dlls/d3dx9_36/math.c

   1 /*
   2  * Mathematical operations specific to D3DX9.
   3  *
   4  * Copyright (C) 2008 David Adam
   5  * Copyright (C) 2008 Luis Busquets
   6  * Copyright (C) 2008 Jérôme Gardou
   7  * Copyright (C) 2008 Philip Nilsson
   8  * Copyright (C) 2008 Henri Verbeet
   9  *
  10  * This library is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * This library is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with this library; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  23  */
  24
  25 #define NONAMELESSUNION
  26
  27 #include "config.h"
  28 #include "wine/port.h"
  29
  30 #include "windef.h"
  31 #include "wingdi.h"
  32 #include "d3dx9_36_private.h"
  33
  34 #include "wine/debug.h"
  35
  36 WINE_DEFAULT_DEBUG_CHANNEL(d3dx);
  37
  38 static const ID3DXMatrixStackVtbl ID3DXMatrixStack_Vtbl;
  39
  40 typedef struct ID3DXMatrixStackImpl
  41 {
  42   ID3DXMatrixStack ID3DXMatrixStack_iface;
  43   LONG ref;
  44
  45   unsigned int current;
  46   unsigned int stack_size;
  47   D3DXMATRIX *stack;
  48 } ID3DXMatrixStackImpl;
  49
  50
  51 /*_________________D3DXColor____________________*/
  52
  53 D3DXCOLOR* WINAPI D3DXColorAdjustContrast(D3DXCOLOR *pout, CONST D3DXCOLOR *pc, FLOAT s)
  54 {
  55     pout->r = 0.5f + s * (pc->r - 0.5f);
  56     pout->g = 0.5f + s * (pc->g - 0.5f);
  57     pout->b = 0.5f + s * (pc->b - 0.5f);
  58     pout->a = pc->a;
  59     return pout;
  60 }
  61
  62 D3DXCOLOR* WINAPI D3DXColorAdjustSaturation(D3DXCOLOR *pout, CONST D3DXCOLOR *pc, FLOAT s)
  63 {
  64     FLOAT grey;
  65
  66     grey = pc->r * 0.2125f + pc->g * 0.7154f + pc->b * 0.0721f;
  67     pout->r = grey + s * (pc->r - grey);
  68     pout->g = grey + s * (pc->g - grey);
  69     pout->b = grey + s * (pc->b - grey);
  70     pout->a = pc->a;
  71     return pout;
  72 }
  73
  74 /*_________________Misc__________________________*/
  75
  76 FLOAT WINAPI D3DXFresnelTerm(FLOAT costheta, FLOAT refractionindex)
  77 {
  78     FLOAT a, d, g, result;
  79
  80     g = sqrt(refractionindex * refractionindex + costheta * costheta - 1.0f);
  81     a = g + costheta;
  82     d = g - costheta;
  83     result = ( costheta * a - 1.0f ) * ( costheta * a - 1.0f ) / ( ( costheta * d + 1.0f ) * ( costheta * d + 1.0f ) ) + 1.0f;
  84     result = result * 0.5f * d * d / ( a * a );
  85     return result;
  86 }
  87
  88 /*_________________D3DXMatrix____________________*/
  89
  90 D3DXMATRIX* WINAPI D3DXMatrixAffineTransformation(D3DXMATRIX *pout, FLOAT scaling, CONST D3DXVECTOR3 *rotationcenter, CONST D3DXQUATERNION *rotation, CONST D3DXVECTOR3 *translation)
  91 {
  92     D3DXMATRIX m1, m2, m3, m4, m5;
  93
  94     D3DXMatrixScaling(&m1, scaling, scaling, scaling);
  95
  96     if ( !rotationcenter )
  97     {
  98         D3DXMatrixIdentity(&m2);
  99         D3DXMatrixIdentity(&m4);
 100     }
 101     else
 102     {
 103         D3DXMatrixTranslation(&m2, -rotationcenter->x, -rotationcenter->y, -rotationcenter->z);
 104         D3DXMatrixTranslation(&m4, rotationcenter->x, rotationcenter->y, rotationcenter->z);
 105     }
 106
 107     if ( !rotation ) D3DXMatrixIdentity(&m3);
 108     else D3DXMatrixRotationQuaternion(&m3, rotation);
 109
 110     if ( !translation ) D3DXMatrixIdentity(&m5);
 111     else D3DXMatrixTranslation(&m5, translation->x, translation->y, translation->z);
 112
 113     D3DXMatrixMultiply(&m1, &m1, &m2);
 114     D3DXMatrixMultiply(&m1, &m1, &m3);
 115     D3DXMatrixMultiply(&m1, &m1, &m4);
 116     D3DXMatrixMultiply(pout, &m1, &m5);
 117     return pout;
 118 }
 119
 120 D3DXMATRIX* WINAPI D3DXMatrixAffineTransformation2D(D3DXMATRIX *pout, FLOAT scaling, CONST D3DXVECTOR2 *protationcenter, FLOAT rotation, CONST D3DXVECTOR2 *ptranslation)
 121 {
 122     D3DXMATRIX m1, m2, m3, m4, m5;
 123     D3DXQUATERNION rot;
 124     D3DXVECTOR3 rot_center, trans;
 125
 126     rot.w=cos(rotation/2.0f);
 127     rot.x=0.0f;
 128     rot.y=0.0f;
 129     rot.z=sin(rotation/2.0f);
 130
 131     if ( protationcenter )
 132     {
 133         rot_center.x=protationcenter->x;
 134         rot_center.y=protationcenter->y;
 135         rot_center.z=0.0f;
 136     }
 137     else
 138     {
 139         rot_center.x=0.0f;
 140         rot_center.y=0.0f;
 141         rot_center.z=0.0f;
 142     }
 143
 144     if ( ptranslation )
 145     {
 146         trans.x=ptranslation->x;
 147         trans.y=ptranslation->y;
 148         trans.z=0.0f;
 149     }
 150     else
 151     {
 152         trans.x=0.0f;
 153         trans.y=0.0f;
 154         trans.z=0.0f;
 155     }
 156
 157     D3DXMatrixScaling(&m1, scaling, scaling, 1.0f);
 158     D3DXMatrixTranslation(&m2, -rot_center.x, -rot_center.y, -rot_center.z);
 159     D3DXMatrixTranslation(&m4, rot_center.x, rot_center.y, rot_center.z);
 160     D3DXMatrixRotationQuaternion(&m3, &rot);
 161     D3DXMatrixTranslation(&m5, trans.x, trans.y, trans.z);
 162
 163     D3DXMatrixMultiply(&m1, &m1, &m2);
 164     D3DXMatrixMultiply(&m1, &m1, &m3);
 165     D3DXMatrixMultiply(&m1, &m1, &m4);
 166     D3DXMatrixMultiply(pout, &m1, &m5);
 167
 168     return pout;
 169 }
 170
 171 HRESULT WINAPI D3DXMatrixDecompose(D3DXVECTOR3 *poutscale, D3DXQUATERNION *poutrotation, D3DXVECTOR3 *pouttranslation, CONST D3DXMATRIX *pm)
 172 {
 173     D3DXMATRIX normalized;
 174     D3DXVECTOR3 vec;
 175
 176     /*Compute the scaling part.*/
 177     vec.x=pm->u.m[0][0];
 178     vec.y=pm->u.m[0][1];
 179     vec.z=pm->u.m[0][2];
 180     poutscale->x=D3DXVec3Length(&vec);
 181
 182     vec.x=pm->u.m[1][0];
 183     vec.y=pm->u.m[1][1];
 184     vec.z=pm->u.m[1][2];
 185     poutscale->y=D3DXVec3Length(&vec);
 186
 187     vec.x=pm->u.m[2][0];
 188     vec.y=pm->u.m[2][1];
 189     vec.z=pm->u.m[2][2];
 190     poutscale->z=D3DXVec3Length(&vec);
 191
 192     /*Compute the translation part.*/
 193     pouttranslation->x=pm->u.m[3][0];
 194     pouttranslation->y=pm->u.m[3][1];
 195     pouttranslation->z=pm->u.m[3][2];
 196
 197     /*Let's calculate the rotation now*/
 198     if ( (poutscale->x == 0.0f) || (poutscale->y == 0.0f) || (poutscale->z == 0.0f) ) return D3DERR_INVALIDCALL;
 199
 200     normalized.u.m[0][0]=pm->u.m[0][0]/poutscale->x;
 201     normalized.u.m[0][1]=pm->u.m[0][1]/poutscale->x;
 202     normalized.u.m[0][2]=pm->u.m[0][2]/poutscale->x;
 203     normalized.u.m[1][0]=pm->u.m[1][0]/poutscale->y;
 204     normalized.u.m[1][1]=pm->u.m[1][1]/poutscale->y;
 205     normalized.u.m[1][2]=pm->u.m[1][2]/poutscale->y;
 206     normalized.u.m[2][0]=pm->u.m[2][0]/poutscale->z;
 207     normalized.u.m[2][1]=pm->u.m[2][1]/poutscale->z;
 208     normalized.u.m[2][2]=pm->u.m[2][2]/poutscale->z;
 209
 210     D3DXQuaternionRotationMatrix(poutrotation,&normalized);
 211     return S_OK;
 212 }
 213
 214 FLOAT WINAPI D3DXMatrixDeterminant(CONST D3DXMATRIX *pm)
 215 {
 216     D3DXVECTOR4 minor, v1, v2, v3;
 217     FLOAT det;
 218
 219     v1.x = pm->u.m[0][0]; v1.y = pm->u.m[1][0]; v1.z = pm->u.m[2][0]; v1.w = pm->u.m[3][0];
 220     v2.x = pm->u.m[0][1]; v2.y = pm->u.m[1][1]; v2.z = pm->u.m[2][1]; v2.w = pm->u.m[3][1];
 221     v3.x = pm->u.m[0][2]; v3.y = pm->u.m[1][2]; v3.z = pm->u.m[2][2]; v3.w = pm->u.m[3][2];
 222     D3DXVec4Cross(&minor, &v1, &v2, &v3);
 223     det =  - (pm->u.m[0][3] * minor.x + pm->u.m[1][3] * minor.y + pm->u.m[2][3] * minor.z + pm->u.m[3][3] * minor.w);
 224     return det;
 225 }
 226
 227 D3DXMATRIX* WINAPI D3DXMatrixInverse(D3DXMATRIX *pout, FLOAT *pdeterminant, CONST D3DXMATRIX *pm)
 228 {
 229     int a, i, j;
 230     D3DXMATRIX out;
 231     D3DXVECTOR4 v, vec[3];
 232     FLOAT det;
 233
 234     det = D3DXMatrixDeterminant(pm);
 235     if ( !det ) return NULL;
 236     if ( pdeterminant ) *pdeterminant = det;
 237     for (i=0; i<4; i++)
 238     {
 239         for (j=0; j<4; j++)
 240         {
 241             if (j != i )
 242             {
 243                 a = j;
 244                 if ( j > i ) a = a-1;
 245                 vec[a].x = pm->u.m[j][0];
 246                 vec[a].y = pm->u.m[j][1];
 247                 vec[a].z = pm->u.m[j][2];
 248                 vec[a].w = pm->u.m[j][3];
 249             }
 250         }
 251     D3DXVec4Cross(&v, &vec[0], &vec[1], &vec[2]);
 252     out.u.m[0][i] = pow(-1.0f, i) * v.x / det;
 253     out.u.m[1][i] = pow(-1.0f, i) * v.y / det;
 254     out.u.m[2][i] = pow(-1.0f, i) * v.z / det;
 255     out.u.m[3][i] = pow(-1.0f, i) * v.w / det;
 256    }
 257
 258    *pout = out;
 259    return pout;
 260 }
 261
 262 D3DXMATRIX* WINAPI D3DXMatrixLookAtLH(D3DXMATRIX *pout, CONST D3DXVECTOR3 *peye, CONST D3DXVECTOR3 *pat, CONST D3DXVECTOR3 *pup)
 263 {
 264     D3DXVECTOR3 right, rightn, up, upn, vec, vec2;
 265
 266     D3DXVec3Subtract(&vec2, pat, peye);
 267     D3DXVec3Normalize(&vec, &vec2);
 268     D3DXVec3Cross(&right, pup, &vec);
 269     D3DXVec3Cross(&up, &vec, &right);
 270     D3DXVec3Normalize(&rightn, &right);
 271     D3DXVec3Normalize(&upn, &up);
 272     pout->u.m[0][0] = rightn.x;
 273     pout->u.m[1][0] = rightn.y;
 274     pout->u.m[2][0] = rightn.z;
 275     pout->u.m[3][0] = -D3DXVec3Dot(&rightn,peye);
 276     pout->u.m[0][1] = upn.x;
 277     pout->u.m[1][1] = upn.y;
 278     pout->u.m[2][1] = upn.z;
 279     pout->u.m[3][1] = -D3DXVec3Dot(&upn, peye);
 280     pout->u.m[0][2] = vec.x;
 281     pout->u.m[1][2] = vec.y;
 282     pout->u.m[2][2] = vec.z;
 283     pout->u.m[3][2] = -D3DXVec3Dot(&vec, peye);
 284     pout->u.m[0][3] = 0.0f;
 285     pout->u.m[1][3] = 0.0f;
 286     pout->u.m[2][3] = 0.0f;
 287     pout->u.m[3][3] = 1.0f;
 288     return pout;
 289 }
 290
 291 D3DXMATRIX* WINAPI D3DXMatrixLookAtRH(D3DXMATRIX *pout, CONST D3DXVECTOR3 *peye, CONST D3DXVECTOR3 *pat, CONST D3DXVECTOR3 *pup)
 292 {
 293     D3DXVECTOR3 right, rightn, up, upn, vec, vec2;
 294
 295     D3DXVec3Subtract(&vec2, pat, peye);
 296     D3DXVec3Normalize(&vec, &vec2);
 297     D3DXVec3Cross(&right, pup, &vec);
 298     D3DXVec3Cross(&up, &vec, &right);
 299     D3DXVec3Normalize(&rightn, &right);
 300     D3DXVec3Normalize(&upn, &up);
 301     pout->u.m[0][0] = -rightn.x;
 302     pout->u.m[1][0] = -rightn.y;
 303     pout->u.m[2][0] = -rightn.z;
 304     pout->u.m[3][0] = D3DXVec3Dot(&rightn,peye);
 305     pout->u.m[0][1] = upn.x;
 306     pout->u.m[1][1] = upn.y;
 307     pout->u.m[2][1] = upn.z;
 308     pout->u.m[3][1] = -D3DXVec3Dot(&upn, peye);
 309     pout->u.m[0][2] = -vec.x;
 310     pout->u.m[1][2] = -vec.y;
 311     pout->u.m[2][2] = -vec.z;
 312     pout->u.m[3][2] = D3DXVec3Dot(&vec, peye);
 313     pout->u.m[0][3] = 0.0f;
 314     pout->u.m[1][3] = 0.0f;
 315     pout->u.m[2][3] = 0.0f;
 316     pout->u.m[3][3] = 1.0f;
 317     return pout;
 318 }
 319
 320 D3DXMATRIX* WINAPI D3DXMatrixMultiply(D3DXMATRIX *pout, CONST D3DXMATRIX *pm1, CONST D3DXMATRIX *pm2)
 321 {
 322     D3DXMATRIX out;
 323     int i,j;
 324
 325     for (i=0; i<4; i++)
 326     {
 327         for (j=0; j<4; j++)
 328         {
 329             out.u.m[i][j] = pm1->u.m[i][0] * pm2->u.m[0][j] + pm1->u.m[i][1] * pm2->u.m[1][j] + pm1->u.m[i][2] * pm2->u.m[2][j] + pm1->u.m[i][3] * pm2->u.m[3][j];
 330         }
 331     }
 332
 333     *pout = out;
 334     return pout;
 335 }
 336
 337 D3DXMATRIX* WINAPI D3DXMatrixMultiplyTranspose(D3DXMATRIX *pout, CONST D3DXMATRIX *pm1, CONST D3DXMATRIX *pm2)
 338 {
 339     D3DXMatrixMultiply(pout, pm1, pm2);
 340     D3DXMatrixTranspose(pout, pout);
 341     return pout;
 342 }
 343
 344 D3DXMATRIX* WINAPI D3DXMatrixOrthoLH(D3DXMATRIX *pout, FLOAT w, FLOAT h, FLOAT zn, FLOAT zf)
 345 {
 346     D3DXMatrixIdentity(pout);
 347     pout->u.m[0][0] = 2.0f / w;
 348     pout->u.m[1][1] = 2.0f / h;
 349     pout->u.m[2][2] = 1.0f / (zf - zn);
 350     pout->u.m[3][2] = zn / (zn - zf);
 351     return pout;
 352 }
 353
 354 D3DXMATRIX* WINAPI D3DXMatrixOrthoOffCenterLH(D3DXMATRIX *pout, FLOAT l, FLOAT r, FLOAT b, FLOAT t, FLOAT zn, FLOAT zf)
 355 {
 356     D3DXMatrixIdentity(pout);
 357     pout->u.m[0][0] = 2.0f / (r - l);
 358     pout->u.m[1][1] = 2.0f / (t - b);
 359     pout->u.m[2][2] = 1.0f / (zf -zn);
 360     pout->u.m[3][0] = -1.0f -2.0f *l / (r - l);
 361     pout->u.m[3][1] = 1.0f + 2.0f * t / (b - t);
 362     pout->u.m[3][2] = zn / (zn -zf);
 363     return pout;
 364 }
 365
 366 D3DXMATRIX* WINAPI D3DXMatrixOrthoOffCenterRH(D3DXMATRIX *pout, FLOAT l, FLOAT r, FLOAT b, FLOAT t, FLOAT zn, FLOAT zf)
 367 {
 368     D3DXMatrixIdentity(pout);
 369     pout->u.m[0][0] = 2.0f / (r - l);
 370     pout->u.m[1][1] = 2.0f / (t - b);
 371     pout->u.m[2][2] = 1.0f / (zn -zf);
 372     pout->u.m[3][0] = -1.0f -2.0f *l / (r - l);
 373     pout->u.m[3][1] = 1.0f + 2.0f * t / (b - t);
 374     pout->u.m[3][2] = zn / (zn -zf);
 375     return pout;
 376 }
 377
 378 D3DXMATRIX* WINAPI D3DXMatrixOrthoRH(D3DXMATRIX *pout, FLOAT w, FLOAT h, FLOAT zn, FLOAT zf)
 379 {
 380     D3DXMatrixIdentity(pout);
 381     pout->u.m[0][0] = 2.0f / w;
 382     pout->u.m[1][1] = 2.0f / h;
 383     pout->u.m[2][2] = 1.0f / (zn - zf);
 384     pout->u.m[3][2] = zn / (zn - zf);
 385     return pout;
 386 }
 387
 388 D3DXMATRIX* WINAPI D3DXMatrixPerspectiveFovLH(D3DXMATRIX *pout, FLOAT fovy, FLOAT aspect, FLOAT zn, FLOAT zf)
 389 {
 390     D3DXMatrixIdentity(pout);
 391     pout->u.m[0][0] = 1.0f / (aspect * tan(fovy/2.0f));
 392     pout->u.m[1][1] = 1.0f / tan(fovy/2.0f);
 393     pout->u.m[2][2] = zf / (zf - zn);
 394     pout->u.m[2][3] = 1.0f;
 395     pout->u.m[3][2] = (zf * zn) / (zn - zf);
 396     pout->u.m[3][3] = 0.0f;
 397     return pout;
 398 }
 399
 400 D3DXMATRIX* WINAPI D3DXMatrixPerspectiveFovRH(D3DXMATRIX *pout, FLOAT fovy, FLOAT aspect, FLOAT zn, FLOAT zf)
 401 {
 402     D3DXMatrixIdentity(pout);
 403     pout->u.m[0][0] = 1.0f / (aspect * tan(fovy/2.0f));
 404     pout->u.m[1][1] = 1.0f / tan(fovy/2.0f);
 405     pout->u.m[2][2] = zf / (zn - zf);
 406     pout->u.m[2][3] = -1.0f;
 407     pout->u.m[3][2] = (zf * zn) / (zn - zf);
 408     pout->u.m[3][3] = 0.0f;
 409     return pout;
 410 }
 411
 412 D3DXMATRIX* WINAPI D3DXMatrixPerspectiveLH(D3DXMATRIX *pout, FLOAT w, FLOAT h, FLOAT zn, FLOAT zf)
 413 {
 414     D3DXMatrixIdentity(pout);
 415     pout->u.m[0][0] = 2.0f * zn / w;
 416     pout->u.m[1][1] = 2.0f * zn / h;
 417     pout->u.m[2][2] = zf / (zf - zn);
 418     pout->u.m[3][2] = (zn * zf) / (zn - zf);
 419     pout->u.m[2][3] = 1.0f;
 420     pout->u.m[3][3] = 0.0f;
 421     return pout;
 422 }
 423
 424 D3DXMATRIX* WINAPI D3DXMatrixPerspectiveOffCenterLH(D3DXMATRIX *pout, FLOAT l, FLOAT r, FLOAT b, FLOAT t, FLOAT zn, FLOAT zf)
 425 {
 426     D3DXMatrixIdentity(pout);
 427     pout->u.m[0][0] = 2.0f * zn / (r - l);
 428     pout->u.m[1][1] = -2.0f * zn / (b - t);
 429     pout->u.m[2][0] = -1.0f - 2.0f * l / (r - l);
 430     pout->u.m[2][1] = 1.0f + 2.0f * t / (b - t);
 431     pout->u.m[2][2] = - zf / (zn - zf);
 432     pout->u.m[3][2] = (zn * zf) / (zn -zf);
 433     pout->u.m[2][3] = 1.0f;
 434     pout->u.m[3][3] = 0.0f;
 435     return pout;
 436 }
 437
 438 D3DXMATRIX* WINAPI D3DXMatrixPerspectiveOffCenterRH(D3DXMATRIX *pout, FLOAT l, FLOAT r, FLOAT b, FLOAT t, FLOAT zn, FLOAT zf)
 439 {
 440     D3DXMatrixIdentity(pout);
 441     pout->u.m[0][0] = 2.0f * zn / (r - l);
 442     pout->u.m[1][1] = -2.0f * zn / (b - t);
 443     pout->u.m[2][0] = 1.0f + 2.0f * l / (r - l);
 444     pout->u.m[2][1] = -1.0f -2.0f * t / (b - t);
 445     pout->u.m[2][2] = zf / (zn - zf);
 446     pout->u.m[3][2] = (zn * zf) / (zn -zf);
 447     pout->u.m[2][3] = -1.0f;
 448     pout->u.m[3][3] = 0.0f;
 449     return pout;
 450 }
 451
 452 D3DXMATRIX* WINAPI D3DXMatrixPerspectiveRH(D3DXMATRIX *pout, FLOAT w, FLOAT h, FLOAT zn, FLOAT zf)
 453 {
 454     D3DXMatrixIdentity(pout);
 455     pout->u.m[0][0] = 2.0f * zn / w;
 456     pout->u.m[1][1] = 2.0f * zn / h;
 457     pout->u.m[2][2] = zf / (zn - zf);
 458     pout->u.m[3][2] = (zn * zf) / (zn - zf);
 459     pout->u.m[2][3] = -1.0f;
 460     pout->u.m[3][3] = 0.0f;
 461     return pout;
 462 }
 463
 464 D3DXMATRIX* WINAPI D3DXMatrixReflect(D3DXMATRIX *pout, CONST D3DXPLANE *pplane)
 465 {
 466     D3DXPLANE Nplane;
 467
 468     D3DXPlaneNormalize(&Nplane, pplane);
 469     D3DXMatrixIdentity(pout);
 470     pout->u.m[0][0] = 1.0f - 2.0f * Nplane.a * Nplane.a;
 471     pout->u.m[0][1] = -2.0f * Nplane.a * Nplane.b;
 472     pout->u.m[0][2] = -2.0f * Nplane.a * Nplane.c;
 473     pout->u.m[1][0] = -2.0f * Nplane.a * Nplane.b;
 474     pout->u.m[1][1] = 1.0f - 2.0f * Nplane.b * Nplane.b;
 475     pout->u.m[1][2] = -2.0f * Nplane.b * Nplane.c;
 476     pout->u.m[2][0] = -2.0f * Nplane.c * Nplane.a;
 477     pout->u.m[2][1] = -2.0f * Nplane.c * Nplane.b;
 478     pout->u.m[2][2] = 1.0f - 2.0f * Nplane.c * Nplane.c;
 479     pout->u.m[3][0] = -2.0f * Nplane.d * Nplane.a;
 480     pout->u.m[3][1] = -2.0f * Nplane.d * Nplane.b;
 481     pout->u.m[3][2] = -2.0f * Nplane.d * Nplane.c;
 482     return pout;
 483 }
 484
 485 D3DXMATRIX* WINAPI D3DXMatrixRotationAxis(D3DXMATRIX *pout, CONST D3DXVECTOR3 *pv, FLOAT angle)
 486 {
 487     D3DXVECTOR3 v;
 488
 489     D3DXVec3Normalize(&v,pv);
 490     D3DXMatrixIdentity(pout);
 491     pout->u.m[0][0] = (1.0f - cos(angle)) * v.x * v.x + cos(angle);
 492     pout->u.m[1][0] = (1.0f - cos(angle)) * v.x * v.y - sin(angle) * v.z;
 493     pout->u.m[2][0] = (1.0f - cos(angle)) * v.x * v.z + sin(angle) * v.y;
 494     pout->u.m[0][1] = (1.0f - cos(angle)) * v.y * v.x + sin(angle) * v.z;
 495     pout->u.m[1][1] = (1.0f - cos(angle)) * v.y * v.y + cos(angle);
 496     pout->u.m[2][1] = (1.0f - cos(angle)) * v.y * v.z - sin(angle) * v.x;
 497     pout->u.m[0][2] = (1.0f - cos(angle)) * v.z * v.x - sin(angle) * v.y;
 498     pout->u.m[1][2] = (1.0f - cos(angle)) * v.z * v.y + sin(angle) * v.x;
 499     pout->u.m[2][2] = (1.0f - cos(angle)) * v.z * v.z + cos(angle);
 500     return pout;
 501 }
 502
 503 D3DXMATRIX* WINAPI D3DXMatrixRotationQuaternion(D3DXMATRIX *pout, CONST D3DXQUATERNION *pq)
 504 {
 505     D3DXMatrixIdentity(pout);
 506     pout->u.m[0][0] = 1.0f - 2.0f * (pq->y * pq->y + pq->z * pq->z);
 507     pout->u.m[0][1] = 2.0f * (pq->x *pq->y + pq->z * pq->w);
 508     pout->u.m[0][2] = 2.0f * (pq->x * pq->z - pq->y * pq->w);
 509     pout->u.m[1][0] = 2.0f * (pq->x * pq->y - pq->z * pq->w);
 510     pout->u.m[1][1] = 1.0f - 2.0f * (pq->x * pq->x + pq->z * pq->z);
 511     pout->u.m[1][2] = 2.0f * (pq->y *pq->z + pq->x *pq->w);
 512     pout->u.m[2][0] = 2.0f * (pq->x * pq->z + pq->y * pq->w);
 513     pout->u.m[2][1] = 2.0f * (pq->y *pq->z - pq->x *pq->w);
 514     pout->u.m[2][2] = 1.0f - 2.0f * (pq->x * pq->x + pq->y * pq->y);
 515     return pout;
 516 }
 517
 518 D3DXMATRIX* WINAPI D3DXMatrixRotationX(D3DXMATRIX *pout, FLOAT angle)
 519 {
 520     D3DXMatrixIdentity(pout);
 521     pout->u.m[1][1] = cos(angle);
 522     pout->u.m[2][2] = cos(angle);
 523     pout->u.m[1][2] = sin(angle);
 524     pout->u.m[2][1] = -sin(angle);
 525     return pout;
 526 }
 527
 528 D3DXMATRIX* WINAPI D3DXMatrixRotationY(D3DXMATRIX *pout, FLOAT angle)
 529 {
 530     D3DXMatrixIdentity(pout);
 531     pout->u.m[0][0] = cos(angle);
 532     pout->u.m[2][2] = cos(angle);
 533     pout->u.m[0][2] = -sin(angle);
 534     pout->u.m[2][0] = sin(angle);
 535     return pout;
 536 }
 537
 538 D3DXMATRIX* WINAPI D3DXMatrixRotationYawPitchRoll(D3DXMATRIX *pout, FLOAT yaw, FLOAT pitch, FLOAT roll)
 539 {
 540     D3DXMATRIX m;
 541
 542     D3DXMatrixIdentity(pout);
 543     D3DXMatrixRotationZ(&m, roll);
 544     D3DXMatrixMultiply(pout, pout, &m);
 545     D3DXMatrixRotationX(&m, pitch);
 546     D3DXMatrixMultiply(pout, pout, &m);
 547     D3DXMatrixRotationY(&m, yaw);
 548     D3DXMatrixMultiply(pout, pout, &m);
 549     return pout;
 550 }
 551 D3DXMATRIX* WINAPI D3DXMatrixRotationZ(D3DXMATRIX *pout, FLOAT angle)
 552 {
 553     D3DXMatrixIdentity(pout);
 554     pout->u.m[0][0] = cos(angle);
 555     pout->u.m[1][1] = cos(angle);
 556     pout->u.m[0][1] = sin(angle);
 557     pout->u.m[1][0] = -sin(angle);
 558     return pout;
 559 }
 560
 561 D3DXMATRIX* WINAPI D3DXMatrixScaling(D3DXMATRIX *pout, FLOAT sx, FLOAT sy, FLOAT sz)
 562 {
 563     D3DXMatrixIdentity(pout);
 564     pout->u.m[0][0] = sx;
 565     pout->u.m[1][1] = sy;
 566     pout->u.m[2][2] = sz;
 567     return pout;
 568 }
 569
 570 D3DXMATRIX* WINAPI D3DXMatrixShadow(D3DXMATRIX *pout, CONST D3DXVECTOR4 *plight, CONST D3DXPLANE *pplane)
 571 {
 572     D3DXPLANE Nplane;
 573     FLOAT dot;
 574
 575     D3DXPlaneNormalize(&Nplane, pplane);
 576     dot = D3DXPlaneDot(&Nplane, plight);
 577     pout->u.m[0][0] = dot - Nplane.a * plight->x;
 578     pout->u.m[0][1] = -Nplane.a * plight->y;
 579     pout->u.m[0][2] = -Nplane.a * plight->z;
 580     pout->u.m[0][3] = -Nplane.a * plight->w;
 581     pout->u.m[1][0] = -Nplane.b * plight->x;
 582     pout->u.m[1][1] = dot - Nplane.b * plight->y;
 583     pout->u.m[1][2] = -Nplane.b * plight->z;
 584     pout->u.m[1][3] = -Nplane.b * plight->w;
 585     pout->u.m[2][0] = -Nplane.c * plight->x;
 586     pout->u.m[2][1] = -Nplane.c * plight->y;
 587     pout->u.m[2][2] = dot - Nplane.c * plight->z;
 588     pout->u.m[2][3] = -Nplane.c * plight->w;
 589     pout->u.m[3][0] = -Nplane.d * plight->x;
 590     pout->u.m[3][1] = -Nplane.d * plight->y;
 591     pout->u.m[3][2] = -Nplane.d * plight->z;
 592     pout->u.m[3][3] = dot - Nplane.d * plight->w;
 593     return pout;
 594 }
 595
 596 D3DXMATRIX* WINAPI D3DXMatrixTransformation(D3DXMATRIX *pout, CONST D3DXVECTOR3 *pscalingcenter, CONST D3DXQUATERNION *pscalingrotation, CONST D3DXVECTOR3 *pscaling, CONST D3DXVECTOR3 *protationcenter, CONST D3DXQUATERNION *protation, CONST D3DXVECTOR3 *ptranslation)
 597 {
 598     D3DXMATRIX m1, m2, m3, m4, m5, m6, m7;
 599     D3DXQUATERNION prc;
 600     D3DXVECTOR3 psc, pt;
 601
 602     if ( !pscalingcenter )
 603     {
 604         psc.x = 0.0f;
 605         psc.y = 0.0f;
 606         psc.z = 0.0f;
 607     }
 608     else
 609     {
 610         psc.x = pscalingcenter->x;
 611         psc.y = pscalingcenter->y;
 612         psc.z = pscalingcenter->z;
 613     }
 614
 615     if ( !protationcenter )
 616     {
 617         prc.x = 0.0f;
 618         prc.y = 0.0f;
 619         prc.z = 0.0f;
 620     }
 621     else
 622     {
 623         prc.x = protationcenter->x;
 624         prc.y = protationcenter->y;
 625         prc.z = protationcenter->z;
 626     }
 627
 628     if ( !ptranslation )
 629     {
 630         pt.x = 0.0f;
 631         pt.y = 0.0f;
 632         pt.z = 0.0f;
 633     }
 634     else
 635     {
 636         pt.x = ptranslation->x;
 637         pt.y = ptranslation->y;
 638         pt.z = ptranslation->z;
 639     }
 640
 641     D3DXMatrixTranslation(&m1, -psc.x, -psc.y, -psc.z);
 642
 643     if ( !pscalingrotation )
 644     {
 645         D3DXMatrixIdentity(&m2);
 646         D3DXMatrixIdentity(&m4);
 647     }
 648     else
 649     {
 650         D3DXMatrixRotationQuaternion(&m4, pscalingrotation);
 651         D3DXMatrixInverse(&m2, NULL, &m4);
 652     }
 653
 654     if ( !pscaling ) D3DXMatrixIdentity(&m3);
 655     else D3DXMatrixScaling(&m3, pscaling->x, pscaling->y, pscaling->z);
 656
 657     if ( !protation ) D3DXMatrixIdentity(&m6);
 658     else D3DXMatrixRotationQuaternion(&m6, protation);
 659
 660     D3DXMatrixTranslation(&m5, psc.x - prc.x,  psc.y - prc.y,  psc.z - prc.z);
 661     D3DXMatrixTranslation(&m7, prc.x + pt.x, prc.y + pt.y, prc.z + pt.z);
 662     D3DXMatrixMultiply(&m1, &m1, &m2);
 663     D3DXMatrixMultiply(&m1, &m1, &m3);
 664     D3DXMatrixMultiply(&m1, &m1, &m4);
 665     D3DXMatrixMultiply(&m1, &m1, &m5);
 666     D3DXMatrixMultiply(&m1, &m1, &m6);
 667     D3DXMatrixMultiply(pout, &m1, &m7);
 668     return pout;
 669 }
 670 D3DXMATRIX* WINAPI D3DXMatrixTransformation2D(D3DXMATRIX *pout, CONST D3DXVECTOR2 *pscalingcenter, FLOAT scalingrotation, CONST D3DXVECTOR2 *pscaling, CONST D3DXVECTOR2 *protationcenter, FLOAT rotation, CONST D3DXVECTOR2 *ptranslation)
 671 {
 672     D3DXQUATERNION rot, sca_rot;
 673     D3DXVECTOR3 rot_center, sca, sca_center, trans;
 674
 675     if ( pscalingcenter )
 676     {
 677         sca_center.x=pscalingcenter->x;
 678         sca_center.y=pscalingcenter->y;
 679         sca_center.z=0.0f;
 680     }
 681     else
 682     {
 683         sca_center.x=0.0f;
 684         sca_center.y=0.0f;
 685         sca_center.z=0.0f;
 686     }
 687
 688     if ( pscaling )
 689     {
 690         sca.x=pscaling->x;
 691         sca.y=pscaling->y;
 692         sca.z=1.0f;
 693     }
 694     else
 695     {
 696         sca.x=1.0f;
 697         sca.y=1.0f;
 698         sca.z=1.0f;
 699     }
 700
 701     if ( protationcenter )
 702     {
 703         rot_center.x=protationcenter->x;
 704         rot_center.y=protationcenter->y;
 705         rot_center.z=0.0f;
 706     }
 707     else
 708     {
 709         rot_center.x=0.0f;
 710         rot_center.y=0.0f;
 711         rot_center.z=0.0f;
 712     }
 713
 714     if ( ptranslation )
 715     {
 716         trans.x=ptranslation->x;
 717         trans.y=ptranslation->y;
 718         trans.z=0.0f;
 719     }
 720     else
 721     {
 722         trans.x=0.0f;
 723         trans.y=0.0f;
 724         trans.z=0.0f;
 725     }
 726
 727     rot.w=cos(rotation/2.0f);
 728     rot.x=0.0f;
 729     rot.y=0.0f;
 730     rot.z=sin(rotation/2.0f);
 731
 732     sca_rot.w=cos(scalingrotation/2.0f);
 733     sca_rot.x=0.0f;
 734     sca_rot.y=0.0f;
 735     sca_rot.z=sin(scalingrotation/2.0f);
 736
 737     D3DXMatrixTransformation(pout, &sca_center, &sca_rot, &sca, &rot_center, &rot, &trans);
 738
 739     return pout;
 740 }
 741
 742 D3DXMATRIX* WINAPI D3DXMatrixTranslation(D3DXMATRIX *pout, FLOAT x, FLOAT y, FLOAT z)
 743 {
 744     D3DXMatrixIdentity(pout);
 745     pout->u.m[3][0] = x;
 746     pout->u.m[3][1] = y;
 747     pout->u.m[3][2] = z;
 748     return pout;
 749 }
 750
 751 D3DXMATRIX* WINAPI D3DXMatrixTranspose(D3DXMATRIX *pout, CONST D3DXMATRIX *pm)
 752 {
 753     CONST D3DXMATRIX m = *pm;
 754     int i,j;
 755
 756     for (i=0; i<4; i++)
 757         for (j=0; j<4; j++) pout->u.m[i][j] = m.u.m[j][i];
 758
 759     return pout;
 760 }
 761
 762 /*_________________D3DXMatrixStack____________________*/
 763
 764 static const unsigned int INITIAL_STACK_SIZE = 32;
 765
 766 HRESULT WINAPI D3DXCreateMatrixStack(DWORD flags, LPD3DXMATRIXSTACK* ppstack)
 767 {
 768     ID3DXMatrixStackImpl* object;
 769
 770     TRACE("flags %#x, ppstack %p\n", flags, ppstack);
 771
 772     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(ID3DXMatrixStackImpl));
 773     if ( object == NULL )
 774     {
 775      *ppstack = NULL;
 776      return E_OUTOFMEMORY;
 777     }
 778     object->ID3DXMatrixStack_iface.lpVtbl = &ID3DXMatrixStack_Vtbl;
 779     object->ref = 1;
 780
 781     object->stack = HeapAlloc(GetProcessHeap(), 0, INITIAL_STACK_SIZE * sizeof(D3DXMATRIX));
 782     if (!object->stack)
 783     {
 784         HeapFree(GetProcessHeap(), 0, object);
 785         *ppstack = NULL;
 786         return E_OUTOFMEMORY;
 787     }
 788
 789     object->current = 0;
 790     object->stack_size = INITIAL_STACK_SIZE;
 791     D3DXMatrixIdentity(&object->stack[0]);
 792
 793     TRACE("Created matrix stack %p\n", object);
 794
 795     *ppstack = &object->ID3DXMatrixStack_iface;
 796     return D3D_OK;
 797 }
 798
 799 static inline ID3DXMatrixStackImpl *impl_from_ID3DXMatrixStack(ID3DXMatrixStack *iface)
 800 {
 801   return CONTAINING_RECORD(iface, ID3DXMatrixStackImpl, ID3DXMatrixStack_iface);
 802 }
 803
 804 static HRESULT WINAPI ID3DXMatrixStackImpl_QueryInterface(ID3DXMatrixStack *iface, REFIID riid, void **ppobj)
 805 {
 806     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 807     if (IsEqualGUID(riid, &IID_IUnknown) || IsEqualGUID(riid, &IID_ID3DXMatrixStack))
 808     {
 809      ID3DXMatrixStack_AddRef(iface);
 810      *ppobj = This;
 811      return S_OK;
 812     }
 813     *ppobj = NULL;
 814     WARN("(%p)->(%s,%p), not found\n", This, debugstr_guid(riid), ppobj);
 815     return E_NOINTERFACE;
 816 }
 817
 818 static ULONG WINAPI ID3DXMatrixStackImpl_AddRef(ID3DXMatrixStack *iface)
 819 {
 820     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 821     ULONG ref = InterlockedIncrement(&This->ref);
 822     TRACE("(%p) : AddRef from %d\n", This, ref - 1);
 823     return ref;
 824 }
 825
 826 static ULONG WINAPI ID3DXMatrixStackImpl_Release(ID3DXMatrixStack* iface)
 827 {
 828     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 829     ULONG ref = InterlockedDecrement(&This->ref);
 830     if (!ref)
 831     {
 832         HeapFree(GetProcessHeap(), 0, This->stack);
 833         HeapFree(GetProcessHeap(), 0, This);
 834     }
 835     TRACE("(%p) : ReleaseRef to %d\n", This, ref);
 836     return ref;
 837 }
 838
 839 static D3DXMATRIX* WINAPI ID3DXMatrixStackImpl_GetTop(ID3DXMatrixStack *iface)
 840 {
 841     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 842
 843     TRACE("iface %p\n", iface);
 844
 845     return &This->stack[This->current];
 846 }
 847
 848 static HRESULT WINAPI ID3DXMatrixStackImpl_LoadIdentity(ID3DXMatrixStack *iface)
 849 {
 850     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 851
 852     TRACE("iface %p\n", iface);
 853
 854     D3DXMatrixIdentity(&This->stack[This->current]);
 855
 856     return D3D_OK;
 857 }
 858
 859 static HRESULT WINAPI ID3DXMatrixStackImpl_LoadMatrix(ID3DXMatrixStack *iface, CONST D3DXMATRIX *pm)
 860 {
 861     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 862
 863     TRACE("iface %p\n", iface);
 864
 865     This->stack[This->current] = *pm;
 866
 867     return D3D_OK;
 868 }
 869
 870 static HRESULT WINAPI ID3DXMatrixStackImpl_MultMatrix(ID3DXMatrixStack *iface, CONST D3DXMATRIX *pm)
 871 {
 872     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 873
 874     TRACE("iface %p\n", iface);
 875
 876     D3DXMatrixMultiply(&This->stack[This->current], &This->stack[This->current], pm);
 877
 878     return D3D_OK;
 879 }
 880
 881 static HRESULT WINAPI ID3DXMatrixStackImpl_MultMatrixLocal(ID3DXMatrixStack *iface, CONST D3DXMATRIX *pm)
 882 {
 883     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 884
 885     TRACE("iface %p\n", iface);
 886
 887     D3DXMatrixMultiply(&This->stack[This->current], pm, &This->stack[This->current]);
 888
 889     return D3D_OK;
 890 }
 891
 892 static HRESULT WINAPI ID3DXMatrixStackImpl_Pop(ID3DXMatrixStack *iface)
 893 {
 894     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 895
 896     TRACE("iface %p\n", iface);
 897
 898     /* Popping the last element on the stack returns D3D_OK, but does nothing. */
 899     if (!This->current) return D3D_OK;
 900
 901     if (This->current <= This->stack_size / 4 && This->stack_size >= INITIAL_STACK_SIZE * 2)
 902     {
 903         unsigned int new_size;
 904         D3DXMATRIX *new_stack;
 905
 906         new_size = This->stack_size / 2;
 907         new_stack = HeapReAlloc(GetProcessHeap(), 0, This->stack, new_size * sizeof(D3DXMATRIX));
 908         if (new_stack)
 909         {
 910             This->stack_size = new_size;
 911             This->stack = new_stack;
 912         }
 913     }
 914
 915     --This->current;
 916
 917     return D3D_OK;
 918 }
 919
 920 static HRESULT WINAPI ID3DXMatrixStackImpl_Push(ID3DXMatrixStack *iface)
 921 {
 922     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 923
 924     TRACE("iface %p\n", iface);
 925
 926     if (This->current == This->stack_size - 1)
 927     {
 928         unsigned int new_size;
 929         D3DXMATRIX *new_stack;
 930
 931         if (This->stack_size > UINT_MAX / 2) return E_OUTOFMEMORY;
 932
 933         new_size = This->stack_size * 2;
 934         new_stack = HeapReAlloc(GetProcessHeap(), 0, This->stack, new_size * sizeof(D3DXMATRIX));
 935         if (!new_stack) return E_OUTOFMEMORY;
 936
 937         This->stack_size = new_size;
 938         This->stack = new_stack;
 939     }
 940
 941     ++This->current;
 942     This->stack[This->current] = This->stack[This->current - 1];
 943
 944     return D3D_OK;
 945 }
 946
 947 static HRESULT WINAPI ID3DXMatrixStackImpl_RotateAxis(ID3DXMatrixStack *iface, CONST D3DXVECTOR3 *pv, FLOAT angle)
 948 {
 949     D3DXMATRIX temp;
 950     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 951
 952     TRACE("iface %p\n", iface);
 953
 954     D3DXMatrixRotationAxis(&temp, pv, angle);
 955     D3DXMatrixMultiply(&This->stack[This->current], &This->stack[This->current], &temp);
 956
 957     return D3D_OK;
 958 }
 959
 960 static HRESULT WINAPI ID3DXMatrixStackImpl_RotateAxisLocal(ID3DXMatrixStack *iface, CONST D3DXVECTOR3 *pv, FLOAT angle)
 961 {
 962     D3DXMATRIX temp;
 963     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 964
 965     TRACE("iface %p\n", iface);
 966
 967     D3DXMatrixRotationAxis(&temp, pv, angle);
 968     D3DXMatrixMultiply(&This->stack[This->current], &temp, &This->stack[This->current]);
 969
 970     return D3D_OK;
 971 }
 972
 973 static HRESULT WINAPI ID3DXMatrixStackImpl_RotateYawPitchRoll(ID3DXMatrixStack *iface, FLOAT x, FLOAT y, FLOAT z)
 974 {
 975     D3DXMATRIX temp;
 976     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 977
 978     TRACE("iface %p\n", iface);
 979
 980     D3DXMatrixRotationYawPitchRoll(&temp, x, y, z);
 981     D3DXMatrixMultiply(&This->stack[This->current], &This->stack[This->current], &temp);
 982
 983     return D3D_OK;
 984 }
 985
 986 static HRESULT WINAPI ID3DXMatrixStackImpl_RotateYawPitchRollLocal(ID3DXMatrixStack *iface, FLOAT x, FLOAT y, FLOAT z)
 987 {
 988     D3DXMATRIX temp;
 989     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
 990
 991     TRACE("iface %p\n", iface);
 992
 993     D3DXMatrixRotationYawPitchRoll(&temp, x, y, z);
 994     D3DXMatrixMultiply(&This->stack[This->current], &temp, &This->stack[This->current]);
 995
 996     return D3D_OK;
 997 }
 998
 999 static HRESULT WINAPI ID3DXMatrixStackImpl_Scale(ID3DXMatrixStack *iface, FLOAT x, FLOAT y, FLOAT z)
1000 {
1001     D3DXMATRIX temp;
1002     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
1003
1004     TRACE("iface %p\n", iface);
1005
1006     D3DXMatrixScaling(&temp, x, y, z);
1007     D3DXMatrixMultiply(&This->stack[This->current], &This->stack[This->current], &temp);
1008
1009     return D3D_OK;
1010 }
1011
1012 static HRESULT WINAPI ID3DXMatrixStackImpl_ScaleLocal(ID3DXMatrixStack *iface, FLOAT x, FLOAT y, FLOAT z)
1013 {
1014     D3DXMATRIX temp;
1015     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
1016
1017     TRACE("iface %p\n", iface);
1018
1019     D3DXMatrixScaling(&temp, x, y, z);
1020     D3DXMatrixMultiply(&This->stack[This->current], &temp, &This->stack[This->current]);
1021
1022     return D3D_OK;
1023 }
1024
1025 static HRESULT WINAPI ID3DXMatrixStackImpl_Translate(ID3DXMatrixStack *iface, FLOAT x, FLOAT y, FLOAT z)
1026 {
1027     D3DXMATRIX temp;
1028     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
1029
1030     TRACE("iface %p\n", iface);
1031
1032     D3DXMatrixTranslation(&temp, x, y, z);
1033     D3DXMatrixMultiply(&This->stack[This->current], &This->stack[This->current], &temp);
1034
1035     return D3D_OK;
1036 }
1037
1038 static HRESULT WINAPI ID3DXMatrixStackImpl_TranslateLocal(ID3DXMatrixStack *iface, FLOAT x, FLOAT y, FLOAT z)
1039 {
1040     D3DXMATRIX temp;
1041     ID3DXMatrixStackImpl *This = impl_from_ID3DXMatrixStack(iface);
1042
1043     TRACE("iface %p\n", iface);
1044
1045     D3DXMatrixTranslation(&temp, x, y, z);
1046     D3DXMatrixMultiply(&This->stack[This->current], &temp,&This->stack[This->current]);
1047
1048     return D3D_OK;
1049 }
1050
1051 static const ID3DXMatrixStackVtbl ID3DXMatrixStack_Vtbl =
1052 {
1053     ID3DXMatrixStackImpl_QueryInterface,
1054     ID3DXMatrixStackImpl_AddRef,
1055     ID3DXMatrixStackImpl_Release,
1056     ID3DXMatrixStackImpl_Pop,
1057     ID3DXMatrixStackImpl_Push,
1058     ID3DXMatrixStackImpl_LoadIdentity,
1059     ID3DXMatrixStackImpl_LoadMatrix,
1060     ID3DXMatrixStackImpl_MultMatrix,
1061     ID3DXMatrixStackImpl_MultMatrixLocal,
1062     ID3DXMatrixStackImpl_RotateAxis,
1063     ID3DXMatrixStackImpl_RotateAxisLocal,
1064     ID3DXMatrixStackImpl_RotateYawPitchRoll,
1065     ID3DXMatrixStackImpl_RotateYawPitchRollLocal,
1066     ID3DXMatrixStackImpl_Scale,
1067     ID3DXMatrixStackImpl_ScaleLocal,
1068     ID3DXMatrixStackImpl_Translate,
1069     ID3DXMatrixStackImpl_TranslateLocal,
1070     ID3DXMatrixStackImpl_GetTop
1071 };
1072
1073 /*_________________D3DXPLANE________________*/
1074
1075 D3DXPLANE* WINAPI D3DXPlaneFromPointNormal(D3DXPLANE *pout, CONST D3DXVECTOR3 *pvpoint, CONST D3DXVECTOR3 *pvnormal)
1076 {
1077     pout->a = pvnormal->x;
1078     pout->b = pvnormal->y;
1079     pout->c = pvnormal->z;
1080     pout->d = -D3DXVec3Dot(pvpoint, pvnormal);
1081     return pout;
1082 }
1083
1084 D3DXPLANE* WINAPI D3DXPlaneFromPoints(D3DXPLANE *pout, CONST D3DXVECTOR3 *pv1, CONST D3DXVECTOR3 *pv2, CONST D3DXVECTOR3 *pv3)
1085 {
1086     D3DXVECTOR3 edge1, edge2, normal, Nnormal;
1087
1088     edge1.x = 0.0f; edge1.y = 0.0f; edge1.z = 0.0f;
1089     edge2.x = 0.0f; edge2.y = 0.0f; edge2.z = 0.0f;
1090     D3DXVec3Subtract(&edge1, pv2, pv1);
1091     D3DXVec3Subtract(&edge2, pv3, pv1);
1092     D3DXVec3Cross(&normal, &edge1, &edge2);
1093     D3DXVec3Normalize(&Nnormal, &normal);
1094     D3DXPlaneFromPointNormal(pout, pv1, &Nnormal);
1095     return pout;
1096 }
1097
1098 D3DXVECTOR3* WINAPI D3DXPlaneIntersectLine(D3DXVECTOR3 *pout, CONST D3DXPLANE *pp, CONST D3DXVECTOR3 *pv1, CONST D3DXVECTOR3 *pv2)
1099 {
1100     D3DXVECTOR3 direction, normal;
1101     FLOAT dot, temp;
1102
1103     normal.x = pp->a;
1104     normal.y = pp->b;
1105     normal.z = pp->c;
1106     direction.x = pv2->x - pv1->x;
1107     direction.y = pv2->y - pv1->y;
1108     direction.z = pv2->z - pv1->z;
1109     dot = D3DXVec3Dot(&normal, &direction);
1110     if ( !dot ) return NULL;
1111     temp = ( pp->d + D3DXVec3Dot(&normal, pv1) ) / dot;
1112     pout->x = pv1->x - temp * direction.x;
1113     pout->y = pv1->y - temp * direction.y;
1114     pout->z = pv1->z - temp * direction.z;
1115     return pout;
1116 }
1117
1118 D3DXPLANE* WINAPI D3DXPlaneNormalize(D3DXPLANE *pout, CONST D3DXPLANE *pp)
1119 {
1120     D3DXPLANE out;
1121     FLOAT norm;
1122
1123     norm = sqrt(pp->a * pp->a + pp->b * pp->b + pp->c * pp->c);
1124     if ( norm )
1125     {
1126      out.a = pp->a / norm;
1127      out.b = pp->b / norm;
1128      out.c = pp->c / norm;
1129      out.d = pp->d / norm;
1130     }
1131     else
1132     {
1133      out.a = 0.0f;
1134      out.b = 0.0f;
1135      out.c = 0.0f;
1136      out.d = 0.0f;
1137     }
1138     *pout = out;
1139     return pout;
1140 }
1141
1142 D3DXPLANE* WINAPI D3DXPlaneTransform(D3DXPLANE *pout, CONST D3DXPLANE *pplane, CONST D3DXMATRIX *pm)
1143 {
1144     CONST D3DXPLANE plane = *pplane;
1145     pout->a = pm->u.m[0][0] * plane.a + pm->u.m[1][0] * plane.b + pm->u.m[2][0] * plane.c + pm->u.m[3][0] * plane.d;
1146     pout->b = pm->u.m[0][1] * plane.a + pm->u.m[1][1] * plane.b + pm->u.m[2][1] * plane.c + pm->u.m[3][1] * plane.d;
1147     pout->c = pm->u.m[0][2] * plane.a + pm->u.m[1][2] * plane.b + pm->u.m[2][2] * plane.c + pm->u.m[3][2] * plane.d;
1148     pout->d = pm->u.m[0][3] * plane.a + pm->u.m[1][3] * plane.b + pm->u.m[2][3] * plane.c + pm->u.m[3][3] * plane.d;
1149     return pout;
1150 }
1151
1152 D3DXPLANE* WINAPI D3DXPlaneTransformArray(D3DXPLANE* out, UINT outstride, CONST D3DXPLANE* in, UINT instride, CONST D3DXMATRIX* matrix, UINT elements)
1153 {
1154     UINT i;
1155
1156     for (i = 0; i < elements; ++i) {
1157         D3DXPlaneTransform(
1158             (D3DXPLANE*)((char*)out + outstride * i),
1159             (CONST D3DXPLANE*)((const char*)in + instride * i),
1160             matrix);
1161     }
1162     return out;
1163 }
1164
1165 /*_________________D3DXQUATERNION________________*/
1166
1167 D3DXQUATERNION* WINAPI D3DXQuaternionBaryCentric(D3DXQUATERNION *pout, CONST D3DXQUATERNION *pq1, CONST D3DXQUATERNION *pq2, CONST D3DXQUATERNION *pq3, FLOAT f, FLOAT g)
1168 {
1169     D3DXQUATERNION temp1, temp2;
1170     D3DXQuaternionSlerp(pout, D3DXQuaternionSlerp(&temp1, pq1, pq2, f + g), D3DXQuaternionSlerp(&temp2, pq1, pq3, f+g), g / (f + g));
1171     return pout;
1172 }
1173
1174 D3DXQUATERNION* WINAPI D3DXQuaternionExp(D3DXQUATERNION *pout, CONST D3DXQUATERNION *pq)
1175 {
1176     FLOAT norm;
1177
1178     norm = sqrt(pq->x * pq->x + pq->y * pq->y + pq->z * pq->z);
1179     if (norm )
1180     {
1181      pout->x = sin(norm) * pq->x / norm;
1182      pout->y = sin(norm) * pq->y / norm;
1183      pout->z = sin(norm) * pq->z / norm;
1184      pout->w = cos(norm);
1185     }
1186     else
1187     {
1188      pout->x = 0.0f;
1189      pout->y = 0.0f;
1190      pout->z = 0.0f;
1191      pout->w = 1.0f;
1192     }
1193     return pout;
1194 }
1195
1196 D3DXQUATERNION* WINAPI D3DXQuaternionInverse(D3DXQUATERNION *pout, CONST D3DXQUATERNION *pq)
1197 {
1198     D3DXQUATERNION out;
1199     FLOAT norm;
1200
1201     norm = D3DXQuaternionLengthSq(pq);
1202
1203     out.x = -pq->x / norm;
1204     out.y = -pq->y / norm;
1205     out.z = -pq->z / norm;
1206     out.w = pq->w / norm;
1207
1208     *pout =out;
1209     return pout;
1210 }
1211
1212 D3DXQUATERNION* WINAPI D3DXQuaternionLn(D3DXQUATERNION *pout, CONST D3DXQUATERNION *pq)
1213 {
1214     FLOAT norm, normvec, theta;
1215
1216     norm = D3DXQuaternionLengthSq(pq);
1217     if ( norm > 1.0001f )
1218     {
1219      pout->x = pq->x;
1220      pout->y = pq->y;
1221      pout->z = pq->z;
1222      pout->w = 0.0f;
1223     }
1224     else if( norm > 0.99999f)
1225     {
1226      normvec = sqrt( pq->x * pq->x + pq->y * pq->y + pq->z * pq->z );
1227      theta = atan2(normvec, pq->w) / normvec;
1228      pout->x = theta * pq->x;
1229      pout->y = theta * pq->y;
1230      pout->z = theta * pq->z;
1231      pout->w = 0.0f;
1232     }
1233     else
1234     {
1235      FIXME("The quaternion (%f, %f, %f, %f) has a norm <1. This should not happen. Windows returns a result anyway. This case is not implemented yet.\n", pq->x, pq->y, pq->z, pq->w);
1236     }
1237     return pout;
1238 }
1239
1240 D3DXQUATERNION* WINAPI D3DXQuaternionMultiply(D3DXQUATERNION *pout, CONST D3DXQUATERNION *pq1, CONST D3DXQUATERNION *pq2)
1241 {
1242     D3DXQUATERNION out;
1243     out.x = pq2->w * pq1->x + pq2->x * pq1->w + pq2->y * pq1->z - pq2->z * pq1->y;
1244     out.y = pq2->w * pq1->y - pq2->x * pq1->z + pq2->y * pq1->w + pq2->z * pq1->x;
1245     out.z = pq2->w * pq1->z + pq2->x * pq1->y - pq2->y * pq1->x + pq2->z * pq1->w;
1246     out.w = pq2->w * pq1->w - pq2->x * pq1->x - pq2->y * pq1->y - pq2->z * pq1->z;
1247     *pout = out;
1248     return pout;
1249 }
1250
1251 D3DXQUATERNION* WINAPI D3DXQuaternionNormalize(D3DXQUATERNION *pout, CONST D3DXQUATERNION *pq)
1252 {
1253     D3DXQUATERNION out;
1254     FLOAT norm;
1255
1256     norm = D3DXQuaternionLength(pq);
1257
1258     out.x = pq->x / norm;
1259     out.y = pq->y / norm;
1260     out.z = pq->z / norm;
1261     out.w = pq->w / norm;
1262
1263     *pout=out;
1264
1265     return pout;
1266 }
1267
1268 D3DXQUATERNION* WINAPI D3DXQuaternionRotationAxis(D3DXQUATERNION *pout, CONST D3DXVECTOR3 *pv, FLOAT angle)
1269 {
1270     D3DXVECTOR3 temp;
1271
1272     D3DXVec3Normalize(&temp, pv);
1273     pout->x = sin( angle / 2.0f ) * temp.x;
1274     pout->y = sin( angle / 2.0f ) * temp.y;
1275     pout->z = sin( angle / 2.0f ) * temp.z;
1276     pout->w = cos( angle / 2.0f );
1277     return pout;
1278 }
1279
1280 D3DXQUATERNION* WINAPI D3DXQuaternionRotationMatrix(D3DXQUATERNION *pout, CONST D3DXMATRIX *pm)
1281 {
1282     int i, maxi;
1283     FLOAT maxdiag, S, trace;
1284
1285     trace = pm->u.m[0][0] + pm->u.m[1][1] + pm->u.m[2][2] + 1.0f;
1286     if ( trace > 1.0f)
1287     {
1288      pout->x = ( pm->u.m[1][2] - pm->u.m[2][1] ) / ( 2.0f * sqrt(trace) );
1289      pout->y = ( pm->u.m[2][0] - pm->u.m[0][2] ) / ( 2.0f * sqrt(trace) );
1290      pout->z = ( pm->u.m[0][1] - pm->u.m[1][0] ) / ( 2.0f * sqrt(trace) );
1291      pout->w = sqrt(trace) / 2.0f;
1292      return pout;
1293      }
1294     maxi = 0;
1295     maxdiag = pm->u.m[0][0];
1296     for (i=1; i<3; i++)
1297     {
1298      if ( pm->u.m[i][i] > maxdiag )
1299      {
1300       maxi = i;
1301       maxdiag = pm->u.m[i][i];
1302      }
1303     }
1304     switch( maxi )
1305     {
1306      case 0:
1307        S = 2.0f * sqrt(1.0f + pm->u.m[0][0] - pm->u.m[1][1] - pm->u.m[2][2]);
1308        pout->x = 0.25f * S;
1309        pout->y = ( pm->u.m[0][1] + pm->u.m[1][0] ) / S;
1310        pout->z = ( pm->u.m[0][2] + pm->u.m[2][0] ) / S;
1311        pout->w = ( pm->u.m[1][2] - pm->u.m[2][1] ) / S;
1312      break;
1313      case 1:
1314        S = 2.0f * sqrt(1.0f + pm->u.m[1][1] - pm->u.m[0][0] - pm->u.m[2][2]);
1315        pout->x = ( pm->u.m[0][1] + pm->u.m[1][0] ) / S;
1316        pout->y = 0.25f * S;
1317        pout->z = ( pm->u.m[1][2] + pm->u.m[2][1] ) / S;
1318        pout->w = ( pm->u.m[2][0] - pm->u.m[0][2] ) / S;
1319      break;
1320      case 2:
1321        S = 2.0f * sqrt(1.0f + pm->u.m[2][2] - pm->u.m[0][0] - pm->u.m[1][1]);
1322        pout->x = ( pm->u.m[0][2] + pm->u.m[2][0] ) / S;
1323        pout->y = ( pm->u.m[1][2] + pm->u.m[2][1] ) / S;
1324        pout->z = 0.25f * S;
1325        pout->w = ( pm->u.m[0][1] - pm->u.m[1][0] ) / S;
1326      break;
1327     }
1328     return pout;
1329 }
1330
1331 D3DXQUATERNION* WINAPI D3DXQuaternionRotationYawPitchRoll(D3DXQUATERNION *pout, FLOAT yaw, FLOAT pitch, FLOAT roll)
1332 {
1333     pout->x = sin( yaw / 2.0f) * cos(pitch / 2.0f) * sin(roll / 2.0f) + cos(yaw / 2.0f) * sin(pitch / 2.0f) * cos(roll / 2.0f);
1334     pout->y = sin( yaw / 2.0f) * cos(pitch / 2.0f) * cos(roll / 2.0f) - cos(yaw / 2.0f) * sin(pitch / 2.0f) * sin(roll / 2.0f);
1335     pout->z = cos(yaw / 2.0f) * cos(pitch / 2.0f) * sin(roll / 2.0f) - sin( yaw / 2.0f) * sin(pitch / 2.0f) * cos(roll / 2.0f);
1336     pout->w = cos( yaw / 2.0f) * cos(pitch / 2.0f) * cos(roll / 2.0f) + sin(yaw / 2.0f) * sin(pitch / 2.0f) * sin(roll / 2.0f);
1337     return pout;
1338 }
1339
1340 D3DXQUATERNION* WINAPI D3DXQuaternionSlerp(D3DXQUATERNION *pout, CONST D3DXQUATERNION *pq1, CONST D3DXQUATERNION *pq2, FLOAT t)
1341 {
1342     FLOAT dot, epsilon, temp, theta, u;
1343
1344     epsilon = 1.0f;
1345     temp = 1.0f - t;
1346     u = t;
1347     dot = D3DXQuaternionDot(pq1, pq2);
1348     if ( dot < 0.0f )
1349     {
1350         epsilon = -1.0f;
1351         dot = -dot;
1352     }
1353     if( 1.0f - dot > 0.001f )
1354     {
1355         theta = acos(dot);
1356         temp  = sin(theta * temp) / sin(theta);
1357         u = sin(theta * u) / sin(theta);
1358     }
1359     pout->x = temp * pq1->x + epsilon * u * pq2->x;
1360     pout->y = temp * pq1->y + epsilon * u * pq2->y;
1361     pout->z = temp * pq1->z + epsilon * u * pq2->z;
1362     pout->w = temp * pq1->w + epsilon * u * pq2->w;
1363     return pout;
1364 }
1365
1366 D3DXQUATERNION* WINAPI D3DXQuaternionSquad(D3DXQUATERNION *pout, CONST D3DXQUATERNION *pq1, CONST D3DXQUATERNION *pq2, CONST D3DXQUATERNION *pq3, CONST D3DXQUATERNION *pq4, FLOAT t)
1367 {
1368     D3DXQUATERNION temp1, temp2;
1369
1370     D3DXQuaternionSlerp(pout, D3DXQuaternionSlerp(&temp1, pq1, pq4, t), D3DXQuaternionSlerp(&temp2, pq2, pq3, t), 2.0f * t * (1.0f - t));
1371     return pout;
1372 }
1373
1374 void WINAPI D3DXQuaternionToAxisAngle(CONST D3DXQUATERNION *pq, D3DXVECTOR3 *paxis, FLOAT *pangle)
1375 {
1376     paxis->x = pq->x;
1377     paxis->y = pq->y;
1378     paxis->z = pq->z;
1379     *pangle = 2.0f * acos(pq->w);
1380 }
1381
1382 /*_________________D3DXVec2_____________________*/
1383
1384 D3DXVECTOR2* WINAPI D3DXVec2BaryCentric(D3DXVECTOR2 *pout, CONST D3DXVECTOR2 *pv1, CONST D3DXVECTOR2 *pv2, CONST D3DXVECTOR2 *pv3, FLOAT f, FLOAT g)
1385 {
1386     pout->x = (1.0f-f-g) * (pv1->x) + f * (pv2->x) + g * (pv3->x);
1387     pout->y = (1.0f-f-g) * (pv1->y) + f * (pv2->y) + g * (pv3->y);
1388     return pout;
1389 }
1390
1391 D3DXVECTOR2* WINAPI D3DXVec2CatmullRom(D3DXVECTOR2 *pout, CONST D3DXVECTOR2 *pv0, CONST D3DXVECTOR2 *pv1, CONST D3DXVECTOR2 *pv2, CONST D3DXVECTOR2 *pv3, FLOAT s)
1392 {
1393     pout->x = 0.5f * (2.0f * pv1->x + (pv2->x - pv0->x) *s + (2.0f *pv0->x - 5.0f * pv1->x + 4.0f * pv2->x - pv3->x) * s * s + (pv3->x -3.0f * pv2->x + 3.0f * pv1->x - pv0->x) * s * s * s);
1394     pout->y = 0.5f * (2.0f * pv1->y + (pv2->y - pv0->y) *s + (2.0f *pv0->y - 5.0f * pv1->y + 4.0f * pv2->y - pv3->y) * s * s + (pv3->y -3.0f * pv2->y + 3.0f * pv1->y - pv0->y) * s * s * s);
1395     return pout;
1396 }
1397
1398 D3DXVECTOR2* WINAPI D3DXVec2Hermite(D3DXVECTOR2 *pout, CONST D3DXVECTOR2 *pv1, CONST D3DXVECTOR2 *pt1, CONST D3DXVECTOR2 *pv2, CONST D3DXVECTOR2 *pt2, FLOAT s)
1399 {
1400     FLOAT h1, h2, h3, h4;
1401
1402     h1 = 2.0f * s * s * s - 3.0f * s * s + 1.0f;
1403     h2 = s * s * s - 2.0f * s * s + s;
1404     h3 = -2.0f * s * s * s + 3.0f * s * s;
1405     h4 = s * s * s - s * s;
1406
1407     pout->x = h1 * (pv1->x) + h2 * (pt1->x) + h3 * (pv2->x) + h4 * (pt2->x);
1408     pout->y = h1 * (pv1->y) + h2 * (pt1->y) + h3 * (pv2->y) + h4 * (pt2->y);
1409     return pout;
1410 }
1411
1412 D3DXVECTOR2* WINAPI D3DXVec2Normalize(D3DXVECTOR2 *pout, CONST D3DXVECTOR2 *pv)
1413 {
1414     D3DXVECTOR2 out;
1415     FLOAT norm;
1416
1417     norm = D3DXVec2Length(pv);
1418     if ( !norm )
1419     {
1420      out.x = 0.0f;
1421      out.y = 0.0f;
1422     }
1423     else
1424     {
1425      out.x = pv->x / norm;
1426      out.y = pv->y / norm;
1427     }
1428     *pout=out;
1429     return pout;
1430 }
1431
1432 D3DXVECTOR4* WINAPI D3DXVec2Transform(D3DXVECTOR4 *pout, CONST D3DXVECTOR2 *pv, CONST D3DXMATRIX *pm)
1433 {
1434     pout->x = pm->u.m[0][0] * pv->x + pm->u.m[1][0] * pv->y  + pm->u.m[3][0];
1435     pout->y = pm->u.m[0][1] * pv->x + pm->u.m[1][1] * pv->y  + pm->u.m[3][1];
1436     pout->z = pm->u.m[0][2] * pv->x + pm->u.m[1][2] * pv->y  + pm->u.m[3][2];
1437     pout->w = pm->u.m[0][3] * pv->x + pm->u.m[1][3] * pv->y  + pm->u.m[3][3];
1438     return pout;
1439 }
1440
1441 D3DXVECTOR4* WINAPI D3DXVec2TransformArray(D3DXVECTOR4* out, UINT outstride, CONST D3DXVECTOR2* in, UINT instride, CONST D3DXMATRIX* matrix, UINT elements)
1442 {
1443     UINT i;
1444
1445     for (i = 0; i < elements; ++i) {
1446         D3DXVec2Transform(
1447             (D3DXVECTOR4*)((char*)out + outstride * i),
1448             (CONST D3DXVECTOR2*)((const char*)in + instride * i),
1449             matrix);
1450     }
1451     return out;
1452 }
1453
1454 D3DXVECTOR2* WINAPI D3DXVec2TransformCoord(D3DXVECTOR2 *pout, CONST D3DXVECTOR2 *pv, CONST D3DXMATRIX *pm)
1455 {
1456     D3DXVECTOR2 v;
1457     FLOAT norm;
1458
1459     v = *pv;
1460     norm = pm->u.m[0][3] * pv->x + pm->u.m[1][3] * pv->y + pm->u.m[3][3];
1461
1462     pout->x = (pm->u.m[0][0] * v.x + pm->u.m[1][0] * v.y + pm->u.m[3][0]) / norm;
1463     pout->y = (pm->u.m[0][1] * v.x + pm->u.m[1][1] * v.y + pm->u.m[3][1]) / norm;
1464
1465     return pout;
1466 }
1467
1468 D3DXVECTOR2* WINAPI D3DXVec2TransformCoordArray(D3DXVECTOR2* out, UINT outstride, CONST D3DXVECTOR2* in, UINT instride, CONST D3DXMATRIX* matrix, UINT elements)
1469 {
1470     UINT i;
1471
1472     for (i = 0; i < elements; ++i) {
1473         D3DXVec2TransformCoord(
1474             (D3DXVECTOR2*)((char*)out + outstride * i),
1475             (CONST D3DXVECTOR2*)((const char*)in + instride * i),
1476             matrix);
1477     }
1478     return out;
1479 }
1480
1481 D3DXVECTOR2* WINAPI D3DXVec2TransformNormal(D3DXVECTOR2 *pout, CONST D3DXVECTOR2 *pv, CONST D3DXMATRIX *pm)
1482 {
1483     CONST D3DXVECTOR2 v = *pv;
1484     pout->x = pm->u.m[0][0] * v.x + pm->u.m[1][0] * v.y;
1485     pout->y = pm->u.m[0][1] * v.x + pm->u.m[1][1] * v.y;
1486     return pout;
1487 }
1488
1489 D3DXVECTOR2* WINAPI D3DXVec2TransformNormalArray(D3DXVECTOR2* out, UINT outstride, CONST D3DXVECTOR2 *in, UINT instride, CONST D3DXMATRIX *matrix, UINT elements)
1490 {
1491     UINT i;
1492
1493     for (i = 0; i < elements; ++i) {
1494         D3DXVec2TransformNormal(
1495             (D3DXVECTOR2*)((char*)out + outstride * i),
1496             (CONST D3DXVECTOR2*)((const char*)in + instride * i),
1497             matrix);
1498     }
1499     return out;
1500 }
1501
1502 /*_________________D3DXVec3_____________________*/
1503
1504 D3DXVECTOR3* WINAPI D3DXVec3BaryCentric(D3DXVECTOR3 *pout, CONST D3DXVECTOR3 *pv1, CONST D3DXVECTOR3 *pv2, CONST D3DXVECTOR3 *pv3, FLOAT f, FLOAT g)
1505 {
1506     pout->x = (1.0f-f-g) * (pv1->x) + f * (pv2->x) + g * (pv3->x);
1507     pout->y = (1.0f-f-g) * (pv1->y) + f * (pv2->y) + g * (pv3->y);
1508     pout->z = (1.0f-f-g) * (pv1->z) + f * (pv2->z) + g * (pv3->z);
1509     return pout;
1510 }
1511
1512 D3DXVECTOR3* WINAPI D3DXVec3CatmullRom( D3DXVECTOR3 *pout, CONST D3DXVECTOR3 *pv0, CONST D3DXVECTOR3 *pv1, CONST D3DXVECTOR3 *pv2, CONST D3DXVECTOR3 *pv3, FLOAT s)
1513 {
1514     pout->x = 0.5f * (2.0f * pv1->x + (pv2->x - pv0->x) *s + (2.0f *pv0->x - 5.0f * pv1->x + 4.0f * pv2->x - pv3->x) * s * s + (pv3->x -3.0f * pv2->x + 3.0f * pv1->x - pv0->x) * s * s * s);
1515     pout->y = 0.5f * (2.0f * pv1->y + (pv2->y - pv0->y) *s + (2.0f *pv0->y - 5.0f * pv1->y + 4.0f * pv2->y - pv3->y) * s * s + (pv3->y -3.0f * pv2->y + 3.0f * pv1->y - pv0->y) * s * s * s);
1516     pout->z = 0.5f * (2.0f * pv1->z + (pv2->z - pv0->z) *s + (2.0f *pv0->z - 5.0f * pv1->z + 4.0f * pv2->z - pv3->z) * s * s + (pv3->z -3.0f * pv2->z + 3.0f * pv1->z - pv0->z) * s * s * s);
1517     return pout;
1518 }
1519
1520 D3DXVECTOR3* WINAPI D3DXVec3Hermite(D3DXVECTOR3 *pout, CONST D3DXVECTOR3 *pv1, CONST D3DXVECTOR3 *pt1, CONST D3DXVECTOR3 *pv2, CONST D3DXVECTOR3 *pt2, FLOAT s)
1521 {
1522     FLOAT h1, h2, h3, h4;
1523
1524     h1 = 2.0f * s * s * s - 3.0f * s * s + 1.0f;
1525     h2 = s * s * s - 2.0f * s * s + s;
1526     h3 = -2.0f * s * s * s + 3.0f * s * s;
1527     h4 = s * s * s - s * s;
1528
1529     pout->x = h1 * (pv1->x) + h2 * (pt1->x) + h3 * (pv2->x) + h4 * (pt2->x);
1530     pout->y = h1 * (pv1->y) + h2 * (pt1->y) + h3 * (pv2->y) + h4 * (pt2->y);
1531     pout->z = h1 * (pv1->z) + h2 * (pt1->z) + h3 * (pv2->z) + h4 * (pt2->z);
1532     return pout;
1533 }
1534
1535 D3DXVECTOR3* WINAPI D3DXVec3Normalize(D3DXVECTOR3 *pout, CONST D3DXVECTOR3 *pv)
1536 {
1537     D3DXVECTOR3 out;
1538     FLOAT norm;
1539
1540     norm = D3DXVec3Length(pv);
1541     if ( !norm )
1542     {
1543      out.x = 0.0f;
1544      out.y = 0.0f;
1545      out.z = 0.0f;
1546     }
1547     else
1548     {
1549      out.x = pv->x / norm;
1550      out.y = pv->y / norm;
1551      out.z = pv->z / norm;
1552     }
1553     *pout = out;
1554     return pout;
1555 }
1556
1557 D3DXVECTOR3* WINAPI D3DXVec3Project(D3DXVECTOR3 *pout, CONST D3DXVECTOR3 *pv, CONST D3DVIEWPORT9 *pviewport, CONST D3DXMATRIX *pprojection, CONST D3DXMATRIX *pview, CONST D3DXMATRIX *pworld)
1558 {
1559     D3DXMATRIX m;
1560     D3DXVECTOR3 out;
1561
1562     D3DXMatrixMultiply(&m, pworld, pview);
1563     D3DXMatrixMultiply(&m, &m, pprojection);
1564     D3DXVec3TransformCoord(&out, pv, &m);
1565     out.x = pviewport->X +  ( 1.0f + out.x ) * pviewport->Width / 2.0f;
1566     out.y = pviewport->Y +  ( 1.0f - out.y ) * pviewport->Height / 2.0f;
1567     out.z = pviewport->MinZ + out.z * ( pviewport->MaxZ - pviewport->MinZ );
1568     *pout = out;
1569     return pout;
1570 }
1571
1572 D3DXVECTOR3* WINAPI D3DXVec3ProjectArray(D3DXVECTOR3* out, UINT outstride, CONST D3DXVECTOR3* in, UINT instride, CONST D3DVIEWPORT9* viewport, CONST D3DXMATRIX* projection, CONST D3DXMATRIX* view, CONST D3DXMATRIX* world, UINT elements)
1573 {
1574     UINT i;
1575
1576     for (i = 0; i < elements; ++i) {
1577         D3DXVec3Project(
1578             (D3DXVECTOR3*)((char*)out + outstride * i),
1579             (CONST D3DXVECTOR3*)((const char*)in + instride * i),
1580             viewport, projection, view, world);
1581     }
1582     return out;
1583 }
1584
1585 D3DXVECTOR4* WINAPI D3DXVec3Transform(D3DXVECTOR4 *pout, CONST D3DXVECTOR3 *pv, CONST D3DXMATRIX *pm)
1586 {
1587     pout->x = pm->u.m[0][0] * pv->x + pm->u.m[1][0] * pv->y + pm->u.m[2][0] * pv->z + pm->u.m[3][0];
1588     pout->y = pm->u.m[0][1] * pv->x + pm->u.m[1][1] * pv->y + pm->u.m[2][1] * pv->z + pm->u.m[3][1];
1589     pout->z = pm->u.m[0][2] * pv->x + pm->u.m[1][2] * pv->y + pm->u.m[2][2] * pv->z + pm->u.m[3][2];
1590     pout->w = pm->u.m[0][3] * pv->x + pm->u.m[1][3] * pv->y + pm->u.m[2][3] * pv->z + pm->u.m[3][3];
1591     return pout;
1592 }
1593
1594 D3DXVECTOR4* WINAPI D3DXVec3TransformArray(D3DXVECTOR4* out, UINT outstride, CONST D3DXVECTOR3* in, UINT instride, CONST D3DXMATRIX* matrix, UINT elements)
1595 {
1596     UINT i;
1597
1598     for (i = 0; i < elements; ++i) {
1599         D3DXVec3Transform(
1600             (D3DXVECTOR4*)((char*)out + outstride * i),
1601             (CONST D3DXVECTOR3*)((const char*)in + instride * i),
1602             matrix);
1603     }
1604     return out;
1605 }
1606
1607 D3DXVECTOR3* WINAPI D3DXVec3TransformCoord(D3DXVECTOR3 *pout, CONST D3DXVECTOR3 *pv, CONST D3DXMATRIX *pm)
1608 {
1609     D3DXVECTOR3 out;
1610     FLOAT norm;
1611
1612     norm = pm->u.m[0][3] * pv->x + pm->u.m[1][3] * pv->y + pm->u.m[2][3] *pv->z + pm->u.m[3][3];
1613
1614     out.x = (pm->u.m[0][0] * pv->x + pm->u.m[1][0] * pv->y + pm->u.m[2][0] * pv->z + pm->u.m[3][0]) / norm;
1615     out.y = (pm->u.m[0][1] * pv->x + pm->u.m[1][1] * pv->y + pm->u.m[2][1] * pv->z + pm->u.m[3][1]) / norm;
1616     out.z = (pm->u.m[0][2] * pv->x + pm->u.m[1][2] * pv->y + pm->u.m[2][2] * pv->z + pm->u.m[3][2]) / norm;
1617
1618     *pout = out;
1619
1620     return pout;
1621 }
1622
1623 D3DXVECTOR3* WINAPI D3DXVec3TransformCoordArray(D3DXVECTOR3* out, UINT outstride, CONST D3DXVECTOR3* in, UINT instride, CONST D3DXMATRIX* matrix, UINT elements)
1624 {
1625     UINT i;
1626
1627     for (i = 0; i < elements; ++i) {
1628         D3DXVec3TransformCoord(
1629             (D3DXVECTOR3*)((char*)out + outstride * i),
1630             (CONST D3DXVECTOR3*)((const char*)in + instride * i),
1631             matrix);
1632     }
1633     return out;
1634 }
1635
1636 D3DXVECTOR3* WINAPI D3DXVec3TransformNormal(D3DXVECTOR3 *pout, CONST D3DXVECTOR3 *pv, CONST D3DXMATRIX *pm)
1637 {
1638     CONST D3DXVECTOR3 v = *pv;
1639     pout->x = pm->u.m[0][0] * v.x + pm->u.m[1][0] * v.y + pm->u.m[2][0] * v.z;
1640     pout->y = pm->u.m[0][1] * v.x + pm->u.m[1][1] * v.y + pm->u.m[2][1] * v.z;
1641     pout->z = pm->u.m[0][2] * v.x + pm->u.m[1][2] * v.y + pm->u.m[2][2] * v.z;
1642     return pout;
1643
1644 }
1645
1646 D3DXVECTOR3* WINAPI D3DXVec3TransformNormalArray(D3DXVECTOR3* out, UINT outstride, CONST D3DXVECTOR3* in, UINT instride, CONST D3DXMATRIX* matrix, UINT elements)
1647 {
1648     UINT i;
1649
1650     for (i = 0; i < elements; ++i) {
1651         D3DXVec3TransformNormal(
1652             (D3DXVECTOR3*)((char*)out + outstride * i),
1653             (CONST D3DXVECTOR3*)((const char*)in + instride * i),
1654             matrix);
1655     }
1656     return out;
1657 }
1658
1659 D3DXVECTOR3* WINAPI D3DXVec3Unproject(D3DXVECTOR3 *pout, CONST D3DXVECTOR3 *pv, CONST D3DVIEWPORT9 *pviewport, CONST D3DXMATRIX *pprojection, CONST D3DXMATRIX *pview, CONST D3DXMATRIX *pworld)
1660 {
1661     D3DXMATRIX m;
1662     D3DXVECTOR3 out;
1663
1664     D3DXMatrixMultiply(&m, pworld, pview);
1665     D3DXMatrixMultiply(&m, &m, pprojection);
1666     D3DXMatrixInverse(&m, NULL, &m);
1667     out.x = 2.0f * ( pv->x - pviewport->X ) / pviewport->Width - 1.0f;
1668     out.y = 1.0f - 2.0f * ( pv->y - pviewport->Y ) / pviewport->Height;
1669     out.z = ( pv->z - pviewport->MinZ) / ( pviewport->MaxZ - pviewport->MinZ );
1670     D3DXVec3TransformCoord(&out, &out, &m);
1671     *pout = out;
1672     return pout;
1673 }
1674
1675 D3DXVECTOR3* WINAPI D3DXVec3UnprojectArray(D3DXVECTOR3* out, UINT outstride, CONST D3DXVECTOR3* in, UINT instride, CONST D3DVIEWPORT9* viewport, CONST D3DXMATRIX* projection, CONST D3DXMATRIX* view, CONST D3DXMATRIX* world, UINT elements)
1676 {
1677     UINT i;
1678
1679     for (i = 0; i < elements; ++i) {
1680         D3DXVec3Unproject(
1681             (D3DXVECTOR3*)((char*)out + outstride * i),
1682             (CONST D3DXVECTOR3*)((const char*)in + instride * i),
1683             viewport, projection, view, world);
1684     }
1685     return out;
1686 }
1687
1688 /*_________________D3DXVec4_____________________*/
1689
1690 D3DXVECTOR4* WINAPI D3DXVec4BaryCentric(D3DXVECTOR4 *pout, CONST D3DXVECTOR4 *pv1, CONST D3DXVECTOR4 *pv2, CONST D3DXVECTOR4 *pv3, FLOAT f, FLOAT g)
1691 {
1692     pout->x = (1.0f-f-g) * (pv1->x) + f * (pv2->x) + g * (pv3->x);
1693     pout->y = (1.0f-f-g) * (pv1->y) + f * (pv2->y) + g * (pv3->y);
1694     pout->z = (1.0f-f-g) * (pv1->z) + f * (pv2->z) + g * (pv3->z);
1695     pout->w = (1.0f-f-g) * (pv1->w) + f * (pv2->w) + g * (pv3->w);
1696     return pout;
1697 }
1698
1699 D3DXVECTOR4* WINAPI D3DXVec4CatmullRom(D3DXVECTOR4 *pout, CONST D3DXVECTOR4 *pv0, CONST D3DXVECTOR4 *pv1, CONST D3DXVECTOR4 *pv2, CONST D3DXVECTOR4 *pv3, FLOAT s)
1700 {
1701     pout->x = 0.5f * (2.0f * pv1->x + (pv2->x - pv0->x) *s + (2.0f *pv0->x - 5.0f * pv1->x + 4.0f * pv2->x - pv3->x) * s * s + (pv3->x -3.0f * pv2->x + 3.0f * pv1->x - pv0->x) * s * s * s);
1702     pout->y = 0.5f * (2.0f * pv1->y + (pv2->y - pv0->y) *s + (2.0f *pv0->y - 5.0f * pv1->y + 4.0f * pv2->y - pv3->y) * s * s + (pv3->y -3.0f * pv2->y + 3.0f * pv1->y - pv0->y) * s * s * s);
1703     pout->z = 0.5f * (2.0f * pv1->z + (pv2->z - pv0->z) *s + (2.0f *pv0->z - 5.0f * pv1->z + 4.0f * pv2->z - pv3->z) * s * s + (pv3->z -3.0f * pv2->z + 3.0f * pv1->z - pv0->z) * s * s * s);
1704     pout->w = 0.5f * (2.0f * pv1->w + (pv2->w - pv0->w) *s + (2.0f *pv0->w - 5.0f * pv1->w + 4.0f * pv2->w - pv3->w) * s * s + (pv3->w -3.0f * pv2->w + 3.0f * pv1->w - pv0->w) * s * s * s);
1705     return pout;
1706 }
1707
1708 D3DXVECTOR4* WINAPI D3DXVec4Cross(D3DXVECTOR4 *pout, CONST D3DXVECTOR4 *pv1, CONST D3DXVECTOR4 *pv2, CONST D3DXVECTOR4 *pv3)
1709 {
1710     D3DXVECTOR4 out;
1711     out.x = pv1->y * (pv2->z * pv3->w - pv3->z * pv2->w) - pv1->z * (pv2->y * pv3->w - pv3->y * pv2->w) + pv1->w * (pv2->y * pv3->z - pv2->z *pv3->y);
1712     out.y = -(pv1->x * (pv2->z * pv3->w - pv3->z * pv2->w) - pv1->z * (pv2->x * pv3->w - pv3->x * pv2->w) + pv1->w * (pv2->x * pv3->z - pv3->x * pv2->z));
1713     out.z = pv1->x * (pv2->y * pv3->w - pv3->y * pv2->w) - pv1->y * (pv2->x *pv3->w - pv3->x * pv2->w) + pv1->w * (pv2->x * pv3->y - pv3->x * pv2->y);
1714     out.w = -(pv1->x * (pv2->y * pv3->z - pv3->y * pv2->z) - pv1->y * (pv2->x * pv3->z - pv3->x *pv2->z) + pv1->z * (pv2->x * pv3->y - pv3->x * pv2->y));
1715     *pout = out;
1716     return pout;
1717 }
1718
1719 D3DXVECTOR4* WINAPI D3DXVec4Hermite(D3DXVECTOR4 *pout, CONST D3DXVECTOR4 *pv1, CONST D3DXVECTOR4 *pt1, CONST D3DXVECTOR4 *pv2, CONST D3DXVECTOR4 *pt2, FLOAT s)
1720 {
1721     FLOAT h1, h2, h3, h4;
1722
1723     h1 = 2.0f * s * s * s - 3.0f * s * s + 1.0f;
1724     h2 = s * s * s - 2.0f * s * s + s;
1725     h3 = -2.0f * s * s * s + 3.0f * s * s;
1726     h4 = s * s * s - s * s;
1727
1728     pout->x = h1 * (pv1->x) + h2 * (pt1->x) + h3 * (pv2->x) + h4 * (pt2->x);
1729     pout->y = h1 * (pv1->y) + h2 * (pt1->y) + h3 * (pv2->y) + h4 * (pt2->y);
1730     pout->z = h1 * (pv1->z) + h2 * (pt1->z) + h3 * (pv2->z) + h4 * (pt2->z);
1731     pout->w = h1 * (pv1->w) + h2 * (pt1->w) + h3 * (pv2->w) + h4 * (pt2->w);
1732     return pout;
1733 }
1734
1735 D3DXVECTOR4* WINAPI D3DXVec4Normalize(D3DXVECTOR4 *pout, CONST D3DXVECTOR4 *pv)
1736 {
1737     D3DXVECTOR4 out;
1738     FLOAT norm;
1739
1740     norm = D3DXVec4Length(pv);
1741
1742     out.x = pv->x / norm;
1743     out.y = pv->y / norm;
1744     out.z = pv->z / norm;
1745     out.w = pv->w / norm;
1746
1747     *pout = out;
1748     return pout;
1749 }
1750
1751 D3DXVECTOR4* WINAPI D3DXVec4Transform(D3DXVECTOR4 *pout, CONST D3DXVECTOR4 *pv, CONST D3DXMATRIX *pm)
1752 {
1753     D3DXVECTOR4 out;
1754     out.x = pm->u.m[0][0] * pv->x + pm->u.m[1][0] * pv->y + pm->u.m[2][0] * pv->z + pm->u.m[3][0] * pv->w;
1755     out.y = pm->u.m[0][1] * pv->x + pm->u.m[1][1] * pv->y + pm->u.m[2][1] * pv->z + pm->u.m[3][1] * pv->w;
1756     out.z = pm->u.m[0][2] * pv->x + pm->u.m[1][2] * pv->y + pm->u.m[2][2] * pv->z + pm->u.m[3][2] * pv->w;
1757     out.w = pm->u.m[0][3] * pv->x + pm->u.m[1][3] * pv->y + pm->u.m[2][3] * pv->z + pm->u.m[3][3] * pv->w;
1758     *pout = out;
1759     return pout;
1760 }
1761
1762 D3DXVECTOR4* WINAPI D3DXVec4TransformArray(D3DXVECTOR4* out, UINT outstride, CONST D3DXVECTOR4* in, UINT instride, CONST D3DXMATRIX* matrix, UINT elements)
1763 {
1764     UINT i;
1765
1766     for (i = 0; i < elements; ++i) {
1767         D3DXVec4Transform(
1768             (D3DXVECTOR4*)((char*)out + outstride * i),
1769             (CONST D3DXVECTOR4*)((const char*)in + instride * i),
1770             matrix);
1771     }
1772     return out;
1773 }
1774
1775 static inline unsigned short float_32_to_16(const float in)
1776 {
1777     int exp = 0, origexp;
1778     float tmp = fabs(in);
1779     int sign = (copysignf(1, in) < 0);
1780     unsigned int mantissa;
1781     unsigned short ret;
1782
1783     /* Deal with special numbers */
1784     if (isinf(in)) return (sign ? 0xffff : 0x7fff);
1785     if (isnan(in)) return (sign ? 0xffff : 0x7fff);
1786     if (in == 0.0f) return (sign ? 0x8000 : 0x0000);
1787
1788     if (tmp < powf(2, 10))
1789     {
1790         do
1791         {
1792             tmp *= 2.0f;
1793             exp--;
1794         } while (tmp < powf(2, 10));
1795     }
1796     else if (tmp >= powf(2, 11))
1797     {
1798         do
1799         {
1800             tmp /= 2.0f;
1801             exp++;
1802         } while (tmp >= powf(2, 11));
1803     }
1804
1805     exp += 10;  /* Normalize the mantissa */
1806     exp += 15;  /* Exponent is encoded with excess 15 */
1807
1808     origexp = exp;
1809
1810     mantissa = (unsigned int) tmp;
1811     if ((tmp - mantissa == 0.5f && mantissa % 2 == 1) || /* round half to even */
1812         (tmp - mantissa > 0.5f))
1813     {
1814         mantissa++; /* round to nearest, away from zero */
1815     }
1816     if (mantissa == 2048)
1817     {
1818         mantissa = 1024;
1819         exp++;
1820     }
1821
1822     if (exp > 31)
1823     {
1824         /* too big */
1825         ret = 0x7fff; /* INF */
1826     }
1827     else if (exp <= 0)
1828     {
1829         unsigned int rounding = 0;
1830
1831         /* Denormalized half float */
1832
1833         /* return 0x0000 (=0.0) for numbers too small to represent in half floats */
1834         if (exp < -11)
1835             return (sign ? 0x8000 : 0x0000);
1836
1837         exp = origexp;
1838
1839         /* the 13 extra bits from single precision are used for rounding */
1840         mantissa = (unsigned int)(tmp * powf(2, 13));
1841         mantissa >>= 1 - exp; /* denormalize */
1842
1843         mantissa -= ~(mantissa >> 13) & 1; /* round half to even */
1844         /* remove 13 least significant bits to get half float precision */
1845         mantissa >>= 12;
1846         rounding = mantissa & 1;
1847         mantissa >>= 1;
1848
1849         ret = mantissa + rounding;
1850     }
1851     else
1852     {
1853         ret = (exp << 10) | (mantissa & 0x3ff);
1854     }
1855
1856     ret |= ((sign ? 1 : 0) << 15); /* Add the sign */
1857     return ret;
1858 }
1859
1860 D3DXFLOAT16 *WINAPI D3DXFloat32To16Array(D3DXFLOAT16 *pout, CONST FLOAT *pin, UINT n)
1861 {
1862     unsigned int i;
1863
1864     for (i = 0; i < n; ++i)
1865     {
1866         pout[i].value = float_32_to_16(pin[i]);
1867     }
1868
1869     return pout;
1870 }
1871
1872 /* Native d3dx9's D3DXFloat16to32Array lacks support for NaN and Inf. Specifically, e = 16 is treated as a
1873  * regular number - e.g., 0x7fff is converted to 131008.0 and 0xffff to -131008.0. */
1874 static inline float float_16_to_32(const unsigned short in)
1875 {
1876     const unsigned short s = (in & 0x8000);
1877     const unsigned short e = (in & 0x7C00) >> 10;
1878     const unsigned short m = in & 0x3FF;
1879     const float sgn = (s ? -1.0f : 1.0f);
1880
1881     if (e == 0)
1882     {
1883         if (m == 0) return sgn * 0.0f; /* +0.0 or -0.0 */
1884         else return sgn * powf(2, -14.0f) * (m / 1024.0f);
1885     }
1886     else
1887     {
1888         return sgn * powf(2, e - 15.0f) * (1.0f + (m / 1024.0f));
1889     }
1890 }
1891
1892 FLOAT *WINAPI D3DXFloat16To32Array(FLOAT *pout, CONST D3DXFLOAT16 *pin, UINT n)
1893 {
1894     unsigned int i;
1895
1896     for (i = 0; i < n; ++i)
1897     {
1898         pout[i] = float_16_to_32(pin[i].value);
1899     }
1900
1901     return pout;
1902 }
1903
1904 FLOAT* WINAPI D3DXSHAdd(FLOAT *out, UINT order, const FLOAT *a, const FLOAT *b)
1905 {
1906     UINT i;
1907
1908     TRACE("out %p, order %u, a %p, b %p\n", out, order, a, b);
1909
1910     for (i = 0; i < order * order; i++)
1911         out[i] = a[i] + b[i];
1912
1913     return out;
1914 }