera::getMaxView() const void LLCamera::setUserClipPlane(LLPlane plane) { mPlaneCount = 7; - mAgentPlanes[6].p = plane; - mAgentPlanes[6].mask = calcPlaneMask(plane); + mAgentPlanes[6] = plane; + mPlaneMask[6] = calcPlaneMask(plane); } void LLCamera::disableUserClipPlane() @@ -164,129 +182,66 @@ size_t LLCamera::readFrustumFromBuffer(const char *buffer) // ---------------- test methods ---------------- -S32 LLCamera::AABBInFrustum(const LLVector3 ¢er, const LLVector3& radius) -{ - static const LLVector3 scaler[] = { - LLVector3(-1,-1,-1), - LLVector3( 1,-1,-1), - LLVector3(-1, 1,-1), - LLVector3( 1, 1,-1), - LLVector3(-1,-1, 1), - LLVector3( 1,-1, 1), - LLVector3(-1, 1, 1), - LLVector3( 1, 1, 1) +S32 LLCamera::AABBInFrustum(const LLVector4a ¢er, const LLVector4a& radius) +{ + static const LLVector4a scaler[] = { + LLVector4a(-1,-1,-1), + LLVector4a( 1,-1,-1), + LLVector4a(-1, 1,-1), + LLVector4a( 1, 1,-1), + LLVector4a(-1,-1, 1), + LLVector4a( 1,-1, 1), + LLVector4a(-1, 1, 1), + LLVector4a( 1, 1, 1) }; U8 mask = 0; S32 result = 2; - /*if (mFrustumCornerDist > 0.f && radius.magVecSquared() > mFrustumCornerDist * mFrustumCornerDist) - { //box is larger than frustum, check frustum quads against box planes - - static const LLVector3 dir[] = - { - LLVector3(1, 0, 0), - LLVector3(-1, 0, 0), - LLVector3(0, 1, 0), - LLVector3(0, -1, 0), - LLVector3(0, 0, 1), - LLVector3(0, 0, -1) - }; - - U32 quads[] = + for (U32 i = 0; i < mPlaneCount; i++) + { + mask = mPlaneMask[i]; + if (mask == 0xff) { - 0, 1, 2, 3, - 0, 1, 5, 4, - 2, 3, 7, 6, - 3, 0, 7, 4, - 1, 2, 6, 4, - 4, 5, 6, 7 - }; - - result = 0; - - BOOL total_inside = TRUE; - for (U32 i = 0; i < 6; i++) - { - LLVector3 p = center + radius.scaledVec(dir[i]); - F32 d = -p*dir[i]; - - for (U32 j = 0; j < 6; j++) - { //for each quad - F32 dist = mAgentFrustum[quads[j*4+0]]*dir[i] + d; - if (dist > 0) - { //at least one frustum point is outside the AABB - total_inside = FALSE; - for (U32 k = 1; k < 4; k++) - { //for each other point on quad - if ( mAgentFrustum[quads[j*4+k]]*dir[i]+d <= 0.f) - { //quad is straddling some plane of AABB - return 1; - } - } - } - else - { - for (U32 k = 1; k < 4; k++) - { - if (mAgentFrustum[quads[j*4+k]]*dir[i]+d > 0.f) - { - return 1; - } - } - } - } + continue; } - if (total_inside) + const LLPlane& p = mAgentPlanes[i]; + const LLVector4a& n = reinterpret_cast<const LLVector4a&>(p); + float d = p.mV[3]; + LLVector4a rscale; + rscale.setMul(radius, scaler[mask]); + + LLVector4a minp, maxp; + minp.setSub(center, rscale); + maxp.setAdd(center, rscale); + + if (n.dot3(minp) > -d) { - result = 1; + return 0; } - } - else*/ - { - for (U32 i = 0; i < mPlaneCount; i++) + + if (n.dot3(maxp) > -d) { - mask = mAgentPlanes[i].mask; - if (mask == 0xff) - { - continue; - } - LLPlane p = mAgentPlanes[i].p; - LLVector3 n = LLVector3(p); - float d = p.mV[3]; - LLVector3 rscale = radius.scaledVec(scaler[mask]); - - LLVector3 minp = center - rscale; - LLVector3 maxp = center + rscale; - - if (n * minp > -d) - { - return 0; - } - - if (n * maxp > -d) - { - result = 1; - } + result = 1; } } - return result; } -S32 LLCamera::AABBInFrustumNoFarClip(const LLVector3 ¢er, const LLVector3& radius) -{ - static const LLVector3 scaler[] = { - LLVector3(-1,-1,-1), - LLVector3( 1,-1,-1), - LLVector3(-1, 1,-1), - LLVector3( 1, 1,-1), - LLVector3(-1,-1, 1), - LLVector3( 1,-1, 1), - LLVector3(-1, 1, 1), - LLVector3( 1, 1, 1) + +S32 LLCamera::AABBInFrustumNoFarClip(const LLVector4a& center, const LLVector4a& radius) +{ + static const LLVector4a scaler[] = { + LLVector4a(-1,-1,-1), + LLVector4a( 1,-1,-1), + LLVector4a(-1, 1,-1), + LLVector4a( 1, 1,-1), + LLVector4a(-1,-1, 1), + LLVector4a( 1,-1, 1), + LLVector4a(-1, 1, 1), + LLVector4a( 1, 1, 1) }; U8 mask = 0; @@ -299,25 +254,28 @@ S32 LLCamera::AABBInFrustumNoFarClip(const LLVector3 ¢er, const LLVector3& r continue; } - mask = mAgentPlanes[i].mask; + mask = mPlaneMask[i]; if (mask == 0xff) { continue; } - LLPlane p = mAgentPlanes[i].p; - LLVector3 n = LLVector3(p); + + const LLPlane& p = mAgentPlanes[i]; + const LLVector4a& n = reinterpret_cast<const LLVector4a&>(p); float d = p.mV[3]; - LLVector3 rscale = radius.scaledVec(scaler[mask]); + LLVector4a rscale; + rscale.setMul(radius, scaler[mask]); - LLVector3 minp = center - rscale; - LLVector3 maxp = center + rscale; + LLVector4a minp, maxp; + minp.setSub(center, rscale); + maxp.setAdd(center, rscale); - if (n * minp > -d) + if (n.dot3(minp) > -d) { return 0; } - if (n * maxp > -d) + if (n.dot3(maxp) > -d) { result = 1; } @@ -447,12 +405,12 @@ int LLCamera::sphereInFrustum(const LLVector3 &sphere_center, const F32 radius) int res = 2; for (int i = 0; i < 6; i++) { - if (mAgentPlanes[i].mask == 0xff) + if (mPlaneMask[i] == 0xff) { continue; } - float d = mAgentPlanes[i].p.dist(sphere_center); + float d = mAgentPlanes[i].dist(sphere_center); if (d > radius) { @@ -644,12 +602,14 @@ void LLCamera::ignoreAgentFrustumPlane(S32 idx) return; } - mAgentPlanes[idx].mask = 0xff; - mAgentPlanes[idx].p.clearVec(); + mPlaneMask[idx] = 0xff; + mAgentPlanes[idx].clearVec(); } void LLCamera::calcAgentFrustumPlanes(LLVector3* frust) { + alignPlanes(); + for (int i = 0; i < 8; i++) { mAgentFrustum[i] = frust[i]; @@ -662,27 +622,27 @@ void LLCamera::calcAgentFrustumPlanes(LLVector3* frust) //order of planes is important, keep most likely to fail in the front of the list //near - frust[0], frust[1], frust[2] - mAgentPlanes[2].p = planeFromPoints(frust[0], frust[1], frust[2]); + mAgentPlanes[2] = planeFromPoints(frust[0], frust[1], frust[2]); //far - mAgentPlanes[5].p = planeFromPoints(frust[5], frust[4], frust[6]); + mAgentPlanes[5] = planeFromPoints(frust[5], frust[4], frust[6]); //left - mAgentPlanes[0].p = planeFromPoints(frust[4], frust[0], frust[7]); + mAgentPlanes[0] = planeFromPoints(frust[4], frust[0], frust[7]); //right - mAgentPlanes[1].p = planeFromPoints(frust[1], frust[5], frust[6]); + mAgentPlanes[1] = planeFromPoints(frust[1], frust[5], frust[6]); //top - mAgentPlanes[4].p = planeFromPoints(frust[3], frust[2], frust[6]); + mAgentPlanes[4] = planeFromPoints(frust[3], frust[2], frust[6]); //bottom - mAgentPlanes[3].p = planeFromPoints(frust[1], frust[0], frust[4]); + mAgentPlanes[3] = planeFromPoints(frust[1], frust[0], frust[4]); //cache plane octant facing mask for use in AABBInFrustum for (U32 i = 0; i < mPlaneCount; i++) { - mAgentPlanes[i].mask = calcPlaneMask(mAgentPlanes[i].p); + mPlaneMask[i] = calcPlaneMask(mAgentPlanes[i]); } } diff --git a/indra/llmath/llcamera.h b/indra/llmath/llcamera.h index 0c81067919..c40e819dcf 100644 --- a/indra/llmath/llcamera.h +++ b/indra/llmath/llcamera.h @@ -37,6 +37,7 @@ #include "llmath.h" #include "llcoordframe.h" #include "llplane.h" +#include "llvector4a.h" const F32 DEFAULT_FIELD_OF_VIEW = 60.f * DEG_TO_RAD; const F32 DEFAULT_ASPECT_RATIO = 640.f / 480.f; @@ -79,6 +80,14 @@ class LLCamera : public LLCoordFrame { public: + + LLCamera(const LLCamera& rhs) + { + *this = rhs; + } + + const LLCamera& operator=(const LLCamera& rhs); + enum { PLANE_LEFT = 0, PLANE_RIGHT = 1, @@ -129,13 +138,9 @@ private: LLPlane mWorldPlanes[PLANE_NUM]; LLPlane mHorizPlanes[HORIZ_PLANE_NUM]; - struct frustum_plane - { - frustum_plane() : mask(0) {} - LLPlane p; - U8 mask; - }; - frustum_plane mAgentPlanes[7]; //frustum planes in agent space a la gluUnproject (I'm a bastard, I know) - DaveP + LLPlane* mAgentPlanes; //frustum planes in agent space a la gluUnproject (I'm a bastard, I know) - DaveP + U8 mAgentPlaneBuffer[sizeof(LLPlane)*8]; + U8 mPlaneMask[7]; U32 mPlaneCount; //defaults to 6, if setUserClipPlane is called, uses user supplied clip plane in @@ -143,12 +148,14 @@ private: public: LLVector3 mAgentFrustum[8]; //8 corners of 6-plane frustum F32 mFrustumCornerDist; //distance to corner of frustum against far clip plane - LLPlane getAgentPlane(U32 idx) { return mAgentPlanes[idx].p; } + LLPlane& getAgentPlane(U32 idx) { return mAgentPlanes[idx]; } public: LLCamera(); LLCamera(F32 vertical_fov_rads, F32 aspect_ratio, S32 view_height_in_pixels, F32 near_plane, F32 far_plane); - virtual ~LLCamera(){} // no-op virtual destructor + virtual ~LLCamera(); + + void alignPlanes(); void setUserClipPlane(LLPlane plane); void disableUserClipPlane(); @@ -199,8 +206,8 @@ public: S32 sphereInFrustum(const LLVector3 ¢er, const F32 radius) const; S32 pointInFrustum(const LLVector3 &point) const { return sphereInFrustum(point, 0.0f); } S32 sphereInFrustumFull(const LLVector3 ¢er, const F32 radius) const { return sphereInFrustum(center, radius); } - S32 AABBInFrustum(const LLVector3 ¢er, const LLVector3& radius); - S32 AABBInFrustumNoFarClip(const LLVector3 ¢er, const LLVector3& radius); + S32 AABBInFrustum(const LLVector4a& center, const LLVector4a& radius); + S32 AABBInFrustumNoFarClip(const LLVector4a& center, const LLVector4a& radius); //does a quick 'n dirty sphere-sphere check S32 sphereInFrustumQuick(const LLVector3 &sphere_center, const F32 radius); diff --git a/indra/llmath/llmath.h b/indra/llmath/llmath.h index 209b506c30..e572381b1a 100644 --- a/indra/llmath/llmath.h +++ b/indra/llmath/llmath.h @@ -35,7 +35,6 @@ #include <cmath> #include <cstdlib> -#include <complex> #include "lldefs.h" //#include "llstl.h" // *TODO: Remove when LLString is gone //#include "llstring.h" // *TODO: Remove when LLString is gone @@ -61,29 +60,11 @@ #endif // Single Precision Floating Point Routines -#ifndef fsqrtf -#define fsqrtf(x) ((F32)sqrt((F64)(x))) -#endif -#ifndef sqrtf -#define sqrtf(x) ((F32)sqrt((F64)(x))) -#endif - -#ifndef cosf -#define cosf(x) ((F32)cos((F64)(x))) -#endif -#ifndef sinf -#define sinf(x) ((F32)sin((F64)(x))) -#endif -#ifndef tanf +// (There used to be more defined here, but they appeared to be redundant and +// were breaking some other includes. Removed by Falcon, reviewed by Andrew, 11/25/09) +/*#ifndef tanf #define tanf(x) ((F32)tan((F64)(x))) -#endif -#ifndef acosf -#define acosf(x) ((F32)acos((F64)(x))) -#endif - -#ifndef powf -#define powf(x,y) ((F32)pow((F64)(x),(F64)(y))) -#endif +#endif*/ const F32 GRAVITY = -9.8f; @@ -203,7 +184,7 @@ inline S32 llfloor( F32 f ) } return result; #else - return (S32)floorf(f); + return (S32)floor(f); #endif } @@ -381,11 +362,14 @@ inline F32 snap_to_sig_figs(F32 foo, S32 sig_figs) bar *= 10.f; } - foo = (F32)llround(foo * bar); + //F32 new_foo = (F32)llround(foo * bar); + // the llround() implementation sucks. Don't us it. - // shift back - foo /= bar; - return foo; + F32 sign = (foo > 0.f) ? 1.f : -1.f; + F32 new_foo = F32( S64(foo * bar + sign * 0.5f)); + new_foo /= bar; + + return new_foo; } inline F32 lerp(F32 a, F32 b, F32 u) @@ -519,4 +503,7 @@ inline F32 llgaussian(F32 x, F32 o) return 1.f/(F_SQRT_TWO_PI*o)*powf(F_E, -(x*x)/(2*o*o)); } +// Include simd math header +#include "llsimdmath.h" + #endif diff --git a/indra/llmath/llmatrix3a.cpp b/indra/llmath/llmatrix3a.cpp new file mode 100644 index 0000000000..b7468f4914 --- /dev/null +++ b/indra/llmath/llmatrix3a.cpp @@ -0,0 +1,140 @@ +/** + * @file llvector4a.cpp + * @brief SIMD vector implementation + * + * $LicenseInfo:firstyear=2010&license=viewergpl$ + * + * Copyright (c) 2007-2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#include "llmath.h" + +static LL_ALIGN_16(const F32 M_IDENT_3A[12]) = + { 1.f, 0.f, 0.f, 0.f, // Column 1 + 0.f, 1.f, 0.f, 0.f, // Column 2 + 0.f, 0.f, 1.f, 0.f }; // Column 3 + +extern const LLMatrix3a LL_M3A_IDENTITY = *reinterpret_cast<const LLMatrix3a*> (M_IDENT_3A); + +void LLMatrix3a::setMul( const LLMatrix3a& lhs, const LLMatrix3a& rhs ) +{ + const LLVector4a col0 = lhs.getColumn(0); + const LLVector4a col1 = lhs.getColumn(1); + const LLVector4a col2 = lhs.getColumn(2); + + for ( int i = 0; i < 3; i++ ) + { + LLVector4a xxxx = _mm_load_ss( rhs.mColumns[i].getF32ptr() ); + xxxx.splat<0>( xxxx ); + xxxx.mul( col0 ); + + { + LLVector4a yyyy = _mm_load_ss( rhs.mColumns[i].getF32ptr() + 1 ); + yyyy.splat<0>( yyyy ); + yyyy.mul( col1 ); + xxxx.add( yyyy ); + } + + { + LLVector4a zzzz = _mm_load_ss( rhs.mColumns[i].getF32ptr() + 2 ); + zzzz.splat<0>( zzzz ); + zzzz.mul( col2 ); + xxxx.add( zzzz ); + } + + xxxx.store4a( mColumns[i].getF32ptr() ); + } + +} + +/*static */void LLMatrix3a::batchTransform( const LLMatrix3a& xform, const LLVector4a* src, int numVectors, LLVector4a* dst ) +{ + const LLVector4a col0 = xform.getColumn(0); + const LLVector4a col1 = xform.getColumn(1); + const LLVector4a col2 = xform.getColumn(2); + const LLVector4a* maxAddr = src + numVectors; + + if ( numVectors & 0x1 ) + { + LLVector4a xxxx = _mm_load_ss( (const F32*)src ); + LLVector4a yyyy = _mm_load_ss( (const F32*)src + 1 ); + LLVector4a zzzz = _mm_load_ss( (const F32*)src + 2 ); + xxxx.splat<0>( xxxx ); + yyyy.splat<0>( yyyy ); + zzzz.splat<0>( zzzz ); + xxxx.mul( col0 ); + yyyy.mul( col1 ); + zzzz.mul( col2 ); + xxxx.add( yyyy ); + xxxx.add( zzzz ); + xxxx.store4a( (F32*)dst ); + src++; + dst++; + } + + + numVectors >>= 1; + while ( src < maxAddr ) + { + _mm_prefetch( (const char*)(src + 32 ), _MM_HINT_NTA ); + _mm_prefetch( (const char*)(dst + 32), _MM_HINT_NTA ); + LLVector4a xxxx = _mm_load_ss( (const F32*)src ); + LLVector4a xxxx1= _mm_load_ss( (const F32*)(src + 1) ); + + xxxx.splat<0>( xxxx ); + xxxx1.splat<0>( xxxx1 ); + xxxx.mul( col0 ); + xxxx1.mul( col0 ); + + { + LLVector4a yyyy = _mm_load_ss( (const F32*)src + 1 ); + LLVector4a yyyy1 = _mm_load_ss( (const F32*)(src + 1) + 1); + yyyy.splat<0>( yyyy ); + yyyy1.splat<0>( yyyy1 ); + yyyy.mul( col1 ); + yyyy1.mul( col1 ); + xxxx.add( yyyy ); + xxxx1.add( yyyy1 ); + } + + { + LLVector4a zzzz = _mm_load_ss( (const F32*)(src) + 2 ); + LLVector4a zzzz1 = _mm_load_ss( (const F32*)(++src) + 2 ); + zzzz.splat<0>( zzzz ); + zzzz1.splat<0>( zzzz1 ); + zzzz.mul( col2 ); + zzzz1.mul( col2 ); + xxxx.add( zzzz ); + xxxx1.add( zzzz1 ); + } + + xxxx.store4a(dst->getF32ptr()); + src++; + dst++; + + xxxx1.store4a((F32*)dst++); + } +} diff --git a/indra/llmath/llmatrix3a.h b/indra/llmath/llmatrix3a.h new file mode 100644 index 0000000000..56327f9f6d --- /dev/null +++ b/indra/llmath/llmatrix3a.h @@ -0,0 +1,134 @@ +/** + * @file llmatrix3a.h + * @brief LLMatrix3a class header file - memory aligned and vectorized 3x3 matrix + * + * $LicenseInfo:firstyear=2010&license=viewergpl$ + * + * Copyright (c) 2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#ifndef LL_LLMATRIX3A_H +#define LL_LLMATRIX3A_H + +///////////////////////////// +// LLMatrix3a, LLRotation +///////////////////////////// +// This class stores a 3x3 (technically 4x3) matrix in column-major order +///////////////////////////// +///////////////////////////// +// These classes are intentionally minimal right now. If you need additional +// functionality, please contact someone with SSE experience (e.g., Falcon or +// Huseby). +///////////////////////////// + +// LLMatrix3a is the base class for LLRotation, which should be used instead any time you're dealing with a +// rotation matrix. +class LLMatrix3a +{ +public: + + // Utility function for quickly transforming an array of LLVector4a's + // For transforming a single LLVector4a, see LLVector4a::setRotated + static void batchTransform( const LLMatrix3a& xform, const LLVector4a* src, int numVectors, LLVector4a* dst ); + + // Utility function to obtain the identity matrix + static inline const LLMatrix3a& getIdentity(); + + ////////////////////////// + // Ctors + ////////////////////////// + + // Ctor + LLMatrix3a() {} + + // Ctor for setting by columns + inline LLMatrix3a( const LLVector4a& c0, const LLVector4a& c1, const LLVector4a& c2 ); + + ////////////////////////// + // Get/Set + ////////////////////////// + + // Loads from an LLMatrix3 + inline void loadu(const LLMatrix3& src); + + // Set rows + inline void setRows(const LLVector4a& r0, const LLVector4a& r1, const LLVector4a& r2); + + // Set columns + inline void setColumns(const LLVector4a& c0, const LLVector4a& c1, const LLVector4a& c2); + + // Get the read-only access to a specified column. Valid columns are 0-2, but the + // function is unchecked. You've been warned. + inline const LLVector4a& getColumn(const U32 column) const; + + ///////////////////////// + // Matrix modification + ///////////////////////// + + // Set this matrix to the product of lhs and rhs ( this = lhs * rhs ) + void setMul( const LLMatrix3a& lhs, const LLMatrix3a& rhs ); + + // Set this matrix to the transpose of src + inline void setTranspose(const LLMatrix3a& src); + + // Set this matrix to a*w + b*(1-w) + inline void setLerp(const LLMatrix3a& a, const LLMatrix3a& b, F32 w); + + ///////////////////////// + // Matrix inspection + ///////////////////////// + + // Sets all 4 elements in 'dest' to the determinant of this matrix. + // If you will be using the determinant in subsequent ops with LLVector4a, use this version + inline void getDeterminant( LLVector4a& dest ) const; + + // Returns the determinant as an LLSimdScalar. Use this if you will be using the determinant + // primary for scalar operations. + inline LLSimdScalar getDeterminant() const; + + // Returns nonzero if rows 0-2 and colums 0-2 contain no NaN or INF values. Row 3 is ignored + inline LLBool32 isFinite() const; + + // Returns true if this matrix is equal to 'rhs' up to 'tolerance' + inline bool isApproximatelyEqual( const LLMatrix3a& rhs, F32 tolerance = F_APPROXIMATELY_ZERO ) const; + +protected: + + LLVector4a mColumns[3]; + +}; + +class LLRotation : public LLMatrix3a +{ +public: + + LLRotation() {} + + // Returns true if this rotation is orthonormal with det ~= 1 + inline bool isOkRotation() const; +}; + +#endif diff --git a/indra/llmath/llmatrix3a.inl b/indra/llmath/llmatrix3a.inl new file mode 100644 index 0000000000..65fd949f78 --- /dev/null +++ b/indra/llmath/llmatrix3a.inl @@ -0,0 +1,125 @@ +/** + * @file llmatrix3a.inl + * @brief LLMatrix3a inline definitions + * + * $LicenseInfo:firstyear=2010&license=viewergpl$ + * + * Copyright (c) 2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#include "llmatrix3a.h" +#include "m3math.h" + +inline LLMatrix3a::LLMatrix3a( const LLVector4a& c0, const LLVector4a& c1, const LLVector4a& c2 ) +{ + setColumns( c0, c1, c2 ); +} + +inline void LLMatrix3a::loadu(const LLMatrix3& src) +{ + mColumns[0].load3(src.mMatrix[0]); + mColumns[1].load3(src.mMatrix[1]); + mColumns[2].load3(src.mMatrix[2]); +} + +inline void LLMatrix3a::setRows(const LLVector4a& r0, const LLVector4a& r1, const LLVector4a& r2) +{ + mColumns[0] = r0; + mColumns[1] = r1; + mColumns[2] = r2; + setTranspose( *this ); +} + +inline void LLMatrix3a::setColumns(const LLVector4a& c0, const LLVector4a& c1, const LLVector4a& c2) +{ + mColumns[0] = c0; + mColumns[1] = c1; + mColumns[2] = c2; +} + +inline void LLMatrix3a::setTranspose(const LLMatrix3a& src) +{ + const LLQuad srcCol0 = src.mColumns[0]; + const LLQuad srcCol1 = src.mColumns[1]; + const LLQuad unpacklo = _mm_unpacklo_ps( srcCol0, srcCol1 ); + mColumns[0] = _mm_movelh_ps( unpacklo, src.mColumns[2] ); + mColumns[1] = _mm_shuffle_ps( _mm_movehl_ps( srcCol0, unpacklo ), src.mColumns[2], _MM_SHUFFLE(0, 1, 1, 0) ); + mColumns[2] = _mm_shuffle_ps( _mm_unpackhi_ps( srcCol0, srcCol1 ), src.mColumns[2], _MM_SHUFFLE(0, 2, 1, 0) ); +} + +inline const LLVector4a& LLMatrix3a::getColumn(const U32 column) const +{ + llassert( column < 3 ); + return mColumns[column]; +} + +inline void LLMatrix3a::setLerp(const LLMatrix3a& a, const LLMatrix3a& b, F32 w) +{ + mColumns[0].setLerp( a.mColumns[0], b.mColumns[0], w ); + mColumns[1].setLerp( a.mColumns[1], b.mColumns[1], w ); + mColumns[2].setLerp( a.mColumns[2], b.mColumns[2], w ); +} + +inline LLBool32 LLMatrix3a::isFinite() const +{ + return mColumns[0].isFinite3() && mColumns[1].isFinite3() && mColumns[2].isFinite3(); +} + +inline void LLMatrix3a::getDeterminant( LLVector4a& dest ) const +{ + LLVector4a col1xcol2; col1xcol2.setCross3( mColumns[1], mColumns[2] ); + dest.setAllDot3( col1xcol2, mColumns[0] ); +} + +inline LLSimdScalar LLMatrix3a::getDeterminant() const +{ + LLVector4a col1xcol2; col1xcol2.setCross3( mColumns[1], mColumns[2] ); + return col1xcol2.dot3( mColumns[0] ); +} + +inline bool LLMatrix3a::isApproximatelyEqual( const LLMatrix3a& rhs, F32 tolerance /*= F_APPROXIMATELY_ZERO*/ ) const +{ + return rhs.getColumn(0).equals3(mColumns[0], tolerance) + && rhs.getColumn(1).equals3(mColumns[1], tolerance) + && rhs.getColumn(2).equals3(mColumns[2], tolerance); +} + +inline const LLMatrix3a& LLMatrix3a::getIdentity() +{ + extern const LLMatrix3a LL_M3A_IDENTITY; + return LL_M3A_IDENTITY; +} + +inline bool LLRotation::isOkRotation() const +{ + LLMatrix3a transpose; transpose.setTranspose( *this ); + LLMatrix3a product; product.setMul( *this, transpose ); + + LLSimdScalar detMinusOne = getDeterminant() - 1.f; + + return product.isApproximatelyEqual( LLMatrix3a::getIdentity() ) && (detMinusOne.getAbs() < F_APPROXIMATELY_ZERO); +} + diff --git a/indra/llmath/llmatrix4a.h b/indra/llmath/llmatrix4a.h new file mode 100644 index 0000000000..0ead045d04 --- /dev/null +++ b/indra/llmath/llmatrix4a.h @@ -0,0 +1,149 @@ +/** + * @file llmatrix4a.h + * @brief LLMatrix4a class header file - memory aligned and vectorized 4x4 matrix + * + * $LicenseInfo:firstyear=2007&license=viewergpl$ + * + * Copyright (c) 2007-2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#ifndef LL_LLMATRIX4A_H +#define LL_LLMATRIX4A_H + +#include "llvector4a.h" +#include "m4math.h" +#include "m3math.h" + +class LLMatrix4a +{ +public: + LLVector4a mMatrix[4]; + + inline void clear() + { + mMatrix[0].clear(); + mMatrix[1].clear(); + mMatrix[2].clear(); + mMatrix[3].clear(); + } + + inline void loadu(const LLMatrix4& src) + { + mMatrix[0] = _mm_loadu_ps(src.mMatrix[0]); + mMatrix[1] = _mm_loadu_ps(src.mMatrix[1]); + mMatrix[2] = _mm_loadu_ps(src.mMatrix[2]); + mMatrix[3] = _mm_loadu_ps(src.mMatrix[3]); + + } + + inline void loadu(const LLMatrix3& src) + { + mMatrix[0].load3(src.mMatrix[0]); + mMatrix[1].load3(src.mMatrix[1]); + mMatrix[2].load3(src.mMatrix[2]); + mMatrix[3].set(0,0,0,1.f); + } + + inline void add(const LLMatrix4a& rhs) + { + mMatrix[0].add(rhs.mMatrix[0]); + mMatrix[1].add(rhs.mMatrix[1]); + mMatrix[2].add(rhs.mMatrix[2]); + mMatrix[3].add(rhs.mMatrix[3]); + } + + inline void setRows(const LLVector4a& r0, const LLVector4a& r1, const LLVector4a& r2) + { + mMatrix[0] = r0; + mMatrix[1] = r1; + mMatrix[2] = r2; + } + + inline void setMul(const LLMatrix4a& m, const F32 s) + { + mMatrix[0].setMul(m.mMatrix[0], s); + mMatrix[1].setMul(m.mMatrix[1], s); + mMatrix[2].setMul(m.mMatrix[2], s); + mMatrix[3].setMul(m.mMatrix[3], s); + } + + inline void setLerp(const LLMatrix4a& a, const LLMatrix4a& b, F32 w) + { + LLVector4a d0,d1,d2,d3; + d0.setSub(b.mMatrix[0], a.mMatrix[0]); + d1.setSub(b.mMatrix[1], a.mMatrix[1]); + d2.setSub(b.mMatrix[2], a.mMatrix[2]); + d3.setSub(b.mMatrix[3], a.mMatrix[3]); + + // this = a + d*w + + d0.mul(w); + d1.mul(w); + d2.mul(w); + d3.mul(w); + + mMatrix[0].setAdd(a.mMatrix[0],d0); + mMatrix[1].setAdd(a.mMatrix[1],d1); + mMatrix[2].setAdd(a.mMatrix[2],d2); + mMatrix[3].setAdd(a.mMatrix[3],d3); + } + + inline void rotate(const LLVector4a& v, LLVector4a& res) + { + res = _mm_shuffle_ps(v, v, _MM_SHUFFLE(0, 0, 0, 0)); + res.mul(mMatrix[0]); + + LLVector4a y; + y = _mm_shuffle_ps(v, v, _MM_SHUFFLE(1, 1, 1, 1)); + y.mul(mMatrix[1]); + + LLVector4a z; + z = _mm_shuffle_ps(v, v, _MM_SHUFFLE(2, 2, 2, 2)); + z.mul(mMatrix[2]); + + res.add(y); + res.add(z); + } + + inline void affineTransform(const LLVector4a& v, LLVector4a& res) + { + LLVector4a x,y,z; + + x = _mm_shuffle_ps(v, v, _MM_SHUFFLE(0, 0, 0, 0)); + y = _mm_shuffle_ps(v, v, _MM_SHUFFLE(1, 1, 1, 1)); + z = _mm_shuffle_ps(v, v, _MM_SHUFFLE(2, 2, 2, 2)); + + x.mul(mMatrix[0]); + y.mul(mMatrix[1]); + z.mul(mMatrix[2]); + + x.add(y); + z.add(mMatrix[3]); + res.setAdd(x,z); + } +}; + +#endif diff --git a/indra/llmath/lloctree.h b/indra/llmath/lloctree.h index 2f34fb1bb0..432e9fbcd8 100644 --- a/indra/llmath/lloctree.h +++ b/indra/llmath/lloctree.h @@ -35,6 +35,7 @@ #include "lltreenode.h" #include "v3math.h" +#include "llvector4a.h" #include <vector> #include <set> @@ -73,6 +74,13 @@ public: }; template <class T> +class LLOctreeTravelerDepthFirst : public LLOctreeTraveler<T> +{ +public: + virtual void traverse(const LLOctreeNode<T>* node); +}; + +template <class T> class LLOctreeNode : public LLTreeNode<T> { public: @@ -87,23 +95,22 @@ public: typedef LLOctreeNode<T> oct_node; typedef LLOctreeListener<T> oct_listener; - static const U8 OCTANT_POSITIVE_X = 0x01; - static const U8 OCTANT_POSITIVE_Y = 0x02; - static const U8 OCTANT_POSITIVE_Z = 0x04; - - LLOctreeNode( LLVector3d center, - LLVector3d size, + LLOctreeNode( const LLVector4a& center, + const LLVector4a& size, BaseType* parent, - U8 octant = 255) + S32 octant = -1) : mParent((oct_node*)parent), - mCenter(center), - mSize(size), mOctant(octant) { + mD = (LLVector4a*) ll_aligned_malloc_16(sizeof(LLVector4a)*4); + + mD[CENTER] = center; + mD[SIZE] = size; + updateMinMax(); - if ((mOctant == 255) && mParent) + if ((mOctant == -1) && mParent) { - mOctant = ((oct_node*) mParent)->getOctant(mCenter.mdV); + mOctant = ((oct_node*) mParent)->getOctant(mD[CENTER]); } clearChildren(); @@ -117,43 +124,30 @@ public: { delete getChild(i); } + + ll_aligned_free_16(mD); } inline const BaseType* getParent() const { return mParent; } - inline void setParent(BaseType* parent) { mParent = (oct_node*) parent; } - inline const LLVector3d& getCenter() const { return mCenter; } - inline const LLVector3d& getSize() const { return mSize; } - inline void setCenter(LLVector3d center) { mCenter = center; } - inline void setSize(LLVector3d size) { mSize = size; } - inline oct_node* getNodeAt(T* data) { return getNodeAt(data->getPositionGroup(), data->getBinRadius()); } - inline U8 getOctant() const { return mOctant; } - inline void setOctant(U8 octant) { mOctant = octant; } + inline void setParent(BaseType* parent) { mParent = (oct_node*) parent; } + inline const LLVector4a& getCenter() const { return mD[CENTER]; } + inline const LLVector4a& getSize() const { return mD[SIZE]; } + inline void setCenter(const LLVector4a& center) { mD[CENTER] = center; } + inline void setSize(const LLVector4a& size) { mD[SIZE] = size; } + inline oct_node* getNodeAt(T* data) { return getNodeAt(data->getPositionGroup(), data->getBinRadius()); } + inline S32 getOctant() const { return mOctant; } + inline void setOctant(S32 octant) { mOctant = octant; } inline const oct_node* getOctParent() const { return (const oct_node*) getParent(); } inline oct_node* getOctParent() { return (oct_node*) getParent(); } - U8 getOctant(const F64 pos[]) const //get the octant pos is in + S32 getOctant(const LLVector4a& pos) const //get the octant pos is in { - U8 ret = 0; - - if (pos[0] > mCenter.mdV[0]) - { - ret |= OCTANT_POSITIVE_X; - } - if (pos[1] > mCenter.mdV[1]) - { - ret |= OCTANT_POSITIVE_Y; - } - if (pos[2] > mCenter.mdV[2]) - { - ret |= OCTANT_POSITIVE_Z; - } - - return ret; + return pos.greaterThan(mD[CENTER]).getGatheredBits() & 0x7; } - inline bool isInside(const LLVector3d& pos, const F64& rad) const + inline bool isInside(const LLVector4a& pos, const F32& rad) const { - return rad <= mSize.mdV[0]*2.0 && isInside(pos); + return rad <= mD[SIZE][0]*2.f && isInside(pos); } inline bool isInside(T* data) const @@ -161,29 +155,27 @@ public: return isInside(data->getPositionGroup(), data->getBinRadius()); } - bool isInside(const LLVector3d& pos) const + bool isInside(const LLVector4a& pos) const { - const F64& x = pos.mdV[0]; - const F64& y = pos.mdV[1]; - const F64& z = pos.mdV[2]; - - if (x > mMax.mdV[0] || x <= mMin.mdV[0] || - y > mMax.mdV[1] || y <= mMin.mdV[1] || - z > mMax.mdV[2] || z <= mMin.mdV[2]) + S32 gt = pos.greaterThan(mD[MAX]).getGatheredBits() & 0x7; + if (gt) { return false; } - + + S32 lt = pos.lessEqual(mD[MIN]).getGatheredBits() & 0x7; + if (lt) + { + return false; + } + return true; } void updateMinMax() { - for (U32 i = 0; i < 3; i++) - { - mMax.mdV[i] = mCenter.mdV[i] + mSize.mdV[i]; - mMin.mdV[i] = mCenter.mdV[i] - mSize.mdV[i]; - } + mD[MAX].setAdd(mD[CENTER], mD[SIZE]); + mD[MIN].setSub(mD[CENTER], mD[SIZE]); } inline oct_listener* getOctListener(U32 index) @@ -196,34 +188,34 @@ public: return contains(xform->getBinRadius()); } - bool contains(F64 radius) + bool contains(F32 radius) { if (mParent == NULL) { //root node contains nothing return false; } - F64 size = mSize.mdV[0]; - F64 p_size = size * 2.0; + F32 size = mD[SIZE][0]; + F32 p_size = size * 2.f; - return (radius <= 0.001 && size <= 0.001) || + return (radius <= 0.001f && size <= 0.001f) || (radius <= p_size && radius > size); } - static void pushCenter(LLVector3d ¢er, const LLVector3d &size, const T* data) + static void pushCenter(LLVector4a ¢er, const LLVector4a &size, const T* data) { - const LLVector3d& pos = data->getPositionGroup(); - for (U32 i = 0; i < 3; i++) - { - if (pos.mdV[i] > center.mdV[i]) - { - center.mdV[i] += size.mdV[i]; - } - else - { - center.mdV[i] -= size.mdV[i]; - } - } + const LLVector4a& pos = data->getPositionGroup(); + + LLVector4a gt = pos.greaterThan(center); + + LLVector4a up; + up = _mm_and_ps(size, gt); + + LLVector4a down; + down = _mm_andnot_ps(gt, size); + + center.add(up); + center.sub(down); } void accept(oct_traveler* visitor) { visitor->visit(this); } @@ -242,21 +234,21 @@ public: void accept(tree_traveler* visitor) const { visitor->visit(this); } void accept(oct_traveler* visitor) const { visitor->visit(this); } - oct_node* getNodeAt(const LLVector3d& pos, const F64& rad) + oct_node* getNodeAt(const LLVector4a& pos, const F32& rad) { LLOctreeNode<T>* node = this; if (node->isInside(pos, rad)) { //do a quick search by octant - U8 octant = node->getOctant(pos.mdV); + S32 octant = node->getOctant(pos); BOOL keep_going = TRUE; //traverse the tree until we find a node that has no node //at the appropriate octant or is smaller than the object. //by definition, that node is the smallest node that contains // the data - while (keep_going && node->getSize().mdV[0] >= rad) + while (keep_going && node->getSize()[0] >= rad) { keep_going = FALSE; for (U32 i = 0; i < node->getChildCount() && !keep_going; i++) @@ -264,7 +256,7 @@ public: if (node->getChild(i)->getOctant() == octant) { node = node->getChild(i); - octant = node->getOctant(pos.mdV); + octant = node->getOctant(pos); keep_going = TRUE; } } @@ -282,7 +274,7 @@ public: { if (data == NULL) { - //OCT_ERRS << "!!! INVALID ELEMENT ADDED TO OCTREE BRANCH !!!" << llendl; + OCT_ERRS << "!!! INVALID ELEMENT ADDED TO OCTREE BRANCH !!!" << llendl; return false; } LLOctreeNode<T>* parent = getOctParent(); @@ -292,7 +284,7 @@ public: { if (getElementCount() < LL_OCTREE_MAX_CAPACITY && (contains(data->getBinRadius()) || - (data->getBinRadius() > getSize().mdV[0] && + (data->getBinRadius() > getSize()[0] && parent && parent->getElementCount() >= LL_OCTREE_MAX_CAPACITY))) { //it belongs here #if LL_OCTREE_PARANOIA_CHECK @@ -323,16 +315,21 @@ public: } //it's here, but no kids are in the right place, make a new kid - LLVector3d center(getCenter()); - LLVector3d size(getSize()*0.5); + LLVector4a center = getCenter(); + LLVector4a size = getSize(); + size.mul(0.5f); //push center in direction of data LLOctreeNode<T>::pushCenter(center, size, data); // handle case where floating point number gets too small - if( llabs(center.mdV[0] - getCenter().mdV[0]) < F_APPROXIMATELY_ZERO && - llabs(center.mdV[1] - getCenter().mdV[1]) < F_APPROXIMATELY_ZERO && - llabs(center.mdV[2] - getCenter().mdV[2]) < F_APPROXIMATELY_ZERO) + LLVector4a val; + val.setSub(center, getCenter()); + val.setAbs(val); + + S32 lt = val.lessThan(LLVector4a::getEpsilon()).getGatheredBits() & 0x7; + + if( lt == 0x7 ) { mData.insert(data); BaseType::insert(data); @@ -350,7 +347,7 @@ public: //make sure no existing node matches this position for (U32 i = 0; i < getChildCount(); i++) { - if (mChild[i]->getCenter() == center) + if (mChild[i]->getCenter().equal3(center)) { OCT_ERRS << "Octree detected duplicate child center and gave up." << llendl; return false; @@ -368,7 +365,7 @@ public: else { //it's not in here, give it to the root - //OCT_ERRS << "Octree insertion failed, starting over from root!" << llendl; + OCT_ERRS << "Octree insertion failed, starting over from root!" << llendl; oct_node* node = this; @@ -475,13 +472,19 @@ public: void addChild(oct_node* child, BOOL silent = FALSE) { #if LL_OCTREE_PARANOIA_CHECK + + if (child->getSize().equal3(getSize())) + { + OCT_ERRS << "Child size is same as parent size!" << llendl; + } + for (U32 i = 0; i < getChildCount(); i++) { - if(mChild[i]->getSize() != child->getSize()) + if(!mChild[i]->getSize().equal3(child->getSize())) { OCT_ERRS <<"Invalid octree child size." << llendl; } - if (mChild[i]->getCenter() == child->getCenter()) + if (mChild[i]->getCenter().equal3(child->getCenter())) { OCT_ERRS <<"Duplicate octree child position." << llendl; } @@ -506,7 +509,7 @@ public: } } - void removeChild(U8 index, BOOL destroy = FALSE) + void removeChild(S32 index, BOOL destroy = FALSE) { for (U32 i = 0; i < this->getListenerCount(); i++) { @@ -547,18 +550,26 @@ public: } } - //OCT_ERRS << "Octree failed to delete requested child." << llendl; + OCT_ERRS << "Octree failed to delete requested child." << llendl; } protected: + typedef enum + { + CENTER = 0, + SIZE = 1, + MAX = 2, + MIN = 3 + } eDName; + + LLVector4a* mD; + + oct_node* mParent; + S32 mOctant; + child_list mChild; element_list mData; - oct_node* mParent; - LLVector3d mCenter; - LLVector3d mSize; - LLVector3d mMax; - LLVector3d mMin; - U8 mOctant; + }; //just like a regular node, except it might expand on insert and compress on balance @@ -569,9 +580,9 @@ public: typedef LLOctreeNode<T> BaseType; typedef LLOctreeNode<T> oct_node; - LLOctreeRoot( LLVector3d center, - LLVector3d size, - BaseType* parent) + LLOctreeRoot(const LLVector4a& center, + const LLVector4a& size, + BaseType* parent) : BaseType(center, size, parent) { } @@ -612,28 +623,33 @@ public: { if (data == NULL) { - //OCT_ERRS << "!!! INVALID ELEMENT ADDED TO OCTREE ROOT !!!" << llendl; + OCT_ERRS << "!!! INVALID ELEMENT ADDED TO OCTREE ROOT !!!" << llendl; return false; } if (data->getBinRadius() > 4096.0) { - //OCT_ERRS << "!!! ELEMENT EXCEEDS MAXIMUM SIZE IN OCTREE ROOT !!!" << llendl; + OCT_ERRS << "!!! ELEMENT EXCEEDS MAXIMUM SIZE IN OCTREE ROOT !!!" << llendl; return false; } - const F64 MAX_MAG = 1024.0*1024.0; + LLVector4a MAX_MAG; + MAX_MAG.splat(1024.f*1024.f); + + const LLVector4a& v = data->getPositionGroup(); - const LLVector3d& v = data->getPositionGroup(); - if (!(fabs(v.mdV[0]-this->mCenter.mdV[0]) < MAX_MAG && - fabs(v.mdV[1]-this->mCenter.mdV[1]) < MAX_MAG && - fabs(v.mdV[2]-this->mCenter.mdV[2]) < MAX_MAG)) + LLVector4a val; + val.setSub(v, BaseType::mD[BaseType::CENTER]); + val.setAbs(val); + S32 lt = val.lessThan(MAX_MAG).getGatheredBits() & 0x7; + + if (lt != 0x7) { - //OCT_ERRS << "!!! ELEMENT EXCEEDS RANGE OF SPATIAL PARTITION !!!" << llendl; + OCT_ERRS << "!!! ELEMENT EXCEEDS RANGE OF SPATIAL PARTITION !!!" << llendl; return false; } - if (this->getSize().mdV[0] > data->getBinRadius() && isInside(data->getPositionGroup())) + if (this->getSize()[0] > data->getBinRadius() && isInside(data->getPositionGroup())) { //we got it, just act like a branch oct_node* node = getNodeAt(data); @@ -649,31 +665,34 @@ public: else if (this->getChildCount() == 0) { //first object being added, just wrap it up - while (!(this->getSize().mdV[0] > data->getBinRadius() && isInside(data->getPositionGroup()))) + while (!(this->getSize()[0] > data->getBinRadius() && isInside(data->getPositionGroup()))) { - LLVector3d center, size; + LLVector4a center, size; center = this->getCenter(); size = this->getSize(); LLOctreeNode<T>::pushCenter(center, size, data); this->setCenter(center); - this->setSize(size*2); + size.mul(2.f); + this->setSize(size); this->updateMinMax(); } LLOctreeNode<T>::insert(data); } else { - while (!(this->getSize().mdV[0] > data->getBinRadius() && isInside(data->getPositionGroup()))) + while (!(this->getSize()[0] > data->getBinRadius() && isInside(data->getPositionGroup()))) { //the data is outside the root node, we need to grow - LLVector3d center(this->getCenter()); - LLVector3d size(this->getSize()); + LLVector4a center(this->getCenter()); + LLVector4a size(this->getSize()); //expand this node - LLVector3d newcenter(center); + LLVector4a newcenter(center); LLOctreeNode<T>::pushCenter(newcenter, size, data); this->setCenter(newcenter); - this->setSize(size*2); + LLVector4a size2 = size; + size2.mul(2.f); + this->setSize(size2); this->updateMinMax(); //copy our children to a new branch @@ -710,4 +729,15 @@ void LLOctreeTraveler<T>::traverse(const LLOctreeNode<T>* node) traverse(node->getChild(i)); } } + +template <class T> +void LLOctreeTravelerDepthFirst<T>::traverse(const LLOctreeNode<T>* node) +{ + for (U32 i = 0; i < node->getChildCount(); i++) + { + traverse(node->getChild(i)); + } + node->accept(this); +} + #endif diff --git a/indra/llmath/llquantize.h b/indra/llmath/llquantize.h index 2192427f07..c043f7f752 100644 --- a/indra/llmath/llquantize.h +++ b/indra/llmath/llquantize.h @@ -35,10 +35,16 @@ #define LL_LLQUANTIZE_H const U16 U16MAX = 65535; +LL_ALIGN_16( const F32 F_U16MAX_4A[4] ) = { 65535.f, 65535.f, 65535.f, 65535.f }; + const F32 OOU16MAX = 1.f/(F32)(U16MAX); +LL_ALIGN_16( const F32 F_OOU16MAX_4A[4] ) = { OOU16MAX, OOU16MAX, OOU16MAX, OOU16MAX }; const U8 U8MAX = 255; +LL_ALIGN_16( const F32 F_U8MAX_4A[4] ) = { 255.f, 255.f, 255.f, 255.f }; + const F32 OOU8MAX = 1.f/(F32)(U8MAX); +LL_ALIGN_16( const F32 F_OOU8MAX_4A[4] ) = { OOU8MAX, OOU8MAX, OOU8MAX, OOU8MAX }; const U8 FIRSTVALIDCHAR = 54; const U8 MAXSTRINGVAL = U8MAX - FIRSTVALIDCHAR; //we don't allow newline or null diff --git a/indra/llmath/llquaternion.cpp b/indra/llmath/llquaternion.cpp index fdcc19d657..73c5f4505e 100644 --- a/indra/llmath/llquaternion.cpp +++ b/indra/llmath/llquaternion.cpp @@ -32,9 +32,10 @@ #include "linden_common.h" +#include "llmath.h" // for F_PI + #include "llquaternion.h" -#include "llmath.h" // for F_PI //#include "vmath.h" #include "v3math.h" #include "v3dmath.h" diff --git a/indra/llmath/llquaternion.h b/indra/llmath/llquaternion.h index 0769f29f23..a7bb09fae3 100644 --- a/indra/llmath/llquaternion.h +++ b/indra/llmath/llquaternion.h @@ -33,7 +33,11 @@ #ifndef LLQUATERNION_H #define LLQUATERNION_H -#include "llmath.h" +#include <iostream> + +#ifndef LLMATH_H //enforce specific include order to avoid tangling inline dependencies +#error "Please include llmath.h first." +#endif class LLVector4; class LLVector3; diff --git a/indra/llmath/llquaternion2.h b/indra/llmath/llquaternion2.h new file mode 100644 index 0000000000..dbb4afe312 --- /dev/null +++ b/indra/llmath/llquaternion2.h @@ -0,0 +1,111 @@ +/** + * @file llquaternion2.h + * @brief LLQuaternion2 class header file - SIMD-enabled quaternion class + * + * $LicenseInfo:firstyear=2010&license=viewergpl$ + * + * Copyright (c) 2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#ifndef LL_QUATERNION2_H +#define LL_QUATERNION2_H + +///////////////////////////// +// LLQuaternion2 +///////////////////////////// +// This class stores a quaternion x*i + y*j + z*k + w in <x, y, z, w> order +// (i.e., w in high order element of vector) +///////////////////////////// +///////////////////////////// +// These classes are intentionally minimal right now. If you need additional +// functionality, please contact someone with SSE experience (e.g., Falcon or +// Huseby). +///////////////////////////// +#include "llquaternion.h" + +class LLQuaternion2 +{ +public: + + ////////////////////////// + // Ctors + ////////////////////////// + + // Ctor + LLQuaternion2() {} + + // Ctor from LLQuaternion + explicit LLQuaternion2( const class LLQuaternion& quat ); + + ////////////////////////// + // Get/Set + ////////////////////////// + + // Load from an LLQuaternion + inline void operator=( const LLQuaternion& quat ) + { + mQ.loadua( quat.mQ ); + } + + // Return the internal LLVector4a representation of the quaternion + inline const LLVector4a& getVector4a() const; + inline LLVector4a& getVector4aRw(); + + ///////////////////////// + // Quaternion modification + ///////////////////////// + + // Set this quaternion to the conjugate of src + inline void setConjugate(const LLQuaternion2& src); + + // Renormalizes the quaternion. Assumes it has nonzero length. + inline void normalize(); + + // Quantize this quaternion to 8 bit precision + inline void quantize8(); + + // Quantize this quaternion to 16 bit precision + inline void quantize16(); + + ///////////////////////// + // Quaternion inspection + ///////////////////////// + + // Return true if this quaternion is equal to 'rhs'. + // Note! Quaternions exhibit "double-cover", so any rotation has two equally valid + // quaternion representations and they will NOT compare equal. + inline bool equals(const LLQuaternion2& rhs, F32 tolerance = F_APPROXIMATELY_ZERO ) const; + + // Return true if all components are finite and the quaternion is normalized + inline bool isOkRotation() const; + +protected: + + LLVector4a mQ; + +}; + +#endif diff --git a/indra/llmath/llquaternion2.inl b/indra/llmath/llquaternion2.inl new file mode 100644 index 0000000000..9a4274d6a4 --- /dev/null +++ b/indra/llmath/llquaternion2.inl @@ -0,0 +1,108 @@ +/** + * @file llquaternion2.inl + * @brief LLQuaternion2 inline definitions + * + * $LicenseInfo:firstyear=2010&license=viewergpl$ + * + * Copyright (c) 2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#include "llquaternion2.h" + +static const LLQuad LL_V4A_PLUS_ONE = {1.f, 1.f, 1.f, 1.f}; +static const LLQuad LL_V4A_MINUS_ONE = {-1.f, -1.f, -1.f, -1.f}; + +// Ctor from LLQuaternion +inline LLQuaternion2::LLQuaternion2( const LLQuaternion& quat ) +{ + mQ.set(quat.mQ[VX], quat.mQ[VY], quat.mQ[VZ], quat.mQ[VW]); +} + +////////////////////////// +// Get/Set +////////////////////////// + +// Return the internal LLVector4a representation of the quaternion +inline const LLVector4a& LLQuaternion2::getVector4a() const +{ + return mQ; +} + +inline LLVector4a& LLQuaternion2::getVector4aRw() +{ + return mQ; +} + +///////////////////////// +// Quaternion modification +///////////////////////// + +// Set this quaternion to the conjugate of src +inline void LLQuaternion2::setConjugate(const LLQuaternion2& src) +{ + static LL_ALIGN_16( const U32 F_QUAT_INV_MASK_4A[4] ) = { 0x80000000, 0x80000000, 0x80000000, 0x00000000 }; + mQ = _mm_xor_ps(src.mQ, *reinterpret_cast<const LLQuad*>(&F_QUAT_INV_MASK_4A)); +} + +// Renormalizes the quaternion. Assumes it has nonzero length. +inline void LLQuaternion2::normalize() +{ + mQ.normalize4(); +} + +// Quantize this quaternion to 8 bit precision +inline void LLQuaternion2::quantize8() +{ + mQ.quantize8( LL_V4A_MINUS_ONE, LL_V4A_PLUS_ONE ); + normalize(); +} + +// Quantize this quaternion to 16 bit precision +inline void LLQuaternion2::quantize16() +{ + mQ.quantize16( LL_V4A_MINUS_ONE, LL_V4A_PLUS_ONE ); + normalize(); +} + + +///////////////////////// +// Quaternion inspection +///////////////////////// + +// Return true if this quaternion is equal to 'rhs'. +// Note! Quaternions exhibit "double-cover", so any rotation has two equally valid +// quaternion representations and they will NOT compare equal. +inline bool LLQuaternion2::equals(const LLQuaternion2 &rhs, F32 tolerance/* = F_APPROXIMATELY_ZERO*/) const +{ + return mQ.equals4(rhs.mQ, tolerance); +} + +// Return true if all components are finite and the quaternion is normalized +inline bool LLQuaternion2::isOkRotation() const +{ + return mQ.isFinite4() && mQ.isNormalized4(); +} + diff --git a/indra/llmath/llsimdmath.h b/indra/llmath/llsimdmath.h new file mode 100644 index 0000000000..9377bfdb53 --- /dev/null +++ b/indra/llmath/llsimdmath.h @@ -0,0 +1,95 @@ +/** + * @file llsimdmath.h + * @brief Common header for SIMD-based math library (llvector4a, llmatrix3a, etc.) + * + * $LicenseInfo:firstyear=2010&license=viewergpl$ + * + * Copyright (c) 2007-2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#ifndef LL_SIMD_MATH_H +#define LL_SIMD_MATH_H + +#ifndef LLMATH_H +#error "Please include llmath.h before this file." +#endif + +#if ( ( LL_DARWIN || LL_LINUX ) && !(__SSE2__) ) || ( LL_WINDOWS && ( _M_IX86_FP < 2 ) ) +#error SSE2 not enabled. LLVector4a and related class will not compile. +#endif + +template <typename T> T* LL_NEXT_ALIGNED_ADDRESS(T* address) +{ + return reinterpret_cast<T*>( + (reinterpret_cast<U32>(address) + 0xF) & ~0xF); +} + +template <typename T> T* LL_NEXT_ALIGNED_ADDRESS_64(T* address) +{ + return reinterpret_cast<T*>( + (reinterpret_cast<U32>(address) + 0x3F) & ~0x3F); +} + +#if LL_LINUX || LL_DARWIN + +#define LL_ALIGN_PREFIX(x) +#define LL_ALIGN_POSTFIX(x) __attribute__((aligned(x))) + +#elif LL_WINDOWS + +#define LL_ALIGN_PREFIX(x) __declspec(align(x)) +#define LL_ALIGN_POSTFIX(x) + +#else +#error "LL_ALIGN_PREFIX and LL_ALIGN_POSTFIX undefined" +#endif + +#define LL_ALIGN_16(var) LL_ALIGN_PREFIX(16) var LL_ALIGN_POSTFIX(16) + + + +#include <xmmintrin.h> +#include <emmintrin.h> + +#include "llsimdtypes.h" +#include "llsimdtypes.inl" + +class LLMatrix3a; +class LLRotation; +class LLMatrix3; + +#include "llquaternion.h" + +#include "llvector4logical.h" +#include "llvector4a.h" +#include "llmatrix3a.h" +#include "llquaternion2.h" +#include "llvector4a.inl" +#include "llmatrix3a.inl" +#include "llquaternion2.inl" + + +#endif //LL_SIMD_MATH_H diff --git a/indra/llmath/llsimdtypes.h b/indra/llmath/llsimdtypes.h new file mode 100644 index 0000000000..82e318c8bf --- /dev/null +++ b/indra/llmath/llsimdtypes.h @@ -0,0 +1,130 @@ +/** + * @file llsimdtypes.h + * @brief Declaration of basic SIMD math related types + * + * $LicenseInfo:firstyear=2010&license=viewergpl$ + * + * Copyright (c) 2007-2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#ifndef LL_SIMD_TYPES_H +#define LL_SIMD_TYPES_H + +#ifndef LL_SIMD_MATH_H +#error "Please include llmath.h before this file." +#endif + +typedef __m128 LLQuad; + + +#if LL_WINDOWS +#pragma warning(push) +#pragma warning( disable : 4800 3 ) // Disable warning about casting int to bool for this class. +#if defined(_MSC_VER) && (_MSC_VER < 1500) +// VC++ 2005 is missing these intrinsics +// __forceinline is MSVC specific and attempts to override compiler inlining judgment. This is so +// even in debug builds this call is a NOP. +__forceinline const __m128 _mm_castsi128_ps( const __m128i a ) { return reinterpret_cast<const __m128&>(a); } +__forceinline const __m128i _mm_castps_si128( const __m128 a ) { return reinterpret_cast<const __m128i&>(a); } +#endif // _MSC_VER + +#endif // LL_WINDOWS + +class LLBool32 +{ +public: + inline LLBool32() {} + inline LLBool32(int rhs) : m_bool(rhs) {} + inline LLBool32(unsigned int rhs) : m_bool(rhs) {} + inline LLBool32(bool rhs) { m_bool = static_cast<const int>(rhs); } + inline LLBool32& operator= (bool rhs) { m_bool = (int)rhs; return *this; } + inline bool operator== (bool rhs) const { return static_cast<const bool&>(m_bool) == rhs; } + inline bool operator!= (bool rhs) const { return !operator==(rhs); } + inline operator bool() const { return static_cast<const bool&>(m_bool); } + +private: + int m_bool; +}; + +#if LL_WINDOWS +#pragma warning(pop) +#endif + +class LLSimdScalar +{ +public: + inline LLSimdScalar() {} + inline LLSimdScalar(LLQuad q) + { + mQ = q; + } + + inline LLSimdScalar(F32 f) + { + mQ = _mm_set_ss(f); + } + + static inline const LLSimdScalar& getZero() + { + extern const LLQuad F_ZERO_4A; + return reinterpret_cast<const LLSimdScalar&>(F_ZERO_4A); + } + + inline F32 getF32() const; + + inline LLBool32 isApproximatelyEqual(const LLSimdScalar& rhs, F32 tolerance = F_APPROXIMATELY_ZERO) const; + + inline LLSimdScalar getAbs() const; + + inline void setMax( const LLSimdScalar& a, const LLSimdScalar& b ); + + inline void setMin( const LLSimdScalar& a, const LLSimdScalar& b ); + + inline LLSimdScalar& operator=(F32 rhs); + + inline LLSimdScalar& operator+=(const LLSimdScalar& rhs); + + inline LLSimdScalar& operator-=(const LLSimdScalar& rhs); + + inline LLSimdScalar& operator*=(const LLSimdScalar& rhs); + + inline LLSimdScalar& operator/=(const LLSimdScalar& rhs); + + inline operator LLQuad() const + { + return mQ; + } + + inline const LLQuad& getQuad() const + { + return mQ; + } + +private: + LLQuad mQ; +}; + +#endif //LL_SIMD_TYPES_H diff --git a/indra/llmath/llsimdtypes.inl b/indra/llmath/llsimdtypes.inl new file mode 100644 index 0000000000..69c858e310 --- /dev/null +++ b/indra/llmath/llsimdtypes.inl @@ -0,0 +1,163 @@ +/** + * @file llsimdtypes.inl + * @brief Inlined definitions of basic SIMD math related types + * + * $LicenseInfo:firstyear=2010&license=viewergpl$ + * + * Copyright (c) 2007-2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + + + + +////////////////// +// LLSimdScalar +////////////////// + +inline LLSimdScalar operator+(const LLSimdScalar& a, const LLSimdScalar& b) +{ + LLSimdScalar t(a); + t += b; + return t; +} + +inline LLSimdScalar operator-(const LLSimdScalar& a, const LLSimdScalar& b) +{ + LLSimdScalar t(a); + t -= b; + return t; +} + +inline LLSimdScalar operator*(const LLSimdScalar& a, const LLSimdScalar& b) +{ + LLSimdScalar t(a); + t *= b; + return t; +} + +inline LLSimdScalar operator/(const LLSimdScalar& a, const LLSimdScalar& b) +{ + LLSimdScalar t(a); + t /= b; + return t; +} + +inline LLSimdScalar operator-(const LLSimdScalar& a) +{ + static LL_ALIGN_16(const U32 signMask[4]) = {0x80000000, 0x80000000, 0x80000000, 0x80000000 }; + return _mm_xor_ps(*reinterpret_cast<const LLQuad*>(signMask), a); +} + +inline LLBool32 operator==(const LLSimdScalar& a, const LLSimdScalar& b) +{ + return _mm_comieq_ss(a, b); +} + +inline LLBool32 operator!=(const LLSimdScalar& a, const LLSimdScalar& b) +{ + return _mm_comineq_ss(a, b); +} + +inline LLBool32 operator<(const LLSimdScalar& a, const LLSimdScalar& b) +{ + return _mm_comilt_ss(a, b); +} + +inline LLBool32 operator<=(const LLSimdScalar& a, const LLSimdScalar& b) +{ + return _mm_comile_ss(a, b); +} + +inline LLBool32 operator>(const LLSimdScalar& a, const LLSimdScalar& b) +{ + return _mm_comigt_ss(a, b); +} + +inline LLBool32 operator>=(const LLSimdScalar& a, const LLSimdScalar& b) +{ + return _mm_comige_ss(a, b); +} + +inline LLBool32 LLSimdScalar::isApproximatelyEqual(const LLSimdScalar& rhs, F32 tolerance /* = F_APPROXIMATELY_ZERO */) const +{ + const LLSimdScalar tol( tolerance ); + const LLSimdScalar diff = _mm_sub_ss( mQ, rhs.mQ ); + const LLSimdScalar absDiff = diff.getAbs(); + return absDiff <= tol; +} + +inline void LLSimdScalar::setMax( const LLSimdScalar& a, const LLSimdScalar& b ) +{ + mQ = _mm_max_ss( a, b ); +} + +inline void LLSimdScalar::setMin( const LLSimdScalar& a, const LLSimdScalar& b ) +{ + mQ = _mm_min_ss( a, b ); +} + +inline LLSimdScalar& LLSimdScalar::operator=(F32 rhs) +{ + mQ = _mm_set_ss(rhs); + return *this; +} + +inline LLSimdScalar& LLSimdScalar::operator+=(const LLSimdScalar& rhs) +{ + mQ = _mm_add_ss( mQ, rhs ); + return *this; +} + +inline LLSimdScalar& LLSimdScalar::operator-=(const LLSimdScalar& rhs) +{ + mQ = _mm_sub_ss( mQ, rhs ); + return *this; +} + +inline LLSimdScalar& LLSimdScalar::operator*=(const LLSimdScalar& rhs) +{ + mQ = _mm_mul_ss( mQ, rhs ); + return *this; +} + +inline LLSimdScalar& LLSimdScalar::operator/=(const LLSimdScalar& rhs) +{ + mQ = _mm_div_ss( mQ, rhs ); + return *this; +} + +inline LLSimdScalar LLSimdScalar::getAbs() const +{ + static const LL_ALIGN_16(U32 F_ABS_MASK_4A[4]) = { 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF }; + return _mm_and_ps( mQ, *reinterpret_cast<const LLQuad*>(F_ABS_MASK_4A)); +} + +inline F32 LLSimdScalar::getF32() const +{ + F32 ret; + _mm_store_ss(&ret, mQ); + return ret; +} diff --git a/indra/llmath/lltreenode.h b/indra/llmath/lltreenode.h index ee9836241a..e6d2521b2a 100644 --- a/indra/llmath/lltreenode.h +++ b/indra/llmath/lltreenode.h @@ -34,6 +34,9 @@ #include "stdtypes.h" #include "xform.h" +#include "llpointer.h" +#include "llrefcount.h" + #include <vector> template <class T> class LLTreeNode; diff --git a/indra/llmath/llvector4a.cpp b/indra/llmath/llvector4a.cpp new file mode 100644 index 0000000000..b62c17302f --- /dev/null +++ b/indra/llmath/llvector4a.cpp @@ -0,0 +1,228 @@ +/** + * @file llvector4a.cpp + * @brief SIMD vector implementation + * + * $LicenseInfo:firstyear=2010&license=viewergpl$ + * + * Copyright (c) 2007-2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#include "llmath.h" +#include "llquantize.h" + +extern const LLQuad F_ZERO_4A = { 0, 0, 0, 0 }; +extern const LLQuad F_APPROXIMATELY_ZERO_4A = { + F_APPROXIMATELY_ZERO, + F_APPROXIMATELY_ZERO, + F_APPROXIMATELY_ZERO, + F_APPROXIMATELY_ZERO +}; + +extern const LLVector4a LL_V4A_ZERO = reinterpret_cast<const LLVector4a&> ( F_ZERO_4A ); +extern const LLVector4a LL_V4A_EPSILON = reinterpret_cast<const LLVector4a&> ( F_APPROXIMATELY_ZERO_4A ); + +/*static */void LLVector4a::memcpyNonAliased16(F32* __restrict dst, const F32* __restrict src, size_t bytes) +{ + assert(src != NULL); + assert(dst != NULL); + assert(bytes > 0); + assert((bytes % sizeof(F32))== 0); + + F32* end = dst + (bytes / sizeof(F32) ); + + if (bytes > 64) + { + F32* begin_64 = LL_NEXT_ALIGNED_ADDRESS_64(dst); + + //at least 64 (16*4) bytes before the end of the destination, switch to 16 byte copies + F32* end_64 = end-16; + + _mm_prefetch((char*)begin_64, _MM_HINT_NTA); + _mm_prefetch((char*)begin_64 + 64, _MM_HINT_NTA); + _mm_prefetch((char*)begin_64 + 128, _MM_HINT_NTA); + _mm_prefetch((char*)begin_64 + 192, _MM_HINT_NTA); + + while (dst < begin_64) + { + copy4a(dst, src); + dst += 4; + src += 4; + } + + while (dst < end_64) + { + _mm_prefetch((char*)src + 512, _MM_HINT_NTA); + _mm_prefetch((char*)dst + 512, _MM_HINT_NTA); + copy4a(dst, src); + copy4a(dst+4, src+4); + copy4a(dst+8, src+8); + copy4a(dst+12, src+12); + + dst += 16; + src += 16; + } + } + + while (dst < end) + { + copy4a(dst, src); + dst += 4; + src += 4; + } +} + +void LLVector4a::setRotated( const LLRotation& rot, const LLVector4a& vec ) +{ + const LLVector4a col0 = rot.getColumn(0); + const LLVector4a col1 = rot.getColumn(1); + const LLVector4a col2 = rot.getColumn(2); + + LLVector4a result = _mm_load_ss( vec.getF32ptr() ); + result.splat<0>( result ); + result.mul( col0 ); + + { + LLVector4a yyyy = _mm_load_ss( vec.getF32ptr() + 1 ); + yyyy.splat<0>( yyyy ); + yyyy.mul( col1 ); + result.add( yyyy ); + } + + { + LLVector4a zzzz = _mm_load_ss( vec.getF32ptr() + 2 ); + zzzz.splat<0>( zzzz ); + zzzz.mul( col2 ); + result.add( zzzz ); + } + + *this = result; +} + +void LLVector4a::setRotated( const LLQuaternion2& quat, const LLVector4a& vec ) +{ + const LLVector4a& quatVec = quat.getVector4a(); + LLVector4a temp; temp.setCross3(quatVec, vec); + temp.add( temp ); + + const LLVector4a realPart( quatVec.getScalarAt<3>() ); + LLVector4a tempTimesReal; tempTimesReal.setMul( temp, realPart ); + + mQ = vec; + add( tempTimesReal ); + + LLVector4a imagCrossTemp; imagCrossTemp.setCross3( quatVec, temp ); + add(imagCrossTemp); +} + +void LLVector4a::quantize8( const LLVector4a& low, const LLVector4a& high ) +{ + LLVector4a val(mQ); + LLVector4a delta; delta.setSub( high, low ); + + { + val.clamp(low, high); + val.sub(low); + + // 8-bit quantization means we can do with just 12 bits of reciprocal accuracy + const LLVector4a oneOverDelta = _mm_rcp_ps(delta.mQ); +// { +// static LL_ALIGN_16( const F32 F_TWO_4A[4] ) = { 2.f, 2.f, 2.f, 2.f }; +// LLVector4a two; two.load4a( F_TWO_4A ); +// +// // Here we use _mm_rcp_ps plus one round of newton-raphson +// // We wish to find 'x' such that x = 1/delta +// // As a first approximation, we take x0 = _mm_rcp_ps(delta) +// // Then x1 = 2 * x0 - a * x0^2 or x1 = x0 * ( 2 - a * x0 ) +// // See Intel AP-803 http://ompf.org/!/Intel_application_note_AP-803.pdf +// const LLVector4a recipApprox = _mm_rcp_ps(delta.mQ); +// oneOverDelta.setMul( delta, recipApprox ); +// oneOverDelta.setSub( two, oneOverDelta ); +// oneOverDelta.mul( recipApprox ); +// } + + val.mul(oneOverDelta); + val.mul(*reinterpret_cast<const LLVector4a*>(F_U8MAX_4A)); + } + + val = _mm_cvtepi32_ps(_mm_cvtps_epi32( val.mQ )); + + { + val.mul(*reinterpret_cast<const LLVector4a*>(F_OOU8MAX_4A)); + val.mul(delta); + val.add(low); + } + + { + LLVector4a maxError; maxError.setMul(delta, *reinterpret_cast<const LLVector4a*>(F_OOU8MAX_4A)); + LLVector4a absVal; absVal.setAbs( val ); + setSelectWithMask( absVal.lessThan( maxError ), F_ZERO_4A, val ); + } +} + +void LLVector4a::quantize16( const LLVector4a& low, const LLVector4a& high ) +{ + LLVector4a val(mQ); + LLVector4a delta; delta.setSub( high, low ); + + { + val.clamp(low, high); + val.sub(low); + + // 16-bit quantization means we need a round of Newton-Raphson + LLVector4a oneOverDelta; + { + static LL_ALIGN_16( const F32 F_TWO_4A[4] ) = { 2.f, 2.f, 2.f, 2.f }; + LLVector4a two; two.load4a( F_TWO_4A ); + + // Here we use _mm_rcp_ps plus one round of newton-raphson + // We wish to find 'x' such that x = 1/delta + // As a first approximation, we take x0 = _mm_rcp_ps(delta) + // Then x1 = 2 * x0 - a * x0^2 or x1 = x0 * ( 2 - a * x0 ) + // See Intel AP-803 http://ompf.org/!/Intel_application_note_AP-803.pdf + const LLVector4a recipApprox = _mm_rcp_ps(delta.mQ); + oneOverDelta.setMul( delta, recipApprox ); + oneOverDelta.setSub( two, oneOverDelta ); + oneOverDelta.mul( recipApprox ); + } + + val.mul(oneOverDelta); + val.mul(*reinterpret_cast<const LLVector4a*>(F_U16MAX_4A)); + } + + val = _mm_cvtepi32_ps(_mm_cvtps_epi32( val.mQ )); + + { + val.mul(*reinterpret_cast<const LLVector4a*>(F_OOU16MAX_4A)); + val.mul(delta); + val.add(low); + } + + { + LLVector4a maxError; maxError.setMul(delta, *reinterpret_cast<const LLVector4a*>(F_OOU16MAX_4A)); + LLVector4a absVal; absVal.setAbs( val ); + setSelectWithMask( absVal.lessThan( maxError ), F_ZERO_4A, val ); + } +} diff --git a/indra/llmath/llvector4a.h b/indra/llmath/llvector4a.h new file mode 100644 index 0000000000..76a3e999ce --- /dev/null +++ b/indra/llmath/llvector4a.h @@ -0,0 +1,331 @@ +/** + * @file llvector4a.h + * @brief LLVector4a class header file - memory aligned and vectorized 4 component vector + * + * $LicenseInfo:firstyear=2010&license=viewergpl$ + * + * Copyright (c) 2007-2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#ifndef LL_LLVECTOR4A_H +#define LL_LLVECTOR4A_H + + +class LLRotation; + +#include <assert.h> +#include "llpreprocessor.h" + +/////////////////////////////////// +// FIRST TIME USERS PLEASE READ +////////////////////////////////// +// This is just the beginning of LLVector4a. There are many more useful functions +// yet to be implemented. For example, setNeg to negate a vector, rotate() to apply +// a matrix rotation, various functions to manipulate only the X, Y, and Z elements +// and many others (including a whole variety of accessors). So if you don't see a +// function here that you need, please contact Falcon or someone else with SSE +// experience (Richard, I think, has some and davep has a little as of the time +// of this writing, July 08, 2010) about getting it implemented before you resort to +// LLVector3/LLVector4. +///////////////////////////////// + +class LLVector4a +{ +public: + + /////////////////////////////////// + // STATIC METHODS + /////////////////////////////////// + + // Call initClass() at startup to avoid 15,000+ cycle penalties from denormalized numbers + static void initClass() + { + _MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON); + _MM_SET_ROUNDING_MODE(_MM_ROUND_NEAREST); + } + + // Return a vector of all zeros + static inline const LLVector4a& getZero() + { + extern const LLVector4a LL_V4A_ZERO; + return LL_V4A_ZERO; + } + + // Return a vector of all epsilon, where epsilon is a small float suitable for approximate equality checks + static inline const LLVector4a& getEpsilon() + { + extern const LLVector4a LL_V4A_EPSILON; + return LL_V4A_EPSILON; + } + + // Copy 16 bytes from src to dst. Source and destination must be 16-byte aligned + static inline void copy4a(F32* dst, const F32* src) + { + _mm_store_ps(dst, _mm_load_ps(src)); + } + + // Copy words 16-byte blocks from src to dst. Source and destination must not overlap. + static void memcpyNonAliased16(F32* __restrict dst, const F32* __restrict src, size_t bytes); + + //////////////////////////////////// + // CONSTRUCTORS + //////////////////////////////////// + + LLVector4a() + { //DO NOT INITIALIZE -- The overhead is completely unnecessary + } + + LLVector4a(F32 x, F32 y, F32 z, F32 w = 0.f) + { + set(x,y,z,w); + } + + LLVector4a(F32 x) + { + splat(x); + } + + LLVector4a(const LLSimdScalar& x) + { + splat(x); + } + + LLVector4a(LLQuad q) + { + mQ = q; + } + + //////////////////////////////////// + // LOAD/STORE + //////////////////////////////////// + + // Load from 16-byte aligned src array (preferred method of loading) + inline void load4a(const F32* src); + + // Load from unaligned src array (NB: Significantly slower than load4a) + inline void loadua(const F32* src); + + // Load only three floats beginning at address 'src'. Slowest method. + inline void load3(const F32* src); + + // Store to a 16-byte aligned memory address + inline void store4a(F32* dst) const; + + //////////////////////////////////// + // BASIC GET/SET + //////////////////////////////////// + + // Return a "this" as an F32 pointer. Do not use unless you have a very good reason. (Not sure? Ask Falcon) + inline F32* getF32ptr(); + + // Return a "this" as a const F32 pointer. Do not use unless you have a very good reason. (Not sure? Ask Falcon) + inline const F32* const getF32ptr() const; + + // Read-only access a single float in this vector. Do not use in proximity to any function call that manipulates + // the data at the whole vector level or you will incur a substantial penalty. Consider using the splat functions instead + inline F32 operator[](const S32 idx) const; + + // Prefer this method for read-only access to a single element. Prefer the templated version if the elem is known at compile time. + inline LLSimdScalar getScalarAt(const S32 idx) const; + + // Prefer this method for read-only access to a single element. Prefer the templated version if the elem is known at compile time. + template <int N> LL_FORCE_INLINE LLSimdScalar getScalarAt() const; + template <> LL_FORCE_INLINE LLSimdScalar getScalarAt<0>() const; + + // Set to an x, y, z and optional w provided + inline void set(F32 x, F32 y, F32 z, F32 w = 0.f); + + // Set to all zeros. This is preferred to using ::getZero() + inline void clear(); + + // Set all elements to 'x' + inline void splat(const F32 x); + + // Set all elements to 'x' + inline void splat(const LLSimdScalar& x); + + // Set all 4 elements to element N of src, with N known at compile time + template <int N> void splat(const LLVector4a& src); + + // Set all 4 elements to element i of v, with i NOT known at compile time + inline void splat(const LLVector4a& v, U32 i); + + // Select bits from sourceIfTrue and sourceIfFalse according to bits in mask + inline void setSelectWithMask( const LLVector4Logical& mask, const LLVector4a& sourceIfTrue, const LLVector4a& sourceIfFalse ); + + //////////////////////////////////// + // ALGEBRAIC + //////////////////////////////////// + + // Set this to the element-wise (a + b) + inline void setAdd(const LLVector4a& a, const LLVector4a& b); + + // Set this to element-wise (a - b) + inline void setSub(const LLVector4a& a, const LLVector4a& b); + + // Set this to element-wise multiply (a * b) + inline void setMul(const LLVector4a& a, const LLVector4a& b); + + // Set this to element-wise quotient (a / b) + inline void setDiv(const LLVector4a& a, const LLVector4a& b); + + // Set this to the element-wise absolute value of src + inline void setAbs(const LLVector4a& src); + + // Add to each component in this vector the corresponding component in rhs + inline void add(const LLVector4a& rhs); + + // Subtract from each component in this vector the corresponding component in rhs + inline void sub(const LLVector4a& rhs); + + // Multiply each component in this vector by the corresponding component in rhs + inline void mul(const LLVector4a& rhs); + + // Divide each component in this vector by the corresponding component in rhs + inline void div(const LLVector4a& rhs); + + // Multiply this vector by x in a scalar fashion + inline void mul(const F32 x); + + // Set this to (a x b) (geometric cross-product) + inline void setCross3(const LLVector4a& a, const LLVector4a& b); + + // Set all elements to the dot product of the x, y, and z elements in a and b + inline void setAllDot3(const LLVector4a& a, const LLVector4a& b); + + // Set all elements to the dot product of the x, y, z, and w elements in a and b + inline void setAllDot4(const LLVector4a& a, const LLVector4a& b); + + // Return the 3D dot product of this vector and b + inline LLSimdScalar dot3(const LLVector4a& b) const; + + // Return the 4D dot product of this vector and b + inline LLSimdScalar dot4(const LLVector4a& b) const; + + // Normalize this vector with respect to the x, y, and z components only. Accurate to 22 bites of precision. W component is destroyed + // Note that this does not consider zero length vectors! + inline void normalize3(); + + // Same as normalize3() but with respect to all 4 components + inline void normalize4(); + + // Same as normalize3(), but returns length as a SIMD scalar + inline LLSimdScalar normalize3withLength(); + + // Normalize this vector with respect to the x, y, and z components only. Accurate only to 10-12 bits of precision. W component is destroyed + // Note that this does not consider zero length vectors! + inline void normalize3fast(); + + // Return true if this vector is normalized with respect to x,y,z up to tolerance + inline LLBool32 isNormalized3( F32 tolerance = 1e-3 ) const; + + // Return true if this vector is normalized with respect to all components up to tolerance + inline LLBool32 isNormalized4( F32 tolerance = 1e-3 ) const; + + // Set all elements to the length of vector 'v' + inline void setAllLength3( const LLVector4a& v ); + + // Get this vector's length + inline LLSimdScalar getLength3() const; + + // Set the components of this vector to the minimum of the corresponding components of lhs and rhs + inline void setMin(const LLVector4a& lhs, const LLVector4a& rhs); + + // Set the components of this vector to the maximum of the corresponding components of lhs and rhs + inline void setMax(const LLVector4a& lhs, const LLVector4a& rhs); + + // Clamps this vector to be within the component-wise range low to high (inclusive) + inline void clamp( const LLVector4a& low, const LLVector4a& high ); + + // Set this to (c * lhs) + rhs * ( 1 - c) + inline void setLerp(const LLVector4a& lhs, const LLVector4a& rhs, F32 c); + + // Return true (nonzero) if x, y, z (and w for Finite4) are all finite floats + inline LLBool32 isFinite3() const; + inline LLBool32 isFinite4() const; + + // Set this vector to 'vec' rotated by the LLRotation or LLQuaternion2 provided + void setRotated( const LLRotation& rot, const LLVector4a& vec ); + void setRotated( const class LLQuaternion2& quat, const LLVector4a& vec ); + + // Set this vector to 'vec' rotated by the INVERSE of the LLRotation or LLQuaternion2 provided + inline void setRotatedInv( const LLRotation& rot, const LLVector4a& vec ); + inline void setRotatedInv( const class LLQuaternion2& quat, const LLVector4a& vec ); + + // Quantize this vector to 8 or 16 bit precision + void quantize8( const LLVector4a& low, const LLVector4a& high ); + void quantize16( const LLVector4a& low, const LLVector4a& high ); + + //////////////////////////////////// + // LOGICAL + //////////////////////////////////// + // The functions in this section will compare the elements in this vector + // to those in rhs and return an LLVector4Logical with all bits set in elements + // where the comparison was true and all bits unset in elements where the comparison + // was false. See llvector4logica.h + //////////////////////////////////// + // WARNING: Other than equals3 and equals4, these functions do NOT account + // for floating point tolerance. You should include the appropriate tolerance + // in the inputs. + //////////////////////////////////// + + inline LLVector4Logical greaterThan(const LLVector4a& rhs) const; + + inline LLVector4Logical lessThan(const LLVector4a& rhs) const; + + inline LLVector4Logical greaterEqual(const LLVector4a& rhs) const; + + inline LLVector4Logical lessEqual(const LLVector4a& rhs) const; + + inline LLVector4Logical equal(const LLVector4a& rhs) const; + + // Returns true if this and rhs are componentwise equal up to the specified absolute tolerance + inline bool equals4(const LLVector4a& rhs, F32 tolerance = F_APPROXIMATELY_ZERO ) const; + + inline bool equals3(const LLVector4a& rhs, F32 tolerance = F_APPROXIMATELY_ZERO ) const; + + //////////////////////////////////// + // OPERATORS + //////////////////////////////////// + + // Do NOT add aditional operators without consulting someone with SSE experience + inline const LLVector4a& operator= ( const LLVector4a& rhs ); + + inline const LLVector4a& operator= ( const LLQuad& rhs ); + + inline operator LLQuad() const; + +private: + LLQuad mQ; +}; + +inline void update_min_max(LLVector4a& min, LLVector4a& max, const LLVector4a& p) +{ + min.setMin(min, p); + max.setMax(max, p); +} + +#endif diff --git a/indra/llmath/llvector4a.inl b/indra/llmath/llvector4a.inl new file mode 100644 index 0000000000..e52b550883 --- /dev/null +++ b/indra/llmath/llvector4a.inl @@ -0,0 +1,599 @@ +/** + * @file llvector4a.inl + * @brief LLVector4a inline function implementations + * + * $LicenseInfo:firstyear=2010&license=viewergpl$ + * + * Copyright (c) 2007-2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +//////////////////////////////////// +// LOAD/STORE +//////////////////////////////////// + +// Load from 16-byte aligned src array (preferred method of loading) +inline void LLVector4a::load4a(const F32* src) +{ + mQ = _mm_load_ps(src); +} + +// Load from unaligned src array (NB: Significantly slower than load4a) +inline void LLVector4a::loadua(const F32* src) +{ + mQ = _mm_loadu_ps(src); +} + +// Load only three floats beginning at address 'src'. Slowest method. +inline void LLVector4a::load3(const F32* src) +{ + // mQ = { 0.f, src[2], src[1], src[0] } = { W, Z, Y, X } + // NB: This differs from the convention of { Z, Y, X, W } + mQ = _mm_set_ps(0.f, src[2], src[1], src[0]); +} + +// Store to a 16-byte aligned memory address +inline void LLVector4a::store4a(F32* dst) const +{ + _mm_store_ps(dst, mQ); +} + +//////////////////////////////////// +// BASIC GET/SET +//////////////////////////////////// + +// Return a "this" as an F32 pointer. Do not use unless you have a very good reason. (Not sure? Ask Falcon) +F32* LLVector4a::getF32ptr() +{ + return (F32*) &mQ; +} + +// Return a "this" as a const F32 pointer. Do not use unless you have a very good reason. (Not sure? Ask Falcon) +const F32* const LLVector4a::getF32ptr() const +{ + return (const F32* const) &mQ; +} + +// Read-only access a single float in this vector. Do not use in proximity to any function call that manipulates +// the data at the whole vector level or you will incur a substantial penalty. Consider using the splat functions instead +inline F32 LLVector4a::operator[](const S32 idx) const +{ + return ((F32*)&mQ)[idx]; +} + +// Prefer this method for read-only access to a single element. Prefer the templated version if the elem is known at compile time. +inline LLSimdScalar LLVector4a::getScalarAt(const S32 idx) const +{ + // Return appropriate LLQuad. It will be cast to LLSimdScalar automatically (should be effectively a nop) + switch (idx) + { + case 0: + return mQ; + case 1: + return _mm_shuffle_ps(mQ, mQ, _MM_SHUFFLE(1, 1, 1, 1)); + case 2: + return _mm_shuffle_ps(mQ, mQ, _MM_SHUFFLE(2, 2, 2, 2)); + case 3: + default: + return _mm_shuffle_ps(mQ, mQ, _MM_SHUFFLE(3, 3, 3, 3)); + } +} + +// Prefer this method for read-only access to a single element. Prefer the templated version if the elem is known at compile time. +template <int N> LL_FORCE_INLINE LLSimdScalar LLVector4a::getScalarAt() const +{ + return _mm_shuffle_ps(mQ, mQ, _MM_SHUFFLE(N, N, N, N)); +} + +template<> LL_FORCE_INLINE LLSimdScalar LLVector4a::getScalarAt<0>() const +{ + return mQ; +} + +// Set to an x, y, z and optional w provided +inline void LLVector4a::set(F32 x, F32 y, F32 z, F32 w) +{ + mQ = _mm_set_ps(w, z, y, x); +} + +// Set to all zeros +inline void LLVector4a::clear() +{ + mQ = LLVector4a::getZero().mQ; +} + +inline void LLVector4a::splat(const F32 x) +{ + mQ = _mm_set1_ps(x); +} + +inline void LLVector4a::splat(const LLSimdScalar& x) +{ + mQ = _mm_shuffle_ps( x.getQuad(), x.getQuad(), _MM_SHUFFLE(0,0,0,0) ); +} + +// Set all 4 elements to element N of src, with N known at compile time +template <int N> void LLVector4a::splat(const LLVector4a& src) +{ + mQ = _mm_shuffle_ps(src.mQ, src.mQ, _MM_SHUFFLE(N, N, N, N) ); +} + +// Set all 4 elements to element i of v, with i NOT known at compile time +inline void LLVector4a::splat(const LLVector4a& v, U32 i) +{ + switch (i) + { + case 0: + mQ = _mm_shuffle_ps(v.mQ, v.mQ, _MM_SHUFFLE(0, 0, 0, 0)); + break; + case 1: + mQ = _mm_shuffle_ps(v.mQ, v.mQ, _MM_SHUFFLE(1, 1, 1, 1)); + break; + case 2: + mQ = _mm_shuffle_ps(v.mQ, v.mQ, _MM_SHUFFLE(2, 2, 2, 2)); + break; + case 3: + mQ = _mm_shuffle_ps(v.mQ, v.mQ, _MM_SHUFFLE(3, 3, 3, 3)); + break; + } +} + +// Select bits from sourceIfTrue and sourceIfFalse according to bits in mask +inline void LLVector4a::setSelectWithMask( const LLVector4Logical& mask, const LLVector4a& sourceIfTrue, const LLVector4a& sourceIfFalse ) +{ + // ((( sourceIfTrue ^ sourceIfFalse ) & mask) ^ sourceIfFalse ) + // E.g., sourceIfFalse = 1010b, sourceIfTrue = 0101b, mask = 1100b + // (sourceIfTrue ^ sourceIfFalse) = 1111b --> & mask = 1100b --> ^ sourceIfFalse = 0110b, + // as expected (01 from sourceIfTrue, 10 from sourceIfFalse) + // Courtesy of Mark++, http://markplusplus.wordpress.com/2007/03/14/fast-sse-select-operation/ + mQ = _mm_xor_ps( sourceIfFalse, _mm_and_ps( mask, _mm_xor_ps( sourceIfTrue, sourceIfFalse ) ) ); +} + +//////////////////////////////////// +// ALGEBRAIC +//////////////////////////////////// + +// Set this to the element-wise (a + b) +inline void LLVector4a::setAdd(const LLVector4a& a, const LLVector4a& b) +{ + mQ = _mm_add_ps(a.mQ, b.mQ); +} + +// Set this to element-wise (a - b) +inline void LLVector4a::setSub(const LLVector4a& a, const LLVector4a& b) +{ + mQ = _mm_sub_ps(a.mQ, b.mQ); +} + +// Set this to element-wise multiply (a * b) +inline void LLVector4a::setMul(const LLVector4a& a, const LLVector4a& b) +{ + mQ = _mm_mul_ps(a.mQ, b.mQ); +} + +// Set this to element-wise quotient (a / b) +inline void LLVector4a::setDiv(const LLVector4a& a, const LLVector4a& b) +{ + mQ = _mm_div_ps( a.mQ, b.mQ ); +} + +// Set this to the element-wise absolute value of src +inline void LLVector4a::setAbs(const LLVector4a& src) +{ + static const LL_ALIGN_16(U32 F_ABS_MASK_4A[4]) = { 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF }; + mQ = _mm_and_ps(src.mQ, *reinterpret_cast<const LLQuad*>(F_ABS_MASK_4A)); +} + +// Add to each component in this vector the corresponding component in rhs +inline void LLVector4a::add(const LLVector4a& rhs) +{ + mQ = _mm_add_ps(mQ, rhs.mQ); +} + +// Subtract from each component in this vector the corresponding component in rhs +inline void LLVector4a::sub(const LLVector4a& rhs) +{ + mQ = _mm_sub_ps(mQ, rhs.mQ); +} + +// Multiply each component in this vector by the corresponding component in rhs +inline void LLVector4a::mul(const LLVector4a& rhs) +{ + mQ = _mm_mul_ps(mQ, rhs.mQ); +} + +// Divide each component in this vector by the corresponding component in rhs +inline void LLVector4a::div(const LLVector4a& rhs) +{ + // TODO: Check accuracy, maybe add divFast + mQ = _mm_div_ps(mQ, rhs.mQ); +} + +// Multiply this vector by x in a scalar fashion +inline void LLVector4a::mul(const F32 x) +{ + LLVector4a t; + t.splat(x); + + mQ = _mm_mul_ps(mQ, t.mQ); +} + +// Set this to (a x b) (geometric cross-product) +inline void LLVector4a::setCross3(const LLVector4a& a, const LLVector4a& b) +{ + // Vectors are stored in memory in w, z, y, x order from high to low + // Set vector1 = { a[W], a[X], a[Z], a[Y] } + const LLQuad vector1 = _mm_shuffle_ps( a.mQ, a.mQ, _MM_SHUFFLE( 3, 0, 2, 1 )); + // Set vector2 = { b[W], b[Y], b[X], b[Z] } + const LLQuad vector2 = _mm_shuffle_ps( b.mQ, b.mQ, _MM_SHUFFLE( 3, 1, 0, 2 )); + // mQ = { a[W]*b[W], a[X]*b[Y], a[Z]*b[X], a[Y]*b[Z] } + mQ = _mm_mul_ps( vector1, vector2 ); + // vector3 = { a[W], a[Y], a[X], a[Z] } + const LLQuad vector3 = _mm_shuffle_ps( a.mQ, a.mQ, _MM_SHUFFLE( 3, 1, 0, 2 )); + // vector4 = { b[W], b[X], b[Z], b[Y] } + const LLQuad vector4 = _mm_shuffle_ps( b.mQ, b.mQ, _MM_SHUFFLE( 3, 0, 2, 1 )); + // mQ = { 0, a[X]*b[Y] - a[Y]*b[X], a[Z]*b[X] - a[X]*b[Z], a[Y]*b[Z] - a[Z]*b[Y] } + mQ = _mm_sub_ps( mQ, _mm_mul_ps( vector3, vector4 )); +} + +/* This function works, but may be slightly slower than the one below on older machines + inline void LLVector4a::setAllDot3(const LLVector4a& a, const LLVector4a& b) + { + // ab = { a[W]*b[W], a[Z]*b[Z], a[Y]*b[Y], a[X]*b[X] } + const LLQuad ab = _mm_mul_ps( a.mQ, b.mQ ); + // yzxw = { a[W]*b[W], a[Z]*b[Z], a[X]*b[X], a[Y]*b[Y] } + const LLQuad wzxy = _mm_shuffle_ps( ab, ab, _MM_SHUFFLE(3, 2, 0, 1 )); + // xPlusY = { 2*a[W]*b[W], 2 * a[Z] * b[Z], a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y] } + const LLQuad xPlusY = _mm_add_ps(ab, wzxy); + // xPlusYSplat = { a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y], a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y] } + const LLQuad xPlusYSplat = _mm_movelh_ps(xPlusY, xPlusY); + // zSplat = { a[Z]*b[Z], a[Z]*b[Z], a[Z]*b[Z], a[Z]*b[Z] } + const LLQuad zSplat = _mm_shuffle_ps( ab, ab, _MM_SHUFFLE( 2, 2, 2, 2 )); + // mQ = { a[Z] * b[Z] + a[Y] * b[Y] + a[X] * b[X], same, same, same } + mQ = _mm_add_ps(zSplat, xPlusYSplat); + }*/ + +// Set all elements to the dot product of the x, y, and z elements in a and b +inline void LLVector4a::setAllDot3(const LLVector4a& a, const LLVector4a& b) +{ + // ab = { a[W]*b[W], a[Z]*b[Z], a[Y]*b[Y], a[X]*b[X] } + const LLQuad ab = _mm_mul_ps( a.mQ, b.mQ ); + // yzxw = { a[W]*b[W], a[Z]*b[Z], a[X]*b[X], a[Y]*b[Y] } + const __m128i wzxy = _mm_shuffle_epi32(_mm_castps_si128(ab), _MM_SHUFFLE(3, 2, 0, 1 )); + // xPlusY = { 2*a[W]*b[W], 2 * a[Z] * b[Z], a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y] } + const LLQuad xPlusY = _mm_add_ps(ab, _mm_castsi128_ps(wzxy)); + // xPlusYSplat = { a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y], a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y] } + const LLQuad xPlusYSplat = _mm_movelh_ps(xPlusY, xPlusY); + // zSplat = { a[Z]*b[Z], a[Z]*b[Z], a[Z]*b[Z], a[Z]*b[Z] } + const __m128i zSplat = _mm_shuffle_epi32(_mm_castps_si128(ab), _MM_SHUFFLE( 2, 2, 2, 2 )); + // mQ = { a[Z] * b[Z] + a[Y] * b[Y] + a[X] * b[X], same, same, same } + mQ = _mm_add_ps(_mm_castsi128_ps(zSplat), xPlusYSplat); +} + +// Set all elements to the dot product of the x, y, z, and w elements in a and b +inline void LLVector4a::setAllDot4(const LLVector4a& a, const LLVector4a& b) +{ + // ab = { a[W]*b[W], a[Z]*b[Z], a[Y]*b[Y], a[X]*b[X] } + const LLQuad ab = _mm_mul_ps( a.mQ, b.mQ ); + // yzxw = { a[W]*b[W], a[Z]*b[Z], a[X]*b[X], a[Y]*b[Y] } + const __m128i zwxy = _mm_shuffle_epi32(_mm_castps_si128(ab), _MM_SHUFFLE(2, 3, 0, 1 )); + // zPlusWandXplusY = { a[W]*b[W] + a[Z]*b[Z], a[Z] * b[Z] + a[W]*b[W], a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y] } + const LLQuad zPlusWandXplusY = _mm_add_ps(ab, _mm_castsi128_ps(zwxy)); + // xPlusYSplat = { a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y], a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y] } + const LLQuad xPlusYSplat = _mm_movelh_ps(zPlusWandXplusY, zPlusWandXplusY); + const LLQuad zPlusWSplat = _mm_movehl_ps(zPlusWandXplusY, zPlusWandXplusY); + + // mQ = { a[W]*b[W] + a[Z] * b[Z] + a[Y] * b[Y] + a[X] * b[X], same, same, same } + mQ = _mm_add_ps(xPlusYSplat, zPlusWSplat); +} + +// Return the 3D dot product of this vector and b +inline LLSimdScalar LLVector4a::dot3(const LLVector4a& b) const +{ + const LLQuad ab = _mm_mul_ps( mQ, b.mQ ); + const LLQuad splatY = _mm_castsi128_ps( _mm_shuffle_epi32( _mm_castps_si128(ab), _MM_SHUFFLE(1, 1, 1, 1) ) ); + const LLQuad splatZ = _mm_castsi128_ps( _mm_shuffle_epi32( _mm_castps_si128(ab), _MM_SHUFFLE(2, 2, 2, 2) ) ); + const LLQuad xPlusY = _mm_add_ps( ab, splatY ); + return _mm_add_ps( xPlusY, splatZ ); +} + +// Return the 4D dot product of this vector and b +inline LLSimdScalar LLVector4a::dot4(const LLVector4a& b) const +{ + // ab = { w, z, y, x } + const LLQuad ab = _mm_mul_ps( mQ, b.mQ ); + // upperProdsInLowerElems = { y, x, y, x } + const LLQuad upperProdsInLowerElems = _mm_movehl_ps( ab, ab ); + // sumOfPairs = { w+y, z+x, 2y, 2x } + const LLQuad sumOfPairs = _mm_add_ps( upperProdsInLowerElems, ab ); + // shuffled = { z+x, z+x, z+x, z+x } + const LLQuad shuffled = _mm_castsi128_ps( _mm_shuffle_epi32( _mm_castps_si128( sumOfPairs ), _MM_SHUFFLE(1, 1, 1, 1) ) ); + return _mm_add_ss( sumOfPairs, shuffled ); +} + +// Normalize this vector with respect to the x, y, and z components only. Accurate to 22 bites of precision. W component is destroyed +// Note that this does not consider zero length vectors! +inline void LLVector4a::normalize3() +{ + // lenSqrd = a dot a + LLVector4a lenSqrd; lenSqrd.setAllDot3( *this, *this ); + // rsqrt = approximate reciprocal square (i.e., { ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2 } + const LLQuad rsqrt = _mm_rsqrt_ps(lenSqrd.mQ); + static const LLQuad half = { 0.5f, 0.5f, 0.5f, 0.5f }; + static const LLQuad three = {3.f, 3.f, 3.f, 3.f }; + // Now we do one round of Newton-Raphson approximation to get full accuracy + // According to the Newton-Raphson method, given a first 'w' for the root of f(x) = 1/x^2 - a (i.e., x = 1/sqrt(a)) + // the next better approximation w[i+1] = w - f(w)/f'(w) = w - (1/w^2 - a)/(-2*w^(-3)) + // w[i+1] = w + 0.5 * (1/w^2 - a) * w^3 = w + 0.5 * (w - a*w^3) = 1.5 * w - 0.5 * a * w^3 + // = 0.5 * w * (3 - a*w^2) + // Our first approx is w = rsqrt. We need out = a * w[i+1] (this is the input vector 'a', not the 'a' from the above formula + // which is actually lenSqrd). So out = a * [0.5*rsqrt * (3 - lenSqrd*rsqrt*rsqrt)] + const LLQuad AtimesRsqrt = _mm_mul_ps( lenSqrd.mQ, rsqrt ); + const LLQuad AtimesRsqrtTimesRsqrt = _mm_mul_ps( AtimesRsqrt, rsqrt ); + const LLQuad threeMinusAtimesRsqrtTimesRsqrt = _mm_sub_ps(three, AtimesRsqrtTimesRsqrt ); + const LLQuad nrApprox = _mm_mul_ps(half, _mm_mul_ps(rsqrt, threeMinusAtimesRsqrtTimesRsqrt)); + mQ = _mm_mul_ps( mQ, nrApprox ); +} + +// Normalize this vector with respect to all components. Accurate to 22 bites of precision. +// Note that this does not consider zero length vectors! +inline void LLVector4a::normalize4() +{ + // lenSqrd = a dot a + LLVector4a lenSqrd; lenSqrd.setAllDot4( *this, *this ); + // rsqrt = approximate reciprocal square (i.e., { ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2 } + const LLQuad rsqrt = _mm_rsqrt_ps(lenSqrd.mQ); + static const LLQuad half = { 0.5f, 0.5f, 0.5f, 0.5f }; + static const LLQuad three = {3.f, 3.f, 3.f, 3.f }; + // Now we do one round of Newton-Raphson approximation to get full accuracy + // According to the Newton-Raphson method, given a first 'w' for the root of f(x) = 1/x^2 - a (i.e., x = 1/sqrt(a)) + // the next better approximation w[i+1] = w - f(w)/f'(w) = w - (1/w^2 - a)/(-2*w^(-3)) + // w[i+1] = w + 0.5 * (1/w^2 - a) * w^3 = w + 0.5 * (w - a*w^3) = 1.5 * w - 0.5 * a * w^3 + // = 0.5 * w * (3 - a*w^2) + // Our first approx is w = rsqrt. We need out = a * w[i+1] (this is the input vector 'a', not the 'a' from the above formula + // which is actually lenSqrd). So out = a * [0.5*rsqrt * (3 - lenSqrd*rsqrt*rsqrt)] + const LLQuad AtimesRsqrt = _mm_mul_ps( lenSqrd.mQ, rsqrt ); + const LLQuad AtimesRsqrtTimesRsqrt = _mm_mul_ps( AtimesRsqrt, rsqrt ); + const LLQuad threeMinusAtimesRsqrtTimesRsqrt = _mm_sub_ps(three, AtimesRsqrtTimesRsqrt ); + const LLQuad nrApprox = _mm_mul_ps(half, _mm_mul_ps(rsqrt, threeMinusAtimesRsqrtTimesRsqrt)); + mQ = _mm_mul_ps( mQ, nrApprox ); +} + +// Normalize this vector with respect to the x, y, and z components only. Accurate to 22 bites of precision. W component is destroyed +// Note that this does not consider zero length vectors! +inline LLSimdScalar LLVector4a::normalize3withLength() +{ + // lenSqrd = a dot a + LLVector4a lenSqrd; lenSqrd.setAllDot3( *this, *this ); + // rsqrt = approximate reciprocal square (i.e., { ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2 } + const LLQuad rsqrt = _mm_rsqrt_ps(lenSqrd.mQ); + static const LLQuad half = { 0.5f, 0.5f, 0.5f, 0.5f }; + static const LLQuad three = {3.f, 3.f, 3.f, 3.f }; + // Now we do one round of Newton-Raphson approximation to get full accuracy + // According to the Newton-Raphson method, given a first 'w' for the root of f(x) = 1/x^2 - a (i.e., x = 1/sqrt(a)) + // the next better approximation w[i+1] = w - f(w)/f'(w) = w - (1/w^2 - a)/(-2*w^(-3)) + // w[i+1] = w + 0.5 * (1/w^2 - a) * w^3 = w + 0.5 * (w - a*w^3) = 1.5 * w - 0.5 * a * w^3 + // = 0.5 * w * (3 - a*w^2) + // Our first approx is w = rsqrt. We need out = a * w[i+1] (this is the input vector 'a', not the 'a' from the above formula + // which is actually lenSqrd). So out = a * [0.5*rsqrt * (3 - lenSqrd*rsqrt*rsqrt)] + const LLQuad AtimesRsqrt = _mm_mul_ps( lenSqrd.mQ, rsqrt ); + const LLQuad AtimesRsqrtTimesRsqrt = _mm_mul_ps( AtimesRsqrt, rsqrt ); + const LLQuad threeMinusAtimesRsqrtTimesRsqrt = _mm_sub_ps(three, AtimesRsqrtTimesRsqrt ); + const LLQuad nrApprox = _mm_mul_ps(half, _mm_mul_ps(rsqrt, threeMinusAtimesRsqrtTimesRsqrt)); + mQ = _mm_mul_ps( mQ, nrApprox ); + return _mm_sqrt_ss(lenSqrd); +} + +// Normalize this vector with respect to the x, y, and z components only. Accurate only to 10-12 bits of precision. W component is destroyed +// Note that this does not consider zero length vectors! +inline void LLVector4a::normalize3fast() +{ + LLVector4a lenSqrd; lenSqrd.setAllDot3( *this, *this ); + const LLQuad approxRsqrt = _mm_rsqrt_ps(lenSqrd.mQ); + mQ = _mm_mul_ps( mQ, approxRsqrt ); +} + +// Return true if this vector is normalized with respect to x,y,z up to tolerance +inline LLBool32 LLVector4a::isNormalized3( F32 tolerance ) const +{ + static LL_ALIGN_16(const U32 ones[4]) = { 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 }; + LLSimdScalar tol = _mm_load_ss( &tolerance ); + tol = _mm_mul_ss( tol, tol ); + LLVector4a lenSquared; lenSquared.setAllDot3( *this, *this ); + lenSquared.sub( *reinterpret_cast<const LLVector4a*>(ones) ); + lenSquared.setAbs(lenSquared); + return _mm_comile_ss( lenSquared, tol ); +} + +// Return true if this vector is normalized with respect to all components up to tolerance +inline LLBool32 LLVector4a::isNormalized4( F32 tolerance ) const +{ + static LL_ALIGN_16(const U32 ones[4]) = { 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 }; + LLSimdScalar tol = _mm_load_ss( &tolerance ); + tol = _mm_mul_ss( tol, tol ); + LLVector4a lenSquared; lenSquared.setAllDot4( *this, *this ); + lenSquared.sub( *reinterpret_cast<const LLVector4a*>(ones) ); + lenSquared.setAbs(lenSquared); + return _mm_comile_ss( lenSquared, tol ); +} + +// Set all elements to the length of vector 'v' +inline void LLVector4a::setAllLength3( const LLVector4a& v ) +{ + LLVector4a lenSqrd; + lenSqrd.setAllDot3(v, v); + + mQ = _mm_sqrt_ps(lenSqrd.mQ); +} + +// Get this vector's length +inline LLSimdScalar LLVector4a::getLength3() const +{ + return _mm_sqrt_ss( dot3( (const LLVector4a)mQ ) ); +} + +// Set the components of this vector to the minimum of the corresponding components of lhs and rhs +inline void LLVector4a::setMin(const LLVector4a& lhs, const LLVector4a& rhs) +{ + mQ = _mm_min_ps(lhs.mQ, rhs.mQ); +} + +// Set the components of this vector to the maximum of the corresponding components of lhs and rhs +inline void LLVector4a::setMax(const LLVector4a& lhs, const LLVector4a& rhs) +{ + mQ = _mm_max_ps(lhs.mQ, rhs.mQ); +} + +// Set this to (c * lhs) + rhs * ( 1 - c) +inline void LLVector4a::setLerp(const LLVector4a& lhs, const LLVector4a& rhs, F32 c) +{ + LLVector4a a = lhs; + a.mul(c); + + LLVector4a b = rhs; + b.mul(1.f-c); + + setAdd(a, b); +} + +inline LLBool32 LLVector4a::isFinite3() const +{ + static LL_ALIGN_16(const U32 nanOrInfMask[4]) = { 0x7f800000, 0x7f800000, 0x7f800000, 0x7f800000 }; + const __m128i nanOrInfMaskV = *reinterpret_cast<const __m128i*> (nanOrInfMask); + const __m128i maskResult = _mm_and_si128( _mm_castps_si128(mQ), nanOrInfMaskV ); + const LLVector4Logical equalityCheck = _mm_castsi128_ps(_mm_cmpeq_epi32( maskResult, nanOrInfMaskV )); + return !equalityCheck.areAnySet( LLVector4Logical::MASK_XYZ ); +} + +inline LLBool32 LLVector4a::isFinite4() const +{ + static LL_ALIGN_16(const U32 nanOrInfMask[4]) = { 0x7f800000, 0x7f800000, 0x7f800000, 0x7f800000 }; + const __m128i nanOrInfMaskV = *reinterpret_cast<const __m128i*> (nanOrInfMask); + const __m128i maskResult = _mm_and_si128( _mm_castps_si128(mQ), nanOrInfMaskV ); + const LLVector4Logical equalityCheck = _mm_castsi128_ps(_mm_cmpeq_epi32( maskResult, nanOrInfMaskV )); + return !equalityCheck.areAnySet( LLVector4Logical::MASK_XYZW ); +} + +inline void LLVector4a::setRotatedInv( const LLRotation& rot, const LLVector4a& vec ) +{ + LLRotation inv; inv.setTranspose( rot ); + setRotated( inv, vec ); +} + +inline void LLVector4a::setRotatedInv( const LLQuaternion2& quat, const LLVector4a& vec ) +{ + LLQuaternion2 invRot; invRot.setConjugate( quat ); + setRotated(invRot, vec); +} + +inline void LLVector4a::clamp( const LLVector4a& low, const LLVector4a& high ) +{ + const LLVector4Logical highMask = greaterThan( high ); + const LLVector4Logical lowMask = lessThan( low ); + + setSelectWithMask( highMask, high, *this ); + setSelectWithMask( lowMask, low, *this ); +} + + +//////////////////////////////////// +// LOGICAL +//////////////////////////////////// +// The functions in this section will compare the elements in this vector +// to those in rhs and return an LLVector4Logical with all bits set in elements +// where the comparison was true and all bits unset in elements where the comparison +// was false. See llvector4logica.h +//////////////////////////////////// +// WARNING: Other than equals3 and equals4, these functions do NOT account +// for floating point tolerance. You should include the appropriate tolerance +// in the inputs. +//////////////////////////////////// + +inline LLVector4Logical LLVector4a::greaterThan(const LLVector4a& rhs) const +{ + return _mm_cmpgt_ps(mQ, rhs.mQ); +} + +inline LLVector4Logical LLVector4a::lessThan(const LLVector4a& rhs) const +{ + return _mm_cmplt_ps(mQ, rhs.mQ); +} + +inline LLVector4Logical LLVector4a::greaterEqual(const LLVector4a& rhs) const +{ + return _mm_cmpge_ps(mQ, rhs.mQ); +} + +inline LLVector4Logical LLVector4a::lessEqual(const LLVector4a& rhs) const +{ + return _mm_cmple_ps(mQ, rhs.mQ); +} + +inline LLVector4Logical LLVector4a::equal(const LLVector4a& rhs) const +{ + return _mm_cmpeq_ps(mQ, rhs.mQ); +} + +// Returns true if this and rhs are componentwise equal up to the specified absolute tolerance +inline bool LLVector4a::equals4(const LLVector4a& rhs, F32 tolerance ) const +{ + LLVector4a diff; diff.setSub( *this, rhs ); + diff.setAbs( diff ); + const LLQuad tol = _mm_set1_ps( tolerance ); + const LLQuad cmp = _mm_cmplt_ps( diff, tol ); + return (_mm_movemask_ps( cmp ) & LLVector4Logical::MASK_XYZW) == LLVector4Logical::MASK_XYZW; +} + +inline bool LLVector4a::equals3(const LLVector4a& rhs, F32 tolerance ) const +{ + LLVector4a diff; diff.setSub( *this, rhs ); + diff.setAbs( diff ); + const LLQuad tol = _mm_set1_ps( tolerance ); + const LLQuad t = _mm_cmplt_ps( diff, tol ); + return (_mm_movemask_ps( t ) & LLVector4Logical::MASK_XYZ) == LLVector4Logical::MASK_XYZ; + +} + +//////////////////////////////////// +// OPERATORS +//////////////////////////////////// + +// Do NOT add aditional operators without consulting someone with SSE experience +inline const LLVector4a& LLVector4a::operator= ( const LLVector4a& rhs ) +{ + mQ = rhs.mQ; + return *this; +} + +inline const LLVector4a& LLVector4a::operator= ( const LLQuad& rhs ) +{ + mQ = rhs; + return *this; +} + +inline LLVector4a::operator LLQuad() const +{ + return mQ; +} diff --git a/indra/llmath/llvector4logical.h b/indra/llmath/llvector4logical.h new file mode 100644 index 0000000000..1c7ee1d79f --- /dev/null +++ b/indra/llmath/llvector4logical.h @@ -0,0 +1,130 @@ +/** + * @file llvector4logical.h + * @brief LLVector4Logical class header file - Companion class to LLVector4a for logical and bit-twiddling operations + * + * $LicenseInfo:firstyear=2010&license=viewergpl$ + * + * Copyright (c) 2007-2010, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#ifndef LL_VECTOR4LOGICAL_H +#define LL_VECTOR4LOGICAL_H + + +//////////////////////////// +// LLVector4Logical +//////////////////////////// +// This class is incomplete. If you need additional functionality, +// for example setting/unsetting particular elements or performing +// other boolean operations, feel free to implement. If you need +// assistance in determining the most optimal implementation, +// contact someone with SSE experience (Falcon, Richard, Davep, e.g.) +//////////////////////////// + +static LL_ALIGN_16(const U32 S_V4LOGICAL_MASK_TABLE[4*4]) = +{ + 0xFFFFFFFF, 0x00000000, 0x00000000, 0x00000000, + 0x00000000, 0xFFFFFFFF, 0x00000000, 0x00000000, + 0x00000000, 0x00000000, 0xFFFFFFFF, 0x00000000, + 0x00000000, 0x00000000, 0x00000000, 0xFFFFFFFF +}; + +class LLVector4Logical +{ +public: + + enum { + MASK_X = 1, + MASK_Y = 1 << 1, + MASK_Z = 1 << 2, + MASK_W = 1 << 3, + MASK_XYZ = MASK_X | MASK_Y | MASK_Z, + MASK_XYZW = MASK_XYZ | MASK_W + }; + + // Empty default ctor + LLVector4Logical() {} + + LLVector4Logical( const LLQuad& quad ) + { + mQ = quad; + } + + // Create and return a mask consisting of the lowest order bit of each element + inline U32 getGatheredBits() const + { + return _mm_movemask_ps(mQ); + }; + + // Invert this mask + inline LLVector4Logical& invert() + { + static const LL_ALIGN_16(U32 allOnes[4]) = { 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF }; + mQ = _mm_andnot_ps( mQ, *(LLQuad*)(allOnes) ); + return *this; + } + + inline LLBool32 areAllSet( U32 mask ) const + { + return ( getGatheredBits() & mask) == mask; + } + + inline LLBool32 areAllSet() const + { + return areAllSet( MASK_XYZW ); + } + + inline LLBool32 areAnySet( U32 mask ) const + { + return getGatheredBits() & mask; + } + + inline LLBool32 areAnySet() const + { + return areAnySet( MASK_XYZW ); + } + + inline operator LLQuad() const + { + return mQ; + } + + inline void clear() + { + mQ = _mm_setzero_ps(); + } + + template<int N> void setElement() + { + mQ = _mm_or_ps( mQ, *reinterpret_cast<const LLQuad*>(S_V4LOGICAL_MASK_TABLE + 4*N) ); + } + +private: + + LLQuad mQ; +}; + +#endif //LL_VECTOR4ALOGICAL_H diff --git a/indra/llmath/llvolume.cpp b/indra/llmath/llvolume.cpp index c563af592f..ab9f8c4c24 100644 --- a/indra/llmath/llvolume.cpp +++ b/indra/llmath/llvolume.cpp @@ -1,4 +1,5 @@ /** + * @file llvolume.cpp * * $LicenseInfo:firstyear=2002&license=viewergpl$ @@ -30,6 +31,7 @@ */ #include "linden_common.h" +#include "llmemory.h" #include "llmath.h" #include <set> @@ -43,10 +45,15 @@ #include "v4math.h" #include "m4math.h" #include "m3math.h" +#include "llmatrix3a.h" +#include "lloctree.h" #include "lldarray.h" #include "llvolume.h" +#include "llvolumeoctree.h" #include "llstl.h" #include "llsdserialize.h" +#include "llvector4a.h" +#include "llmatrix4a.h" #define DEBUG_SILHOUETTE_BINORMALS 0 #define DEBUG_SILHOUETTE_NORMALS 0 // TomY: Use this to display normals using the silhouette @@ -104,127 +111,264 @@ BOOL check_same_clock_dir( const LLVector3& pt1, const LLVector3& pt2, const LLV BOOL LLLineSegmentBoxIntersect(const LLVector3& start, const LLVector3& end, const LLVector3& center, const LLVector3& size) { - float fAWdU[3]; - LLVector3 dir; - LLVector3 diff; + return LLLineSegmentBoxIntersect(start.mV, end.mV, center.mV, size.mV); +} + +BOOL LLLineSegmentBoxIntersect(const F32* start, const F32* end, const F32* center, const F32* size) +{ + F32 fAWdU[3]; + F32 dir[3]; + F32 diff[3]; for (U32 i = 0; i < 3; i++) { - dir.mV[i] = 0.5f * (end.mV[i] - start.mV[i]); - diff.mV[i] = (0.5f * (end.mV[i] + start.mV[i])) - center.mV[i]; - fAWdU[i] = fabsf(dir.mV[i]); - if(fabsf(diff.mV[i])>size.mV[i] + fAWdU[i]) return false; + dir[i] = 0.5f * (end[i] - start[i]); + diff[i] = (0.5f * (end[i] + start[i])) - center[i]; + fAWdU[i] = fabsf(dir[i]); + if(fabsf(diff[i])>size[i] + fAWdU[i]) return false; } float f; - f = dir.mV[1] * diff.mV[2] - dir.mV[2] * diff.mV[1]; if(fabsf(f)>size.mV[1]*fAWdU[2] + size.mV[2]*fAWdU[1]) return false; - f = dir.mV[2] * diff.mV[0] - dir.mV[0] * diff.mV[2]; if(fabsf(f)>size.mV[0]*fAWdU[2] + size.mV[2]*fAWdU[0]) return false; - f = dir.mV[0] * diff.mV[1] - dir.mV[1] * diff.mV[0]; if(fabsf(f)>size.mV[0]*fAWdU[1] + size.mV[1]*fAWdU[0]) return false; + f = dir[1] * diff[2] - dir[2] * diff[1]; if(fabsf(f)>size[1]*fAWdU[2] + size[2]*fAWdU[1]) return false; + f = dir[2] * diff[0] - dir[0] * diff[2]; if(fabsf(f)>size[0]*fAWdU[2] + size[2]*fAWdU[0]) return false; + f = dir[0] * diff[1] - dir[1] * diff[0]; if(fabsf(f)>size[0]*fAWdU[1] + size[1]*fAWdU[0]) return false; return true; } + // intersect test between triangle vert0, vert1, vert2 and a ray from orig in direction dir. // returns TRUE if intersecting and returns barycentric coordinates in intersection_a, intersection_b, // and returns the intersection point along dir in intersection_t. // Moller-Trumbore algorithm -BOOL LLTriangleRayIntersect(const LLVector3& vert0, const LLVector3& vert1, const LLVector3& vert2, const LLVector3& orig, const LLVector3& dir, - F32* intersection_a, F32* intersection_b, F32* intersection_t, BOOL two_sided) +BOOL LLTriangleRayIntersect(const LLVector4a& vert0, const LLVector4a& vert1, const LLVector4a& vert2, const LLVector4a& orig, const LLVector4a& dir, + F32& intersection_a, F32& intersection_b, F32& intersection_t) { - F32 u, v, t; /* find vectors for two edges sharing vert0 */ - LLVector3 edge1 = vert1 - vert0; + LLVector4a edge1; + edge1.setSub(vert1, vert0); - LLVector3 edge2 = vert2 - vert0;; + LLVector4a edge2; + edge2.setSub(vert2, vert0); /* begin calculating determinant - also used to calculate U parameter */ - LLVector3 pvec = dir % edge2; - - /* if determinant is near zero, ray lies in plane of triangle */ - F32 det = edge1 * pvec; + LLVector4a pvec; + pvec.setCross3(dir, edge2); - if (!two_sided) + /* if determinant is near zero, ray lies in plane of triangle */ + LLVector4a det; + det.setAllDot3(edge1, pvec); + + if (det.greaterEqual(LLVector4a::getEpsilon()).getGatheredBits() & 0x7) { - if (det < F_APPROXIMATELY_ZERO) - { - return FALSE; - } - /* calculate distance from vert0 to ray origin */ - LLVector3 tvec = orig - vert0; + LLVector4a tvec; + tvec.setSub(orig, vert0); /* calculate U parameter and test bounds */ - u = tvec * pvec; + LLVector4a u; + u.setAllDot3(tvec,pvec); - if (u < 0.f || u > det) + if ((u.greaterEqual(LLVector4a::getZero()).getGatheredBits() & 0x7) && + (u.lessEqual(det).getGatheredBits() & 0x7)) { - return FALSE; + /* prepare to test V parameter */ + LLVector4a qvec; + qvec.setCross3(tvec, edge1); + + /* calculate V parameter and test bounds */ + LLVector4a v; + v.setAllDot3(dir, qvec); + + + //if (!(v < 0.f || u + v > det)) + + LLVector4a sum_uv; + sum_uv.setAdd(u, v); + + S32 v_gequal = v.greaterEqual(LLVector4a::getZero()).getGatheredBits() & 0x7; + S32 sum_lequal = sum_uv.lessEqual(det).getGatheredBits() & 0x7; + + if (v_gequal && sum_lequal) + { + /* calculate t, scale parameters, ray intersects triangle */ + LLVector4a t; + t.setAllDot3(edge2,qvec); + + t.div(det); + u.div(det); + v.div(det); + + intersection_a = u[0]; + intersection_b = v[0]; + intersection_t = t[0]; + return TRUE; + } } - - /* prepare to test V parameter */ - LLVector3 qvec = tvec % edge1; + } - /* calculate V parameter and test bounds */ - v = dir * qvec; - if (v < 0.f || u + v > det) - { - return FALSE; - } + return FALSE; +} + +BOOL LLTriangleRayIntersectTwoSided(const LLVector4a& vert0, const LLVector4a& vert1, const LLVector4a& vert2, const LLVector4a& orig, const LLVector4a& dir, + F32& intersection_a, F32& intersection_b, F32& intersection_t) +{ + F32 u, v, t; + + /* find vectors for two edges sharing vert0 */ + LLVector4a edge1; + edge1.setSub(vert1, vert0); + + + LLVector4a edge2; + edge2.setSub(vert2, vert0); + + /* begin calculating determinant - also used to calculate U parameter */ + LLVector4a pvec; + pvec.setCross3(dir, edge2); + + /* if determinant is near zero, ray lies in plane of triangle */ + F32 det = edge1.dot3(pvec).getF32(); - /* calculate t, scale parameters, ray intersects triangle */ - t = edge2 * qvec; - F32 inv_det = 1.0 / det; - t *= inv_det; - u *= inv_det; - v *= inv_det; + + if (det > -F_APPROXIMATELY_ZERO && det < F_APPROXIMATELY_ZERO) + { + return FALSE; } + + F32 inv_det = 1.f / det; + + /* calculate distance from vert0 to ray origin */ + LLVector4a tvec; + tvec.setSub(orig, vert0); - else // two sided - { - if (det > -F_APPROXIMATELY_ZERO && det < F_APPROXIMATELY_ZERO) - { - return FALSE; - } - F32 inv_det = 1.0 / det; + /* calculate U parameter and test bounds */ + u = (tvec.dot3(pvec).getF32()) * inv_det; + if (u < 0.f || u > 1.f) + { + return FALSE; + } - /* calculate distance from vert0 to ray origin */ - LLVector3 tvec = orig - vert0; + /* prepare to test V parameter */ + tvec.sub(edge1); - /* calculate U parameter and test bounds */ - u = (tvec * pvec) * inv_det; - if (u < 0.f || u > 1.f) + /* calculate V parameter and test bounds */ + v = (dir.dot3(tvec).getF32()) * inv_det; + + if (v < 0.f || u + v > 1.f) + { + return FALSE; + } + + /* calculate t, ray intersects triangle */ + t = (edge2.dot3(tvec).getF32()) * inv_det; + + intersection_a = u; + intersection_b = v; + intersection_t = t; + + + return TRUE; +} + +//helper for non-aligned vectors +BOOL LLTriangleRayIntersect(const LLVector3& vert0, const LLVector3& vert1, const LLVector3& vert2, const LLVector3& orig, const LLVector3& dir, + F32& intersection_a, F32& intersection_b, F32& intersection_t, BOOL two_sided) +{ + LLVector4a vert0a, vert1a, vert2a, origa, dira; + vert0a.load3(vert0.mV); + vert1a.load3(vert1.mV); + vert2a.load3(vert2.mV); + origa.load3(orig.mV); + dira.load3(dir.mV); + + if (two_sided) + { + return LLTriangleRayIntersectTwoSided(vert0a, vert1a, vert2a, origa, dira, + intersection_a, intersection_b, intersection_t); + } + else + { + return LLTriangleRayIntersect(vert0a, vert1a, vert2a, origa, dira, + intersection_a, intersection_b, intersection_t); + } +} + +class LLVolumeOctreeRebound : public LLOctreeTravelerDepthFirst<LLVolumeTriangle> +{ +public: + const LLVolumeFace* mFace; + + LLVolumeOctreeRebound(const LLVolumeFace* face) + { + mFace = face; + } + + virtual void visit(const LLOctreeNode<LLVolumeTriangle>* branch) + { + LLVolumeOctreeListener* node = (LLVolumeOctreeListener*) branch->getListener(0); + + LLVector4a& min = node->mExtents[0]; + LLVector4a& max = node->mExtents[1]; + + if (branch->getElementCount() != 0) { - return FALSE; + const LLVolumeTriangle* tri = *(branch->getData().begin()); + + min = *(tri->mV[0]); + max = *(tri->mV[0]); + + for (LLOctreeNode<LLVolumeTriangle>::const_element_iter iter = + branch->getData().begin(); iter != branch->getData().end(); ++iter) + { + //stretch by triangles in node + tri = *iter; + + min.setMin(min, *tri->mV[0]); + min.setMin(min, *tri->mV[1]); + min.setMin(min, *tri->mV[2]); + + max.setMax(max, *tri->mV[0]); + max.setMax(max, *tri->mV[1]); + max.setMax(max, *tri->mV[2]); } - /* prepare to test V parameter */ - LLVector3 qvec = tvec - edge1; - - /* calculate V parameter and test bounds */ - v = (dir * qvec) * inv_det; - - if (v < 0.f || u + v > 1.f) + for (S32 i = 0; i < branch->getChildCount(); ++i) + { //stretch by child extents + LLVolumeOctreeListener* child = (LLVolumeOctreeListener*) branch->getChild(i)->getListener(0); + min.setMin(min, child->mExtents[0]); + max.setMax(min, child->mExtents[1]); + } + } + else if (branch->getChildCount() != 0) { - return FALSE; + LLVolumeOctreeListener* child = (LLVolumeOctreeListener*) branch->getChild(0)->getListener(0); + + min = child->mExtents[0]; + max = child->mExtents[1]; + + for (S32 i = 1; i < branch->getChildCount(); ++i) + { //stretch by child extents + child = (LLVolumeOctreeListener*) branch->getChild(i)->getListener(0); + min.setMin(min, child->mExtents[0]); + max.setMax(max, child->mExtents[1]); + } + } + else + { + llerrs << "WTF? Empty leaf" << llendl; } + + node->mBounds[0].setAdd(min, max); + node->mBounds[0].mul(0.5f); - /* calculate t, ray intersects triangle */ - t = (edge2 * qvec) * inv_det; + node->mBounds[1].setSub(max,min); + node->mBounds[1].mul(0.5f); } - - if (intersection_a != NULL) - *intersection_a = u; - if (intersection_b != NULL) - *intersection_b = v; - if (intersection_t != NULL) - *intersection_t = t; - - - return TRUE; -} +}; //------------------------------------------------------------------- @@ -1843,50 +1987,138 @@ BOOL LLVolume::generate() return FALSE; } -bool LLVolumeFace::VertexData::operator<(const LLVolumeFace::VertexData& rhs)const +void LLVolumeFace::VertexData::init() { - const U8* l = (const U8*) this; - const U8* r = (const U8*) &rhs; - - for (U32 i = 0; i < sizeof(VertexData); ++i) + if (!mData) { - if (l[i] != r[i]) - { - return r[i] < l[i]; - } + mData = (LLVector4a*) ll_aligned_malloc_16(32); } +} + +LLVolumeFace::VertexData::VertexData() +{ + mData = NULL; + init(); +} - return false; +LLVolumeFace::VertexData::VertexData(const VertexData& rhs) +{ + mData = NULL; + *this = rhs; } -bool LLVolumeFace::VertexData::operator==(const LLVolumeFace::VertexData& rhs)const +const LLVolumeFace::VertexData& LLVolumeFace::VertexData::operator=(const LLVolumeFace::VertexData& rhs) +{ + if (this != &rhs) + { + init(); + LLVector4a::memcpyNonAliased16((F32*) mData, (F32*) rhs.mData, 8*sizeof(F32)); + mTexCoord = rhs.mTexCoord; + } + return *this; +} + +LLVolumeFace::VertexData::~VertexData() { - const U8* l = (const U8*) this; - const U8* r = (const U8*) &rhs; + ll_aligned_free_16(mData); +} + +LLVector4a& LLVolumeFace::VertexData::getPosition() +{ + return mData[POSITION]; +} + +LLVector4a& LLVolumeFace::VertexData::getNormal() +{ + return mData[NORMAL]; +} + +const LLVector4a& LLVolumeFace::VertexData::getPosition() const +{ + return mData[POSITION]; +} - for (U32 i = 0; i < sizeof(VertexData); ++i) +const LLVector4a& LLVolumeFace::VertexData::getNormal() const +{ + return mData[NORMAL]; +} + + +void LLVolumeFace::VertexData::setPosition(const LLVector4a& pos) +{ + mData[POSITION] = pos; +} + +void LLVolumeFace::VertexData::setNormal(const LLVector4a& norm) +{ + mData[NORMAL] = norm; +} + +bool LLVolumeFace::VertexData::operator<(const LLVolumeFace::VertexData& rhs)const +{ + const F32* lp = this->getPosition().getF32ptr(); + const F32* rp = rhs.getPosition().getF32ptr(); + + if (lp[0] != rp[0]) { - if (l[i] != r[i]) - { - return false; - } + return lp[0] < rp[0]; } - - return true; + + if (rp[1] != lp[1]) + { + return lp[1] < rp[1]; + } + + if (rp[2] != lp[2]) + { + return lp[2] < rp[2]; + } + + lp = getNormal().getF32ptr(); + rp = rhs.getNormal().getF32ptr(); + + if (lp[0] != rp[0]) + { + return lp[0] < rp[0]; + } + + if (rp[1] != lp[1]) + { + return lp[1] < rp[1]; + } + + if (rp[2] != lp[2]) + { + return lp[2] < rp[2]; + } + + if (mTexCoord.mV[0] != rhs.mTexCoord.mV[0]) + { + return mTexCoord.mV[0] < rhs.mTexCoord.mV[0]; + } + + return mTexCoord.mV[1] < rhs.mTexCoord.mV[1]; +} + +bool LLVolumeFace::VertexData::operator==(const LLVolumeFace::VertexData& rhs)const +{ + return mData[POSITION].equals3(rhs.getPosition()) && + mData[NORMAL].equals3(rhs.getNormal()) && + mTexCoord == rhs.mTexCoord; } bool LLVolumeFace::VertexData::compareNormal(const LLVolumeFace::VertexData& rhs, F32 angle_cutoff) const { bool retval = false; - if (rhs.mPosition == mPosition && rhs.mTexCoord == mTexCoord) + if (rhs.mData[POSITION].equals3(mData[POSITION]) && rhs.mTexCoord == mTexCoord) { if (angle_cutoff > 1.f) { - retval = (mNormal == rhs.mNormal); + retval = (mData[NORMAL].equals3(rhs.mData[NORMAL])); } else { - F32 cur_angle = rhs.mNormal*mNormal; + F32 cur_angle = rhs.mData[NORMAL].dot3(mData[NORMAL]).getF32(); retval = cur_angle > angle_cutoff; } } @@ -1990,11 +2222,10 @@ bool LLVolume::unpackVolumeFaces(std::istream& is, S32 size) LLVolumeFace& face = mVolumeFaces[i]; - face.mHasBinormals = false; - //copy out indices - face.mIndices.resize(idx.size()/2); - if (idx.empty() || face.mIndices.size() < 3) + face.resizeIndices(idx.size()/2); + + if (idx.empty() || face.mNumIndices < 3) { //why is there an empty index list? llerrs <<"WTF?" << llendl; continue; @@ -2008,30 +2239,32 @@ bool LLVolume::unpackVolumeFaces(std::istream& is, S32 size) //copy out vertices U32 num_verts = pos.size()/(3*2); - face.mVertices.resize(num_verts); + face.resizeVertices(num_verts); if (mdl[i].has("Weights")) { - face.mWeights.resize(num_verts); + face.allocateWeights(num_verts); + LLSD::Binary weights = mdl[i]["Weights"]; - LLSD::Binary::iterator iter = weights.begin(); + U32 idx = 0; U32 cur_vertex = 0; - while (iter != weights.end()) + while (idx < weights.size() && cur_vertex < num_verts) { - const S32 END_INFLUENCES = 0xFF; - U8 joint = *(iter++); + const U8 END_INFLUENCES = 0xFF; + U8 joint = weights[idx++]; U32 cur_influence = 0; + LLVector4 wght(0,0,0,0); + while (joint != END_INFLUENCES) { - U16 influence = *(iter++); - influence = influence << 8; - influence |= *(iter++); + U16 influence = weights[idx++]; + influence |= ((U16) weights[idx++] << 8); F32 w = llmin((F32) influence / 65535.f, 0.99999f); - face.mWeights[cur_vertex].mV[cur_influence++] = (F32) joint + w; + wght.mV[cur_influence++] = (F32) joint + w; if (cur_influence >= 4) { @@ -2039,64 +2272,86 @@ bool LLVolume::unpackVolumeFaces(std::istream& is, S32 size) } else { - joint = *(iter++); + joint = weights[idx++]; } } + face.mWeights[cur_vertex].loadua(wght.mV); + cur_vertex++; - iter++; } + + if (cur_vertex != num_verts || idx != weights.size()) + { + llwarns << "Vertex weight count does not match vertex count!" << llendl; + } + } - LLVector3 min_pos; - LLVector3 max_pos; + LLVector3 minp; + LLVector3 maxp; LLVector2 min_tc; LLVector2 max_tc; - - min_pos.setValue(mdl[i]["PositionDomain"]["Min"]); - max_pos.setValue(mdl[i]["PositionDomain"]["Max"]); + minp.setValue(mdl[i]["PositionDomain"]["Min"]); + maxp.setValue(mdl[i]["PositionDomain"]["Max"]); + LLVector4a min_pos, max_pos; + min_pos.load3(minp.mV); + max_pos.load3(maxp.mV); + min_tc.setValue(mdl[i]["TexCoord0Domain"]["Min"]); max_tc.setValue(mdl[i]["TexCoord0Domain"]["Max"]); - LLVector3 pos_range = max_pos - min_pos; + LLVector4a pos_range; + pos_range.setSub(max_pos, min_pos); LLVector2 tc_range = max_tc - min_tc; - LLVector3& min = face.mExtents[0]; - LLVector3& max = face.mExtents[1]; + LLVector4a& min = face.mExtents[0]; + LLVector4a& max = face.mExtents[1]; - min = max = LLVector3(0,0,0); + min.clear(); + max.clear(); + + LLVector4a* pos_out = face.mPositions; + LLVector4a* norm_out = face.mNormals; + LLVector2* tc_out = face.mTexCoords; for (U32 j = 0; j < num_verts; ++j) { U16* v = (U16*) &(pos[j*3*2]); - face.mVertices[j].mPosition.setVec( - (F32) v[0] / 65535.f * pos_range.mV[0] + min_pos.mV[0], - (F32) v[1] / 65535.f * pos_range.mV[1] + min_pos.mV[1], - (F32) v[2] / 65535.f * pos_range.mV[2] + min_pos.mV[2]); + pos_out->set((F32) v[0], (F32) v[1], (F32) v[2]); + pos_out->div(65535.f); + pos_out->mul(pos_range); + pos_out->add(min_pos); if (j == 0) { - min = max = face.mVertices[j].mPosition; + min = *pos_out; + max = min; } else { - update_min_max(min,max,face.mVertices[j].mPosition); + min.setMin(min, *pos_out); + max.setMax(max, *pos_out); } + pos_out++; + U16* n = (U16*) &(norm[j*3*2]); - face.mVertices[j].mNormal.setVec( - (F32) n[0] / 65535.f * 2.f - 1.f, - (F32) n[1] / 65535.f * 2.f - 1.f, - (F32) n[2] / 65535.f * 2.f - 1.f); + norm_out->set((F32) n[0], (F32) n[1], (F32) n[2]); + norm_out->div(65535.f); + norm_out->mul(2.f); + norm_out->sub(1.f); + norm_out++; U16* t = (U16*) &(tc[j*2*2]); - face.mVertices[j].mTexCoord.setVec( - (F32) t[0] / 65535.f * tc_range.mV[0] + min_tc.mV[0], - (F32) t[1] / 65535.f * tc_range.mV[1] + min_tc.mV[1]); + tc_out->mV[0] = (F32) t[0] / 65535.f * tc_range.mV[0] + min_tc.mV[0]; + tc_out->mV[1] = (F32) t[1] / 65535.f * tc_range.mV[1] + min_tc.mV[1]; + + tc_out++; } @@ -2126,24 +2381,29 @@ bool LLVolume::unpackVolumeFaces(std::istream& is, S32 size) if (do_reflect_x) { - for (S32 i = 0; i < face.mVertices.size(); i++) + LLVector4a* p = (LLVector4a*) face.mPositions; + LLVector4a* n = (LLVector4a*) face.mNormals; + + for (S32 i = 0; i < face.mNumVertices; i++) { - face.mVertices[i].mPosition.mV[VX] *= -1.0f; - face.mVertices[i].mNormal.mV[VX] *= -1.0f; + p[i].mul(-1.0f); + n[i].mul(-1.0f); } } if (do_invert_normals) { - for (S32 i = 0; i < face.mVertices.size(); i++) + LLVector4a* n = (LLVector4a*) face.mNormals; + + for (S32 i = 0; i < face.mNumVertices; i++) { - face.mVertices[i].mNormal *= -1.0f; + n[i].mul(-1.0f); } } if (do_reverse_triangles) { - for (U32 j = 0; j < face.mIndices.size(); j += 3) + for (U32 j = 0; j < face.mNumIndices; j += 3) { // swap the 2nd and 3rd index S32 swap = face.mIndices[j+1]; @@ -2161,13 +2421,15 @@ bool LLVolume::unpackVolumeFaces(std::istream& is, S32 size) void tetrahedron_set_normal(LLVolumeFace::VertexData* cv) { - LLVector3 nrm = (cv[1].mPosition-cv[0].mPosition)%(cv[2].mPosition-cv[0].mPosition); - - nrm.normVec(); - - cv[0].mNormal = nrm; - cv[1].mNormal = nrm; - cv[2].mNormal = nrm; + LLVector4a v0; + v0.setSub(cv[1].getPosition(), cv[0].getNormal()); + LLVector4a v1; + v1.setSub(cv[2].getNormal(), cv[0].getPosition()); + + cv[0].getNormal().setCross3(v0,v1); + cv[0].getNormal().normalize3fast(); + cv[1].setNormal(cv[0].getNormal()); + cv[2].setNormal(cv[1].getNormal()); } BOOL LLVolume::isTetrahedron() @@ -2182,16 +2444,16 @@ void LLVolume::makeTetrahedron() LLVolumeFace face; F32 x = 0.25f; - LLVector3 p[] = + LLVector4a p[] = { //unit tetrahedron corners - LLVector3(x,x,x), - LLVector3(-x,-x,x), - LLVector3(-x,x,-x), - LLVector3(x,-x,-x) + LLVector4a(x,x,x), + LLVector4a(-x,-x,x), + LLVector4a(-x,x,-x), + LLVector4a(x,-x,-x) }; - face.mExtents[0].setVec(-x,-x,-x); - face.mExtents[1].setVec(x,x,x); + face.mExtents[0].splat(-x); + face.mExtents[1].splat(x); LLVolumeFace::VertexData cv[3]; @@ -2202,53 +2464,105 @@ void LLVolume::makeTetrahedron() //side 1 - cv[0].mPosition = p[1]; - cv[1].mPosition = p[0]; - cv[2].mPosition = p[2]; + cv[0].setPosition(p[1]); + cv[1].setPosition(p[0]); + cv[2].setPosition(p[2]); tetrahedron_set_normal(cv); - face.mVertices.push_back(cv[0]); - face.mVertices.push_back(cv[1]); - face.mVertices.push_back(cv[2]); + face.resizeVertices(12); + face.resizeIndices(12); + + LLVector4a* v = (LLVector4a*) face.mPositions; + LLVector4a* n = (LLVector4a*) face.mNormals; + LLVector2* tc = (LLVector2*) face.mTexCoords; + + v[0] = cv[0].getPosition(); + v[1] = cv[1].getPosition(); + v[2] = cv[2].getPosition(); + v += 3; + + n[0] = cv[0].getNormal(); + n[1] = cv[1].getNormal(); + n[2] = cv[2].getNormal(); + n += 3; + + tc[0] = cv[0].mTexCoord; + tc[1] = cv[1].mTexCoord; + tc[2] = cv[2].mTexCoord; + tc += 3; + //side 2 - cv[0].mPosition = p[3]; - cv[1].mPosition = p[0]; - cv[2].mPosition = p[1]; + cv[0].setPosition(p[3]); + cv[1].setPosition(p[0]); + cv[2].setPosition(p[1]); tetrahedron_set_normal(cv); - face.mVertices.push_back(cv[0]); - face.mVertices.push_back(cv[1]); - face.mVertices.push_back(cv[2]); + v[0] = cv[0].getPosition(); + v[1] = cv[1].getPosition(); + v[2] = cv[2].getPosition(); + v += 3; + + n[0] = cv[0].getNormal(); + n[1] = cv[1].getNormal(); + n[2] = cv[2].getNormal(); + n += 3; + + tc[0] = cv[0].mTexCoord; + tc[1] = cv[1].mTexCoord; + tc[2] = cv[2].mTexCoord; + tc += 3; //side 3 - cv[0].mPosition = p[3]; - cv[1].mPosition = p[1]; - cv[2].mPosition = p[2]; + cv[0].setPosition(p[3]); + cv[1].setPosition(p[1]); + cv[2].setPosition(p[2]); tetrahedron_set_normal(cv); - face.mVertices.push_back(cv[0]); - face.mVertices.push_back(cv[1]); - face.mVertices.push_back(cv[2]); + v[0] = cv[0].getPosition(); + v[1] = cv[1].getPosition(); + v[2] = cv[2].getPosition(); + v += 3; + + n[0] = cv[0].getNormal(); + n[1] = cv[1].getNormal(); + n[2] = cv[2].getNormal(); + n += 3; + + tc[0] = cv[0].mTexCoord; + tc[1] = cv[1].mTexCoord; + tc[2] = cv[2].mTexCoord; + tc += 3; //side 4 - cv[0].mPosition = p[2]; - cv[1].mPosition = p[0]; - cv[2].mPosition = p[3]; + cv[0].setPosition(p[2]); + cv[1].setPosition(p[0]); + cv[2].setPosition(p[3]); tetrahedron_set_normal(cv); - face.mVertices.push_back(cv[0]); - face.mVertices.push_back(cv[1]); - face.mVertices.push_back(cv[2]); + v[0] = cv[0].getPosition(); + v[1] = cv[1].getPosition(); + v[2] = cv[2].getPosition(); + v += 3; + + n[0] = cv[0].getNormal(); + n[1] = cv[1].getNormal(); + n[2] = cv[2].getNormal(); + n += 3; + + tc[0] = cv[0].mTexCoord; + tc[1] = cv[1].mTexCoord; + tc[2] = cv[2].mTexCoord; + tc += 3; //set index buffer - for (U32 i = 0; i < 12; i++) + for (U16 i = 0; i < 12; i++) { - face.mIndices.push_back(i); + face.mIndices[i] = i; } mVolumeFaces.push_back(face); @@ -2266,12 +2580,14 @@ void LLVolume::copyVolumeFaces(LLVolume* volume) S32 LLVolume::getNumFaces() const { +#if LL_MESH_ENABLED U8 sculpt_type = (mParams.getSculptType() & LL_SCULPT_TYPE_MASK); if (sculpt_type == LL_SCULPT_TYPE_MESH) { return LL_SCULPT_MESH_MAX_FACES; } +#endif return (S32)mProfilep->mFaces.size(); } @@ -2629,7 +2945,7 @@ void sculpt_calc_mesh_resolution(U16 width, U16 height, U8 type, F32 detail, S32 ratio = (F32) width / (F32) height; - s = (S32)fsqrtf(((F32)vertices / ratio)); + s = (S32)(F32) sqrt(((F32)vertices / ratio)); s = llmax(s, 4); // no degenerate sizes, please t = vertices / s; @@ -2644,11 +2960,6 @@ void LLVolume::sculpt(U16 sculpt_width, U16 sculpt_height, S8 sculpt_components, LLMemType m1(LLMemType::MTYPE_VOLUME); U8 sculpt_type = mParams.getSculptType(); - if (sculpt_type & LL_SCULPT_TYPE_MASK == LL_SCULPT_TYPE_MESH) - { - llerrs << "WTF?" << llendl; - } - BOOL data_is_empty = FALSE; if (sculpt_width == 0 || sculpt_height == 0 || sculpt_components < 3 || sculpt_data == NULL) @@ -3824,7 +4135,7 @@ S32 LLVolume::getNumTriangles() const for (S32 i = 0; i < getNumVolumeFaces(); ++i) { - triangle_count += getVolumeFace(i).mIndices.size()/3; + triangle_count += getVolumeFace(i).mNumIndices/3; } return triangle_count; @@ -3837,21 +4148,32 @@ S32 LLVolume::getNumTriangles() const void LLVolume::generateSilhouetteVertices(std::vector<LLVector3> &vertices, std::vector<LLVector3> &normals, std::vector<S32> &segments, - const LLVector3& obj_cam_vec, - const LLMatrix4& mat, - const LLMatrix3& norm_mat, + const LLVector3& obj_cam_vec_in, + const LLMatrix4& mat_in, + const LLMatrix3& norm_mat_in, S32 face_mask) { LLMemType m1(LLMemType::MTYPE_VOLUME); + LLMatrix4a mat; + mat.loadu(mat_in); + + LLMatrix4a norm_mat; + norm_mat.loadu(norm_mat_in); + + LLVector4a obj_cam_vec; + obj_cam_vec.load3(obj_cam_vec_in.mV); + vertices.clear(); normals.clear(); segments.clear(); +#if LL_MESH_ENABLED if ((mParams.getSculptType() & LL_SCULPT_TYPE_MASK) == LL_SCULPT_TYPE_MESH) { return; } +#endif S32 cur_index = 0; //for each face @@ -3861,7 +4183,7 @@ void LLVolume::generateSilhouetteVertices(std::vector<LLVector3> &vertices, LLVolumeFace& face = *iter; if (!(face_mask & (0x1 << cur_index++)) || - face.mIndices.empty() || face.mEdge.empty()) + face.mNumIndices == 0 || face.mEdge.empty()) { continue; } @@ -3878,7 +4200,7 @@ void LLVolume::generateSilhouetteVertices(std::vector<LLVector3> &vertices, #if DEBUG_SILHOUETTE_EDGE_MAP //for each triangle - U32 count = face.mIndices.size(); + U32 count = face.mNumIndices; for (U32 j = 0; j < count/3; j++) { //get vertices S32 v1 = face.mIndices[j*3+0]; @@ -3886,9 +4208,9 @@ void LLVolume::generateSilhouetteVertices(std::vector<LLVector3> &vertices, S32 v3 = face.mIndices[j*3+2]; //get current face center - LLVector3 cCenter = (face.mVertices[v1].mPosition + - face.mVertices[v2].mPosition + - face.mVertices[v3].mPosition) / 3.0f; + LLVector3 cCenter = (face.mVertices[v1].getPosition() + + face.mVertices[v2].getPosition() + + face.mVertices[v3].getPosition()) / 3.0f; //for each edge for (S32 k = 0; k < 3; k++) { @@ -3906,9 +4228,9 @@ void LLVolume::generateSilhouetteVertices(std::vector<LLVector3> &vertices, v3 = face.mIndices[nIndex*3+2]; //get neighbor face center - LLVector3 nCenter = (face.mVertices[v1].mPosition + - face.mVertices[v2].mPosition + - face.mVertices[v3].mPosition) / 3.0f; + LLVector3 nCenter = (face.mVertices[v1].getPosition() + + face.mVertices[v2].getPosition() + + face.mVertices[v3].getPosition()) / 3.0f; //draw line vertices.push_back(cCenter); @@ -3931,15 +4253,15 @@ void LLVolume::generateSilhouetteVertices(std::vector<LLVector3> &vertices, #elif DEBUG_SILHOUETTE_NORMALS //for each vertex - for (U32 j = 0; j < face.mVertices.size(); j++) { - vertices.push_back(face.mVertices[j].mPosition); - vertices.push_back(face.mVertices[j].mPosition + face.mVertices[j].mNormal*0.1f); + for (U32 j = 0; j < face.mNumVertices; j++) { + vertices.push_back(face.mVertices[j].getPosition()); + vertices.push_back(face.mVertices[j].getPosition() + face.mVertices[j].getNormal()*0.1f); normals.push_back(LLVector3(0,0,1)); normals.push_back(LLVector3(0,0,1)); segments.push_back(vertices.size()); #if DEBUG_SILHOUETTE_BINORMALS - vertices.push_back(face.mVertices[j].mPosition); - vertices.push_back(face.mVertices[j].mPosition + face.mVertices[j].mBinormal*0.1f); + vertices.push_back(face.mVertices[j].getPosition()); + vertices.push_back(face.mVertices[j].getPosition() + face.mVertices[j].mBinormal*0.1f); normals.push_back(LLVector3(0,0,1)); normals.push_back(LLVector3(0,0,1)); segments.push_back(vertices.size()); @@ -3957,26 +4279,36 @@ void LLVolume::generateSilhouetteVertices(std::vector<LLVector3> &vertices, //for each triangle std::vector<U8> fFacing; - vector_append(fFacing, face.mIndices.size()/3); - for (U32 j = 0; j < face.mIndices.size()/3; j++) + vector_append(fFacing, face.mNumIndices/3); + + LLVector4a* v = (LLVector4a*) face.mPositions; + LLVector4a* n = (LLVector4a*) face.mNormals; + + for (U32 j = 0; j < face.mNumIndices/3; j++) { //approximate normal S32 v1 = face.mIndices[j*3+0]; S32 v2 = face.mIndices[j*3+1]; S32 v3 = face.mIndices[j*3+2]; - LLVector3 norm = (face.mVertices[v1].mPosition - face.mVertices[v2].mPosition) % - (face.mVertices[v2].mPosition - face.mVertices[v3].mPosition); - - if (norm.magVecSquared() < 0.00000001f) + LLVector4a c1,c2; + c1.setSub(v[v1], v[v2]); + c2.setSub(v[v2], v[v3]); + + LLVector4a norm; + + norm.setCross3(c1, c2); + + if (norm.dot3(norm) < 0.00000001f) { fFacing[j] = AWAY | TOWARDS; } else { //get view vector - LLVector3 view = (obj_cam_vec-face.mVertices[v1].mPosition); - bool away = view * norm > 0.0f; + LLVector4a view; + view.setSub(obj_cam_vec, v[v1]); + bool away = view.dot3(norm) > 0.0f; if (away) { fFacing[j] = AWAY; @@ -3989,7 +4321,7 @@ void LLVolume::generateSilhouetteVertices(std::vector<LLVector3> &vertices, } //for each triangle - for (U32 j = 0; j < face.mIndices.size()/3; j++) + for (U32 j = 0; j < face.mNumIndices/3; j++) { if (fFacing[j] == (AWAY | TOWARDS)) { //this is a degenerate triangle @@ -4022,15 +4354,21 @@ void LLVolume::generateSilhouetteVertices(std::vector<LLVector3> &vertices, S32 v1 = face.mIndices[j*3+k]; S32 v2 = face.mIndices[j*3+((k+1)%3)]; - vertices.push_back(face.mVertices[v1].mPosition*mat); - LLVector3 norm1 = face.mVertices[v1].mNormal * norm_mat; - norm1.normVec(); - normals.push_back(norm1); + LLVector4a t; + mat.affineTransform(v[v1], t); + vertices.push_back(LLVector3(t[0], t[1], t[2])); - vertices.push_back(face.mVertices[v2].mPosition*mat); - LLVector3 norm2 = face.mVertices[v2].mNormal * norm_mat; - norm2.normVec(); - normals.push_back(norm2); + norm_mat.rotate(n[v1], t); + + t.normalize3fast(); + normals.push_back(LLVector3(t[0], t[1], t[2])); + + mat.affineTransform(v[v2], t); + vertices.push_back(LLVector3(t[0], t[1], t[2])); + + norm_mat.rotate(n[v2], t); + t.normalize3fast(); + normals.push_back(LLVector3(t[0], t[1], t[2])); segments.push_back(vertices.size()); } @@ -4045,6 +4383,19 @@ S32 LLVolume::lineSegmentIntersect(const LLVector3& start, const LLVector3& end, S32 face, LLVector3* intersection,LLVector2* tex_coord, LLVector3* normal, LLVector3* bi_normal) { + LLVector4a starta, enda; + starta.load3(start.mV); + enda.load3(end.mV); + + return lineSegmentIntersect(starta, enda, face, intersection, tex_coord, normal, bi_normal); + +} + + +S32 LLVolume::lineSegmentIntersect(const LLVector4a& start, const LLVector4a& end, + S32 face, + LLVector3* intersection,LLVector2* tex_coord, LLVector3* normal, LLVector3* bi_normal) +{ S32 hit_face = -1; S32 start_face; @@ -4061,7 +4412,8 @@ S32 LLVolume::lineSegmentIntersect(const LLVector3& start, const LLVector3& end, end_face = face; } - LLVector3 dir = end - start; + LLVector4a dir; + dir.setSub(end, start); F32 closest_t = 2.f; // must be larger than 1 @@ -4069,10 +4421,14 @@ S32 LLVolume::lineSegmentIntersect(const LLVector3& start, const LLVector3& end, for (S32 i = start_face; i <= end_face; i++) { - const LLVolumeFace &face = getVolumeFace((U32)i); + LLVolumeFace &face = mVolumeFaces[i]; - LLVector3 box_center = (face.mExtents[0] + face.mExtents[1]) / 2.f; - LLVector3 box_size = face.mExtents[1] - face.mExtents[0]; + LLVector4a box_center; + box_center.setAdd(face.mExtents[0], face.mExtents[1]); + box_center.mul(0.5f); + + LLVector4a box_size; + box_size.setSub(face.mExtents[1], face.mExtents[0]); if (LLLineSegmentBoxIntersect(start, end, box_center, box_size)) { @@ -4080,56 +4436,19 @@ S32 LLVolume::lineSegmentIntersect(const LLVector3& start, const LLVector3& end, { genBinormals(i); } - - for (U32 tri = 0; tri < face.mIndices.size()/3; tri++) - { - S32 index1 = face.mIndices[tri*3+0]; - S32 index2 = face.mIndices[tri*3+1]; - S32 index3 = face.mIndices[tri*3+2]; - F32 a, b, t; + if (!face.mOctree) + { + face.createOctree(); + } - if (LLTriangleRayIntersect(face.mVertices[index1].mPosition, - face.mVertices[index2].mPosition, - face.mVertices[index3].mPosition, - start, dir, &a, &b, &t, FALSE)) - { - if ((t >= 0.f) && // if hit is after start - (t <= 1.f) && // and before end - (t < closest_t)) // and this hit is closer - { - closest_t = t; - hit_face = i; + //LLVector4a* p = (LLVector4a*) face.mPositions; - if (intersection != NULL) - { - *intersection = start + dir * closest_t; - } - - if (tex_coord != NULL) + LLOctreeTriangleRayIntersect intersect(start, dir, &face, &closest_t, intersection, tex_coord, normal, bi_normal); + intersect.traverse(face.mOctree); + if (intersect.mHitFace) { - *tex_coord = ((1.f - a - b) * face.mVertices[index1].mTexCoord + - a * face.mVertices[index2].mTexCoord + - b * face.mVertices[index3].mTexCoord); - - } - - if (normal != NULL) - { - *normal = ((1.f - a - b) * face.mVertices[index1].mNormal + - a * face.mVertices[index2].mNormal + - b * face.mVertices[index3].mNormal); - } - - if (bi_normal != NULL) - { - *bi_normal = ((1.f - a - b) * face.mVertices[index1].mBinormal + - a * face.mVertices[index2].mBinormal + - b * face.mVertices[index3].mBinormal); - } - - } - } + hit_face = i; } } } @@ -4896,9 +5215,154 @@ std::ostream& operator<<(std::ostream &s, const LLVolume *volumep) return s; } +LLVolumeFace::LLVolumeFace() : + mID(0), + mTypeMask(0), + mBeginS(0), + mBeginT(0), + mNumS(0), + mNumT(0), + mNumVertices(0), + mNumIndices(0), + mPositions(NULL), + mNormals(NULL), + mBinormals(NULL), + mTexCoords(NULL), + mIndices(NULL), + mWeights(NULL), + mOctree(NULL) +{ + mExtents = (LLVector4a*) ll_aligned_malloc_16(48); + mCenter = mExtents+2; +} + +LLVolumeFace::LLVolumeFace(const LLVolumeFace& src) +: mID(0), + mTypeMask(0), + mBeginS(0), + mBeginT(0), + mNumS(0), + mNumT(0), + mNumVertices(0), + mNumIndices(0), + mPositions(NULL), + mNormals(NULL), + mBinormals(NULL), + mTexCoords(NULL), + mIndices(NULL), + mWeights(NULL), + mOctree(NULL) +{ + mExtents = (LLVector4a*) ll_aligned_malloc_16(48); + mCenter = mExtents+2; + *this = src; +} + +LLVolumeFace& LLVolumeFace::operator=(const LLVolumeFace& src) +{ + if (&src == this) + { //self assignment, do nothing + return *this; + } + + mID = src.mID; + mTypeMask = src.mTypeMask; + mBeginS = src.mBeginS; + mBeginT = src.mBeginT; + mNumS = src.mNumS; + mNumT = src.mNumT; + + mExtents[0] = src.mExtents[0]; + mExtents[1] = src.mExtents[1]; + *mCenter = *src.mCenter; + + mNumVertices = 0; + mNumIndices = 0; + + freeData(); + + LLVector4a::memcpyNonAliased16((F32*) mExtents, (F32*) src.mExtents, 12*sizeof(F32)); + + resizeVertices(src.mNumVertices); + resizeIndices(src.mNumIndices); + + if (mNumVertices) + { + S32 vert_size = mNumVertices*4*sizeof(F32); + S32 tc_size = (mNumVertices*8+0xF) & ~0xF; + + LLVector4a::memcpyNonAliased16((F32*) mPositions, (F32*) src.mPositions, vert_size); + LLVector4a::memcpyNonAliased16((F32*) mNormals, (F32*) src.mNormals, vert_size); + LLVector4a::memcpyNonAliased16((F32*) mTexCoords, (F32*) src.mTexCoords, tc_size); + + + if (src.mBinormals) + { + allocateBinormals(src.mNumVertices); + LLVector4a::memcpyNonAliased16((F32*) mBinormals, (F32*) src.mBinormals, vert_size); + } + else + { + ll_aligned_free_16(mBinormals); + mBinormals = NULL; + } + + if (src.mWeights) + { + allocateWeights(src.mNumVertices); + LLVector4a::memcpyNonAliased16((F32*) mWeights, (F32*) src.mWeights, vert_size); + } + else + { + ll_aligned_free_16(mWeights); + mWeights = NULL; + } + } + + if (mNumIndices) + { + S32 idx_size = (mNumIndices*2+0xF) & ~0xF; + + LLVector4a::memcpyNonAliased16((F32*) mIndices, (F32*) src.mIndices, idx_size); + } + + //delete + return *this; +} + +LLVolumeFace::~LLVolumeFace() +{ + ll_aligned_free_16(mExtents); + mExtents = NULL; + + freeData(); +} + +void LLVolumeFace::freeData() +{ + ll_aligned_free_16(mPositions); + mPositions = NULL; + ll_aligned_free_16(mNormals); + mNormals = NULL; + ll_aligned_free_16(mTexCoords); + mTexCoords = NULL; + ll_aligned_free_16(mIndices); + mIndices = NULL; + ll_aligned_free_16(mBinormals); + mBinormals = NULL; + ll_aligned_free_16(mWeights); + mWeights = NULL; + + delete mOctree; + mOctree = NULL; +} BOOL LLVolumeFace::create(LLVolume* volume, BOOL partial_build) { + //tree for this face is no longer valid + delete mOctree; + mOctree = NULL; + if (mTypeMask & CAP_MASK) { return createCap(volume, partial_build); @@ -4914,6 +5378,35 @@ BOOL LLVolumeFace::create(LLVolume* volume, BOOL partial_build) } } +void LLVolumeFace::getVertexData(U16 index, LLVolumeFace::VertexData& cv) +{ + cv.setPosition(mPositions[index]); + cv.setNormal(mNormals[index]); + cv.mTexCoord = mTexCoords[index]; +} + +bool LLVolumeFace::VertexMapData::operator==(const LLVolumeFace::VertexData& rhs) const +{ + return getPosition().equals3(rhs.getPosition()) && + mTexCoord == rhs.mTexCoord && + getNormal().equals3(rhs.getNormal()); +} + +bool LLVolumeFace::VertexMapData::ComparePosition::operator()(const LLVector3& a, const LLVector3& b) const +{ + if (a.mV[0] != b.mV[0]) + { + return a.mV[0] < b.mV[0]; + } + + if (a.mV[1] != b.mV[1]) + { + return a.mV[1] < b.mV[1]; + } + + return a.mV[2] < b.mV[2]; +} + void LLVolumeFace::optimize(F32 angle_cutoff) { LLVolumeFace new_face; @@ -4921,14 +5414,15 @@ void LLVolumeFace::optimize(F32 angle_cutoff) VertexMapData::PointMap point_map; //remove redundant vertices - for (U32 i = 0; i < mIndices.size(); ++i) + for (U32 i = 0; i < mNumIndices; ++i) { U16 index = mIndices[i]; - LLVolumeFace::VertexData cv = mVertices[index]; - + LLVolumeFace::VertexData cv; + getVertexData(index, cv); + BOOL found = FALSE; - VertexMapData::PointMap::iterator point_iter = point_map.find(cv.mPosition); + VertexMapData::PointMap::iterator point_iter = point_map.find(LLVector3(cv.getPosition().getF32ptr())); if (point_iter != point_map.end()) { //duplicate point might exist for (U32 j = 0; j < point_iter->second.size(); ++j) @@ -4937,7 +5431,7 @@ void LLVolumeFace::optimize(F32 angle_cutoff) if (tv.compareNormal(cv, angle_cutoff)) { found = TRUE; - new_face.mIndices.push_back((point_iter->second)[j].mIndex); + new_face.pushIndex((point_iter->second)[j].mIndex); break; } } @@ -4945,14 +5439,14 @@ void LLVolumeFace::optimize(F32 angle_cutoff) if (!found) { - new_face.mVertices.push_back(cv); - U16 index = (U16) new_face.mVertices.size()-1; - new_face.mIndices.push_back(index); + new_face.pushVertex(cv); + U16 index = (U16) new_face.mNumVertices-1; + new_face.pushIndex(index); VertexMapData d; - d.mPosition = cv.mPosition; + d.setPosition(cv.getPosition()); d.mTexCoord = cv.mTexCoord; - d.mNormal = cv.mNormal; + d.setNormal(cv.getNormal()); d.mIndex = index; if (point_iter != point_map.end()) { @@ -4960,13 +5454,77 @@ void LLVolumeFace::optimize(F32 angle_cutoff) } else { - point_map[d.mPosition].push_back(d); + point_map[LLVector3(d.getPosition().getF32ptr())].push_back(d); } } } - mVertices = new_face.mVertices; - mIndices = new_face.mIndices; + swapData(new_face); +} + + +void LLVolumeFace::createOctree() +{ + LLVector4a center; + LLVector4a size; + center.splat(0.f); + size.splat(1.f); + + mOctree = new LLOctreeRoot<LLVolumeTriangle>(center, size, NULL); + new LLVolumeOctreeListener(mOctree); + + for (U32 i = 0; i < mNumIndices; i+= 3) + { + LLPointer<LLVolumeTriangle> tri = new LLVolumeTriangle(); + + const LLVector4a& v0 = mPositions[mIndices[i]]; + const LLVector4a& v1 = mPositions[mIndices[i+1]]; + const LLVector4a& v2 = mPositions[mIndices[i+2]]; + + tri->mV[0] = &v0; + tri->mV[1] = &v1; + tri->mV[2] = &v2; + + tri->mIndex[0] = mIndices[i]; + tri->mIndex[1] = mIndices[i+1]; + tri->mIndex[2] = mIndices[i+2]; + + LLVector4a min = v0; + min.setMin(min, v1); + min.setMin(min, v2); + + LLVector4a max = v0; + max.setMax(max, v1); + max.setMax(max, v2); + + LLVector4a center; + center.setAdd(min, max); + center.mul(0.5f); + + *tri->mPositionGroup = center; + + LLVector4a size; + size.setSub(max,min); + + tri->mRadius = size.getLength3().getF32() * 0.5f; + + mOctree->insert(tri); + } + + LLVolumeOctreeRebound rebound(this); + rebound.traverse(mOctree); +} + + +void LLVolumeFace::swapData(LLVolumeFace& rhs) +{ + llswap(rhs.mPositions, mPositions); + llswap(rhs.mNormals, mNormals); + llswap(rhs.mBinormals, mBinormals); + llswap(rhs.mTexCoords, mTexCoords); + llswap(rhs.mIndices,mIndices); + llswap(rhs.mNumVertices, mNumVertices); + llswap(rhs.mNumIndices, mNumIndices); } void LerpPlanarVertex(LLVolumeFace::VertexData& v0, @@ -4976,10 +5534,21 @@ void LerpPlanarVertex(LLVolumeFace::VertexData& v0, F32 coef01, F32 coef02) { - vout.mPosition = v0.mPosition + ((v1.mPosition-v0.mPosition)*coef01)+((v2.mPosition-v0.mPosition)*coef02); + + LLVector4a lhs; + lhs.setSub(v1.getPosition(), v0.getPosition()); + lhs.mul(coef01); + LLVector4a rhs; + rhs.setSub(v2.getPosition(), v0.getPosition()); + rhs.mul(coef02); + + rhs.add(lhs); + rhs.add(v0.getPosition()); + + vout.setPosition(rhs); + vout.mTexCoord = v0.mTexCoord + ((v1.mTexCoord-v0.mTexCoord)*coef01)+((v2.mTexCoord-v0.mTexCoord)*coef02); - vout.mNormal = v0.mNormal; - vout.mBinormal = v0.mBinormal; + vout.setNormal(v0.getNormal()); } BOOL LLVolumeFace::createUnCutCubeCap(LLVolume* volume, BOOL partial_build) @@ -4999,82 +5568,113 @@ BOOL LLVolumeFace::createUnCutCubeCap(LLVolume* volume, BOOL partial_build) num_vertices = (grid_size+1)*(grid_size+1); num_indices = quad_count * 4; - LLVector3& min = mExtents[0]; - LLVector3& max = mExtents[1]; + LLVector4a& min = mExtents[0]; + LLVector4a& max = mExtents[1]; S32 offset = 0; if (mTypeMask & TOP_MASK) + { offset = (max_t-1) * max_s; + } else + { offset = mBeginS; + } - VertexData corners[4]; - VertexData baseVert; - for(int t = 0; t < 4; t++){ - corners[t].mPosition = mesh[offset + (grid_size*t)].mPos; - corners[t].mTexCoord.mV[0] = profile[grid_size*t].mV[0]+0.5f; - corners[t].mTexCoord.mV[1] = 0.5f - profile[grid_size*t].mV[1]; - } - baseVert.mNormal = - ((corners[1].mPosition-corners[0].mPosition) % - (corners[2].mPosition-corners[1].mPosition)); - baseVert.mNormal.normVec(); - if(!(mTypeMask & TOP_MASK)){ - baseVert.mNormal *= -1.0f; - }else{ - //Swap the UVs on the U(X) axis for top face - LLVector2 swap; - swap = corners[0].mTexCoord; - corners[0].mTexCoord=corners[3].mTexCoord; - corners[3].mTexCoord=swap; - swap = corners[1].mTexCoord; - corners[1].mTexCoord=corners[2].mTexCoord; - corners[2].mTexCoord=swap; - } - baseVert.mBinormal = calc_binormal_from_triangle( - corners[0].mPosition, corners[0].mTexCoord, - corners[1].mPosition, corners[1].mTexCoord, - corners[2].mPosition, corners[2].mTexCoord); - for(int t = 0; t < 4; t++){ - corners[t].mBinormal = baseVert.mBinormal; - corners[t].mNormal = baseVert.mNormal; - } - mHasBinormals = TRUE; - - if (partial_build) { - mVertices.clear(); - } + VertexData corners[4]; + VertexData baseVert; + for(S32 t = 0; t < 4; t++) + { + corners[t].getPosition().load3( mesh[offset + (grid_size*t)].mPos.mV); + corners[t].mTexCoord.mV[0] = profile[grid_size*t].mV[0]+0.5f; + corners[t].mTexCoord.mV[1] = 0.5f - profile[grid_size*t].mV[1]; + } - S32 vtop = mVertices.size(); - for(int gx = 0;gx<grid_size+1;gx++){ - for(int gy = 0;gy<grid_size+1;gy++){ - VertexData newVert; - LerpPlanarVertex( - corners[0], - corners[1], - corners[3], - newVert, - (F32)gx/(F32)grid_size, - (F32)gy/(F32)grid_size); - mVertices.push_back(newVert); + { + LLVector4a lhs; + lhs.setSub(corners[1].getPosition(), corners[0].getPosition()); + LLVector4a rhs; + rhs.setSub(corners[2].getPosition(), corners[1].getPosition()); + baseVert.getNormal().setCross3(lhs, rhs); + baseVert.getNormal().normalize3fast(); + } - if (gx == 0 && gy == 0) - { - min = max = newVert.mPosition; - } - else + if(!(mTypeMask & TOP_MASK)) + { + baseVert.getNormal().mul(-1.0f); + } + else + { + //Swap the UVs on the U(X) axis for top face + LLVector2 swap; + swap = corners[0].mTexCoord; + corners[0].mTexCoord=corners[3].mTexCoord; + corners[3].mTexCoord=swap; + swap = corners[1].mTexCoord; + corners[1].mTexCoord=corners[2].mTexCoord; + corners[2].mTexCoord=swap; + } + + LLVector4a binormal; + + calc_binormal_from_triangle( binormal, + corners[0].getPosition(), corners[0].mTexCoord, + corners[1].getPosition(), corners[1].mTexCoord, + corners[2].getPosition(), corners[2].mTexCoord); + + binormal.normalize3fast(); + + S32 size = (grid_size+1)*(grid_size+1); + resizeVertices(size); + allocateBinormals(size); + + LLVector4a* pos = (LLVector4a*) mPositions; + LLVector4a* norm = (LLVector4a*) mNormals; + LLVector4a* binorm = (LLVector4a*) mBinormals; + LLVector2* tc = (LLVector2*) mTexCoords; + + for(int gx = 0;gx<grid_size+1;gx++) + { + for(int gy = 0;gy<grid_size+1;gy++) { - update_min_max(min,max,newVert.mPosition); + VertexData newVert; + LerpPlanarVertex( + corners[0], + corners[1], + corners[3], + newVert, + (F32)gx/(F32)grid_size, + (F32)gy/(F32)grid_size); + + *pos++ = newVert.getPosition(); + *norm++ = baseVert.getNormal(); + *tc++ = newVert.mTexCoord; + *binorm++ = binormal; + + if (gx == 0 && gy == 0) + { + min = newVert.getPosition(); + max = min; + } + else + { + min.setMin(min, newVert.getPosition()); + max.setMax(max, newVert.getPosition()); + } } } - } - mCenter = (min + max) * 0.5f; + mCenter->setAdd(min, max); + mCenter->mul(0.5f); + } if (!partial_build) { - mTriStrip.clear(); + resizeIndices(grid_size*grid_size*6); + + U16* out = mIndices; + S32 idxs[] = {0,1,(grid_size+1)+1,(grid_size+1)+1,(grid_size+1),0}; for(S32 gx = 0;gx<grid_size;gx++) { @@ -5085,54 +5685,17 @@ BOOL LLVolumeFace::createUnCutCubeCap(LLVolume* volume, BOOL partial_build) { for(S32 i=5;i>=0;i--) { - mIndices.push_back(vtop+(gy*(grid_size+1))+gx+idxs[i]); - } - - if (gy == 0) - { - mTriStrip.push_back((gx+1)*(grid_size+1)); - mTriStrip.push_back((gx+1)*(grid_size+1)); - mTriStrip.push_back(gx*(grid_size+1)); - } - - mTriStrip.push_back(gy+1+(gx+1)*(grid_size+1)); - mTriStrip.push_back(gy+1+gx*(grid_size+1)); - - - if (gy == grid_size-1) - { - mTriStrip.push_back(gy+1+gx*(grid_size+1)); - } + *out++ = ((gy*(grid_size+1))+gx+idxs[i]); + } } else { for(S32 i=0;i<6;i++) { - mIndices.push_back(vtop+(gy*(grid_size+1))+gx+idxs[i]); - } - - if (gy == 0) - { - mTriStrip.push_back(gx*(grid_size+1)); - mTriStrip.push_back(gx*(grid_size+1)); - mTriStrip.push_back((gx+1)*(grid_size+1)); - } - - mTriStrip.push_back(gy+1+gx*(grid_size+1)); - mTriStrip.push_back(gy+1+(gx+1)*(grid_size+1)); - - if (gy == grid_size-1) - { - mTriStrip.push_back(gy+1+(gx+1)*(grid_size+1)); + *out++ = ((gy*(grid_size+1))+gx+idxs[i]); } } - } - - } - - if (mTriStrip.size()%2 == 1) - { - mTriStrip.push_back(mTriStrip[mTriStrip.size()-1]); + } } } @@ -5163,17 +5726,31 @@ BOOL LLVolumeFace::createCap(LLVolume* volume, BOOL partial_build) num_vertices = profile.size(); num_indices = (profile.size() - 2)*3; - mVertices.resize(num_vertices); + if (!(mTypeMask & HOLLOW_MASK) && !(mTypeMask & OPEN_MASK)) + { + resizeVertices(num_vertices+1); + allocateBinormals(num_vertices+1); - if (!partial_build) + if (!partial_build) + { + resizeIndices(num_indices+3); + } + } + else { - mIndices.resize(num_indices); + resizeVertices(num_vertices); + allocateBinormals(num_vertices); + + if (!partial_build) + { + resizeIndices(num_indices); + } } S32 max_s = volume->getProfile().getTotal(); S32 max_t = volume->getPath().mPath.size(); - mCenter.clearVec(); + mCenter->clear(); S32 offset = 0; if (mTypeMask & TOP_MASK) @@ -5191,82 +5768,91 @@ BOOL LLVolumeFace::createCap(LLVolume* volume, BOOL partial_build) LLVector2 cuv; LLVector2 min_uv, max_uv; - LLVector3& min = mExtents[0]; - LLVector3& max = mExtents[1]; + LLVector4a& min = mExtents[0]; + LLVector4a& max = mExtents[1]; + + LLVector2* tc = (LLVector2*) mTexCoords; + LLVector4a* pos = (LLVector4a*) mPositions; + LLVector4a* norm = (LLVector4a*) mNormals; + LLVector4a* binorm = (LLVector4a*) mBinormals; // Copy the vertices into the array for (S32 i = 0; i < num_vertices; i++) { if (mTypeMask & TOP_MASK) { - mVertices[i].mTexCoord.mV[0] = profile[i].mV[0]+0.5f; - mVertices[i].mTexCoord.mV[1] = profile[i].mV[1]+0.5f; + tc[i].mV[0] = profile[i].mV[0]+0.5f; + tc[i].mV[1] = profile[i].mV[1]+0.5f; } else { // Mirror for underside. - mVertices[i].mTexCoord.mV[0] = profile[i].mV[0]+0.5f; - mVertices[i].mTexCoord.mV[1] = 0.5f - profile[i].mV[1]; + tc[i].mV[0] = profile[i].mV[0]+0.5f; + tc[i].mV[1] = 0.5f - profile[i].mV[1]; } - mVertices[i].mPosition = mesh[i + offset].mPos; + pos[i].load3(mesh[i + offset].mPos.mV); if (i == 0) { - min = max = mVertices[i].mPosition; - min_uv = max_uv = mVertices[i].mTexCoord; + max = pos[i]; + min = max; + min_uv = max_uv = tc[i]; } else { - update_min_max(min,max, mVertices[i].mPosition); - update_min_max(min_uv, max_uv, mVertices[i].mTexCoord); + update_min_max(min,max,pos[i]); + update_min_max(min_uv, max_uv, tc[i]); } } - mCenter = (min+max)*0.5f; + mCenter->setAdd(min, max); + mCenter->mul(0.5f); + cuv = (min_uv + max_uv)*0.5f; - LLVector3 binormal = calc_binormal_from_triangle( - mCenter, cuv, - mVertices[0].mPosition, mVertices[0].mTexCoord, - mVertices[1].mPosition, mVertices[1].mTexCoord); - binormal.normVec(); + LLVector4a binormal; + calc_binormal_from_triangle(binormal, + *mCenter, cuv, + pos[0], tc[0], + pos[1], tc[1]); + binormal.normalize3fast(); - LLVector3 d0; - LLVector3 d1; - LLVector3 normal; + LLVector4a normal; + LLVector4a d0, d1; + - d0 = mCenter-mVertices[0].mPosition; - d1 = mCenter-mVertices[1].mPosition; + d0.setSub(*mCenter, pos[0]); + d1.setSub(*mCenter, pos[1]); - normal = (mTypeMask & TOP_MASK) ? (d0%d1) : (d1%d0); - normal.normVec(); + if (mTypeMask & TOP_MASK) + { + normal.setCross3(d0, d1); + } + else + { + normal.setCross3(d1, d0); + } + + normal.normalize3fast(); VertexData vd; - vd.mPosition = mCenter; - vd.mNormal = normal; - vd.mBinormal = binormal; + vd.setPosition(*mCenter); vd.mTexCoord = cuv; if (!(mTypeMask & HOLLOW_MASK) && !(mTypeMask & OPEN_MASK)) { - mVertices.push_back(vd); + pos[num_vertices] = *mCenter; + tc[num_vertices] = cuv; num_vertices++; - if (!partial_build) - { - vector_append(mIndices, 3); - } } - for (S32 i = 0; i < num_vertices; i++) { - mVertices[i].mBinormal = binormal; - mVertices[i].mNormal = normal; + binorm[i].load4a(binormal.getF32ptr()); + norm[i].load4a(normal.getF32ptr()); } - mHasBinormals = TRUE; - if (partial_build) { return TRUE; @@ -5374,8 +5960,6 @@ BOOL LLVolumeFace::createCap(LLVolume* volume, BOOL partial_build) pt2--; } } - - makeTriStrip(); } else { @@ -5480,8 +6064,6 @@ BOOL LLVolumeFace::createCap(LLVolume* volume, BOOL partial_build) pt2--; } } - - makeTriStrip(); } } else @@ -5503,163 +6085,320 @@ BOOL LLVolumeFace::createCap(LLVolume* volume, BOOL partial_build) mIndices[3*i+v2] = i + 1; } - //make tri strip - if (mTypeMask & OPEN_MASK) - { - makeTriStrip(); - } - else - { - S32 j = num_vertices-2; - if (mTypeMask & TOP_MASK) - { - mTriStrip.push_back(0); - for (S32 i = 0; i <= j; ++i) - { - mTriStrip.push_back(i); - if (i != j) - { - mTriStrip.push_back(j); - } - --j; - } - } - else - { - mTriStrip.push_back(j); - for (S32 i = 0; i <= j; ++i) - { - if (i != j) - { - mTriStrip.push_back(j); - } - mTriStrip.push_back(i); - --j; - } - } - - mTriStrip.push_back(mTriStrip[mTriStrip.size()-1]); - if (mTriStrip.size()%2 == 1) - { - mTriStrip.push_back(mTriStrip[mTriStrip.size()-1]); - } - } } return TRUE; } -void LLVolumeFace::makeTriStrip() -{ - for (U32 i = 0; i < mIndices.size(); i+=3) - { - U16 i0 = mIndices[i]; - U16 i1 = mIndices[i+1]; - U16 i2 = mIndices[i+2]; - - if ((i/3)%2 == 1) - { - mTriStrip.push_back(i0); - mTriStrip.push_back(i0); - mTriStrip.push_back(i1); - mTriStrip.push_back(i2); - mTriStrip.push_back(i2); - } - else - { - mTriStrip.push_back(i2); - mTriStrip.push_back(i2); - mTriStrip.push_back(i1); - mTriStrip.push_back(i0); - mTriStrip.push_back(i0); - } - } - - if (mTriStrip.size()%2 == 1) - { - mTriStrip.push_back(mTriStrip[mTriStrip.size()-1]); - } -} - void LLVolumeFace::createBinormals() { LLMemType m1(LLMemType::MTYPE_VOLUME); - if (!mHasBinormals) + if (!mBinormals) { + allocateBinormals(mNumVertices); + //generate binormals - for (U32 i = 0; i < mIndices.size()/3; i++) + LLVector4a* pos = mPositions; + LLVector2* tc = (LLVector2*) mTexCoords; + LLVector4a* binorm = (LLVector4a*) mBinormals; + + for (U32 i = 0; i < mNumIndices/3; i++) { //for each triangle - const VertexData& v0 = mVertices[mIndices[i*3+0]]; - const VertexData& v1 = mVertices[mIndices[i*3+1]]; - const VertexData& v2 = mVertices[mIndices[i*3+2]]; + const U16& i0 = mIndices[i*3+0]; + const U16& i1 = mIndices[i*3+1]; + const U16& i2 = mIndices[i*3+2]; //calculate binormal - LLVector3 binorm = calc_binormal_from_triangle(v0.mPosition, v0.mTexCoord, - v1.mPosition, v1.mTexCoord, - v2.mPosition, v2.mTexCoord); + LLVector4a binormal; + calc_binormal_from_triangle(binormal, + pos[i0], tc[i0], + pos[i1], tc[i1], + pos[i2], tc[i2]); - for (U32 j = 0; j < 3; j++) - { //add triangle normal to vertices - mVertices[mIndices[i*3+j]].mBinormal += binorm; // * (weight_sum - d[j])/weight_sum; - } + + //add triangle normal to vertices + binorm[i0].add(binormal); + binorm[i1].add(binormal); + binorm[i2].add(binormal); //even out quad contributions if (i % 2 == 0) { - mVertices[mIndices[i*3+2]].mBinormal += binorm; + binorm[i2].add(binormal); } else { - mVertices[mIndices[i*3+1]].mBinormal += binorm; + binorm[i1].add(binormal); } } //normalize binormals - for (U32 i = 0; i < mVertices.size(); i++) + for (U32 i = 0; i < mNumVertices; i++) + { + binorm[i].normalize3fast(); + //bump map/planar projection code requires normals to be normalized + mNormals[i].normalize3fast(); + } + } +} + +void LLVolumeFace::resizeVertices(S32 num_verts) +{ + ll_aligned_free_16(mPositions); + ll_aligned_free_16(mNormals); + ll_aligned_free_16(mBinormals); + ll_aligned_free_16(mTexCoords); + + mBinormals = NULL; + + if (num_verts) + { + mPositions = (LLVector4a*) ll_aligned_malloc_16(num_verts*16); + mNormals = (LLVector4a*) ll_aligned_malloc_16(num_verts*16); + + //pad texture coordinate block end to allow for QWORD reads + S32 size = ((num_verts*8) + 0xF) & ~0xF; + mTexCoords = (LLVector2*) ll_aligned_malloc_16(size); + } + else + { + mPositions = NULL; + mNormals = NULL; + mTexCoords = NULL; + } + + mNumVertices = num_verts; +} + +void LLVolumeFace::pushVertex(const LLVolumeFace::VertexData& cv) +{ + pushVertex(cv.getPosition(), cv.getNormal(), cv.mTexCoord); +} + +void LLVolumeFace::pushVertex(const LLVector4a& pos, const LLVector4a& norm, const LLVector2& tc) +{ + S32 new_verts = mNumVertices+1; + S32 new_size = new_verts*16; + S32 old_size = mNumVertices*16; + + //positions + LLVector4a* dst = (LLVector4a*) ll_aligned_malloc_16(new_size); + if (mPositions) + { + LLVector4a::memcpyNonAliased16((F32*) dst, (F32*) mPositions, old_size); + ll_aligned_free_16(mPositions); + } + mPositions = dst; + + //normals + dst = (LLVector4a*) ll_aligned_malloc_16(new_size); + if (mNormals) + { + LLVector4a::memcpyNonAliased16((F32*) dst, (F32*) mNormals, old_size); + ll_aligned_free_16(mNormals); + } + mNormals = dst; + + //tex coords + new_size = ((new_verts*8)+0xF) & ~0xF; + old_size = ((mNumVertices*8)+0xF) & ~0xF; + + dst = (LLVector4a*) ll_aligned_malloc_16(new_size); + { + LLVector2* dst = (LLVector2*) ll_aligned_malloc_16(new_size); + if (mTexCoords) + { + LLVector4a::memcpyNonAliased16((F32*) dst, (F32*) mTexCoords, old_size); + ll_aligned_free_16(mTexCoords); + } + } + mTexCoords = (LLVector2*) dst; + + //just clear binormals + ll_aligned_free_16(mBinormals); + mBinormals = NULL; + + mPositions[mNumVertices] = pos; + mNormals[mNumVertices] = norm; + mTexCoords[mNumVertices] = tc; + + mNumVertices++; +} + +void LLVolumeFace::allocateBinormals(S32 num_verts) +{ + ll_aligned_free_16(mBinormals); + mBinormals = (LLVector4a*) ll_aligned_malloc_16(num_verts*16); +} + +void LLVolumeFace::allocateWeights(S32 num_verts) +{ + ll_aligned_free_16(mWeights); + mWeights = (LLVector4a*) ll_aligned_malloc_16(num_verts*16); +} + +void LLVolumeFace::resizeIndices(S32 num_indices) +{ + ll_aligned_free_16(mIndices); + + if (num_indices) + { + //pad index block end to allow for QWORD reads + S32 size = ((num_indices*2) + 0xF) & ~0xF; + + mIndices = (U16*) ll_aligned_malloc_16(size); + } + else + { + mIndices = NULL; + } + + mNumIndices = num_indices; +} + +void LLVolumeFace::pushIndex(const U16& idx) +{ + S32 new_count = mNumIndices + 1; + S32 new_size = ((new_count*2)+0xF) & ~0xF; + + S32 old_size = ((mNumIndices*2)+0xF) & ~0xF; + if (new_size != old_size) + { + U16* dst = (U16*) ll_aligned_malloc_16(new_size); + if (mIndices) { - mVertices[i].mBinormal.normVec(); - mVertices[i].mNormal.normVec(); + LLVector4a::memcpyNonAliased16((F32*) dst, (F32*) mIndices, old_size); + ll_aligned_free_16(mIndices); } + mIndices = dst; + } + + mIndices[mNumIndices++] = idx; +} + +void LLVolumeFace::fillFromLegacyData(std::vector<LLVolumeFace::VertexData>& v, std::vector<U16>& idx) +{ + resizeVertices(v.size()); + resizeIndices(idx.size()); - mHasBinormals = TRUE; + for (U32 i = 0; i < v.size(); ++i) + { + mPositions[i] = v[i].getPosition(); + mNormals[i] = v[i].getNormal(); + mTexCoords[i] = v[i].mTexCoord; + } + + for (U32 i = 0; i < idx.size(); ++i) + { + mIndices[i] = idx[i]; } } -void LLVolumeFace::appendFace(const LLVolumeFace& face, LLMatrix4& mat, LLMatrix4& norm_mat) +void LLVolumeFace::appendFace(const LLVolumeFace& face, LLMatrix4& mat_in, LLMatrix4& norm_mat_in) { - U16 offset = mVertices.size(); + U16 offset = mNumVertices; + + S32 new_count = face.mNumVertices + mNumVertices; - if (face.mVertices.size() + mVertices.size() > 65536) + if (new_count > 65536) { llerrs << "Cannot append face -- 16-bit overflow will occur." << llendl; } - for (U32 i = 0; i < face.mVertices.size(); ++i) + if (face.mNumVertices == 0) { - VertexData v = face.mVertices[i]; - v.mPosition = v.mPosition*mat; - v.mNormal = v.mNormal * norm_mat; + llerrs << "Cannot append empty face." << llendl; + } + + //allocate new buffer space + LLVector4a* new_pos = (LLVector4a*) ll_aligned_malloc_16(new_count*16); + LLVector4a* new_norm = (LLVector4a*) ll_aligned_malloc_16(new_count*16); + LLVector2* new_tc = (LLVector2*) ll_aligned_malloc_16((new_count*8+0xF) & ~0xF); + + + if (mNumVertices > 0) + { //copy old buffers + LLVector4a::memcpyNonAliased16((F32*) new_pos, (F32*) mPositions, mNumVertices*4*sizeof(F32)); + LLVector4a::memcpyNonAliased16((F32*) new_norm, (F32*) mNormals, mNumVertices*4*sizeof(F32)); + LLVector4a::memcpyNonAliased16((F32*) new_tc, (F32*) mTexCoords, mNumVertices*2*sizeof(F32)); + } + + //free old buffer space + ll_aligned_free_16(mPositions); + ll_aligned_free_16(mNormals); + ll_aligned_free_16(mTexCoords); + + //point to new buffers + mPositions = new_pos; + mNormals = new_norm; + mTexCoords = new_tc; + + mNumVertices = new_count; + + //get destination address of appended face + LLVector4a* dst_pos = mPositions+offset; + LLVector2* dst_tc = mTexCoords+offset; + LLVector4a* dst_norm = mNormals+offset; - v.mNormal.normalize(); + //get source addresses of appended face + const LLVector4a* src_pos = face.mPositions; + const LLVector2* src_tc = face.mTexCoords; + const LLVector4a* src_norm = face.mNormals; - mVertices.push_back(v); + //load aligned matrices + LLMatrix4a mat, norm_mat; + mat.loadu(mat_in); + norm_mat.loadu(norm_mat_in); + + for (U32 i = 0; i < face.mNumVertices; ++i) + { + //transform appended face position and store + mat.affineTransform(src_pos[i], dst_pos[i]); + + //transform appended face normal and store + norm_mat.rotate(src_norm[i], dst_norm[i]); + dst_norm[i].normalize3fast(); + + //copy appended face texture coordinate + dst_tc[i] = src_tc[i]; if (offset == 0 && i == 0) - { - mExtents[0] = mExtents[1] = v.mPosition; + { //initialize bounding box + mExtents[0] = mExtents[1] = dst_pos[i]; } else { - update_min_max(mExtents[0], mExtents[1], v.mPosition); + //stretch bounding box + update_min_max(mExtents[0], mExtents[1], dst_pos[i]); } } - - for (U32 i = 0; i < face.mIndices.size(); ++i) - { - mIndices.push_back(face.mIndices[i]+offset); + + new_count = mNumIndices + face.mNumIndices; + + //allocate new index buffer + U16* new_indices = (U16*) ll_aligned_malloc_16((new_count*2+0xF) & ~0xF); + if (mNumIndices > 0) + { //copy old index buffer + S32 old_size = (mNumIndices*2+0xF) & ~0xF; + LLVector4a::memcpyNonAliased16((F32*) new_indices, (F32*) mIndices, old_size); + } + + //free old index buffer + ll_aligned_free_16(mIndices); + + //point to new index buffer + mIndices = new_indices; + + //get destination address into new index buffer + U16* dst_idx = mIndices+mNumIndices; + mNumIndices = new_count; + + for (U32 i = 0; i < face.mNumIndices; ++i) + { //copy indices, offsetting by old vertex count + dst_idx[i] = face.mIndices[i]+offset; } } @@ -5689,28 +6428,24 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build) num_vertices = mNumS*mNumT; num_indices = (mNumS-1)*(mNumT-1)*6; - mVertices.resize(num_vertices); - if (!partial_build) { - mIndices.resize(num_indices); + resizeVertices(num_vertices); + resizeIndices(num_indices); +#if LL_MESH_ENABLED if ((volume->getParams().getSculptType() & LL_SCULPT_TYPE_MASK) != LL_SCULPT_TYPE_MESH) { mEdge.resize(num_indices); } - } - else - { - mHasBinormals = FALSE; +#else + mEdge.resize(num_indices); +#endif } - - LLVector3& face_min = mExtents[0]; - LLVector3& face_max = mExtents[1]; - - mCenter.clearVec(); - + LLVector4a* pos = (LLVector4a*) mPositions; + LLVector4a* norm = (LLVector4a*) mNormals; + LLVector2* tc = (LLVector2*) mTexCoords; S32 begin_stex = llfloor( profile[mBeginS].mV[2] ); S32 num_s = ((mTypeMask & INNER_MASK) && (mTypeMask & FLAT_MASK) && mNumS > 2) ? mNumS/2 : mNumS; @@ -5761,30 +6496,20 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build) i = mBeginS + s + max_s*t; } - mVertices[cur_vertex].mPosition = mesh[i].mPos; - mVertices[cur_vertex].mTexCoord = LLVector2(ss,tt); + pos[cur_vertex].load3(mesh[i].mPos.mV); + tc[cur_vertex] = LLVector2(ss,tt); - mVertices[cur_vertex].mNormal = LLVector3(0,0,0); - mVertices[cur_vertex].mBinormal = LLVector3(0,0,0); - - if (cur_vertex == 0) - { - face_min = face_max = mesh[i].mPos; - } - else - { - update_min_max(face_min, face_max, mesh[i].mPos); - } - + norm[cur_vertex].clear(); cur_vertex++; if ((mTypeMask & INNER_MASK) && (mTypeMask & FLAT_MASK) && mNumS > 2 && s > 0) { - mVertices[cur_vertex].mPosition = mesh[i].mPos; - mVertices[cur_vertex].mTexCoord = LLVector2(ss,tt); + + pos[cur_vertex].load3(mesh[i].mPos.mV); + tc[cur_vertex] = LLVector2(ss,tt); - mVertices[cur_vertex].mNormal = LLVector3(0,0,0); - mVertices[cur_vertex].mBinormal = LLVector3(0,0,0); + norm[cur_vertex].clear(); + cur_vertex++; } } @@ -5802,19 +6527,29 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build) i = mBeginS + s + max_s*t; ss = profile[mBeginS + s].mV[2] - begin_stex; - mVertices[cur_vertex].mPosition = mesh[i].mPos; - mVertices[cur_vertex].mTexCoord = LLVector2(ss,tt); - - mVertices[cur_vertex].mNormal = LLVector3(0,0,0); - mVertices[cur_vertex].mBinormal = LLVector3(0,0,0); - - update_min_max(face_min,face_max,mesh[i].mPos); - + pos[cur_vertex].load3(mesh[i].mPos.mV); + tc[cur_vertex] = LLVector2(ss,tt); + norm[cur_vertex].clear(); + cur_vertex++; } } - mCenter = (face_min + face_max) * 0.5f; + + //get bounding box for this side + LLVector4a& face_min = mExtents[0]; + LLVector4a& face_max = mExtents[1]; + mCenter->clear(); + + face_min = face_max = pos[0]; + + for (U32 i = 1; i < mNumVertices; ++i) + { + update_min_max(face_min, face_max, pos[i]); + } + + mCenter->setAdd(face_min, face_max); + mCenter->mul(0.5f); S32 cur_index = 0; S32 cur_edge = 0; @@ -5822,14 +6557,9 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build) if (!partial_build) { - mTriStrip.clear(); - // Now we generate the indices. for (t = 0; t < (mNumT-1); t++) { - //prepend terminating index to strip - mTriStrip.push_back(mNumS*t); - for (s = 0; s < (mNumS-1); s++) { mIndices[cur_index++] = s + mNumS*t; //bottom left @@ -5839,14 +6569,6 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build) mIndices[cur_index++] = s+1 + mNumS*t; //bottom right mIndices[cur_index++] = s+1 + mNumS*(t+1); //top right - if (s == 0) - { - mTriStrip.push_back(s+mNumS*t); - mTriStrip.push_back(s+mNumS*(t+1)); - } - mTriStrip.push_back(s+1+mNumS*t); - mTriStrip.push_back(s+1+mNumS*(t+1)); - mEdge[cur_edge++] = (mNumS-1)*2*t+s*2+1; //bottom left/top right neighbor face if (t < mNumT-2) { //top right/top left neighbor face mEdge[cur_edge++] = (mNumS-1)*2*(t+1)+s*2+1; @@ -5887,59 +6609,55 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build) } mEdge[cur_edge++] = (mNumS-1)*2*t+s*2; //top right/bottom left neighbor face } - //append terminating vertex to strip - mTriStrip.push_back(mNumS-1+mNumS*(t+1)); - } - - if (mTriStrip.size()%2 == 1) - { - mTriStrip.push_back(mTriStrip[mTriStrip.size()-1]); } } //generate normals - for (U32 i = 0; i < mIndices.size()/3; i++) //for each triangle - { - const S32 i0 = mIndices[i*3+0]; - const S32 i1 = mIndices[i*3+1]; - const S32 i2 = mIndices[i*3+2]; - const VertexData& v0 = mVertices[i0]; - const VertexData& v1 = mVertices[i1]; - const VertexData& v2 = mVertices[i2]; - - //calculate triangle normal - LLVector3 norm = (v0.mPosition-v1.mPosition) % (v0.mPosition-v2.mPosition); + for (U32 i = 0; i < mNumIndices/3; i++) //for each triangle + { + const U16* idx = &(mIndices[i*3]); + - for (U32 j = 0; j < 3; j++) - { //add triangle normal to vertices - const S32 idx = mIndices[i*3+j]; - mVertices[idx].mNormal += norm; // * (weight_sum - d[j])/weight_sum; - } + LLVector4a* v[] = + { pos+idx[0], pos+idx[1], pos+idx[2] }; + + LLVector4a* n[] = + { norm+idx[0], norm+idx[1], norm+idx[2] }; + + //calculate triangle normal + LLVector4a a, b, c; + + a.setSub(*v[0], *v[1]); + b.setSub(*v[0], *v[2]); + c.setCross3(a,b); + n[0]->add(c); + n[1]->add(c); + n[2]->add(c); + //even out quad contributions - if ((i & 1) == 0) - { - mVertices[i2].mNormal += norm; - } - else - { - mVertices[i1].mNormal += norm; - } + n[i%2+1]->add(c); } // adjust normals based on wrapping and stitching - BOOL s_bottom_converges = ((mVertices[0].mPosition - mVertices[mNumS*(mNumT-2)].mPosition).magVecSquared() < 0.000001f); - BOOL s_top_converges = ((mVertices[mNumS-1].mPosition - mVertices[mNumS*(mNumT-2)+mNumS-1].mPosition).magVecSquared() < 0.000001f); + LLVector4a top; + top.setSub(pos[0], pos[mNumS*(mNumT-2)]); + BOOL s_bottom_converges = (top.dot3(top) < 0.000001f); + + top.setSub(pos[mNumS-1], pos[mNumS*(mNumT-2)+mNumS-1]); + BOOL s_top_converges = (top.dot3(top) < 0.000001f); + if (sculpt_stitching == LL_SCULPT_TYPE_NONE) // logic for non-sculpt volumes { if (volume->getPath().isOpen() == FALSE) { //wrap normals on T for (S32 i = 0; i < mNumS; i++) { - LLVector3 norm = mVertices[i].mNormal + mVertices[mNumS*(mNumT-1)+i].mNormal; - mVertices[i].mNormal = norm; - mVertices[mNumS*(mNumT-1)+i].mNormal = norm; + LLVector4a n; + n.setAdd(norm[i], norm[mNumS*(mNumT-1)+i]); + norm[i] = n; + norm[mNumS*(mNumT-1)+i] = n; } } @@ -5947,9 +6665,10 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build) { //wrap normals on S for (S32 i = 0; i < mNumT; i++) { - LLVector3 norm = mVertices[mNumS*i].mNormal + mVertices[mNumS*i+mNumS-1].mNormal; - mVertices[mNumS * i].mNormal = norm; - mVertices[mNumS * i+mNumS-1].mNormal = norm; + LLVector4a n; + n.setAdd(norm[mNumS*i], norm[mNumS*i+mNumS-1]); + norm[mNumS * i] = n; + norm[mNumS * i+mNumS-1] = n; } } @@ -5960,7 +6679,7 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build) { //all lower S have same normal for (S32 i = 0; i < mNumT; i++) { - mVertices[mNumS*i].mNormal = LLVector3(1,0,0); + norm[mNumS*i].set(1,0,0); } } @@ -5968,12 +6687,11 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build) { //all upper S have same normal for (S32 i = 0; i < mNumT; i++) { - mVertices[mNumS*i+mNumS-1].mNormal = LLVector3(-1,0,0); + norm[mNumS*i+mNumS-1].set(-1,0,0); } } } } - else // logic for sculpt volumes { BOOL average_poles = FALSE; @@ -5996,30 +6714,33 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build) { // average normals for north pole - LLVector3 average(0.0, 0.0, 0.0); + LLVector4a average; + average.clear(); + for (S32 i = 0; i < mNumS; i++) { - average += mVertices[i].mNormal; + average.add(norm[i]); } // set average for (S32 i = 0; i < mNumS; i++) { - mVertices[i].mNormal = average; + norm[i] = average; } // average normals for south pole - average = LLVector3(0.0, 0.0, 0.0); + average.clear(); + for (S32 i = 0; i < mNumS; i++) { - average += mVertices[i + mNumS * (mNumT - 1)].mNormal; + average.add(norm[i + mNumS * (mNumT - 1)]); } // set average for (S32 i = 0; i < mNumS; i++) { - mVertices[i + mNumS * (mNumT - 1)].mNormal = average; + norm[i + mNumS * (mNumT - 1)] = average; } } @@ -6029,23 +6750,22 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build) { for (S32 i = 0; i < mNumT; i++) { - LLVector3 norm = mVertices[mNumS*i].mNormal + mVertices[mNumS*i+mNumS-1].mNormal; - mVertices[mNumS * i].mNormal = norm; - mVertices[mNumS * i+mNumS-1].mNormal = norm; + LLVector4a n; + n.setAdd(norm[mNumS*i], norm[mNumS*i+mNumS-1]); + norm[mNumS * i] = n; + norm[mNumS * i+mNumS-1] = n; } } - - if (wrap_t) { for (S32 i = 0; i < mNumS; i++) { - LLVector3 norm = mVertices[i].mNormal + mVertices[mNumS*(mNumT-1)+i].mNormal; - mVertices[i].mNormal = norm; - mVertices[mNumS*(mNumT-1)+i].mNormal = norm; + LLVector4a n; + n.setAdd(norm[i], norm[mNumS*(mNumT-1)+i]); + norm[i] = n; + norm[mNumS*(mNumT-1)+i] = n; } - } } @@ -6055,41 +6775,51 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build) // Finds binormal based on three vertices with texture coordinates. // Fills in dummy values if the triangle has degenerate texture coordinates. -LLVector3 calc_binormal_from_triangle( - const LLVector3& pos0, +void calc_binormal_from_triangle(LLVector4a& binormal, + + const LLVector4a& pos0, const LLVector2& tex0, - const LLVector3& pos1, + const LLVector4a& pos1, const LLVector2& tex1, - const LLVector3& pos2, + const LLVector4a& pos2, const LLVector2& tex2) { - LLVector3 rx0( pos0.mV[VX], tex0.mV[VX], tex0.mV[VY] ); - LLVector3 rx1( pos1.mV[VX], tex1.mV[VX], tex1.mV[VY] ); - LLVector3 rx2( pos2.mV[VX], tex2.mV[VX], tex2.mV[VY] ); + LLVector4a rx0( pos0[VX], tex0.mV[VX], tex0.mV[VY] ); + LLVector4a rx1( pos1[VX], tex1.mV[VX], tex1.mV[VY] ); + LLVector4a rx2( pos2[VX], tex2.mV[VX], tex2.mV[VY] ); - LLVector3 ry0( pos0.mV[VY], tex0.mV[VX], tex0.mV[VY] ); - LLVector3 ry1( pos1.mV[VY], tex1.mV[VX], tex1.mV[VY] ); - LLVector3 ry2( pos2.mV[VY], tex2.mV[VX], tex2.mV[VY] ); + LLVector4a ry0( pos0[VY], tex0.mV[VX], tex0.mV[VY] ); + LLVector4a ry1( pos1[VY], tex1.mV[VX], tex1.mV[VY] ); + LLVector4a ry2( pos2[VY], tex2.mV[VX], tex2.mV[VY] ); - LLVector3 rz0( pos0.mV[VZ], tex0.mV[VX], tex0.mV[VY] ); - LLVector3 rz1( pos1.mV[VZ], tex1.mV[VX], tex1.mV[VY] ); - LLVector3 rz2( pos2.mV[VZ], tex2.mV[VX], tex2.mV[VY] ); + LLVector4a rz0( pos0[VZ], tex0.mV[VX], tex0.mV[VY] ); + LLVector4a rz1( pos1[VZ], tex1.mV[VX], tex1.mV[VY] ); + LLVector4a rz2( pos2[VZ], tex2.mV[VX], tex2.mV[VY] ); - LLVector3 r0 = (rx0 - rx1) % (rx0 - rx2); - LLVector3 r1 = (ry0 - ry1) % (ry0 - ry2); - LLVector3 r2 = (rz0 - rz1) % (rz0 - rz2); + LLVector4a lhs, rhs; + + LLVector4a r0; + lhs.setSub(rx0, rx1); rhs.setSub(rx0, rx2); + r0.setCross3(lhs, rhs); + + LLVector4a r1; + lhs.setSub(ry0, ry1); rhs.setSub(ry0, ry2); + r1.setCross3(lhs, rhs); + + LLVector4a r2; + lhs.setSub(rz0, rz1); rhs.setSub(rz0, rz2); + r2.setCross3(lhs, rhs); - if( r0.mV[VX] && r1.mV[VX] && r2.mV[VX] ) + if( r0[VX] && r1[VX] && r2[VX] ) { - LLVector3 binormal( - -r0.mV[VZ] / r0.mV[VX], - -r1.mV[VZ] / r1.mV[VX], - -r2.mV[VZ] / r2.mV[VX]); + binormal.set( + -r0[VZ] / r0[VX], + -r1[VZ] / r1[VX], + -r2[VZ] / r2[VX]); // binormal.normVec(); - return binormal; } else { - return LLVector3( 0, 1 , 0 ); + binormal.set( 0, 1 , 0 ); } } diff --git a/indra/llmath/llvolume.h b/indra/llmath/llvolume.h index c6a156ae37..af28337f57 100644 --- a/indra/llmath/llvolume.h +++ b/indra/llmath/llvolume.h @@ -40,8 +40,15 @@ class LLPathParams; class LLVolumeParams; class LLProfile; class LLPath; + +#define LL_MESH_ENABLED 1 + +template <class T> class LLOctreeNode; + +class LLVector4a; class LLVolumeFace; class LLVolume; +class LLVolumeTriangle; #include "lldarray.h" #include "lluuid.h" @@ -49,6 +56,7 @@ class LLVolume; //#include "vmath.h" #include "v2math.h" #include "v3math.h" +#include "v3dmath.h" #include "v4math.h" #include "llquaternion.h" #include "llstrider.h" @@ -184,10 +192,15 @@ const U8 LL_SCULPT_TYPE_SPHERE = 1; const U8 LL_SCULPT_TYPE_TORUS = 2; const U8 LL_SCULPT_TYPE_PLANE = 3; const U8 LL_SCULPT_TYPE_CYLINDER = 4; +#if LL_MESH_ENABLED const U8 LL_SCULPT_TYPE_MESH = 5; const U8 LL_SCULPT_TYPE_MASK = LL_SCULPT_TYPE_SPHERE | LL_SCULPT_TYPE_TORUS | LL_SCULPT_TYPE_PLANE | LL_SCULPT_TYPE_CYLINDER | LL_SCULPT_TYPE_MESH; +#else +const U8 LL_SCULPT_TYPE_MASK = LL_SCULPT_TYPE_SPHERE | LL_SCULPT_TYPE_TORUS | LL_SCULPT_TYPE_PLANE | + LL_SCULPT_TYPE_CYLINDER; +#endif const U8 LL_SCULPT_FLAG_INVERT = 64; const U8 LL_SCULPT_FLAG_MIRROR = 128; @@ -791,69 +804,86 @@ public: class LLVolumeFace { public: - LLVolumeFace() : - mID(0), - mTypeMask(0), - mHasBinormals(FALSE), - mBeginS(0), - mBeginT(0), - mNumS(0), - mNumT(0) - { - } - - BOOL create(LLVolume* volume, BOOL partial_build = FALSE); - void createBinormals(); - void makeTriStrip(); - - void appendFace(const LLVolumeFace& face, LLMatrix4& transform, LLMatrix4& normal_tranform); - class VertexData { + enum + { + POSITION = 0, + NORMAL = 1 + }; + + private: + void init(); public: - LLVector3 mPosition; - LLVector3 mNormal; - LLVector3 mBinormal; + VertexData(); + VertexData(const VertexData& rhs); + const VertexData& operator=(const VertexData& rhs); + + ~VertexData(); + LLVector4a& getPosition(); + LLVector4a& getNormal(); + const LLVector4a& getPosition() const; + const LLVector4a& getNormal() const; + void setPosition(const LLVector4a& pos); + void setNormal(const LLVector4a& norm); + + LLVector2 mTexCoord; bool operator<(const VertexData& rhs) const; bool operator==(const VertexData& rhs) const; bool compareNormal(const VertexData& rhs, F32 angle_cutoff) const; + + private: + LLVector4a* mData; }; + LLVolumeFace(); + LLVolumeFace(const LLVolumeFace& src); + LLVolumeFace& operator=(const LLVolumeFace& rhs); + + ~LLVolumeFace(); +private: + void freeData(); +public: + + BOOL create(LLVolume* volume, BOOL partial_build = FALSE); + void createBinormals(); + + void appendFace(const LLVolumeFace& face, LLMatrix4& transform, LLMatrix4& normal_tranform); + + void resizeVertices(S32 num_verts); + void allocateBinormals(S32 num_verts); + void allocateWeights(S32 num_verts); + void resizeIndices(S32 num_indices); + void fillFromLegacyData(std::vector<LLVolumeFace::VertexData>& v, std::vector<U16>& idx); + + void pushVertex(const VertexData& cv); + void pushVertex(const LLVector4a& pos, const LLVector4a& norm, const LLVector2& tc); + void pushIndex(const U16& idx); + + void swapData(LLVolumeFace& rhs); + + void getVertexData(U16 indx, LLVolumeFace::VertexData& cv); + class VertexMapData : public LLVolumeFace::VertexData { public: U16 mIndex; - bool operator==(const LLVolumeFace::VertexData& rhs) const - { - return mPosition == rhs.mPosition && - mTexCoord == rhs.mTexCoord && - mNormal == rhs.mNormal; - } + bool operator==(const LLVolumeFace::VertexData& rhs) const; struct ComparePosition { - bool operator()(const LLVector3& a, const LLVector3& b) const - { - if (a.mV[0] != b.mV[0]) - { - return a.mV[0] < b.mV[0]; - } - if (a.mV[1] != b.mV[1]) - { - return a.mV[1] < b.mV[1]; - } - return a.mV[2] < b.mV[2]; - } + bool operator()(const LLVector3& a, const LLVector3& b) const; }; typedef std::map<LLVector3, std::vector<VertexMapData>, VertexMapData::ComparePosition > PointMap; }; void optimize(F32 angle_cutoff = 2.f); - + void createOctree(); + enum { SINGLE_MASK = 0x0001, @@ -872,26 +902,33 @@ public: public: S32 mID; U32 mTypeMask; - LLVector3 mCenter; - BOOL mHasBinormals; - + // Only used for INNER/OUTER faces S32 mBeginS; S32 mBeginT; S32 mNumS; S32 mNumT; - LLVector3 mExtents[2]; //minimum and maximum point of face + LLVector4a* mExtents; //minimum and maximum point of face + LLVector4a* mCenter; + + S32 mNumVertices; + S32 mNumIndices; + + LLVector4a* mPositions; + LLVector4a* mNormals; + LLVector4a* mBinormals; + LLVector2* mTexCoords; + U16* mIndices; - std::vector<VertexData> mVertices; - std::vector<U16> mIndices; - std::vector<U16> mTriStrip; std::vector<S32> mEdge; //list of skin weights for rigged volumes // format is mWeights[vertex_index].mV[influence] = <joint_index>.<weight> // mWeights.size() should be empty or match mVertices.size() - std::vector<LLVector4> mWeights; + LLVector4a* mWeights; + + LLOctreeNode<LLVolumeTriangle>* mOctree; private: BOOL createUnCutCubeCap(LLVolume* volume, BOOL partial_build = FALSE); @@ -974,6 +1011,13 @@ public: LLVector3* normal = NULL, // return the surface normal at the intersection point LLVector3* bi_normal = NULL // return the surface bi-normal at the intersection point ); + + S32 lineSegmentIntersect(const LLVector4a& start, const LLVector4a& end, + S32 face = 1, + LLVector3* intersection = NULL, + LLVector2* tex_coord = NULL, + LLVector3* normal = NULL, + LLVector3* bi_normal = NULL); // The following cleans up vertices and triangles, // getting rid of degenerate triangles and duplicate vertices, @@ -1038,17 +1082,26 @@ public: std::ostream& operator<<(std::ostream &s, const LLVolumeParams &volume_params); -LLVector3 calc_binormal_from_triangle( - const LLVector3& pos0, +void calc_binormal_from_triangle( + LLVector4a& binormal, + const LLVector4a& pos0, const LLVector2& tex0, - const LLVector3& pos1, + const LLVector4a& pos1, const LLVector2& tex1, - const LLVector3& pos2, + const LLVector4a& pos2, const LLVector2& tex2); +BOOL LLLineSegmentBoxIntersect(const F32* start, const F32* end, const F32* center, const F32* size); BOOL LLLineSegmentBoxIntersect(const LLVector3& start, const LLVector3& end, const LLVector3& center, const LLVector3& size); +BOOL LLLineSegmentBoxIntersect(const LLVector4a& start, const LLVector4a& end, const LLVector4a& center, const LLVector4a& size); + BOOL LLTriangleRayIntersect(const LLVector3& vert0, const LLVector3& vert1, const LLVector3& vert2, const LLVector3& orig, const LLVector3& dir, - F32* intersection_a, F32* intersection_b, F32* intersection_t, BOOL two_sided); + F32& intersection_a, F32& intersection_b, F32& intersection_t, BOOL two_sided); + +BOOL LLTriangleRayIntersect(const LLVector4a& vert0, const LLVector4a& vert1, const LLVector4a& vert2, const LLVector4a& orig, const LLVector4a& dir, + F32& intersection_a, F32& intersection_b, F32& intersection_t); +BOOL LLTriangleRayIntersectTwoSided(const LLVector4a& vert0, const LLVector4a& vert1, const LLVector4a& vert2, const LLVector4a& orig, const LLVector4a& dir, + F32& intersection_a, F32& intersection_b, F32& intersection_t); diff --git a/indra/llmath/llvolumeoctree.cpp b/indra/llmath/llvolumeoctree.cpp new file mode 100644 index 0000000000..194b1faf81 --- /dev/null +++ b/indra/llmath/llvolumeoctree.cpp @@ -0,0 +1,208 @@ +/** + + * @file llvolumeoctree.cpp + * + * $LicenseInfo:firstyear=2002&license=viewergpl$ + * + * Copyright (c) 2002-2009, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#include "llvolumeoctree.h" +#include "llvector4a.h" + +BOOL LLLineSegmentBoxIntersect(const LLVector4a& start, const LLVector4a& end, const LLVector4a& center, const LLVector4a& size) +{ + LLVector4a fAWdU; + LLVector4a dir; + LLVector4a diff; + + dir.setSub(end, start); + dir.mul(0.5f); + + diff.setAdd(end,start); + diff.mul(0.5f); + diff.sub(center); + fAWdU.setAbs(dir); + + LLVector4a rhs; + rhs.setAdd(size, fAWdU); + + LLVector4a lhs; + lhs.setAbs(diff); + + U32 grt = lhs.greaterThan(rhs).getGatheredBits(); + + if (grt & 0x7) + { + return false; + } + + LLVector4a f; + f.setCross3(dir, diff); + f.setAbs(f); + + LLVector4a v0, v1; + + v0 = _mm_shuffle_ps(size, size,_MM_SHUFFLE(3,0,0,1)); + v1 = _mm_shuffle_ps(fAWdU, fAWdU, _MM_SHUFFLE(3,1,2,2)); + lhs.setMul(v0, v1); + + v0 = _mm_shuffle_ps(size, size, _MM_SHUFFLE(3,1,2,2)); + v1 = _mm_shuffle_ps(fAWdU, fAWdU, _MM_SHUFFLE(3,0,0,1)); + rhs.setMul(v0, v1); + rhs.add(lhs); + + grt = f.greaterThan(rhs).getGatheredBits(); + + return (grt & 0x7) ? false : true; +} + + +LLVolumeOctreeListener::LLVolumeOctreeListener(LLOctreeNode<LLVolumeTriangle>* node) +{ + node->addListener(this); + + mBounds = (LLVector4a*) ll_aligned_malloc_16(sizeof(LLVector4a)*4); + mExtents = mBounds+2; +} + +LLVolumeOctreeListener::~LLVolumeOctreeListener() +{ + ll_aligned_free_16(mBounds); +} + +void LLVolumeOctreeListener::handleChildAddition(const LLOctreeNode<LLVolumeTriangle>* parent, + LLOctreeNode<LLVolumeTriangle>* child) +{ + new LLVolumeOctreeListener(child); +} + + +LLOctreeTriangleRayIntersect::LLOctreeTriangleRayIntersect(const LLVector4a& start, const LLVector4a& dir, + const LLVolumeFace* face, F32* closest_t, + LLVector3* intersection,LLVector2* tex_coord, LLVector3* normal, LLVector3* bi_normal) + : mFace(face), + mStart(start), + mDir(dir), + mIntersection(intersection), + mTexCoord(tex_coord), + mNormal(normal), + mBinormal(bi_normal), + mClosestT(closest_t), + mHitFace(false) +{ + mEnd.setAdd(mStart, mDir); +} + +void LLOctreeTriangleRayIntersect::traverse(const LLOctreeNode<LLVolumeTriangle>* node) +{ + LLVolumeOctreeListener* vl = (LLVolumeOctreeListener*) node->getListener(0); + + /*const F32* start = mStart.getF32(); + const F32* end = mEnd.getF32(); + const F32* center = vl->mBounds[0].getF32(); + const F32* size = vl->mBounds[1].getF32();*/ + + //if (LLLineSegmentBoxIntersect(mStart.getF32(), mEnd.getF32(), vl->mBounds[0].getF32(), vl->mBounds[1].getF32())) + if (LLLineSegmentBoxIntersect(mStart, mEnd, vl->mBounds[0], vl->mBounds[1])) + { + node->accept(this); + for (S32 i = 0; i < node->getChildCount(); ++i) + { + traverse(node->getChild(i)); + } + } +} + +void LLOctreeTriangleRayIntersect::visit(const LLOctreeNode<LLVolumeTriangle>* node) +{ + for (LLOctreeNode<LLVolumeTriangle>::const_element_iter iter = + node->getData().begin(); iter != node->getData().end(); ++iter) + { + const LLVolumeTriangle* tri = *iter; + + F32 a, b, t; + + if (LLTriangleRayIntersect(*tri->mV[0], *tri->mV[1], *tri->mV[2], + mStart, mDir, a, b, t)) + { + if ((t >= 0.f) && // if hit is after start + (t <= 1.f) && // and before end + (t < *mClosestT)) // and this hit is closer + { + *mClosestT = t; + mHitFace = true; + + if (mIntersection != NULL) + { + LLVector4a intersect = mDir; + intersect.mul(*mClosestT); + intersect.add(mStart); + mIntersection->set(intersect.getF32ptr()); + } + + + if (mTexCoord != NULL) + { + LLVector2* tc = (LLVector2*) mFace->mTexCoords; + *mTexCoord = ((1.f - a - b) * tc[tri->mIndex[0]] + + a * tc[tri->mIndex[1]] + + b * tc[tri->mIndex[2]]); + + } + + if (mNormal != NULL) + { + LLVector4* norm = (LLVector4*) mFace->mNormals; + + *mNormal = ((1.f - a - b) * LLVector3(norm[tri->mIndex[0]]) + + a * LLVector3(norm[tri->mIndex[1]]) + + b * LLVector3(norm[tri->mIndex[2]])); + } + + if (mBinormal != NULL) + { + LLVector4* binormal = (LLVector4*) mFace->mBinormals; + *mBinormal = ((1.f - a - b) * LLVector3(binormal[tri->mIndex[0]]) + + a * LLVector3(binormal[tri->mIndex[1]]) + + b * LLVector3(binormal[tri->mIndex[2]])); + } + } + } + } +} + +const LLVector4a& LLVolumeTriangle::getPositionGroup() const +{ + return *mPositionGroup; +} + +const F32& LLVolumeTriangle::getBinRadius() const +{ + return mRadius; +} + + diff --git a/indra/llmath/llvolumeoctree.h b/indra/llmath/llvolumeoctree.h new file mode 100644 index 0000000000..0031626498 --- /dev/null +++ b/indra/llmath/llvolumeoctree.h @@ -0,0 +1,138 @@ +/** + * @file llvolumeoctree.h + * @brief LLVolume octree classes. + * + * $LicenseInfo:firstyear=2002&license=viewergpl$ + * + * Copyright (c) 2002-2009, Linden Research, Inc. + * + * Second Life Viewer Source Code + * The source code in this file ("Source Code") is provided by Linden Lab + * to you under the terms of the GNU General Public License, version 2.0 + * ("GPL"), unless you have obtained a separate licensing agreement + * ("Other License"), formally executed by you and Linden Lab. Terms of + * the GPL can be found in doc/GPL-license.txt in this distribution, or + * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2 + * + * There are special exceptions to the terms and conditions of the GPL as + * it is applied to this Source Code. View the full text of the exception + * in the file doc/FLOSS-exception.txt in this software distribution, or + * online at + * http://secondlifegrid.net/programs/open_source/licensing/flossexception + * + * By copying, modifying or distributing this software, you acknowledge + * that you have read and understood your obligations described above, + * and agree to abide by those obligations. + * + * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO + * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY, + * COMPLETENESS OR PERFORMANCE. + * $/LicenseInfo$ + */ + +#ifndef LL_LLVOLUME_OCTREE_H +#define LL_LLVOLUME_OCTREE_H + +#include "linden_common.h" +#include "llmemory.h" + +#include "lloctree.h" +#include "llvolume.h" +#include "llvector4a.h" + +class LLVolumeOctreeListener : public LLOctreeListener<LLVolumeTriangle> +{ +public: + + LLVolumeOctreeListener(LLOctreeNode<LLVolumeTriangle>* node); + ~LLVolumeOctreeListener(); + + LLVolumeOctreeListener(const LLVolumeOctreeListener& rhs) + { + *this = rhs; + } + + const LLVolumeOctreeListener& operator=(const LLVolumeOctreeListener& rhs) + { + llerrs << "Illegal operation!" << llendl; + return *this; + } + + //LISTENER FUNCTIONS + virtual void handleChildAddition(const LLOctreeNode<LLVolumeTriangle>* parent, + LLOctreeNode<LLVolumeTriangle>* child); + virtual void handleStateChange(const LLTreeNode<LLVolumeTriangle>* node) { } + virtual void handleChildRemoval(const LLOctreeNode<LLVolumeTriangle>* parent, + const LLOctreeNode<LLVolumeTriangle>* child) { } + virtual void handleInsertion(const LLTreeNode<LLVolumeTriangle>* node, LLVolumeTriangle* tri) { } + virtual void handleRemoval(const LLTreeNode<LLVolumeTriangle>* node, LLVolumeTriangle* tri) { } + virtual void handleDestruction(const LLTreeNode<LLVolumeTriangle>* node) { } + + +public: + LLVector4a* mBounds; // bounding box (center, size) of this node and all its children (tight fit to objects) + LLVector4a* mExtents; // extents (min, max) of this node and all its children +}; + +class LLOctreeTriangleRayIntersect : public LLOctreeTraveler<LLVolumeTriangle> +{ +public: + const LLVolumeFace* mFace; + LLVector4a mStart; + LLVector4a mDir; + LLVector4a mEnd; + LLVector3* mIntersection; + LLVector2* mTexCoord; + LLVector3* mNormal; + LLVector3* mBinormal; + F32* mClosestT; + bool mHitFace; + + LLOctreeTriangleRayIntersect() { }; + + LLOctreeTriangleRayIntersect(const LLVector4a& start, const LLVector4a& dir, + const LLVolumeFace* face, F32* closest_t, + LLVector3* intersection,LLVector2* tex_coord, LLVector3* normal, LLVector3* bi_normal); + + void traverse(const LLOctreeNode<LLVolumeTriangle>* node); + + virtual void visit(const LLOctreeNode<LLVolumeTriangle>* node); +}; + +class LLVolumeTriangle : public LLRefCount +{ +public: + LLVolumeTriangle() + { + mPositionGroup = (LLVector4a*) ll_aligned_malloc_16(16); + } + + LLVolumeTriangle(const LLVolumeTriangle& rhs) + { + *this = rhs; + } + + const LLVolumeTriangle& operator=(const LLVolumeTriangle& rhs) + { + llerrs << "Illegal operation!" << llendl; + return *this; + } + + ~LLVolumeTriangle() + { + ll_aligned_free_16(mPositionGroup); + } + + const LLVector4a* mV[3]; + U16 mIndex[3]; + + LLVector4a* mPositionGroup; + + F32 mRadius; + + virtual const LLVector4a& getPositionGroup() const; + virtual const F32& getBinRadius() const; +}; + + +#endif diff --git a/indra/llmath/tests/mathmisc_test.cpp b/indra/llmath/tests/mathmisc_test.cpp index ea42f6e001..68d9ddc0fe 100644 --- a/indra/llmath/tests/mathmisc_test.cpp +++ b/indra/llmath/tests/mathmisc_test.cpp @@ -334,6 +334,8 @@ namespace tut template<> template<> void sphere_object::test<2>() { + skip("See SNOW-620. Neither the test nor the code being tested seem good. Also sim-only."); + // test LLSphere::getBoundingSphere() S32 number_of_tests = 100; S32 number_of_spheres = 10; diff --git a/indra/llmath/tests/v2math_test.cpp b/indra/llmath/tests/v2math_test.cpp index 4660fcb955..c745b9989e 100644 --- a/indra/llmath/tests/v2math_test.cpp +++ b/indra/llmath/tests/v2math_test.cpp @@ -91,7 +91,7 @@ namespace tut F32 x = 2.2345f, y = 3.5678f ; LLVector2 vec2(x,y); ensure("magVecSquared:Fail ", is_approx_equal(vec2.magVecSquared(), (x*x + y*y))); - ensure("magVec:Fail ", is_approx_equal(vec2.magVec(), fsqrtf(x*x + y*y))); + ensure("magVec:Fail ", is_approx_equal(vec2.magVec(), (F32) sqrt(x*x + y*y))); } template<> template<> @@ -413,7 +413,7 @@ namespace tut ensure_equals("dist_vec_squared values are not equal",val2, val1); val1 = dist_vec(vec2, vec3); - val2 = fsqrtf((x1 - x2)*(x1 - x2) + (y1 - y2)* (y1 - y2)); + val2 = (F32) sqrt((x1 - x2)*(x1 - x2) + (y1 - y2)* (y1 - y2)); ensure_equals("dist_vec values are not equal",val2, val1); } @@ -437,7 +437,7 @@ namespace tut LLVector2 vec2(x1, y1); F32 vecMag = vec2.normVec(); - F32 mag = fsqrtf(x1*x1 + y1*y1); + F32 mag = (F32) sqrt(x1*x1 + y1*y1); F32 oomag = 1.f / mag; val1 = x1 * oomag; diff --git a/indra/llmath/tests/v3color_test.cpp b/indra/llmath/tests/v3color_test.cpp index 316b6e392f..0efba8e9f3 100644 --- a/indra/llmath/tests/v3color_test.cpp +++ b/indra/llmath/tests/v3color_test.cpp @@ -99,7 +99,7 @@ namespace tut F32 r = 2.3436212f, g = 1231.f, b = 4.7849321232f; LLColor3 llcolor3(r,g,b); ensure("magVecSquared:Fail ", is_approx_equal(llcolor3.magVecSquared(), (r*r + g*g + b*b))); - ensure("magVec:Fail ", is_approx_equal(llcolor3.magVec(), fsqrtf(r*r + g*g + b*b))); + ensure("magVec:Fail ", is_approx_equal(llcolor3.magVec(), (F32) sqrt(r*r + g*g + b*b))); } template<> template<> @@ -109,7 +109,7 @@ namespace tut F32 val1, val2,val3; LLColor3 llcolor3(r,g,b); F32 vecMag = llcolor3.normVec(); - F32 mag = fsqrtf(r*r + g*g + b*b); + F32 mag = (F32) sqrt(r*r + g*g + b*b); F32 oomag = 1.f / mag; val1 = r * oomag; val2 = g * oomag; @@ -292,7 +292,7 @@ namespace tut F32 r1 =1.f, g1 = 2.f,b1 = 1.2f, r2 = -2.3f, g2 = 1.11f, b2 = 1234.234f; LLColor3 llcolor3(r1,g1,b1),llcolor3a(r2,g2,b2); F32 val = distVec(llcolor3,llcolor3a); - ensure("distVec failed ", is_approx_equal(fsqrtf((r1-r2)*(r1-r2) + (g1-g2)*(g1-g2) + (b1-b2)*(b1-b2)) ,val)); + ensure("distVec failed ", is_approx_equal((F32) sqrt((r1-r2)*(r1-r2) + (g1-g2)*(g1-g2) + (b1-b2)*(b1-b2)) ,val)); F32 val1 = distVec_squared(llcolor3,llcolor3a); ensure("distVec_squared failed ", is_approx_equal(((r1-r2)*(r1-r2) + (g1-g2)*(g1-g2) + (b1-b2)*(b1-b2)) ,val1)); diff --git a/indra/llmath/tests/v3dmath_test.cpp b/indra/llmath/tests/v3dmath_test.cpp index e7c949186c..894b6200f5 100644 --- a/indra/llmath/tests/v3dmath_test.cpp +++ b/indra/llmath/tests/v3dmath_test.cpp @@ -409,7 +409,7 @@ namespace tut LLVector3d vec3D(x,y,z); F64 res = (x*x + y*y + z*z) - vec3D.magVecSquared(); ensure("1:magVecSquared:Fail ", ((-F_APPROXIMATELY_ZERO <= res)&& (res <=F_APPROXIMATELY_ZERO))); - res = fsqrtf(x*x + y*y + z*z) - vec3D.magVec(); + res = (F32) sqrt(x*x + y*y + z*z) - vec3D.magVec(); ensure("2:magVec: Fail ", ((-F_APPROXIMATELY_ZERO <= res)&& (res <=F_APPROXIMATELY_ZERO))); } diff --git a/indra/llmath/tests/v3math_test.cpp b/indra/llmath/tests/v3math_test.cpp index 7faf076243..d5c8dd2f9c 100644 --- a/indra/llmath/tests/v3math_test.cpp +++ b/indra/llmath/tests/v3math_test.cpp @@ -155,7 +155,7 @@ namespace tut F32 x = 2.32f, y = 1.212f, z = -.12f; LLVector3 vec3(x,y,z); ensure("1:magVecSquared:Fail ", is_approx_equal(vec3.magVecSquared(), (x*x + y*y + z*z))); - ensure("2:magVec:Fail ", is_approx_equal(vec3.magVec(), fsqrtf(x*x + y*y + z*z))); + ensure("2:magVec:Fail ", is_approx_equal(vec3.magVec(), (F32) sqrt(x*x + y*y + z*z))); } template<> template<> @@ -515,7 +515,7 @@ namespace tut F32 val1,val2; LLVector3 vec3(x1,y1,z1),vec3a(x2,y2,z2); val1 = dist_vec(vec3,vec3a); - val2 = fsqrtf((x1 - x2)*(x1 - x2) + (y1 - y2)* (y1 - y2) + (z1 - z2)* (z1 -z2)); + val2 = (F32) sqrt((x1 - x2)*(x1 - x2) + (y1 - y2)* (y1 - y2) + (z1 - z2)* (z1 -z2)); ensure_equals("1:dist_vec: Fail ",val2, val1); val1 = dist_vec_squared(vec3,vec3a); val2 =((x1 - x2)*(x1 - x2) + (y1 - y2)* (y1 - y2) + (z1 - z2)* (z1 -z2)); diff --git a/indra/llmath/tests/v4color_test.cpp b/indra/llmath/tests/v4color_test.cpp index 33921e0f0f..636446027a 100644 --- a/indra/llmath/tests/v4color_test.cpp +++ b/indra/llmath/tests/v4color_test.cpp @@ -161,7 +161,7 @@ namespace tut F32 r = 0x20, g = 0xFFFF, b = 0xFF; LLColor4 llcolor4(r,g,b); ensure("magVecSquared:Fail ", is_approx_equal(llcolor4.magVecSquared(), (r*r + g*g + b*b))); - ensure("magVec:Fail ", is_approx_equal(llcolor4.magVec(), fsqrtf(r*r + g*g + b*b))); + ensure("magVec:Fail ", is_approx_equal(llcolor4.magVec(), (F32) sqrt(r*r + g*g + b*b))); } template<> template<> @@ -170,7 +170,7 @@ namespace tut F32 r = 0x20, g = 0xFFFF, b = 0xFF; LLColor4 llcolor4(r,g,b); F32 vecMag = llcolor4.normVec(); - F32 mag = fsqrtf(r*r + g*g + b*b); + F32 mag = (F32) sqrt(r*r + g*g + b*b); F32 oomag = 1.f / mag; F32 val1 = r * oomag, val2 = g * oomag, val3 = b * oomag; ensure("1:normVec failed ", (is_approx_equal(val1, llcolor4.mV[0]) && is_approx_equal(val2, llcolor4.mV[1]) && is_approx_equal(val3, llcolor4.mV[2]) && is_approx_equal(vecMag, mag))); diff --git a/indra/llmath/tests/v4coloru_test.cpp b/indra/llmath/tests/v4coloru_test.cpp index 9f71cfc8cc..b3dbfece34 100644 --- a/indra/llmath/tests/v4coloru_test.cpp +++ b/indra/llmath/tests/v4coloru_test.cpp @@ -141,7 +141,7 @@ namespace tut U8 r = 0x12, g = 0xFF, b = 0xAF; LLColor4U llcolor4u(r,g,b); ensure("magVecSquared:Fail ", is_approx_equal(llcolor4u.magVecSquared(), (F32)(r*r + g*g + b*b))); - ensure("magVec:Fail ", is_approx_equal(llcolor4u.magVec(), fsqrtf(r*r + g*g + b*b))); + ensure("magVec:Fail ", is_approx_equal(llcolor4u.magVec(), (F32) sqrt((F32) (r*r + g*g + b*b)))); } template<> template<> diff --git a/indra/llmath/tests/v4math_test.cpp b/indra/llmath/tests/v4math_test.cpp index fe051c27e9..e919c90efa 100644 --- a/indra/llmath/tests/v4math_test.cpp +++ b/indra/llmath/tests/v4math_test.cpp @@ -102,7 +102,7 @@ namespace tut { F32 x = 10.f, y = -2.3f, z = -.023f; LLVector4 vec4(x,y,z); - ensure("magVec:Fail ", is_approx_equal(vec4.magVec(), fsqrtf(x*x + y*y + z*z))); + ensure("magVec:Fail ", is_approx_equal(vec4.magVec(), (F32) sqrt(x*x + y*y + z*z))); ensure("magVecSquared:Fail ", is_approx_equal(vec4.magVecSquared(), (x*x + y*y + z*z))); } @@ -343,7 +343,7 @@ namespace tut F32 val1,val2; LLVector4 vec4(x1,y1,z1),vec4a(x2,y2,z2); val1 = dist_vec(vec4,vec4a); - val2 = fsqrtf((x1 - x2)*(x1 - x2) + (y1 - y2)* (y1 - y2) + (z1 - z2)* (z1 -z2)); + val2 = (F32) sqrt((x1 - x2)*(x1 - x2) + (y1 - y2)* (y1 - y2) + (z1 - z2)* (z1 -z2)); ensure_equals("dist_vec: Fail ",val2, val1); val1 = dist_vec_squared(vec4,vec4a); val2 =((x1 - x2)*(x1 - x2) + (y1 - y2)* (y1 - y2) + (z1 - z2)* (z1 -z2)); diff --git a/indra/llmath/v2math.cpp b/indra/llmath/v2math.cpp index 220336e0c2..2603127f75 100644 --- a/indra/llmath/v2math.cpp +++ b/indra/llmath/v2math.cpp @@ -92,7 +92,7 @@ F32 dist_vec(const LLVector2 &a, const LLVector2 &b) { F32 x = a.mV[0] - b.mV[0]; F32 y = a.mV[1] - b.mV[1]; - return fsqrtf( x*x + y*y ); + return (F32) sqrt( x*x + y*y ); } F32 dist_vec_squared(const LLVector2 &a, const LLVector2 &b) diff --git a/indra/llmath/v2math.h b/indra/llmath/v2math.h index f9f1c024f2..35fd1b6048 100644 --- a/indra/llmath/v2math.h +++ b/indra/llmath/v2math.h @@ -73,6 +73,8 @@ class LLVector2 void setVec(const LLVector2 &vec); // deprecated void setVec(const F32 *vec); // deprecated + inline bool isFinite() const; // checks to see if all values of LLVector2 are finite + F32 length() const; // Returns magnitude of LLVector2 F32 lengthSquared() const; // Returns magnitude squared of LLVector2 F32 normalize(); // Normalizes and returns the magnitude of LLVector2 @@ -218,11 +220,12 @@ inline void LLVector2::setVec(const F32 *vec) mV[VY] = vec[VY]; } + // LLVector2 Magnitude and Normalization Functions inline F32 LLVector2::length(void) const { - return fsqrtf(mV[0]*mV[0] + mV[1]*mV[1]); + return (F32) sqrt(mV[0]*mV[0] + mV[1]*mV[1]); } inline F32 LLVector2::lengthSquared(void) const @@ -232,7 +235,7 @@ inline F32 LLVector2::lengthSquared(void) const inline F32 LLVector2::normalize(void) { - F32 mag = fsqrtf(mV[0]*mV[0] + mV[1]*mV[1]); + F32 mag = (F32) sqrt(mV[0]*mV[0] + mV[1]*mV[1]); F32 oomag; if (mag > FP_MAG_THRESHOLD) @@ -250,10 +253,16 @@ inline F32 LLVector2::normalize(void) return (mag); } +// checker +inline bool LLVector2::isFinite() const +{ + return (llfinite(mV[VX]) && llfinite(mV[VY])); +} + // deprecated inline F32 LLVector2::magVec(void) const { - return fsqrtf(mV[0]*mV[0] + mV[1]*mV[1]); + return (F32) sqrt(mV[0]*mV[0] + mV[1]*mV[1]); } // deprecated @@ -265,7 +274,7 @@ inline F32 LLVector2::magVecSquared(void) const // deprecated inline F32 LLVector2::normVec(void) { - F32 mag = fsqrtf(mV[0]*mV[0] + mV[1]*mV[1]); + F32 mag = (F32) sqrt(mV[0]*mV[0] + mV[1]*mV[1]); F32 oomag; if (mag > FP_MAG_THRESHOLD) diff --git a/indra/llmath/v3color.h b/indra/llmath/v3color.h index 1915d80502..95a3de8b62 100644 --- a/indra/llmath/v3color.h +++ b/indra/llmath/v3color.h @@ -284,7 +284,7 @@ inline F32 LLColor3::brightness(void) const inline F32 LLColor3::length(void) const { - return fsqrtf(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); + return (F32) sqrt(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); } inline F32 LLColor3::lengthSquared(void) const @@ -294,7 +294,7 @@ inline F32 LLColor3::lengthSquared(void) const inline F32 LLColor3::normalize(void) { - F32 mag = fsqrtf(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); + F32 mag = (F32) sqrt(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); F32 oomag; if (mag) @@ -310,7 +310,7 @@ inline F32 LLColor3::normalize(void) // deprecated inline F32 LLColor3::magVec(void) const { - return fsqrtf(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); + return (F32) sqrt(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); } // deprecated @@ -322,7 +322,7 @@ inline F32 LLColor3::magVecSquared(void) const // deprecated inline F32 LLColor3::normVec(void) { - F32 mag = fsqrtf(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); + F32 mag = (F32) sqrt(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); F32 oomag; if (mag) @@ -444,7 +444,7 @@ inline F32 distVec(const LLColor3 &a, const LLColor3 &b) F32 x = a.mV[0] - b.mV[0]; F32 y = a.mV[1] - b.mV[1]; F32 z = a.mV[2] - b.mV[2]; - return fsqrtf( x*x + y*y + z*z ); + return (F32) sqrt( x*x + y*y + z*z ); } inline F32 distVec_squared(const LLColor3 &a, const LLColor3 &b) diff --git a/indra/llmath/v3dmath.h b/indra/llmath/v3dmath.h index 6ab31e8a41..ab253de064 100644 --- a/indra/llmath/v3dmath.h +++ b/indra/llmath/v3dmath.h @@ -240,7 +240,7 @@ inline const LLVector3d& LLVector3d::setVec(const F64 *vec) inline F64 LLVector3d::normVec(void) { - F64 mag = fsqrtf(mdV[0]*mdV[0] + mdV[1]*mdV[1] + mdV[2]*mdV[2]); + F64 mag = (F32) sqrt(mdV[0]*mdV[0] + mdV[1]*mdV[1] + mdV[2]*mdV[2]); F64 oomag; if (mag > FP_MAG_THRESHOLD) @@ -262,7 +262,7 @@ inline F64 LLVector3d::normVec(void) inline F64 LLVector3d::normalize(void) { - F64 mag = fsqrtf(mdV[0]*mdV[0] + mdV[1]*mdV[1] + mdV[2]*mdV[2]); + F64 mag = (F32) sqrt(mdV[0]*mdV[0] + mdV[1]*mdV[1] + mdV[2]*mdV[2]); F64 oomag; if (mag > FP_MAG_THRESHOLD) @@ -286,7 +286,7 @@ inline F64 LLVector3d::normalize(void) inline F64 LLVector3d::magVec(void) const { - return fsqrtf(mdV[0]*mdV[0] + mdV[1]*mdV[1] + mdV[2]*mdV[2]); + return (F32) sqrt(mdV[0]*mdV[0] + mdV[1]*mdV[1] + mdV[2]*mdV[2]); } inline F64 LLVector3d::magVecSquared(void) const @@ -296,7 +296,7 @@ inline F64 LLVector3d::magVecSquared(void) const inline F64 LLVector3d::length(void) const { - return fsqrtf(mdV[0]*mdV[0] + mdV[1]*mdV[1] + mdV[2]*mdV[2]); + return (F32) sqrt(mdV[0]*mdV[0] + mdV[1]*mdV[1] + mdV[2]*mdV[2]); } inline F64 LLVector3d::lengthSquared(void) const @@ -406,7 +406,7 @@ inline F64 dist_vec(const LLVector3d &a, const LLVector3d &b) F64 x = a.mdV[0] - b.mdV[0]; F64 y = a.mdV[1] - b.mdV[1]; F64 z = a.mdV[2] - b.mdV[2]; - return fsqrtf( x*x + y*y + z*z ); + return (F32) sqrt( x*x + y*y + z*z ); } inline F64 dist_vec_squared(const LLVector3d &a, const LLVector3d &b) diff --git a/indra/llmath/v3math.h b/indra/llmath/v3math.h index 76dd938887..5d483a8753 100644 --- a/indra/llmath/v3math.h +++ b/indra/llmath/v3math.h @@ -36,7 +36,6 @@ #include "llerror.h" #include "llmath.h" - #include "llsd.h" class LLVector2; class LLVector4; @@ -283,7 +282,7 @@ inline void LLVector3::setVec(const F32 *vec) inline F32 LLVector3::normalize(void) { - F32 mag = fsqrtf(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); + F32 mag = (F32) sqrt(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); F32 oomag; if (mag > FP_MAG_THRESHOLD) @@ -306,7 +305,7 @@ inline F32 LLVector3::normalize(void) // deprecated inline F32 LLVector3::normVec(void) { - F32 mag = fsqrtf(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); + F32 mag = (F32) sqrt(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); F32 oomag; if (mag > FP_MAG_THRESHOLD) @@ -330,7 +329,7 @@ inline F32 LLVector3::normVec(void) inline F32 LLVector3::length(void) const { - return fsqrtf(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); + return (F32) sqrt(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); } inline F32 LLVector3::lengthSquared(void) const @@ -340,7 +339,7 @@ inline F32 LLVector3::lengthSquared(void) const inline F32 LLVector3::magVec(void) const { - return fsqrtf(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); + return (F32) sqrt(mV[0]*mV[0] + mV[1]*mV[1] + mV[2]*mV[2]); } inline F32 LLVector3::magVecSquared(void) const @@ -474,7 +473,7 @@ inline F32 dist_vec(const LLVector3 &a, const LLVector3 &b) F32 x = a.mV[0] - b.mV[0]; F32 y = a.mV[1] - b.mV[1]; F32 z = a.mV[2] - b.mV[2]; - return fsqrtf( x*x + y*y + z*z ); + return (F32) sqrt( x*x + y*y + z*z ); } inline F32 dist_vec_squared(const LLVector3 &a, const LLVector3 &b) @@ -532,6 +531,21 @@ inline void update_min_max(LLVector3& min, LLVector3& max, const LLVector3& pos) } } +inline void update_min_max(LLVector3& min, LLVector3& max, const F32* pos) +{ + for (U32 i = 0; i < 3; i++) + { + if (min.mV[i] > pos[i]) + { + min.mV[i] = pos[i]; + } + if (max.mV[i] < pos[i]) + { + max.mV[i] = pos[i]; + } + } +} + inline F32 angle_between(const LLVector3& a, const LLVector3& b) { LLVector3 an = a; diff --git a/indra/llmath/v4color.h b/indra/llmath/v4color.h index 6b63b976b0..dd92e1cc63 100644 --- a/indra/llmath/v4color.h +++ b/indra/llmath/v4color.h @@ -392,7 +392,7 @@ inline const LLColor4& LLColor4::setAlpha(F32 a) inline F32 LLColor4::length(void) const { - return fsqrtf(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); + return (F32) sqrt(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); } inline F32 LLColor4::lengthSquared(void) const @@ -402,7 +402,7 @@ inline F32 LLColor4::lengthSquared(void) const inline F32 LLColor4::normalize(void) { - F32 mag = fsqrtf(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); + F32 mag = (F32) sqrt(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); F32 oomag; if (mag) @@ -418,7 +418,7 @@ inline F32 LLColor4::normalize(void) // deprecated inline F32 LLColor4::magVec(void) const { - return fsqrtf(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); + return (F32) sqrt(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); } // deprecated @@ -430,7 +430,7 @@ inline F32 LLColor4::magVecSquared(void) const // deprecated inline F32 LLColor4::normVec(void) { - F32 mag = fsqrtf(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); + F32 mag = (F32) sqrt(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); F32 oomag; if (mag) diff --git a/indra/llmath/v4coloru.h b/indra/llmath/v4coloru.h index 4ec5a345eb..08245403a1 100644 --- a/indra/llmath/v4coloru.h +++ b/indra/llmath/v4coloru.h @@ -300,7 +300,7 @@ inline const LLColor4U& LLColor4U::setAlpha(U8 a) inline F32 LLColor4U::length(void) const { - return fsqrtf( ((F32)mV[VX]) * mV[VX] + ((F32)mV[VY]) * mV[VY] + ((F32)mV[VZ]) * mV[VZ] ); + return (F32) sqrt( ((F32)mV[VX]) * mV[VX] + ((F32)mV[VY]) * mV[VY] + ((F32)mV[VZ]) * mV[VZ] ); } inline F32 LLColor4U::lengthSquared(void) const @@ -311,7 +311,7 @@ inline F32 LLColor4U::lengthSquared(void) const // deprecated inline F32 LLColor4U::magVec(void) const { - return fsqrtf( ((F32)mV[VX]) * mV[VX] + ((F32)mV[VY]) * mV[VY] + ((F32)mV[VZ]) * mV[VZ] ); + return (F32) sqrt( ((F32)mV[VX]) * mV[VX] + ((F32)mV[VY]) * mV[VY] + ((F32)mV[VZ]) * mV[VZ] ); } // deprecated diff --git a/indra/llmath/v4math.h b/indra/llmath/v4math.h index 4c82e6b629..72a477ed20 100644 --- a/indra/llmath/v4math.h +++ b/indra/llmath/v4math.h @@ -321,7 +321,7 @@ inline void LLVector4::setVec(const F32 *vec) inline F32 LLVector4::length(void) const { - return fsqrtf(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); + return (F32) sqrt(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); } inline F32 LLVector4::lengthSquared(void) const @@ -331,7 +331,7 @@ inline F32 LLVector4::lengthSquared(void) const inline F32 LLVector4::magVec(void) const { - return fsqrtf(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); + return (F32) sqrt(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); } inline F32 LLVector4::magVecSquared(void) const @@ -463,7 +463,7 @@ inline LLVector4 lerp(const LLVector4 &a, const LLVector4 &b, F32 u) inline F32 LLVector4::normalize(void) { - F32 mag = fsqrtf(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); + F32 mag = (F32) sqrt(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); F32 oomag; if (mag > FP_MAG_THRESHOLD) @@ -486,7 +486,7 @@ inline F32 LLVector4::normalize(void) // deprecated inline F32 LLVector4::normVec(void) { - F32 mag = fsqrtf(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); + F32 mag = (F32) sqrt(mV[VX]*mV[VX] + mV[VY]*mV[VY] + mV[VZ]*mV[VZ]); F32 oomag; if (mag > FP_MAG_THRESHOLD) |