dpsoftrast.c

   1
   2 #include <memory.h>
   3 #include "dpsoftrast.h"
   4 #include <stdio.h>
   5 #include <math.h>
   6
   7 #undef true
   8 #undef false
   9 #ifndef __cplusplus
  10 typedef enum bool {false, true} bool;
  11 #endif
  12
  13 #if defined(__GNUC__) || (defined(_MSC_VER) && _MSC_VER >= 1400)
  14 #define RESTRICT __restrict
  15 #else
  16 #define RESTRICT
  17 #endif
  18
  19 #define GL_NONE                                 0
  20 #define GL_FRONT_LEFT                   0x0400
  21 #define GL_FRONT_RIGHT                  0x0401
  22 #define GL_BACK_LEFT                    0x0402
  23 #define GL_BACK_RIGHT                   0x0403
  24 #define GL_FRONT                                0x0404
  25 #define GL_BACK                                 0x0405
  26 #define GL_LEFT                                 0x0406
  27 #define GL_RIGHT                                0x0407
  28 #define GL_FRONT_AND_BACK               0x0408
  29 #define GL_AUX0                                 0x0409
  30 #define GL_AUX1                                 0x040A
  31 #define GL_AUX2                                 0x040B
  32 #define GL_AUX3                                 0x040C
  33
  34 #define GL_NEVER                                0x0200
  35 #define GL_LESS                                 0x0201
  36 #define GL_EQUAL                                0x0202
  37 #define GL_LEQUAL                               0x0203
  38 #define GL_GREATER                              0x0204
  39 #define GL_NOTEQUAL                             0x0205
  40 #define GL_GEQUAL                               0x0206
  41 #define GL_ALWAYS                               0x0207
  42
  43 #define GL_ZERO                                 0x0
  44 #define GL_ONE                                  0x1
  45 #define GL_SRC_COLOR                            0x0300
  46 #define GL_ONE_MINUS_SRC_COLOR                  0x0301
  47 #define GL_DST_COLOR                            0x0306
  48 #define GL_ONE_MINUS_DST_COLOR                  0x0307
  49 #define GL_SRC_ALPHA                            0x0302
  50 #define GL_ONE_MINUS_SRC_ALPHA                  0x0303
  51 #define GL_DST_ALPHA                            0x0304
  52 #define GL_ONE_MINUS_DST_ALPHA                  0x0305
  53 #define GL_SRC_ALPHA_SATURATE                   0x0308
  54 #define GL_CONSTANT_COLOR                       0x8001
  55 #define GL_ONE_MINUS_CONSTANT_COLOR             0x8002
  56 #define GL_CONSTANT_ALPHA                       0x8003
  57 #define GL_ONE_MINUS_CONSTANT_ALPHA             0x8004
  58
  59 typedef enum DPSOFTRAST_ARRAY_e
  60 {
  61         DPSOFTRAST_ARRAY_POSITION,
  62         DPSOFTRAST_ARRAY_COLOR,
  63         DPSOFTRAST_ARRAY_TEXCOORD0,
  64         DPSOFTRAST_ARRAY_TEXCOORD1,
  65         DPSOFTRAST_ARRAY_TEXCOORD2,
  66         DPSOFTRAST_ARRAY_TEXCOORD3,
  67         DPSOFTRAST_ARRAY_TEXCOORD4,
  68         DPSOFTRAST_ARRAY_TEXCOORD5,
  69         DPSOFTRAST_ARRAY_TEXCOORD6,
  70         DPSOFTRAST_ARRAY_TEXCOORD7,
  71         DPSOFTRAST_ARRAY_TOTAL
  72 }
  73 DPSOFTRAST_ARRAY;
  74
  75 typedef struct DPSOFTRAST_Texture_s
  76 {
  77         int flags;
  78         int width;
  79         int height;
  80         int depth;
  81         int sides;
  82         DPSOFTRAST_TEXTURE_FILTER filter;
  83         int mipmaps;
  84         int size;
  85         unsigned char *bytes;
  86         int mipmap[DPSOFTRAST_MAXMIPMAPS][5];
  87 }
  88 DPSOFTRAST_Texture;
  89
  90 typedef struct DPSOFTRAST_State_User_s
  91 {
  92         int colormask[4];
  93         int blendfunc[2];
  94         int blendsubtract;
  95         int depthmask;
  96         int depthtest;
  97         int depthfunc;
  98         int scissortest;
  99         int cullface;
 100         int alphatest;
 101         int alphafunc;
 102         float alphavalue;
 103         int scissor[4];
 104         int viewport[4];
 105         float depthrange[2];
 106         float polygonoffset[2];
 107         float color[4];
 108 }
 109 DPSOFTRAST_State_User;
 110
 111 #define DPSOFTRAST_MAXSUBSPAN 16
 112
 113 typedef struct DPSOFTRAST_State_Draw_Span_s
 114 {
 115         int start; // pixel index
 116         int length; // pixel count
 117         int startx; // usable range (according to pixelmask)
 118         int endx; // usable range (according to pixelmask)
 119         int mip; // which mipmap of the texture(s) to use on this
 120         unsigned char *pixelmask; // true for pixels that passed depth test, false for others
 121         // [0][n][] is start interpolant values (projected)
 122         // [1][n][] is end interpolant values (projected)
 123         // [0][DPSOFTRAST_ARRAY_TOTAL][] is start screencoord4f
 124         // [1][DPSOFTRAST_ARRAY_TOTAL][] is end screencoord4f
 125         // NOTE: screencoord4f[3] is W (basically 1/Z), useful for depthbuffer
 126         float data[2][DPSOFTRAST_ARRAY_TOTAL+1][4];
 127 }
 128 DPSOFTRAST_State_Draw_Span;
 129
 130 #define DPSOFTRAST_DRAW_MAXSPANQUEUE 1024
 131
 132 typedef struct DPSOFTRAST_State_Draw_s
 133 {
 134         int numvertices;
 135         int maxvertices;
 136         float *in_array4f[DPSOFTRAST_ARRAY_TOTAL];
 137         float *post_array4f[DPSOFTRAST_ARRAY_TOTAL];
 138         float *screencoord4f;
 139
 140         // spans are queued in this structure for dispatch to the pixel shader,
 141         // partly to improve cache locality, partly for batching purposes, spans
 142         // are flushed before DrawTriangles returns to caller
 143         int numspans;
 144         DPSOFTRAST_State_Draw_Span spanqueue[DPSOFTRAST_DRAW_MAXSPANQUEUE];
 145 }
 146 DPSOFTRAST_State_Draw;
 147
 148 #define DPSOFTRAST_VALIDATE_FB 1
 149 #define DPSOFTRAST_VALIDATE_DEPTHFUNC 2
 150 #define DPSOFTRAST_VALIDATE_BLENDFUNC 4
 151 #define DPSOFTRAST_VALIDATE_DRAW (DPSOFTRAST_VALIDATE_FB | DPSOFTRAST_VALIDATE_DEPTHFUNC | DPSOFTRAST_VALIDATE_BLENDFUNC)
 152
 153 typedef enum DPSOFTRAST_BLENDMODE_e
 154 {
 155         DPSOFTRAST_BLENDMODE_OPAQUE,
 156         DPSOFTRAST_BLENDMODE_ALPHA,
 157         DPSOFTRAST_BLENDMODE_ADDALPHA,
 158         DPSOFTRAST_BLENDMODE_ADD,
 159         DPSOFTRAST_BLENDMODE_INVMOD,
 160         DPSOFTRAST_BLENDMODE_MUL,
 161         DPSOFTRAST_BLENDMODE_MUL2,
 162         DPSOFTRAST_BLENDMODE_SUBALPHA,
 163         DPSOFTRAST_BLENDMODE_PSEUDOALPHA,
 164         DPSOFTRAST_BLENDMODE_TOTAL
 165 }
 166 DPSOFTRAST_BLENDMODE;
 167
 168 typedef struct DPSOFTRAST_State_s
 169 {
 170         // DPSOFTRAST_VALIDATE_ flags
 171         int validate;
 172
 173         int fb_colormask;
 174         int fb_width;
 175         int fb_height;
 176         unsigned int *fb_depthpixels;
 177         unsigned int *fb_colorpixels[4];
 178
 179         const float *pointer_vertex3f;
 180         const float *pointer_color4f;
 181         const unsigned char *pointer_color4ub;
 182         const float *pointer_texcoordf[DPSOFTRAST_MAXTEXCOORDARRAYS];
 183         int stride_vertex;
 184         int stride_color;
 185         int stride_texcoord[DPSOFTRAST_MAXTEXCOORDARRAYS];
 186         int components_texcoord[DPSOFTRAST_MAXTEXCOORDARRAYS];
 187         DPSOFTRAST_Texture *texbound[DPSOFTRAST_MAXTEXTUREUNITS];
 188
 189         int shader_mode;
 190         int shader_permutation;
 191         float uniform4f[DPSOFTRAST_UNIFORM_TOTAL*4];
 192         int uniform1i[DPSOFTRAST_UNIFORM_TOTAL];
 193
 194         // derived values (DPSOFTRAST_VALIDATE_FB)
 195         int fb_clearscissor[4];
 196         int fb_viewport[4];
 197         int fb_viewportscissor[4];
 198         float fb_viewportcenter[2];
 199         float fb_viewportscale[2];
 200
 201         // derived values (DPSOFTRAST_VALIDATE_DEPTHFUNC)
 202         int fb_depthfunc;
 203
 204         // derived values (DPSOFTRAST_VALIDATE_BLENDFUNC)
 205         int fb_blendmode;
 206
 207         int texture_max;
 208         int texture_end;
 209         int texture_firstfree;
 210         DPSOFTRAST_Texture *texture;
 211
 212         int bigendian;
 213
 214         // error reporting
 215         const char *errorstring;
 216
 217         DPSOFTRAST_State_User user;
 218
 219         DPSOFTRAST_State_Draw draw;
 220 }
 221 DPSOFTRAST_State;
 222
 223 DPSOFTRAST_State dpsoftrast;
 224
 225 #define DPSOFTRAST_DEPTHSCALE (1024.0f*1048576.0f)
 226 #define DPSOFTRAST_BGRA8_FROM_RGBA32F(r,g,b,a) (((int)(r * 255.0f + 0.5f) << 16) | ((int)(g * 255.0f + 0.5f) << 8) | (int)(b * 255.0f + 0.5f) | ((int)(a * 255.0f + 0.5f) << 24))
 227 #define DPSOFTRAST_DEPTH32_FROM_DEPTH32F(d) ((int)(DPSOFTRAST_DEPTHSCALE * (1-d)))
 228 #define DPSOFTRAST_DRAW_MAXSPANLENGTH 256
 229
 230 void DPSOFTRAST_RecalcFB(void)
 231 {
 232         // calculate framebuffer scissor, viewport, viewport clipped by scissor,
 233         // and viewport projection values
 234         int x1, x2, x3, x4, x5, x6;
 235         int y1, y2, y3, y4, y5, y6;
 236         x1 = dpsoftrast.user.scissor[0];
 237         x2 = dpsoftrast.user.scissor[0] + dpsoftrast.user.scissor[2];
 238         x3 = dpsoftrast.user.viewport[0];
 239         x4 = dpsoftrast.user.viewport[0] + dpsoftrast.user.viewport[2];
 240         y1 = dpsoftrast.fb_height - dpsoftrast.user.scissor[1] - dpsoftrast.user.scissor[3];
 241         y2 = dpsoftrast.fb_height - dpsoftrast.user.scissor[1];
 242         y3 = dpsoftrast.fb_height - dpsoftrast.user.viewport[1] - dpsoftrast.user.viewport[3];
 243         y4 = dpsoftrast.fb_height - dpsoftrast.user.viewport[1];
 244         if (!dpsoftrast.user.scissortest) {x1 = 0;y1 = 0;x2 = dpsoftrast.fb_width;y2 = dpsoftrast.fb_height;}
 245         if (x1 < 0) x1 = 0;
 246         if (x2 > dpsoftrast.fb_width) x2 = dpsoftrast.fb_width;
 247         if (x3 < 0) x1 = 0;
 248         if (x4 > dpsoftrast.fb_width) x4 = dpsoftrast.fb_width;
 249         if (y1 < 0) y1 = 0;
 250         if (y2 > dpsoftrast.fb_height) y2 = dpsoftrast.fb_height;
 251         if (y3 < 0) y1 = 0;
 252         if (y4 > dpsoftrast.fb_height) y4 = dpsoftrast.fb_height;
 253         x5 = x1;if (x5 < x3) x5 = x3;
 254         x6 = x2;if (x6 > x4) x4 = x4;
 255         y5 = y1;if (y5 < y3) y5 = y3;
 256         y6 = y2;if (y6 > y4) y6 = y4;
 257         dpsoftrast.fb_clearscissor[0] = x1;
 258         dpsoftrast.fb_clearscissor[1] = y1;
 259         dpsoftrast.fb_clearscissor[2] = x2 - x1;
 260         dpsoftrast.fb_clearscissor[3] = y2 - y1;
 261         dpsoftrast.fb_viewport[0] = x3;
 262         dpsoftrast.fb_viewport[1] = y3;
 263         dpsoftrast.fb_viewport[2] = x4 - x3;
 264         dpsoftrast.fb_viewport[3] = y4 - y3;
 265         dpsoftrast.fb_viewportscissor[0] = x5;
 266         dpsoftrast.fb_viewportscissor[1] = y5;
 267         dpsoftrast.fb_viewportscissor[2] = x6 - x5;
 268         dpsoftrast.fb_viewportscissor[3] = y6 - y5;
 269         dpsoftrast.fb_viewportcenter[0] = dpsoftrast.user.viewport[0] + 0.5f * dpsoftrast.user.viewport[2] - 0.5f;
 270         dpsoftrast.fb_viewportcenter[1] = dpsoftrast.fb_height - dpsoftrast.user.viewport[1] - 0.5f * dpsoftrast.user.viewport[3] - 0.5f;
 271         dpsoftrast.fb_viewportscale[0] = 0.5f * dpsoftrast.user.viewport[2];
 272         dpsoftrast.fb_viewportscale[1] = -0.5f * dpsoftrast.user.viewport[3];
 273 }
 274
 275 void DPSOFTRAST_RecalcDepthFunc(void)
 276 {
 277         dpsoftrast.fb_depthfunc = dpsoftrast.user.depthtest ? dpsoftrast.user.depthfunc : GL_ALWAYS;
 278 }
 279
 280 int blendmodetable[][4] =
 281 {
 282         {DPSOFTRAST_BLENDMODE_OPAQUE, GL_ONE, GL_ZERO, false},
 283         {DPSOFTRAST_BLENDMODE_ALPHA, GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA, false},
 284         {DPSOFTRAST_BLENDMODE_ADDALPHA, GL_SRC_ALPHA, GL_ONE, false},
 285         {DPSOFTRAST_BLENDMODE_ADD, GL_ONE, GL_ONE, false},
 286         {DPSOFTRAST_BLENDMODE_INVMOD, GL_ZERO, GL_ONE_MINUS_SRC_COLOR, false},
 287         {DPSOFTRAST_BLENDMODE_MUL, GL_ZERO, GL_SRC_COLOR, false},
 288         {DPSOFTRAST_BLENDMODE_MUL, GL_DST_COLOR, GL_ZERO, false},
 289         {DPSOFTRAST_BLENDMODE_MUL2, GL_DST_COLOR, GL_SRC_COLOR, false},
 290         {DPSOFTRAST_BLENDMODE_PSEUDOALPHA, GL_ONE, GL_ONE_MINUS_SRC_ALPHA, false},
 291         {DPSOFTRAST_BLENDMODE_SUBALPHA, GL_SRC_COLOR, GL_ONE, true}
 292 };
 293
 294 void DPSOFTRAST_RecalcBlendFunc(void)
 295 {
 296         int i;
 297         dpsoftrast.fb_blendmode = DPSOFTRAST_BLENDMODE_OPAQUE;
 298         for (i = 0;i < (int)(sizeof(blendmodetable) / sizeof(blendmodetable[0]));i++)
 299         {
 300                 if (dpsoftrast.user.blendfunc[0] == blendmodetable[i][1] && dpsoftrast.user.blendfunc[1] == blendmodetable[i][2] && dpsoftrast.user.blendsubtract == blendmodetable[i][3])
 301                 {
 302                         dpsoftrast.fb_blendmode = blendmodetable[i][0];
 303                         break;
 304                 }
 305         }
 306 }
 307
 308 #define DPSOFTRAST_ValidateQuick(f) ((dpsoftrast.validate & (f)) ? (DPSOFTRAST_Validate(f), 0) : 0)
 309
 310 void DPSOFTRAST_Validate(int mask)
 311 {
 312         mask &= dpsoftrast.validate;
 313         if (!mask)
 314                 return;
 315         if (mask & DPSOFTRAST_VALIDATE_FB)
 316         {
 317                 dpsoftrast.validate &= ~DPSOFTRAST_VALIDATE_FB;
 318                 DPSOFTRAST_RecalcFB();
 319         }
 320         if (mask & DPSOFTRAST_VALIDATE_DEPTHFUNC)
 321         {
 322                 dpsoftrast.validate &= ~DPSOFTRAST_VALIDATE_DEPTHFUNC;
 323                 DPSOFTRAST_RecalcDepthFunc();
 324         }
 325         if (mask & DPSOFTRAST_VALIDATE_BLENDFUNC)
 326         {
 327                 dpsoftrast.validate &= ~DPSOFTRAST_VALIDATE_BLENDFUNC;
 328                 DPSOFTRAST_RecalcBlendFunc();
 329         }
 330 }
 331
 332 DPSOFTRAST_Texture *DPSOFTRAST_Texture_GetByIndex(int index)
 333 {
 334         if (index >= 1 && index < dpsoftrast.texture_end && dpsoftrast.texture[index].bytes)
 335                 return &dpsoftrast.texture[index];
 336         return NULL;
 337 }
 338
 339 int DPSOFTRAST_Texture_New(int flags, int width, int height, int depth)
 340 {
 341         int w;
 342         int h;
 343         int d;
 344         int size;
 345         int s;
 346         int texnum;
 347         int mipmaps;
 348         int sides = (flags & DPSOFTRAST_TEXTURE_FLAG_CUBEMAP) ? 6 : 1;
 349         int texformat = flags & DPSOFTRAST_TEXTURE_FORMAT_COMPAREMASK;
 350         DPSOFTRAST_Texture *texture;
 351         if (width*height*depth < 1)
 352         {
 353                 dpsoftrast.errorstring = "DPSOFTRAST_Texture_New: width, height or depth is less than 1";
 354                 return 0;
 355         }
 356         if (width > DPSOFTRAST_TEXTURE_MAXSIZE || height > DPSOFTRAST_TEXTURE_MAXSIZE || depth > DPSOFTRAST_TEXTURE_MAXSIZE)
 357         {
 358                 dpsoftrast.errorstring = "DPSOFTRAST_Texture_New: texture size is too large";
 359                 return 0;
 360         }
 361         switch(texformat)
 362         {
 363         case DPSOFTRAST_TEXTURE_FORMAT_BGRA8:
 364         case DPSOFTRAST_TEXTURE_FORMAT_RGBA8:
 365         case DPSOFTRAST_TEXTURE_FORMAT_ALPHA8:
 366                 break;
 367         case DPSOFTRAST_TEXTURE_FORMAT_DEPTH:
 368                 if (flags & DPSOFTRAST_TEXTURE_FLAG_CUBEMAP)
 369                 {
 370                         dpsoftrast.errorstring = "DPSOFTRAST_Texture_New: DPSOFTRAST_TEXTURE_FORMAT_DEPTH only permitted on 2D textures";
 371                         return 0;
 372                 }
 373                 if (depth != 1)
 374                 {
 375                         dpsoftrast.errorstring = "DPSOFTRAST_Texture_New: DPSOFTRAST_TEXTURE_FORMAT_DEPTH only permitted on 2D textures";
 376                         return 0;
 377                 }
 378                 if ((flags & DPSOFTRAST_TEXTURE_FLAG_MIPMAP) && (texformat == DPSOFTRAST_TEXTURE_FORMAT_DEPTH))
 379                 {
 380                         dpsoftrast.errorstring = "DPSOFTRAST_Texture_New: DPSOFTRAST_TEXTURE_FORMAT_DEPTH does not permit mipmaps";
 381                         return 0;
 382                 }
 383                 break;
 384         }
 385         if (depth != 1 && (flags & DPSOFTRAST_TEXTURE_FLAG_CUBEMAP))
 386         {
 387                 dpsoftrast.errorstring = "DPSOFTRAST_Texture_New: DPSOFTRAST_TEXTURE_FLAG_CUBEMAP can not be used on 3D textures";
 388                 return 0;
 389         }
 390         if (depth != 1 && (flags & DPSOFTRAST_TEXTURE_FLAG_MIPMAP))
 391         {
 392                 dpsoftrast.errorstring = "DPSOFTRAST_Texture_New: DPSOFTRAST_TEXTURE_FLAG_MIPMAP can not be used on 3D textures";
 393                 return 0;
 394         }
 395         if (depth != 1 && (flags & DPSOFTRAST_TEXTURE_FLAG_MIPMAP))
 396         {
 397                 dpsoftrast.errorstring = "DPSOFTRAST_Texture_New: DPSOFTRAST_TEXTURE_FLAG_MIPMAP can not be used on 3D textures";
 398                 return 0;
 399         }
 400         if ((flags & DPSOFTRAST_TEXTURE_FLAG_CUBEMAP) && (flags & DPSOFTRAST_TEXTURE_FLAG_MIPMAP))
 401         {
 402                 dpsoftrast.errorstring = "DPSOFTRAST_Texture_New: DPSOFTRAST_TEXTURE_FLAG_MIPMAP can not be used on cubemap textures";
 403                 return 0;
 404         }
 405         if ((width & (width-1)) || (height & (height-1)) || (depth & (depth-1)))
 406         {
 407                 dpsoftrast.errorstring = "DPSOFTRAST_Texture_New: dimensions are not power of two";
 408                 return 0;
 409         }
 410         // find first empty slot in texture array
 411         for (texnum = dpsoftrast.texture_firstfree;texnum < dpsoftrast.texture_end;texnum++)
 412                 if (!dpsoftrast.texture[texnum].bytes)
 413                         break;
 414         dpsoftrast.texture_firstfree = texnum + 1;
 415         if (dpsoftrast.texture_max <= texnum)
 416         {
 417                 // expand texture array as needed
 418                 if (dpsoftrast.texture_max < 1024)
 419                         dpsoftrast.texture_max = 1024;
 420                 else
 421                         dpsoftrast.texture_max *= 2;
 422                 dpsoftrast.texture = (DPSOFTRAST_Texture *)realloc(dpsoftrast.texture, dpsoftrast.texture_max * sizeof(DPSOFTRAST_Texture));
 423         }
 424         if (dpsoftrast.texture_end <= texnum)
 425                 dpsoftrast.texture_end = texnum + 1;
 426         texture = &dpsoftrast.texture[texnum];
 427         memset(texture, 0, sizeof(*texture));
 428         texture->flags = flags;
 429         texture->width = width;
 430         texture->height = height;
 431         texture->depth = depth;
 432         texture->sides = sides;
 433         w = width;
 434         h = height;
 435         d = depth;
 436         size = 0;
 437         mipmaps = 0;
 438         w = width;
 439         h = height;
 440         d = depth;
 441         for (;;)
 442         {
 443                 s = w * h * d * sides * 4;
 444                 texture->mipmap[mipmaps][0] = size;
 445                 texture->mipmap[mipmaps][1] = s;
 446                 texture->mipmap[mipmaps][2] = w;
 447                 texture->mipmap[mipmaps][3] = h;
 448                 texture->mipmap[mipmaps][4] = d;
 449                 size += s;
 450                 mipmaps++;
 451                 if (w * h * d == 1 || !(flags & DPSOFTRAST_TEXTURE_FLAG_MIPMAP))
 452                         break;
 453                 if (w > 1) w >>= 1;
 454                 if (h > 1) h >>= 1;
 455                 if (d > 1) d >>= 1;
 456         }
 457         texture->mipmaps = mipmaps;
 458         texture->size = size;
 459
 460         // allocate the pixels now
 461         texture->bytes = calloc(1, size);
 462
 463         return texnum;
 464 }
 465 void DPSOFTRAST_Texture_Free(int index)
 466 {
 467         DPSOFTRAST_Texture *texture;
 468         texture = DPSOFTRAST_Texture_GetByIndex(index);if (!texture) return;
 469         if (texture->bytes)
 470                 free(texture->bytes);
 471         texture->bytes = NULL;
 472         memset(texture, 0, sizeof(*texture));
 473         // adjust the free range and used range
 474         if (dpsoftrast.texture_firstfree > index)
 475                 dpsoftrast.texture_firstfree = index;
 476         while (dpsoftrast.texture_end > 0 && dpsoftrast.texture[dpsoftrast.texture_end-1].bytes == NULL)
 477                 dpsoftrast.texture_end--;
 478 }
 479 void DPSOFTRAST_Texture_CalculateMipmaps(int index)
 480 {
 481         int i, x, y, z, w, layer0, layer1, row0, row1;
 482         unsigned char *o, *i0, *i1, *i2, *i3;
 483         DPSOFTRAST_Texture *texture;
 484         texture = DPSOFTRAST_Texture_GetByIndex(index);if (!texture) return;
 485         if (texture->mipmaps <= 1)
 486                 return;
 487         for (i = 1;i < texture->mipmaps;i++)
 488         {
 489                 for (z = 0;z < texture->mipmap[i][4];z++)
 490                 {
 491                         layer0 = z*2;
 492                         layer1 = z*2+1;
 493                         if (layer1 >= texture->mipmap[i-1][4])
 494                                 layer1 = texture->mipmap[i-1][4]-1;
 495                         for (y = 0;y < texture->mipmap[i][3];y++)
 496                         {
 497                                 row0 = y*2;
 498                                 row1 = y*2+1;
 499                                 if (row1 >= texture->mipmap[i-1][3])
 500                                         row1 = texture->mipmap[i-1][3]-1;
 501                                 o =  texture->bytes + texture->mipmap[i  ][0] + 4*(texture->mipmap[i  ][3] * texture->mipmap[i  ][2] * z      + texture->mipmap[i  ][2] * y   );
 502                                 i0 = texture->bytes + texture->mipmap[i-1][0] + 4*(texture->mipmap[i-1][3] * texture->mipmap[i-1][2] * layer0 + texture->mipmap[i-1][2] * row0);
 503                                 i1 = texture->bytes + texture->mipmap[i-1][0] + 4*(texture->mipmap[i-1][3] * texture->mipmap[i-1][2] * layer0 + texture->mipmap[i-1][2] * row1);
 504                                 i2 = texture->bytes + texture->mipmap[i-1][0] + 4*(texture->mipmap[i-1][3] * texture->mipmap[i-1][2] * layer1 + texture->mipmap[i-1][2] * row0);
 505                                 i3 = texture->bytes + texture->mipmap[i-1][0] + 4*(texture->mipmap[i-1][3] * texture->mipmap[i-1][2] * layer1 + texture->mipmap[i-1][2] * row1);
 506                                 w = texture->mipmap[i][2];
 507                                 if (layer1 > layer0)
 508                                 {
 509                                         if (texture->mipmap[i-1][2] > 1)
 510                                         {
 511                                                 // average 3D texture
 512                                                 for (x = 0;x < w;x++)
 513                                                 {
 514                                                         o[0] = (i0[0] + i0[4] + i1[0] + i1[4] + i2[0] + i2[4] + i3[0] + i3[4] + 4) >> 3;
 515                                                         o[1] = (i0[1] + i0[5] + i1[1] + i1[5] + i2[1] + i2[5] + i3[1] + i3[5] + 4) >> 3;
 516                                                         o[2] = (i0[2] + i0[6] + i1[2] + i1[6] + i2[2] + i2[6] + i3[2] + i3[6] + 4) >> 3;
 517                                                         o[3] = (i0[3] + i0[7] + i1[3] + i1[7] + i2[3] + i2[7] + i3[3] + i3[7] + 4) >> 3;
 518                                                 }
 519                                         }
 520                                         else
 521                                         {
 522                                                 // average 3D mipmap with parent width == 1
 523                                                 for (x = 0;x < w;x++)
 524                                                 {
 525                                                         o[0] = (i0[0] + i1[0] + i2[0] + i3[0] + 2) >> 2;
 526                                                         o[1] = (i0[1] + i1[1] + i2[1] + i3[1] + 2) >> 2;
 527                                                         o[2] = (i0[2] + i1[2] + i2[2] + i3[2] + 2) >> 2;
 528                                                         o[3] = (i0[3] + i1[3] + i2[3] + i3[3] + 2) >> 2;
 529                                                 }
 530                                         }
 531                                 }
 532                                 else
 533                                 {
 534                                         if (texture->mipmap[i-1][2] > 1)
 535                                         {
 536                                                 // average 2D texture (common case)
 537                                                 for (x = 0;x < w;x++)
 538                                                 {
 539                                                         o[0] = (i0[0] + i0[4] + i1[0] + i1[4] + 2) >> 2;
 540                                                         o[1] = (i0[1] + i0[5] + i1[1] + i1[5] + 2) >> 2;
 541                                                         o[2] = (i0[2] + i0[6] + i1[2] + i1[6] + 2) >> 2;
 542                                                         o[3] = (i0[3] + i0[7] + i1[3] + i1[7] + 2) >> 2;
 543                                                 }
 544                                         }
 545                                         else
 546                                         {
 547                                                 // 2D texture with parent width == 1
 548                                                 for (x = 0;x < w;x++)
 549                                                 {
 550                                                         o[0] = (i0[0] + i1[0] + 1) >> 1;
 551                                                         o[1] = (i0[1] + i1[1] + 1) >> 1;
 552                                                         o[2] = (i0[2] + i1[2] + 1) >> 1;
 553                                                         o[3] = (i0[3] + i1[3] + 1) >> 1;
 554                                                 }
 555                                         }
 556                                 }
 557                         }
 558                 }
 559         }
 560 }
 561 void DPSOFTRAST_Texture_UpdatePartial(int index, int mip, const unsigned char *pixels, int blockx, int blocky, int blockwidth, int blockheight)
 562 {
 563         DPSOFTRAST_Texture *texture;
 564         texture = DPSOFTRAST_Texture_GetByIndex(index);if (!texture) return;
 565
 566         // FIXME IMPLEMENT
 567
 568         dpsoftrast.errorstring = "DPSOFTRAST_Texture_UpdatePartial: Not implemented.";
 569 }
 570 void DPSOFTRAST_Texture_UpdateFull(int index, const unsigned char *pixels)
 571 {
 572         DPSOFTRAST_Texture *texture;
 573         texture = DPSOFTRAST_Texture_GetByIndex(index);if (!texture) return;
 574
 575         memcpy(texture->bytes, pixels, texture->mipmap[0][1]);
 576         DPSOFTRAST_Texture_CalculateMipmaps(index);
 577 }
 578 int DPSOFTRAST_Texture_GetWidth(int index, int mip)
 579 {
 580         DPSOFTRAST_Texture *texture;
 581         texture = DPSOFTRAST_Texture_GetByIndex(index);if (!texture) return 0;
 582         return texture->width;
 583 }
 584 int DPSOFTRAST_Texture_GetHeight(int index, int mip)
 585 {
 586         DPSOFTRAST_Texture *texture;
 587         texture = DPSOFTRAST_Texture_GetByIndex(index);if (!texture) return 0;
 588         return texture->height;
 589 }
 590 int DPSOFTRAST_Texture_GetDepth(int index, int mip)
 591 {
 592         DPSOFTRAST_Texture *texture;
 593         texture = DPSOFTRAST_Texture_GetByIndex(index);if (!texture) return 0;
 594         return texture->depth;
 595 }
 596 unsigned char *DPSOFTRAST_Texture_GetPixelPointer(int index, int mip)
 597 {
 598         DPSOFTRAST_Texture *texture;
 599         texture = DPSOFTRAST_Texture_GetByIndex(index);if (!texture) return 0;
 600         return texture->bytes;
 601 }
 602 void DPSOFTRAST_Texture_Filter(int index, DPSOFTRAST_TEXTURE_FILTER filter)
 603 {
 604         DPSOFTRAST_Texture *texture;
 605         texture = DPSOFTRAST_Texture_GetByIndex(index);if (!texture) return;
 606         if (!(texture->flags & DPSOFTRAST_TEXTURE_FLAG_MIPMAP) && filter > DPSOFTRAST_TEXTURE_FILTER_LINEAR)
 607         {
 608                 dpsoftrast.errorstring = "DPSOFTRAST_Texture_Filter: requested filter mode requires mipmaps";
 609                 return;
 610         }
 611         texture->filter = filter;
 612 }
 613
 614 void DPSOFTRAST_SetRenderTargets(int width, int height, unsigned int *depthpixels, unsigned int *colorpixels0, unsigned int *colorpixels1, unsigned int *colorpixels2, unsigned int *colorpixels3)
 615 {
 616         dpsoftrast.fb_width = width;
 617         dpsoftrast.fb_height = height;
 618         dpsoftrast.fb_depthpixels = depthpixels;
 619         dpsoftrast.fb_colorpixels[0] = colorpixels0;
 620         dpsoftrast.fb_colorpixels[1] = colorpixels1;
 621         dpsoftrast.fb_colorpixels[2] = colorpixels2;
 622         dpsoftrast.fb_colorpixels[3] = colorpixels3;
 623 }
 624 void DPSOFTRAST_Viewport(int x, int y, int width, int height)
 625 {
 626         dpsoftrast.user.viewport[0] = x;
 627         dpsoftrast.user.viewport[1] = y;
 628         dpsoftrast.user.viewport[2] = width;
 629         dpsoftrast.user.viewport[3] = height;
 630         dpsoftrast.validate |= DPSOFTRAST_VALIDATE_FB;
 631 }
 632 void DPSOFTRAST_ClearColor(float r, float g, float b, float a)
 633 {
 634         int i, x1, y1, x2, y2, w, h, x, y;
 635         unsigned int *p;
 636         unsigned int c;
 637         DPSOFTRAST_Validate(DPSOFTRAST_VALIDATE_FB);
 638         x1 = dpsoftrast.fb_clearscissor[0];
 639         y1 = dpsoftrast.fb_clearscissor[1];
 640         x2 = dpsoftrast.fb_clearscissor[2];
 641         y2 = dpsoftrast.fb_clearscissor[1] + dpsoftrast.fb_clearscissor[3];
 642         w = x2 - x1;
 643         h = y2 - y1;
 644         if (w < 1 || h < 1)
 645                 return;
 646         // FIXME: honor dpsoftrast.fb_colormask?
 647         c = DPSOFTRAST_BGRA8_FROM_RGBA32F(r,g,b,a);
 648         for (i = 0;i < 4;i++)
 649         {
 650                 if (!dpsoftrast.fb_colorpixels[i])
 651                         continue;
 652                 for (y = y1;y < y2;y++)
 653                 {
 654                         p = dpsoftrast.fb_colorpixels[i] + y * dpsoftrast.fb_width;
 655                         for (x = x1;x < x2;x++)
 656                                 p[x] = c;
 657                 }
 658         }
 659 }
 660 void DPSOFTRAST_ClearDepth(float d)
 661 {
 662         int x1, y1, x2, y2, w, h, x, y;
 663         unsigned int *p;
 664         unsigned int c;
 665         DPSOFTRAST_Validate(DPSOFTRAST_VALIDATE_FB);
 666         x1 = dpsoftrast.fb_clearscissor[0];
 667         y1 = dpsoftrast.fb_clearscissor[1];
 668         x2 = dpsoftrast.fb_clearscissor[2];
 669         y2 = dpsoftrast.fb_clearscissor[1] + dpsoftrast.fb_clearscissor[3];
 670         w = x2 - x1;
 671         h = y2 - y1;
 672         if (w < 1 || h < 1)
 673                 return;
 674         c = DPSOFTRAST_DEPTH32_FROM_DEPTH32F(d);
 675         for (y = y1;y < y2;y++)
 676         {
 677                 p = dpsoftrast.fb_depthpixels + y * dpsoftrast.fb_width;
 678                 for (x = x1;x < x2;x++)
 679                         p[x] = c;
 680         }
 681 }
 682 void DPSOFTRAST_ColorMask(int r, int g, int b, int a)
 683 {
 684         dpsoftrast.user.colormask[0] = r != 0;
 685         dpsoftrast.user.colormask[1] = g != 0;
 686         dpsoftrast.user.colormask[2] = b != 0;
 687         dpsoftrast.user.colormask[3] = a != 0;
 688         dpsoftrast.fb_colormask = ((-dpsoftrast.user.colormask[0]) & 0x00FF0000) | ((-dpsoftrast.user.colormask[1]) & 0x0000FF00) | ((-dpsoftrast.user.colormask[2]) & 0x000000FF) | ((-dpsoftrast.user.colormask[3]) & 0xFF000000);
 689 }
 690 void DPSOFTRAST_DepthTest(int enable)
 691 {
 692         dpsoftrast.user.depthtest = enable;
 693         dpsoftrast.validate |= DPSOFTRAST_VALIDATE_DEPTHFUNC;
 694 }
 695 void DPSOFTRAST_ScissorTest(int enable)
 696 {
 697         dpsoftrast.user.scissortest = enable;
 698         dpsoftrast.validate |= DPSOFTRAST_VALIDATE_FB;
 699 }
 700 void DPSOFTRAST_Scissor(float x, float y, float width, float height)
 701 {
 702         dpsoftrast.user.scissor[0] = x;
 703         dpsoftrast.user.scissor[1] = y;
 704         dpsoftrast.user.scissor[2] = width;
 705         dpsoftrast.user.scissor[3] = height;
 706         dpsoftrast.validate |= DPSOFTRAST_VALIDATE_FB;
 707 }
 708
 709 void DPSOFTRAST_BlendFunc(int smodulate, int dmodulate)
 710 {
 711         // FIXME: validate
 712         dpsoftrast.user.blendfunc[0] = smodulate;
 713         dpsoftrast.user.blendfunc[1] = dmodulate;
 714         dpsoftrast.validate |= DPSOFTRAST_VALIDATE_BLENDFUNC;
 715 }
 716 void DPSOFTRAST_BlendSubtract(int enable)
 717 {
 718         dpsoftrast.user.blendsubtract = enable != 0;
 719         dpsoftrast.validate |= DPSOFTRAST_VALIDATE_BLENDFUNC;
 720 }
 721 void DPSOFTRAST_DepthMask(int enable)
 722 {
 723         dpsoftrast.user.depthmask = enable;
 724 }
 725 void DPSOFTRAST_DepthFunc(int comparemode)
 726 {
 727         // FIXME: validate
 728         dpsoftrast.user.depthfunc = comparemode;
 729 }
 730 void DPSOFTRAST_DepthRange(float range0, float range1)
 731 {
 732         dpsoftrast.user.depthrange[0] = range0;
 733         dpsoftrast.user.depthrange[1] = range1;
 734 }
 735 void DPSOFTRAST_PolygonOffset(float alongnormal, float intoview)
 736 {
 737         dpsoftrast.user.polygonoffset[0] = alongnormal;
 738         dpsoftrast.user.polygonoffset[1] = intoview;
 739 }
 740 void DPSOFTRAST_CullFace(int mode)
 741 {
 742         // FIXME: validate
 743         dpsoftrast.user.cullface = mode;
 744 }
 745 void DPSOFTRAST_AlphaTest(float enable)
 746 {
 747         dpsoftrast.user.alphatest = enable;
 748 }
 749 void DPSOFTRAST_AlphaFunc(int alphafunc, float alphavalue)
 750 {
 751         // FIXME: validate
 752         dpsoftrast.user.alphafunc = alphafunc;
 753         dpsoftrast.user.alphavalue = alphavalue;
 754 }
 755 void DPSOFTRAST_Color4f(float r, float g, float b, float a)
 756 {
 757         dpsoftrast.user.color[0] = r;
 758         dpsoftrast.user.color[1] = g;
 759         dpsoftrast.user.color[2] = b;
 760         dpsoftrast.user.color[3] = a;
 761 }
 762 void DPSOFTRAST_GetPixelsBGRA(int blockx, int blocky, int blockwidth, int blockheight, unsigned char *outpixels)
 763 {
 764         int outstride = blockwidth * 4;
 765         int instride = dpsoftrast.fb_width * 4;
 766         int bx1 = blockx;
 767         int by1 = blocky;
 768         int bx2 = blockx + blockwidth;
 769         int by2 = blocky + blockheight;
 770         int bw;
 771         int bh;
 772         int x;
 773         int y;
 774         unsigned char *inpixels;
 775         unsigned char *b;
 776         unsigned char *o;
 777         if (bx1 < 0) bx1 = 0;
 778         if (by1 < 0) by1 = 0;
 779         if (bx2 > dpsoftrast.fb_width) bx2 = dpsoftrast.fb_width;
 780         if (by2 > dpsoftrast.fb_height) by2 = dpsoftrast.fb_height;
 781         bw = bx2 - bx1;
 782         bh = by2 - by1;
 783         inpixels = (unsigned char *)dpsoftrast.fb_colorpixels[0];
 784         if (dpsoftrast.bigendian)
 785         {
 786                 for (y = by1;y < by2;y++)
 787                 {
 788                         b = (unsigned char *)inpixels + (dpsoftrast.fb_height - 1 - y) * instride + 4 * bx1;
 789                         o = (unsigned char *)outpixels + (y - by1) * outstride;
 790                         for (x = bx1;x < bx2;x++)
 791                         {
 792                                 o[0] = b[3];
 793                                 o[1] = b[2];
 794                                 o[2] = b[1];
 795                                 o[3] = b[0];
 796                                 o += 4;
 797                                 b += 4;
 798                         }
 799                 }
 800         }
 801         else
 802         {
 803                 for (y = by1;y < by2;y++)
 804                 {
 805                         b = (unsigned char *)inpixels + (dpsoftrast.fb_height - 1 - y) * instride + 4 * bx1;
 806                         o = (unsigned char *)outpixels + (y - by1) * outstride;
 807                         memcpy(o, b, bw*4);
 808                 }
 809         }
 810
 811 }
 812 void DPSOFTRAST_CopyRectangleToTexture(int index, int mip, int tx, int ty, int sx, int sy, int width, int height)
 813 {
 814         int tx1 = tx;
 815         int ty1 = ty;
 816         int tx2 = tx + width;
 817         int ty2 = ty + height;
 818         int sx1 = sx;
 819         int sy1 = sy;
 820         int sx2 = sx + width;
 821         int sy2 = sy + height;
 822         int swidth;
 823         int sheight;
 824         int twidth;
 825         int theight;
 826         int sw;
 827         int sh;
 828         int tw;
 829         int th;
 830         int y;
 831         unsigned int *spixels;
 832         unsigned int *tpixels;
 833         DPSOFTRAST_Texture *texture;
 834         texture = DPSOFTRAST_Texture_GetByIndex(index);if (!texture) return;
 835         if (mip < 0 || mip >= texture->mipmaps) return;
 836         spixels = dpsoftrast.fb_colorpixels[0];
 837         swidth = dpsoftrast.fb_width;
 838         sheight = dpsoftrast.fb_height;
 839         tpixels = (unsigned int *)(texture->bytes + texture->mipmap[mip][0]);
 840         twidth = texture->mipmap[mip][2];
 841         theight = texture->mipmap[mip][3];
 842         if (tx1 < 0) tx1 = 0;
 843         if (ty1 < 0) ty1 = 0;
 844         if (tx2 > twidth) tx2 = twidth;
 845         if (ty2 > theight) ty2 = theight;
 846         if (sx1 < 0) sx1 = 0;
 847         if (sy1 < 0) sy1 = 0;
 848         if (sx2 > swidth) sx2 = swidth;
 849         if (sy2 > sheight) sy2 = sheight;
 850         tw = tx2 - tx1;
 851         th = ty2 - ty1;
 852         sw = sx2 - sx1;
 853         sh = sy2 - sy1;
 854         if (tw > sw) tw = sw;
 855         if (th > sh) th = sh;
 856         if (tw < 1 || th < 1)
 857                 return;
 858         for (y = 0;y < th;y++)
 859                 memcpy(tpixels + ((ty1 + y) * twidth + tx1), spixels + ((sy1 + y) * swidth + sx1), tw*4);
 860         if (texture->mipmaps > 1)
 861                 DPSOFTRAST_Texture_CalculateMipmaps(index);
 862 }
 863 void DPSOFTRAST_SetTexture(int unitnum, int index)
 864 {
 865         DPSOFTRAST_Texture *texture;
 866         if (unitnum < 0 || unitnum >= DPSOFTRAST_MAXTEXTUREUNITS)
 867         {
 868                 dpsoftrast.errorstring = "DPSOFTRAST_SetTexture: invalid unit number";
 869                 return;
 870         }
 871         texture = DPSOFTRAST_Texture_GetByIndex(index);
 872         if (index && !texture)
 873         {
 874                 dpsoftrast.errorstring = "DPSOFTRAST_SetTexture: invalid texture handle";
 875                 return;
 876         }
 877         dpsoftrast.texbound[unitnum] = texture;
 878 }
 879
 880 void DPSOFTRAST_SetVertexPointer(const float *vertex3f, size_t stride)
 881 {
 882         dpsoftrast.pointer_vertex3f = vertex3f;
 883         dpsoftrast.stride_vertex = stride;
 884 }
 885 void DPSOFTRAST_SetColorPointer(const float *color4f, size_t stride)
 886 {
 887         dpsoftrast.pointer_color4f = color4f;
 888         dpsoftrast.pointer_color4ub = NULL;
 889         dpsoftrast.stride_color = stride;
 890 }
 891 void DPSOFTRAST_SetColorPointer4ub(const unsigned char *color4ub, size_t stride)
 892 {
 893         dpsoftrast.pointer_color4f = NULL;
 894         dpsoftrast.pointer_color4ub = color4ub;
 895         dpsoftrast.stride_color = stride;
 896 }
 897 void DPSOFTRAST_SetTexCoordPointer(int unitnum, int numcomponents, size_t stride, const float *texcoordf)
 898 {
 899         dpsoftrast.pointer_texcoordf[unitnum] = texcoordf;
 900         dpsoftrast.components_texcoord[unitnum] = numcomponents;
 901         dpsoftrast.stride_texcoord[unitnum] = stride;
 902 }
 903
 904 void DPSOFTRAST_SetShader(unsigned int mode, unsigned int permutation)
 905 {
 906         dpsoftrast.shader_mode = mode;
 907         dpsoftrast.shader_permutation = permutation;
 908 }
 909 void DPSOFTRAST_Uniform4fARB(DPSOFTRAST_UNIFORM index, float v0, float v1, float v2, float v3)
 910 {
 911         dpsoftrast.uniform4f[index*4+0] = v0;
 912         dpsoftrast.uniform4f[index*4+1] = v1;
 913         dpsoftrast.uniform4f[index*4+2] = v2;
 914         dpsoftrast.uniform4f[index*4+3] = v3;
 915 }
 916 void DPSOFTRAST_Uniform4fvARB(DPSOFTRAST_UNIFORM index, const float *v)
 917 {
 918         dpsoftrast.uniform4f[index*4+0] = v[0];
 919         dpsoftrast.uniform4f[index*4+1] = v[1];
 920         dpsoftrast.uniform4f[index*4+2] = v[2];
 921         dpsoftrast.uniform4f[index*4+3] = v[3];
 922 }
 923 void DPSOFTRAST_UniformMatrix4fvARB(DPSOFTRAST_UNIFORM index, int arraysize, int transpose, const float *v)
 924 {
 925         dpsoftrast.uniform4f[index*4+0] = v[0];
 926         dpsoftrast.uniform4f[index*4+1] = v[1];
 927         dpsoftrast.uniform4f[index*4+2] = v[2];
 928         dpsoftrast.uniform4f[index*4+3] = v[3];
 929         dpsoftrast.uniform4f[index*4+4] = v[4];
 930         dpsoftrast.uniform4f[index*4+5] = v[5];
 931         dpsoftrast.uniform4f[index*4+6] = v[6];
 932         dpsoftrast.uniform4f[index*4+7] = v[7];
 933         dpsoftrast.uniform4f[index*4+8] = v[8];
 934         dpsoftrast.uniform4f[index*4+9] = v[9];
 935         dpsoftrast.uniform4f[index*4+10] = v[10];
 936         dpsoftrast.uniform4f[index*4+11] = v[11];
 937         dpsoftrast.uniform4f[index*4+12] = v[12];
 938         dpsoftrast.uniform4f[index*4+13] = v[13];
 939         dpsoftrast.uniform4f[index*4+14] = v[14];
 940         dpsoftrast.uniform4f[index*4+15] = v[15];
 941 }
 942 void DPSOFTRAST_Uniform1iARB(DPSOFTRAST_UNIFORM index, int i0)
 943 {
 944         dpsoftrast.uniform1i[index] = i0;
 945 }
 946
 947 void DPSOFTRAST_Draw_LoadVertices(int firstvertex, int numvertices, bool needcolors)
 948 {
 949         int i;
 950         int j;
 951         int stride;
 952         const float *v;
 953         float *p;
 954         float *data;
 955         const unsigned char *b;
 956         dpsoftrast.draw.numvertices = numvertices;
 957         if (dpsoftrast.draw.maxvertices < dpsoftrast.draw.numvertices)
 958         {
 959                 if (dpsoftrast.draw.maxvertices < 4096)
 960                         dpsoftrast.draw.maxvertices = 4096;
 961                 while (dpsoftrast.draw.maxvertices < dpsoftrast.draw.numvertices)
 962                         dpsoftrast.draw.maxvertices *= 2;
 963                 if (dpsoftrast.draw.in_array4f[0])
 964                         free(dpsoftrast.draw.in_array4f[0]);
 965                 data = calloc(1, dpsoftrast.draw.maxvertices * sizeof(float[4])*(DPSOFTRAST_ARRAY_TOTAL*2 + 1));
 966                 for (i = 0;i < DPSOFTRAST_ARRAY_TOTAL;i++, data += dpsoftrast.draw.maxvertices * 4)
 967                         dpsoftrast.draw.in_array4f[i] = data;
 968                 for (i = 0;i < DPSOFTRAST_ARRAY_TOTAL;i++, data += dpsoftrast.draw.maxvertices * 4)
 969                         dpsoftrast.draw.post_array4f[i] = data;
 970                 dpsoftrast.draw.screencoord4f = data;
 971                 data += dpsoftrast.draw.maxvertices * 4;
 972         }
 973         stride = dpsoftrast.stride_vertex;
 974         v = (const float *)((unsigned char *)dpsoftrast.pointer_vertex3f + firstvertex * stride);
 975         p = dpsoftrast.draw.in_array4f[0];
 976         for (i = 0;i < numvertices;i++)
 977         {
 978                 p[0] = v[0];
 979                 p[1] = v[1];
 980                 p[2] = v[2];
 981                 p[3] = 1.0f;
 982                 p += 4;
 983                 v = (const float *)((const unsigned char *)v + stride);
 984         }
 985         if (needcolors)
 986         {
 987                 if (dpsoftrast.pointer_color4f)
 988                 {
 989                         stride = dpsoftrast.stride_color;
 990                         v = (const float *)((const unsigned char *)dpsoftrast.pointer_color4f + firstvertex * stride);
 991                         p = dpsoftrast.draw.in_array4f[1];
 992                         for (i = 0;i < numvertices;i++)
 993                         {
 994                                 p[0] = v[0];
 995                                 p[1] = v[1];
 996                                 p[2] = v[2];
 997                                 p[3] = v[3];
 998                                 p += 4;
 999                                 v = (const float *)((const unsigned char *)v + stride);
1000                         }
1001                 }
1002                 else if (dpsoftrast.pointer_color4ub)
1003                 {
1004                         stride = dpsoftrast.stride_color;
1005                         b = (const unsigned char *)((const unsigned char *)dpsoftrast.pointer_color4ub + firstvertex * stride);
1006                         p = dpsoftrast.draw.in_array4f[1];
1007                         for (i = 0;i < numvertices;i++)
1008                         {
1009                                 p[0] = b[0] * (1.0f / 255.0f);
1010                                 p[1] = b[1] * (1.0f / 255.0f);
1011                                 p[2] = b[2] * (1.0f / 255.0f);
1012                                 p[3] = b[3] * (1.0f / 255.0f);
1013                                 p += 4;
1014                                 b = (const unsigned char *)((const unsigned char *)b + stride);
1015                         }
1016                 }
1017                 else
1018                 {
1019                         v = dpsoftrast.user.color;
1020                         p = dpsoftrast.draw.in_array4f[1];
1021                         for (i = 0;i < numvertices;i++)
1022                         {
1023                                 p[0] = v[0];
1024                                 p[1] = v[1];
1025                                 p[2] = v[2];
1026                                 p[3] = v[3];
1027                                 p += 4;
1028                         }
1029                 }
1030         }
1031         for (j = 0;j < DPSOFTRAST_ARRAY_TOTAL-2;j++)
1032         {
1033                 if (dpsoftrast.pointer_texcoordf[j])
1034                 {
1035                         stride = dpsoftrast.stride_texcoord[j];
1036                         v = (const float *)((const unsigned char *)dpsoftrast.pointer_texcoordf[j] + firstvertex * stride);
1037                         p = dpsoftrast.draw.in_array4f[j+2];
1038                         switch(dpsoftrast.components_texcoord[j])
1039                         {
1040                         case 2:
1041                                 for (i = 0;i < numvertices;i++)
1042                                 {
1043                                         p[0] = v[0];
1044                                         p[1] = v[1];
1045                                         p[2] = 0.0f;
1046                                         p[3] = 1.0f;
1047                                         p += 4;
1048                                         v = (const float *)((const unsigned char *)v + stride);
1049                                 }
1050                                 break;
1051                         case 3:
1052                                 for (i = 0;i < numvertices;i++)
1053                                 {
1054                                         p[0] = v[0];
1055                                         p[1] = v[1];
1056                                         p[2] = v[2];
1057                                         p[3] = 1.0f;
1058                                         p += 4;
1059                                         v = (const float *)((const unsigned char *)v + stride);
1060                                 }
1061                                 break;
1062                         case 4:
1063                                 for (i = 0;i < numvertices;i++)
1064                                 {
1065                                         p[0] = v[0];
1066                                         p[1] = v[1];
1067                                         p[2] = v[2];
1068                                         p[3] = v[3];
1069                                         p += 4;
1070                                         v = (const float *)((const unsigned char *)v + stride);
1071                                 }
1072                                 break;
1073                         }
1074                 }
1075         }
1076 }
1077
1078 void DPSOFTRAST_Array_Transform(float *out4f, const float *in4f, int numitems, const float *inmatrix16f)
1079 {
1080         // TODO: SIMD
1081         float matrix[4][4];
1082         int i;
1083         memcpy(matrix, inmatrix16f, sizeof(float[16]));
1084         for (i = 0;i < numitems;i++, out4f += 4, in4f += 4)
1085         {
1086                 out4f[0] = in4f[0] * matrix[0][0] + in4f[1] * matrix[1][0] + in4f[2] * matrix[2][0] + in4f[3] * matrix[3][0];
1087                 out4f[1] = in4f[0] * matrix[0][1] + in4f[1] * matrix[1][1] + in4f[2] * matrix[2][1] + in4f[3] * matrix[3][1];
1088                 out4f[2] = in4f[0] * matrix[0][2] + in4f[1] * matrix[1][2] + in4f[2] * matrix[2][2] + in4f[3] * matrix[3][2];
1089                 out4f[3] = in4f[0] * matrix[0][3] + in4f[1] * matrix[1][3] + in4f[2] * matrix[2][3] + in4f[3] * matrix[3][3];
1090         }
1091 }
1092
1093 void DPSOFTRAST_Array_Copy(float *out4f, const float *in4f, int numitems)
1094 {
1095         memcpy(out4f, in4f, numitems * sizeof(float[4]));
1096 }
1097
1098 void DPSOFTRAST_Draw_ProjectVertices(float *out4f, const float *in4f, int numitems)
1099 {
1100         // NOTE: this is used both as a whole mesh transform function and a
1101         // per-triangle transform function (for clipped triangles), accordingly
1102         // it should not crash on divide by 0 but the result of divide by 0 is
1103         // unimportant...
1104         // TODO: SIMD
1105         int i;
1106         float w;
1107         float viewportcenter[4];
1108         float viewportscale[4];
1109         viewportscale[0] = dpsoftrast.fb_viewportscale[0];
1110         viewportscale[1] = dpsoftrast.fb_viewportscale[1];
1111         viewportscale[2] = 0.5f;
1112         viewportscale[3] = 0.0f;
1113         viewportcenter[0] = dpsoftrast.fb_viewportcenter[0];
1114         viewportcenter[1] = dpsoftrast.fb_viewportcenter[1];
1115         viewportcenter[2] = 0.5f;
1116         viewportcenter[3] = 0.0f;
1117         for (i = 0;i < numitems;i++)
1118         {
1119                 if (!in4f[3])
1120                 {
1121                         out4f[0] = 0.0f;
1122                         out4f[1] = 0.0f;
1123                         out4f[2] = 0.0f;
1124                         out4f[3] = 0.0f;
1125                         continue;
1126                 }
1127                 w = 1.0f / in4f[3];
1128                 out4f[0] = viewportcenter[0] + viewportscale[0] * in4f[0] * w;
1129                 out4f[1] = viewportcenter[1] + viewportscale[1] * in4f[1] * w;
1130                 out4f[2] = viewportcenter[2] + viewportscale[2] * in4f[2] * w;
1131                 out4f[3] = viewportcenter[3] + viewportscale[3] * in4f[3] * w;
1132                 out4f[3] = w;
1133                 in4f += 4;
1134                 out4f += 4;
1135         }
1136 }
1137
1138 void DPSOFTRAST_Draw_DebugEdgePoints(const float *screen0, const float *screen1)
1139 {
1140         int i;
1141         int x;
1142         int y;
1143         int w = dpsoftrast.fb_width;
1144         int bounds[4];
1145         float v0[2], v1[2];
1146         unsigned int *pixels = dpsoftrast.fb_colorpixels[0];
1147         //const float *c4f;
1148         bounds[0] = dpsoftrast.fb_viewportscissor[0];
1149         bounds[1] = dpsoftrast.fb_viewportscissor[1];
1150         bounds[2] = dpsoftrast.fb_viewportscissor[0] + dpsoftrast.fb_viewportscissor[2];
1151         bounds[3] = dpsoftrast.fb_viewportscissor[1] + dpsoftrast.fb_viewportscissor[3];
1152         v0[0] = screen0[0];
1153         v0[1] = screen0[1];
1154         v1[0] = screen1[0];
1155         v1[1] = screen1[1];
1156         for (i = 0;i <= 128;i++)
1157         {
1158                 // check nearclip
1159                 //if (dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][i*4+3] != 1.0f)
1160                 //      continue;
1161                 x = (int)(v0[0] + (v1[0] - v0[0]) * (i/128.0f));
1162                 y = (int)(v0[1] + (v1[1] - v0[1]) * (i/128.0f));
1163                 if (x < bounds[0] || y < bounds[1] || x >= bounds[2] || y >= bounds[3])
1164                         continue;
1165                 //c4f = dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_COLOR] + element0*4;
1166                 //pixels[y*w+x] = DPSOFTRAST_BGRA8_FROM_RGBA32F(c4f[0], c4f[1], c4f[2], c4f[3]);
1167                 pixels[y*w+x] = 0xFFFFFFFF;
1168         }
1169 }
1170
1171 void DPSOFTRAST_Draw_Span_Begin(const DPSOFTRAST_State_Draw_Span *span, float *zf)
1172 {
1173         int x;
1174         int startx = span->startx;
1175         int endx = span->endx;
1176         float w = span->data[0][DPSOFTRAST_ARRAY_TOTAL][3];
1177         float wslope = span->data[1][DPSOFTRAST_ARRAY_TOTAL][3];
1178         for (x = startx;x < endx;)
1179         {
1180                 int endsub = x + DPSOFTRAST_MAXSUBSPAN-1;
1181                 float z = 1.0f / (w + wslope * x), dz;
1182                 if (endsub >= endx)
1183                 {
1184                         endsub = endx-1;
1185                         dz = endsub > x ? (1.0f / (w + wslope * endsub) - z) / (endsub - x) : 0.0f;
1186                 }
1187                 else
1188                 {
1189                         dz = (1.0f / (w + wslope * endsub) - z) * (1.0f / (DPSOFTRAST_MAXSUBSPAN-1));
1190                 }
1191                 for (; x <= endsub; x++, z += dz)
1192                         zf[x] = z;
1193         }
1194 }
1195
1196 void DPSOFTRAST_Draw_Span_Finish(const DPSOFTRAST_State_Draw_Span *span, const float * RESTRICT in4f)
1197 {
1198         int x;
1199         int startx = span->startx;
1200         int endx = span->endx;
1201         int d[4];
1202         float a, b;
1203         unsigned char * RESTRICT pixelmask = span->pixelmask;
1204         unsigned char * RESTRICT pixel = (unsigned char *)dpsoftrast.fb_colorpixels[0];
1205         if (!pixel)
1206                 return;
1207         pixel += span->start * 4;
1208         // handle alphatest now (this affects depth writes too)
1209         if (dpsoftrast.user.alphatest)
1210                 for (x = startx;x < endx;x++)
1211                         if (in4f[x*4+3] < 0.5f)
1212                                 pixelmask[x] = false;
1213         // FIXME: this does not handle bigendian
1214         switch(dpsoftrast.fb_blendmode)
1215         {
1216         case DPSOFTRAST_BLENDMODE_OPAQUE:
1217                 for (x = startx;x < endx;x++)
1218                 {
1219                         if (!pixelmask[x])
1220                                 continue;
1221                         d[0] = (int)(in4f[x*4+2]*255.0f);if (d[0] > 255) d[0] = 255;
1222                         d[1] = (int)(in4f[x*4+1]*255.0f);if (d[1] > 255) d[1] = 255;
1223                         d[2] = (int)(in4f[x*4+0]*255.0f);if (d[2] > 255) d[2] = 255;
1224                         d[3] = (int)(in4f[x*4+3]*255.0f);if (d[3] > 255) d[3] = 255;
1225                         pixel[x*4+0] = d[0];
1226                         pixel[x*4+1] = d[1];
1227                         pixel[x*4+2] = d[2];
1228                         pixel[x*4+3] = d[3];
1229                 }
1230                 break;
1231         case DPSOFTRAST_BLENDMODE_ALPHA:
1232                 for (x = startx;x < endx;x++)
1233                 {
1234                         if (!pixelmask[x])
1235                                 continue;
1236                         a = in4f[x*4+3] * 255.0f;
1237                         b = 1.0f - in4f[x*4+3];
1238                         d[0] = (int)(in4f[x*4+2]*a+pixel[x*4+0]*b);if (d[0] > 255) d[0] = 255;
1239                         d[1] = (int)(in4f[x*4+1]*a+pixel[x*4+1]*b);if (d[1] > 255) d[1] = 255;
1240                         d[2] = (int)(in4f[x*4+0]*a+pixel[x*4+2]*b);if (d[2] > 255) d[2] = 255;
1241                         d[3] = (int)(in4f[x*4+3]*a+pixel[x*4+3]*b);if (d[3] > 255) d[3] = 255;
1242                         pixel[x*4+0] = d[0];
1243                         pixel[x*4+1] = d[1];
1244                         pixel[x*4+2] = d[2];
1245                         pixel[x*4+3] = d[3];
1246                 }
1247                 break;
1248         case DPSOFTRAST_BLENDMODE_ADDALPHA:
1249                 for (x = startx;x < endx;x++)
1250                 {
1251                         if (!pixelmask[x])
1252                                 continue;
1253                         a = in4f[x*4+3] * 255.0f;
1254                         d[0] = (int)(in4f[x*4+2]*a+pixel[x*4+0]);if (d[0] > 255) d[0] = 255;
1255                         d[1] = (int)(in4f[x*4+1]*a+pixel[x*4+1]);if (d[1] > 255) d[1] = 255;
1256                         d[2] = (int)(in4f[x*4+0]*a+pixel[x*4+2]);if (d[2] > 255) d[2] = 255;
1257                         d[3] = (int)(in4f[x*4+3]*a+pixel[x*4+3]);if (d[3] > 255) d[3] = 255;
1258                         pixel[x*4+0] = d[0];
1259                         pixel[x*4+1] = d[1];
1260                         pixel[x*4+2] = d[2];
1261                         pixel[x*4+3] = d[3];
1262                 }
1263                 break;
1264         case DPSOFTRAST_BLENDMODE_ADD:
1265                 for (x = startx;x < endx;x++)
1266                 {
1267                         if (!pixelmask[x])
1268                                 continue;
1269                         d[0] = (int)(in4f[x*4+2]*255.0f+pixel[x*4+0]);if (d[0] > 255) d[0] = 255;
1270                         d[1] = (int)(in4f[x*4+1]*255.0f+pixel[x*4+1]);if (d[1] > 255) d[1] = 255;
1271                         d[2] = (int)(in4f[x*4+0]*255.0f+pixel[x*4+2]);if (d[2] > 255) d[2] = 255;
1272                         d[3] = (int)(in4f[x*4+3]*255.0f+pixel[x*4+3]);if (d[3] > 255) d[3] = 255;
1273                         pixel[x*4+0] = d[0];
1274                         pixel[x*4+1] = d[1];
1275                         pixel[x*4+2] = d[2];
1276                         pixel[x*4+3] = d[3];
1277                 }
1278                 break;
1279         case DPSOFTRAST_BLENDMODE_INVMOD:
1280                 for (x = startx;x < endx;x++)
1281                 {
1282                         if (!pixelmask[x])
1283                                 continue;
1284                         d[0] = (int)((1.0f-in4f[x*4+2])*pixel[x*4+0]);if (d[0] > 255) d[0] = 255;
1285                         d[1] = (int)((1.0f-in4f[x*4+1])*pixel[x*4+1]);if (d[1] > 255) d[1] = 255;
1286                         d[2] = (int)((1.0f-in4f[x*4+0])*pixel[x*4+2]);if (d[2] > 255) d[2] = 255;
1287                         d[3] = (int)((1.0f-in4f[x*4+3])*pixel[x*4+3]);if (d[3] > 255) d[3] = 255;
1288                         pixel[x*4+0] = d[0];
1289                         pixel[x*4+1] = d[1];
1290                         pixel[x*4+2] = d[2];
1291                         pixel[x*4+3] = d[3];
1292                 }
1293                 break;
1294         case DPSOFTRAST_BLENDMODE_MUL:
1295                 for (x = startx;x < endx;x++)
1296                 {
1297                         if (!pixelmask[x])
1298                                 continue;
1299                         d[0] = (int)(in4f[x*4+2]*pixel[x*4+0]);if (d[0] > 255) d[0] = 255;
1300                         d[1] = (int)(in4f[x*4+1]*pixel[x*4+1]);if (d[1] > 255) d[1] = 255;
1301                         d[2] = (int)(in4f[x*4+0]*pixel[x*4+2]);if (d[2] > 255) d[2] = 255;
1302                         d[3] = (int)(in4f[x*4+3]*pixel[x*4+3]);if (d[3] > 255) d[3] = 255;
1303                         pixel[x*4+0] = d[0];
1304                         pixel[x*4+1] = d[1];
1305                         pixel[x*4+2] = d[2];
1306                         pixel[x*4+3] = d[3];
1307                 }
1308                 break;
1309         case DPSOFTRAST_BLENDMODE_MUL2:
1310                 for (x = startx;x < endx;x++)
1311                 {
1312                         if (!pixelmask[x])
1313                                 continue;
1314                         d[0] = (int)(in4f[x*4+2]*pixel[x*4+0]*2.0f);if (d[0] > 255) d[0] = 255;
1315                         d[1] = (int)(in4f[x*4+1]*pixel[x*4+1]*2.0f);if (d[1] > 255) d[1] = 255;
1316                         d[2] = (int)(in4f[x*4+0]*pixel[x*4+2]*2.0f);if (d[2] > 255) d[2] = 255;
1317                         d[3] = (int)(in4f[x*4+3]*pixel[x*4+3]*2.0f);if (d[3] > 255) d[3] = 255;
1318                         pixel[x*4+0] = d[0];
1319                         pixel[x*4+1] = d[1];
1320                         pixel[x*4+2] = d[2];
1321                         pixel[x*4+3] = d[3];
1322                 }
1323                 break;
1324         case DPSOFTRAST_BLENDMODE_SUBALPHA:
1325                 for (x = startx;x < endx;x++)
1326                 {
1327                         if (!pixelmask[x])
1328                                 continue;
1329                         a = in4f[x*4+3] * -255.0f;
1330                         d[0] = (int)(in4f[x*4+2]*a+pixel[x*4+0]);if (d[0] > 255) d[0] = 255;if (d[0] < 0) d[0] = 0;
1331                         d[1] = (int)(in4f[x*4+1]*a+pixel[x*4+1]);if (d[1] > 255) d[1] = 255;if (d[1] < 0) d[1] = 0;
1332                         d[2] = (int)(in4f[x*4+0]*a+pixel[x*4+2]);if (d[2] > 255) d[2] = 255;if (d[2] < 0) d[2] = 0;
1333                         d[3] = (int)(in4f[x*4+3]*a+pixel[x*4+3]);if (d[3] > 255) d[3] = 255;if (d[3] < 0) d[3] = 0;
1334                         pixel[x*4+0] = d[0];
1335                         pixel[x*4+1] = d[1];
1336                         pixel[x*4+2] = d[2];
1337                         pixel[x*4+3] = d[3];
1338                 }
1339                 break;
1340         case DPSOFTRAST_BLENDMODE_PSEUDOALPHA:
1341                 for (x = startx;x < endx;x++)
1342                 {
1343                         if (!pixelmask[x])
1344                                 continue;
1345                         a = 255.0f;
1346                         b = 1.0f - in4f[x*4+3];
1347                         d[0] = (int)(in4f[x*4+2]*a+pixel[x*4+0]*b);if (d[0] > 255) d[0] = 255;
1348                         d[1] = (int)(in4f[x*4+1]*a+pixel[x*4+1]*b);if (d[1] > 255) d[1] = 255;
1349                         d[2] = (int)(in4f[x*4+0]*a+pixel[x*4+2]*b);if (d[2] > 255) d[2] = 255;
1350                         d[3] = (int)(in4f[x*4+3]*a+pixel[x*4+3]*b);if (d[3] > 255) d[3] = 255;
1351                         pixel[x*4+0] = d[0];
1352                         pixel[x*4+1] = d[1];
1353                         pixel[x*4+2] = d[2];
1354                         pixel[x*4+3] = d[3];
1355                 }
1356                 break;
1357         }
1358 }
1359
1360 void DPSOFTRAST_Draw_Span_Texture2DVarying(const DPSOFTRAST_State_Draw_Span *span, float * RESTRICT out4f, int texunitindex, int arrayindex, const float * RESTRICT zf)
1361 {
1362         int x;
1363         int startx = span->startx;
1364         int endx = span->endx;
1365         int flags;
1366         float c[4];
1367         float data[4];
1368         float slope[4];
1369         float tc[2];
1370         float tcscale[2];
1371         unsigned int tci[2];
1372         unsigned int tci1[2];
1373         unsigned int tcimin[2];
1374         unsigned int tcimax[2];
1375         int tciwrapmask[2];
1376         int tciwidth;
1377         int filter;
1378         int mip;
1379         const unsigned char * RESTRICT pixelbase;
1380         const unsigned char * RESTRICT pixel[4];
1381         DPSOFTRAST_Texture *texture = dpsoftrast.texbound[texunitindex];
1382         // if no texture is bound, just fill it with white
1383         if (!texture)
1384         {
1385                 for (x = startx;x < endx;x++)
1386                 {
1387                         out4f[x*4+0] = 1.0f;
1388                         out4f[x*4+1] = 1.0f;
1389                         out4f[x*4+2] = 1.0f;
1390                         out4f[x*4+3] = 1.0f;
1391                 }
1392                 return;
1393         }
1394         // if this mipmap of the texture is 1 pixel, just fill it with that color
1395         mip = span->mip;
1396         if (mip >= texture->mipmaps)
1397                 mip = texture->mipmaps - 1;
1398         if (texture->mipmap[mip][1] == 4)
1399         {
1400                 c[0] = texture->bytes[2] * (1.0f/255.0f);
1401                 c[1] = texture->bytes[1] * (1.0f/255.0f);
1402                 c[2] = texture->bytes[0] * (1.0f/255.0f);
1403                 c[3] = texture->bytes[3] * (1.0f/255.0f);
1404                 for (x = startx;x < endx;x++)
1405                 {
1406                         out4f[x*4+0] = c[0];
1407                         out4f[x*4+1] = c[1];
1408                         out4f[x*4+2] = c[2];
1409                         out4f[x*4+3] = c[3];
1410                 }
1411                 return;
1412         }
1413         filter = texture->filter & DPSOFTRAST_TEXTURE_FILTER_LINEAR;
1414         data[0] = span->data[0][arrayindex][0];
1415         data[1] = span->data[0][arrayindex][1];
1416         data[2] = span->data[0][arrayindex][2];
1417         data[3] = span->data[0][arrayindex][3];
1418         slope[0] = span->data[1][arrayindex][0];
1419         slope[1] = span->data[1][arrayindex][1];
1420         slope[2] = span->data[1][arrayindex][2];
1421         slope[3] = span->data[1][arrayindex][3];
1422         flags = texture->flags;
1423         pixelbase = (unsigned char *)texture->bytes + texture->mipmap[mip][0];
1424         tcscale[0] = texture->mipmap[mip][2];
1425         tcscale[1] = texture->mipmap[mip][3];
1426         tciwidth = texture->mipmap[mip][2];
1427         tcimin[0] = 0;
1428         tcimin[1] = 0;
1429         tcimax[0] = texture->mipmap[mip][2]-1;
1430         tcimax[1] = texture->mipmap[mip][3]-1;
1431         tciwrapmask[0] = texture->mipmap[mip][2]-1;
1432         tciwrapmask[1] = texture->mipmap[mip][3]-1;
1433         for (x = startx;x < endx;)
1434         {
1435                 float endtc[2];
1436                 unsigned int subtc[2];
1437                 unsigned int substep[2];
1438                 int endsub = x + DPSOFTRAST_MAXSUBSPAN-1;
1439                 float subscale = 4096.0f/(DPSOFTRAST_MAXSUBSPAN-1);
1440                 if (endsub >= endx)
1441                 {
1442                         endsub = endx-1;
1443                         subscale = endsub > x ? 4096.0f / (endsub - x) : 1.0f;
1444                 }
1445                 tc[0] = (data[0] + slope[0]*x) * zf[x] * tcscale[0] - 0.5f;
1446                 tc[1] = (data[1] + slope[1]*x) * zf[x] * tcscale[1] - 0.5f;
1447                 endtc[0] = (data[0] + slope[0]*endsub) * zf[endsub] * tcscale[0] - 0.5f;
1448                 endtc[1] = (data[1] + slope[1]*endsub) * zf[endsub] * tcscale[1] - 0.5f;
1449                 substep[0] = (endtc[0] - tc[0]) * subscale;
1450                 substep[1] = (endtc[1] - tc[1]) * subscale;
1451                 subtc[0] = tc[0] * (1<<12);
1452                 subtc[1] = tc[1] * (1<<12);
1453                 if (!(flags & DPSOFTRAST_TEXTURE_FLAG_CLAMPTOEDGE))
1454                 {
1455                         subtc[0] &= (tciwrapmask[0]<<12)|0xFFF;
1456                         subtc[1] &= (tciwrapmask[1]<<12)|0xFFF;
1457                 }
1458                 if(filter)
1459                 {
1460                         tci[0] = (subtc[0]>>12) - tcimin[0];
1461                         tci[1] = (subtc[1]>>12) - tcimin[0];
1462                         tci1[0] = ((subtc[0] + (endsub - x)*substep[0])>>12) + 1;
1463                         tci1[1] = ((subtc[1] + (endsub - x)*substep[1])>>12) + 1;
1464                         if (tci[0] <= tcimax[0] && tci[1] <= tcimax[1] && tci1[0] <= tcimax[0] && tci1[1] <= tcimax[1])
1465                         {
1466                                 for (; x <= endsub; x++, subtc[0] += substep[0], subtc[1] += substep[1])
1467                                 {
1468                                         unsigned int frac[2] = { subtc[0]&0xFFF, subtc[1]&0xFFF };
1469                                         unsigned int ifrac[2] = { 0x1000 - frac[0], 0x1000 - frac[1] };
1470                                         unsigned int lerp[4] = { ifrac[0]*ifrac[1], frac[0]*ifrac[1], ifrac[0]*frac[1], frac[0]*frac[1] };
1471                                         tci[0] = subtc[0]>>12;
1472                                         tci[1] = subtc[1]>>12;
1473                                         pixel[0] = pixelbase + 4 * (tci[1]*tciwidth+tci[0]);
1474                                         pixel[1] = pixel[0] + 4 * tciwidth;
1475                                         c[0] = (pixel[0][2]*lerp[0]+pixel[0][4+2]*lerp[1]+pixel[1][2]*lerp[2]+pixel[1][4+2]*lerp[3]) * (1.0f / 0xFF000000);
1476                                         c[1] = (pixel[0][1]*lerp[0]+pixel[0][4+1]*lerp[1]+pixel[1][1]*lerp[2]+pixel[1][4+1]*lerp[3]) * (1.0f / 0xFF000000);
1477                                         c[2] = (pixel[0][0]*lerp[0]+pixel[0][4+0]*lerp[1]+pixel[1][0]*lerp[2]+pixel[1][4+0]*lerp[3]) * (1.0f / 0xFF000000);
1478                                         c[3] = (pixel[0][3]*lerp[0]+pixel[0][4+3]*lerp[1]+pixel[1][3]*lerp[2]+pixel[1][4+3]*lerp[3]) * (1.0f / 0xFF000000);
1479                                         out4f[x*4+0] = c[0];
1480                                         out4f[x*4+1] = c[1];
1481                                         out4f[x*4+2] = c[2];
1482                                         out4f[x*4+3] = c[3];
1483                                 }
1484                         }
1485                         else if (flags & DPSOFTRAST_TEXTURE_FLAG_CLAMPTOEDGE)
1486                         {
1487                                 for (; x <= endsub; x++, subtc[0] += substep[0], subtc[1] += substep[1])
1488                                 {
1489                                         unsigned int frac[2] = { subtc[0]&0xFFF, subtc[1]&0xFFF };
1490                                         unsigned int ifrac[2] = { 0x1000 - frac[0], 0x1000 - frac[1] };
1491                                         unsigned int lerp[4] = { ifrac[0]*ifrac[1], frac[0]*ifrac[1], ifrac[0]*frac[1], frac[0]*frac[1] };
1492                                         tci[0] = subtc[0]>>12;
1493                                         tci[1] = subtc[1]>>12;
1494                                         tci1[0] = tci[0] + 1;
1495                                         tci1[1] = tci[1] + 1;
1496                                         tci[0] = tci[0] >= tcimin[0] ? (tci[0] <= tcimax[0] ? tci[0] : tcimax[0]) : tcimin[0];
1497                                         tci[1] = tci[1] >= tcimin[1] ? (tci[1] <= tcimax[1] ? tci[1] : tcimax[1]) : tcimin[1];
1498                                         tci1[0] = tci1[0] >= tcimin[0] ? (tci1[0] <= tcimax[0] ? tci1[0] : tcimax[0]) : tcimin[0];
1499                                         tci1[1] = tci1[1] >= tcimin[1] ? (tci1[1] <= tcimax[1] ? tci1[1] : tcimax[1]) : tcimin[1];
1500                                         pixel[0] = pixelbase + 4 * (tci[1]*tciwidth+tci[0]);
1501                                         pixel[1] = pixelbase + 4 * (tci[1]*tciwidth+tci1[0]);
1502                                         pixel[2] = pixelbase + 4 * (tci1[1]*tciwidth+tci[0]);
1503                                         pixel[3] = pixelbase + 4 * (tci1[1]*tciwidth+tci1[0]);
1504                                         c[0] = (pixel[0][2]*lerp[0]+pixel[1][2]*lerp[1]+pixel[2][2]*lerp[2]+pixel[3][2]*lerp[3]) * (1.0f / 0xFF000000);
1505                                         c[1] = (pixel[0][1]*lerp[0]+pixel[1][1]*lerp[1]+pixel[2][1]*lerp[2]+pixel[3][1]*lerp[3]) * (1.0f / 0xFF000000);
1506                                         c[2] = (pixel[0][0]*lerp[0]+pixel[1][0]*lerp[1]+pixel[2][0]*lerp[2]+pixel[3][0]*lerp[3]) * (1.0f / 0xFF000000);
1507                                         c[3] = (pixel[0][3]*lerp[0]+pixel[1][3]*lerp[1]+pixel[2][3]*lerp[2]+pixel[3][3]*lerp[3]) * (1.0f / 0xFF000000);
1508                                         out4f[x*4+0] = c[0];
1509                                         out4f[x*4+1] = c[1];
1510                                         out4f[x*4+2] = c[2];
1511                                         out4f[x*4+3] = c[3];
1512                                 }
1513                         }
1514                         else
1515                         {
1516                                 for (; x <= endsub; x++, subtc[0] += substep[0], subtc[1] += substep[1])
1517                                 {
1518                                         unsigned int frac[2] = { subtc[0]&0xFFF, subtc[1]&0xFFF };
1519                                         unsigned int ifrac[2] = { 0x1000 - frac[0], 0x1000 - frac[1] };
1520                                         unsigned int lerp[4] = { ifrac[0]*ifrac[1], frac[0]*ifrac[1], ifrac[0]*frac[1], frac[0]*frac[1] };
1521                                         tci[0] = subtc[0]>>12;
1522                                         tci[1] = subtc[1]>>12;
1523                                         tci1[0] = tci[0] + 1;
1524                                         tci1[1] = tci[1] + 1;
1525                                         tci[0] &= tciwrapmask[0];
1526                                         tci[1] &= tciwrapmask[1];
1527                                         tci1[0] &= tciwrapmask[0];
1528                                         tci1[1] &= tciwrapmask[1];
1529                                         pixel[0] = pixelbase + 4 * (tci[1]*tciwidth+tci[0]);
1530                                         pixel[1] = pixelbase + 4 * (tci[1]*tciwidth+tci1[0]);
1531                                         pixel[2] = pixelbase + 4 * (tci1[1]*tciwidth+tci[0]);
1532                                         pixel[3] = pixelbase + 4 * (tci1[1]*tciwidth+tci1[0]);
1533                                         c[0] = (pixel[0][2]*lerp[0]+pixel[1][2]*lerp[1]+pixel[2][2]*lerp[2]+pixel[3][2]*lerp[3]) * (1.0f / 0xFF000000);
1534                                         c[1] = (pixel[0][1]*lerp[0]+pixel[1][1]*lerp[1]+pixel[2][1]*lerp[2]+pixel[3][1]*lerp[3]) * (1.0f / 0xFF000000);
1535                                         c[2] = (pixel[0][0]*lerp[0]+pixel[1][0]*lerp[1]+pixel[2][0]*lerp[2]+pixel[3][0]*lerp[3]) * (1.0f / 0xFF000000);
1536                                         c[3] = (pixel[0][3]*lerp[0]+pixel[1][3]*lerp[1]+pixel[2][3]*lerp[2]+pixel[3][3]*lerp[3]) * (1.0f / 0xFF000000);
1537                                         out4f[x*4+0] = c[0];
1538                                         out4f[x*4+1] = c[1];
1539                                         out4f[x*4+2] = c[2];
1540                                         out4f[x*4+3] = c[3];
1541                                 }
1542                         }
1543                 }
1544                 else if (flags & DPSOFTRAST_TEXTURE_FLAG_CLAMPTOEDGE)
1545                 {
1546                         for (; x <= endsub; x++, subtc[0] += substep[0], subtc[1] += substep[1])
1547                         {
1548                                 tci[0] = subtc[0]>>12;
1549                                 tci[1] = subtc[1]>>12;
1550                                 tci[0] = tci[0] >= tcimin[0] ? (tci[0] <= tcimax[0] ? tci[0] : tcimax[0]) : tcimin[0];
1551                                 tci[1] = tci[1] >= tcimin[1] ? (tci[1] <= tcimax[1] ? tci[1] : tcimax[1]) : tcimin[1];
1552                                 pixel[0] = pixelbase + 4 * (tci[1]*tciwidth+tci[0]);
1553                                 c[0] = pixel[0][2] * (1.0f / 255.0f);
1554                                 c[1] = pixel[0][1] * (1.0f / 255.0f);
1555                                 c[2] = pixel[0][0] * (1.0f / 255.0f);
1556                                 c[3] = pixel[0][3] * (1.0f / 255.0f);
1557                                 out4f[x*4+0] = c[0];
1558                                 out4f[x*4+1] = c[1];
1559                                 out4f[x*4+2] = c[2];
1560                                 out4f[x*4+3] = c[3];
1561                         }
1562                 }
1563                 else
1564                 {
1565                         for (; x <= endsub; x++, subtc[0] += substep[0], subtc[1] += substep[1])
1566                         {
1567                                 tci[0] = subtc[0]>>12;
1568                                 tci[1] = subtc[1]>>12;
1569                                 tci[0] &= tciwrapmask[0];
1570                                 tci[1] &= tciwrapmask[1];
1571                                 pixel[0] = pixelbase + 4 * (tci[1]*tciwidth+tci[0]);
1572                                 c[0] = pixel[0][2] * (1.0f / 255.0f);
1573                                 c[1] = pixel[0][1] * (1.0f / 255.0f);
1574                                 c[2] = pixel[0][0] * (1.0f / 255.0f);
1575                                 c[3] = pixel[0][3] * (1.0f / 255.0f);
1576                                 out4f[x*4+0] = c[0];
1577                                 out4f[x*4+1] = c[1];
1578                                 out4f[x*4+2] = c[2];
1579                                 out4f[x*4+3] = c[3];
1580                         }
1581                 }
1582         }
1583 }
1584
1585 void DPSOFTRAST_Draw_Span_MultiplyVarying(const DPSOFTRAST_State_Draw_Span *span, float *out4f, const float *in4f, int arrayindex, const float *zf)
1586 {
1587         int x;
1588         int startx = span->startx;
1589         int endx = span->endx;
1590         float c[4];
1591         float data[4];
1592         float slope[4];
1593         float z;
1594         data[0] = span->data[0][arrayindex][0];
1595         data[1] = span->data[0][arrayindex][1];
1596         data[2] = span->data[0][arrayindex][2];
1597         data[3] = span->data[0][arrayindex][3];
1598         slope[0] = span->data[1][arrayindex][0];
1599         slope[1] = span->data[1][arrayindex][1];
1600         slope[2] = span->data[1][arrayindex][2];
1601         slope[3] = span->data[1][arrayindex][3];
1602         for (x = startx;x < endx;x++)
1603         {
1604                 z = zf[x];
1605                 c[0] = (data[0] + slope[0]*x) * z;
1606                 c[1] = (data[1] + slope[1]*x) * z;
1607                 c[2] = (data[2] + slope[2]*x) * z;
1608                 c[3] = (data[3] + slope[3]*x) * z;
1609                 out4f[x*4+0] = in4f[x*4+0] * c[0];
1610                 out4f[x*4+1] = in4f[x*4+1] * c[1];
1611                 out4f[x*4+2] = in4f[x*4+2] * c[2];
1612                 out4f[x*4+3] = in4f[x*4+3] * c[3];
1613         }
1614 }
1615
1616 void DPSOFTRAST_Draw_Span_AddBloom(const DPSOFTRAST_State_Draw_Span *span, float *out4f, const float *ina4f, const float *inb4f, const float *subcolor)
1617 {
1618         int x, startx = span->startx, endx = span->endx;
1619         float c[4], localcolor[4];
1620         localcolor[0] = subcolor[0];
1621         localcolor[1] = subcolor[1];
1622         localcolor[2] = subcolor[2];
1623         localcolor[3] = subcolor[3];
1624         for (x = startx;x < endx;x++)
1625         {
1626                 c[0] = inb4f[x*4+0] - localcolor[0];if (c[0] < 0.0f) c[0] = 0.0f;
1627                 c[1] = inb4f[x*4+1] - localcolor[1];if (c[1] < 0.0f) c[1] = 0.0f;
1628                 c[2] = inb4f[x*4+2] - localcolor[2];if (c[2] < 0.0f) c[2] = 0.0f;
1629                 c[3] = inb4f[x*4+3] - localcolor[3];if (c[3] < 0.0f) c[3] = 0.0f;
1630                 out4f[x*4+0] = ina4f[x*4+0] + c[0];
1631                 out4f[x*4+1] = ina4f[x*4+1] + c[1];
1632                 out4f[x*4+2] = ina4f[x*4+2] + c[2];
1633                 out4f[x*4+3] = ina4f[x*4+3] + c[3];
1634         }
1635 }
1636
1637 void DPSOFTRAST_Draw_Span_MixUniformColor(const DPSOFTRAST_State_Draw_Span *span, float *out4f, const float *in4f, const float *color)
1638 {
1639         int x, startx = span->startx, endx = span->endx;
1640         float localcolor[4], ilerp, lerp;
1641         localcolor[0] = color[0];
1642         localcolor[1] = color[1];
1643         localcolor[2] = color[2];
1644         localcolor[3] = color[3];
1645         ilerp = 1.0f - localcolor[3];
1646         lerp = localcolor[3];
1647         for (x = startx;x < endx;x++)
1648         {
1649                 out4f[x*4+0] = in4f[x*4+0] * ilerp + localcolor[0] * lerp;
1650                 out4f[x*4+1] = in4f[x*4+1] * ilerp + localcolor[1] * lerp;
1651                 out4f[x*4+2] = in4f[x*4+2] * ilerp + localcolor[2] * lerp;
1652                 out4f[x*4+3] = in4f[x*4+3] * ilerp + localcolor[3] * lerp;
1653         }
1654 }
1655
1656 void DPSOFTRAST_Draw_Span_Lightmap(const DPSOFTRAST_State_Draw_Span *span, float * RESTRICT out4f, const float * RESTRICT diffuse, const float * RESTRICT lightmap)
1657 {
1658         int x, startx = span->startx, endx = span->endx;
1659         float Color_Ambient[4], Color_Diffuse[4];
1660         Color_Ambient[0] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Ambient*4+0];
1661         Color_Ambient[1] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Ambient*4+1];
1662         Color_Ambient[2] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Ambient*4+2];
1663         Color_Ambient[3] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Alpha*4+0];
1664         Color_Diffuse[0] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Diffuse*4+0];
1665         Color_Diffuse[1] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Diffuse*4+1];
1666         Color_Diffuse[2] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Diffuse*4+2];
1667         Color_Diffuse[3] = 0.0f;
1668         for (x = startx;x < endx;x++)
1669         {
1670                 out4f[x*4+0] = diffuse[x*4+0] * (Color_Ambient[0] + lightmap[x*4+0] * Color_Diffuse[0]);
1671                 out4f[x*4+1] = diffuse[x*4+1] * (Color_Ambient[1] + lightmap[x*4+1] * Color_Diffuse[1]);
1672                 out4f[x*4+2] = diffuse[x*4+2] * (Color_Ambient[2] + lightmap[x*4+2] * Color_Diffuse[2]);
1673                 out4f[x*4+3] = diffuse[x*4+3] * (Color_Ambient[3] + lightmap[x*4+3] * Color_Diffuse[3]);
1674         }
1675 }
1676
1677 void DPSOFTRAST_Draw_Span_Lightmap_Finish(const DPSOFTRAST_State_Draw_Span *span, const float * RESTRICT diffuse, const float * RESTRICT lightmap)
1678 {
1679         int x, startx = span->startx, endx = span->endx;
1680         int d[4];
1681         float Color_Ambient[4], Color_Diffuse[4];
1682         unsigned char * RESTRICT pixelmask = span->pixelmask;
1683         unsigned char * RESTRICT pixel = (unsigned char *)dpsoftrast.fb_colorpixels[0];
1684         if (!pixel)
1685                 return;
1686         pixel += span->start * 4;
1687         Color_Ambient[0] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Ambient*4+0]*255.0f;
1688         Color_Ambient[1] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Ambient*4+1]*255.0f;
1689         Color_Ambient[2] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Ambient*4+2]*255.0f;
1690         Color_Ambient[3] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Alpha*4+0]*255.0f;
1691         Color_Diffuse[0] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Diffuse*4+0]*255.0f;
1692         Color_Diffuse[1] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Diffuse*4+1]*255.0f;
1693         Color_Diffuse[2] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Diffuse*4+2]*255.0f;
1694         Color_Diffuse[3] = 0.0f;
1695         for (x = startx;x < endx;x++)
1696         {
1697                 if (!pixelmask[x])
1698                         continue;
1699                 d[0] = diffuse[x*4+0] * (Color_Ambient[0] + lightmap[x*4+0] * Color_Diffuse[0]);if (d[0] > 255) d[0] = 255;
1700                 d[1] = diffuse[x*4+1] * (Color_Ambient[1] + lightmap[x*4+1] * Color_Diffuse[1]);if (d[1] > 255) d[1] = 255;
1701                 d[2] = diffuse[x*4+2] * (Color_Ambient[2] + lightmap[x*4+2] * Color_Diffuse[2]);if (d[2] > 255) d[2] = 255;
1702                 d[3] = diffuse[x*4+3] * (Color_Ambient[3] + lightmap[x*4+3] * Color_Diffuse[3]);if (d[3] > 255) d[3] = 255;
1703                 pixel[x*4+0] = d[2];
1704                 pixel[x*4+1] = d[1];
1705                 pixel[x*4+2] = d[0];
1706                 pixel[x*4+3] = d[3];
1707         }
1708 }
1709
1710 void DPSOFTRAST_Draw_Span_VertexColor(const DPSOFTRAST_State_Draw_Span *span, float *out4f, const float *diffuse, const float *zf)
1711 {
1712         int x, startx = span->startx, endx = span->endx;
1713         float Color_Ambient[4], Color_Diffuse[4];
1714         float c[4];
1715         float data[4];
1716         float slope[4];
1717         float z;
1718         int arrayindex = DPSOFTRAST_ARRAY_COLOR;
1719         data[0] = span->data[0][arrayindex][0];
1720         data[1] = span->data[0][arrayindex][1];
1721         data[2] = span->data[0][arrayindex][2];
1722         data[3] = span->data[0][arrayindex][3];
1723         slope[0] = span->data[1][arrayindex][0];
1724         slope[1] = span->data[1][arrayindex][1];
1725         slope[2] = span->data[1][arrayindex][2];
1726         slope[3] = span->data[1][arrayindex][3];
1727         Color_Ambient[0] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Ambient*4+0];
1728         Color_Ambient[1] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Ambient*4+1];
1729         Color_Ambient[2] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Ambient*4+2];
1730         Color_Ambient[3] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Alpha*4+0];
1731         Color_Diffuse[0] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Diffuse*4+0];
1732         Color_Diffuse[1] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Diffuse*4+1];
1733         Color_Diffuse[2] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Diffuse*4+2];
1734         Color_Diffuse[3] = 0.0f;
1735         for (x = startx;x < endx;x++)
1736         {
1737                 z = zf[x];
1738                 c[0] = (data[0] + slope[0]*x) * z;
1739                 c[1] = (data[1] + slope[1]*x) * z;
1740                 c[2] = (data[2] + slope[2]*x) * z;
1741                 c[3] = (data[3] + slope[3]*x) * z;
1742                 out4f[x*4+0] = diffuse[x*4+0] * (Color_Ambient[0] + c[0] * Color_Diffuse[0]);
1743                 out4f[x*4+1] = diffuse[x*4+1] * (Color_Ambient[1] + c[1] * Color_Diffuse[1]);
1744                 out4f[x*4+2] = diffuse[x*4+2] * (Color_Ambient[2] + c[2] * Color_Diffuse[2]);
1745                 out4f[x*4+3] = diffuse[x*4+3] * (Color_Ambient[3] + c[3] * Color_Diffuse[3]);
1746         }
1747 }
1748
1749 void DPSOFTRAST_Draw_Span_FlatColor(const DPSOFTRAST_State_Draw_Span *span, float *out4f, const float *diffuse)
1750 {
1751         int x, startx = span->startx, endx = span->endx;
1752         float Color_Ambient[4];
1753         Color_Ambient[0] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Ambient*4+0];
1754         Color_Ambient[1] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Ambient*4+1];
1755         Color_Ambient[2] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Color_Ambient*4+2];
1756         Color_Ambient[3] = dpsoftrast.uniform4f[DPSOFTRAST_UNIFORM_Alpha*4+0];
1757         for (x = startx;x < endx;x++)
1758         {
1759                 out4f[x*4+0] = diffuse[x*4+0] * Color_Ambient[0];
1760                 out4f[x*4+1] = diffuse[x*4+1] * Color_Ambient[1];
1761                 out4f[x*4+2] = diffuse[x*4+2] * Color_Ambient[2];
1762                 out4f[x*4+3] = diffuse[x*4+3] * Color_Ambient[3];
1763         }
1764 }
1765
1766 void DPSOFTRAST_Draw_VertexShader(void)
1767 {
1768         DPSOFTRAST_Array_Transform(dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION], dpsoftrast.draw.in_array4f[DPSOFTRAST_ARRAY_POSITION], dpsoftrast.draw.numvertices, dpsoftrast.uniform4f + 4*DPSOFTRAST_UNIFORM_ModelViewProjectionMatrixM1);
1769         switch(dpsoftrast.shader_mode)
1770         {
1771         case SHADERMODE_GENERIC: ///< (particles/HUD/etc) vertex color: optionally multiplied by one texture
1772                 DPSOFTRAST_Array_Copy(dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_COLOR], dpsoftrast.draw.in_array4f[DPSOFTRAST_ARRAY_COLOR], dpsoftrast.draw.numvertices);
1773                 DPSOFTRAST_Array_Copy(dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_TEXCOORD0], dpsoftrast.draw.in_array4f[DPSOFTRAST_ARRAY_TEXCOORD0], dpsoftrast.draw.numvertices);
1774                 break;
1775         case SHADERMODE_POSTPROCESS: ///< postprocessing shader (r_glsl_postprocess)
1776                 DPSOFTRAST_Array_Copy(dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_TEXCOORD0], dpsoftrast.draw.in_array4f[DPSOFTRAST_ARRAY_TEXCOORD0], dpsoftrast.draw.numvertices);
1777                 DPSOFTRAST_Array_Copy(dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_TEXCOORD1], dpsoftrast.draw.in_array4f[DPSOFTRAST_ARRAY_TEXCOORD1], dpsoftrast.draw.numvertices);
1778                 break;
1779         case SHADERMODE_DEPTH_OR_SHADOW: ///< (depthfirst/shadows) vertex shader only
1780                 break;
1781         case SHADERMODE_FLATCOLOR: ///< (lightmap) modulate texture by uniform color (q1bsp: q3bsp)
1782                 DPSOFTRAST_Array_Copy(dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_TEXCOORD0], dpsoftrast.draw.in_array4f[DPSOFTRAST_ARRAY_TEXCOORD0], dpsoftrast.draw.numvertices);
1783                 break;
1784         case SHADERMODE_VERTEXCOLOR: ///< (lightmap) modulate texture by vertex colors (q3bsp)
1785                 DPSOFTRAST_Array_Copy(dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_COLOR], dpsoftrast.draw.in_array4f[DPSOFTRAST_ARRAY_COLOR], dpsoftrast.draw.numvertices);
1786                 DPSOFTRAST_Array_Copy(dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_TEXCOORD0], dpsoftrast.draw.in_array4f[DPSOFTRAST_ARRAY_TEXCOORD0], dpsoftrast.draw.numvertices);
1787                 break;
1788         case SHADERMODE_LIGHTMAP: ///< (lightmap) modulate texture by lightmap texture (q1bsp: q3bsp)
1789                 DPSOFTRAST_Array_Copy(dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_TEXCOORD0], dpsoftrast.draw.in_array4f[DPSOFTRAST_ARRAY_TEXCOORD0], dpsoftrast.draw.numvertices);
1790                 DPSOFTRAST_Array_Copy(dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_TEXCOORD4], dpsoftrast.draw.in_array4f[DPSOFTRAST_ARRAY_TEXCOORD4], dpsoftrast.draw.numvertices);
1791                 break;
1792         case SHADERMODE_FAKELIGHT: ///< (fakelight) modulate texture by "fake" lighting (no lightmaps: no nothing)
1793                 break;
1794         case SHADERMODE_LIGHTDIRECTIONMAP_MODELSPACE: ///< (lightmap) use directional pixel shading from texture containing modelspace light directions (q3bsp deluxemap)
1795                 break;
1796         case SHADERMODE_LIGHTDIRECTIONMAP_TANGENTSPACE: ///< (lightmap) use directional pixel shading from texture containing tangentspace light directions (q1bsp deluxemap)
1797                 break;
1798         case SHADERMODE_LIGHTDIRECTION: ///< (lightmap) use directional pixel shading from fixed light direction (q3bsp)
1799                 break;
1800         case SHADERMODE_LIGHTSOURCE: ///< (lightsource) use directional pixel shading from light source (rtlight)
1801                 break;
1802         case SHADERMODE_REFRACTION: ///< refract background (the material is rendered normally after this pass)
1803                 break;
1804         case SHADERMODE_WATER: ///< refract background and reflection (the material is rendered normally after this pass)
1805                 break;
1806         case SHADERMODE_SHOWDEPTH: ///< (debugging) renders depth as color
1807                 break;
1808         case SHADERMODE_DEFERREDGEOMETRY: ///< (deferred) render material properties to screenspace geometry buffers
1809                 break;
1810         case SHADERMODE_DEFERREDLIGHTSOURCE: ///< (deferred) use directional pixel shading from light source (rtlight) on screenspace geometry buffers
1811                 break;
1812         case SHADERMODE_COUNT:
1813                 break;
1814         }
1815 }
1816
1817 void DPSOFTRAST_Draw_PixelShaderSpan(const DPSOFTRAST_State_Draw_Span *span)
1818 {
1819         float buffer_z[DPSOFTRAST_DRAW_MAXSPANLENGTH];
1820         float buffer_texture_color[DPSOFTRAST_DRAW_MAXSPANLENGTH*4];
1821         float buffer_texture_lightmap[DPSOFTRAST_DRAW_MAXSPANLENGTH*4];
1822         float buffer_FragColor[DPSOFTRAST_DRAW_MAXSPANLENGTH*4];
1823         switch(dpsoftrast.shader_mode)
1824         {
1825         case SHADERMODE_GENERIC: ///< (particles/HUD/etc) vertex color: optionally multiplied by one texture
1826                 DPSOFTRAST_Draw_Span_Begin(span, buffer_z);
1827                 DPSOFTRAST_Draw_Span_Texture2DVarying(span, buffer_texture_color, GL20TU_FIRST, 2, buffer_z);
1828                 DPSOFTRAST_Draw_Span_MultiplyVarying(span, buffer_FragColor, buffer_texture_color, 1, buffer_z);
1829                 DPSOFTRAST_Draw_Span_Finish(span, buffer_FragColor);
1830                 break;
1831         case SHADERMODE_POSTPROCESS: ///< postprocessing shader (r_glsl_postprocess)
1832                 // TODO: optimize!!  at the very least there is no reason to use texture sampling on the frame texture
1833                 DPSOFTRAST_Draw_Span_Begin(span, buffer_z);
1834                 DPSOFTRAST_Draw_Span_Texture2DVarying(span, buffer_FragColor, GL20TU_FIRST, 2, buffer_z);
1835                 if (dpsoftrast.shader_permutation & SHADERPERMUTATION_BLOOM)
1836                 {
1837                         DPSOFTRAST_Draw_Span_Texture2DVarying(span, buffer_texture_color, GL20TU_SECOND, 3, buffer_z);
1838                         DPSOFTRAST_Draw_Span_AddBloom(span, buffer_FragColor, buffer_FragColor, buffer_texture_color, dpsoftrast.uniform4f + DPSOFTRAST_UNIFORM_BloomColorSubtract * 4);
1839                 }
1840                 DPSOFTRAST_Draw_Span_MixUniformColor(span, buffer_FragColor, buffer_FragColor, dpsoftrast.uniform4f + DPSOFTRAST_UNIFORM_ViewTintColor * 4);
1841                 if (dpsoftrast.shader_permutation & SHADERPERMUTATION_SATURATION)
1842                 {
1843                         // TODO: implement saturation
1844                 }
1845                 if (dpsoftrast.shader_permutation & SHADERPERMUTATION_GAMMARAMPS)
1846                 {
1847                         // TODO: implement gammaramps
1848                 }
1849                 DPSOFTRAST_Draw_Span_Finish(span, buffer_FragColor);
1850                 break;
1851         case SHADERMODE_DEPTH_OR_SHADOW: ///< (depthfirst/shadows) vertex shader only
1852                 break;
1853         case SHADERMODE_FLATCOLOR: ///< (lightmap) modulate texture by uniform color (q1bsp: q3bsp)
1854                 DPSOFTRAST_Draw_Span_Begin(span, buffer_z);
1855                 DPSOFTRAST_Draw_Span_Texture2DVarying(span, buffer_texture_color, GL20TU_COLOR, 2, buffer_z);
1856                 DPSOFTRAST_Draw_Span_FlatColor(span, buffer_FragColor, buffer_texture_color);
1857                 DPSOFTRAST_Draw_Span_Finish(span, buffer_FragColor);
1858                 break;
1859         case SHADERMODE_VERTEXCOLOR: ///< (lightmap) modulate texture by vertex colors (q3bsp)
1860                 DPSOFTRAST_Draw_Span_Begin(span, buffer_z);
1861                 DPSOFTRAST_Draw_Span_Texture2DVarying(span, buffer_texture_color, GL20TU_COLOR, 2, buffer_z);
1862                 DPSOFTRAST_Draw_Span_VertexColor(span, buffer_FragColor, buffer_texture_color, buffer_z);
1863                 DPSOFTRAST_Draw_Span_Finish(span, buffer_FragColor);
1864                 break;
1865         case SHADERMODE_LIGHTMAP: ///< (lightmap) modulate texture by lightmap texture (q1bsp: q3bsp)
1866                 DPSOFTRAST_Draw_Span_Begin(span, buffer_z);
1867                 DPSOFTRAST_Draw_Span_Texture2DVarying(span, buffer_texture_color, GL20TU_COLOR, 2, buffer_z);
1868                 DPSOFTRAST_Draw_Span_Texture2DVarying(span, buffer_texture_lightmap, GL20TU_LIGHTMAP, 6, buffer_z);
1869                 if(!dpsoftrast.user.alphatest && dpsoftrast.fb_blendmode == DPSOFTRAST_BLENDMODE_OPAQUE)
1870                 {
1871                         DPSOFTRAST_Draw_Span_Lightmap_Finish(span, buffer_texture_color, buffer_texture_lightmap);
1872                 }
1873                 else
1874                 {
1875                         DPSOFTRAST_Draw_Span_Lightmap(span, buffer_FragColor, buffer_texture_color, buffer_texture_lightmap);
1876                         DPSOFTRAST_Draw_Span_Finish(span, buffer_FragColor);
1877                 }
1878                 break;
1879         case SHADERMODE_FAKELIGHT: ///< (fakelight) modulate texture by "fake" lighting (no lightmaps: no nothing)
1880                 break;
1881         case SHADERMODE_LIGHTDIRECTIONMAP_MODELSPACE: ///< (lightmap) use directional pixel shading from texture containing modelspace light directions (q3bsp deluxemap)
1882                 break;
1883         case SHADERMODE_LIGHTDIRECTIONMAP_TANGENTSPACE: ///< (lightmap) use directional pixel shading from texture containing tangentspace light directions (q1bsp deluxemap)
1884                 break;
1885         case SHADERMODE_LIGHTDIRECTION: ///< (lightmap) use directional pixel shading from fixed light direction (q3bsp)
1886                 break;
1887         case SHADERMODE_LIGHTSOURCE: ///< (lightsource) use directional pixel shading from light source (rtlight)
1888                 break;
1889         case SHADERMODE_REFRACTION: ///< refract background (the material is rendered normally after this pass)
1890                 break;
1891         case SHADERMODE_WATER: ///< refract background and reflection (the material is rendered normally after this pass)
1892                 break;
1893         case SHADERMODE_SHOWDEPTH: ///< (debugging) renders depth as color
1894                 break;
1895         case SHADERMODE_DEFERREDGEOMETRY: ///< (deferred) render material properties to screenspace geometry buffers
1896                 break;
1897         case SHADERMODE_DEFERREDLIGHTSOURCE: ///< (deferred) use directional pixel shading from light source (rtlight) on screenspace geometry buffers
1898                 break;
1899         case SHADERMODE_COUNT:
1900                 break;
1901         }
1902 }
1903
1904 void DPSOFTRAST_Draw_ProcessSpans(void)
1905 {
1906         int i;
1907         int x;
1908         int startx;
1909         int endx;
1910         int numspans = dpsoftrast.draw.numspans;
1911 //      unsigned int c;
1912 //      unsigned int *colorpixel;
1913         unsigned int *depthpixel;
1914         float w;
1915         float wslope;
1916         int depth;
1917         int depthslope;
1918         unsigned int d;
1919         DPSOFTRAST_State_Draw_Span *span = dpsoftrast.draw.spanqueue;
1920         unsigned char pixelmask[DPSOFTRAST_DRAW_MAXSPANLENGTH];
1921         for (i = 0;i < numspans;i++, span++)
1922         {
1923                 w = span->data[0][DPSOFTRAST_ARRAY_TOTAL][3];
1924                 wslope = span->data[1][DPSOFTRAST_ARRAY_TOTAL][3];
1925                 if (dpsoftrast.user.depthtest && dpsoftrast.fb_depthpixels)
1926                 {
1927                         depth = (int)(w*DPSOFTRAST_DEPTHSCALE);
1928                         depthslope = (int)(wslope*DPSOFTRAST_DEPTHSCALE);
1929                         depthpixel = dpsoftrast.fb_depthpixels + span->start;
1930                         switch(dpsoftrast.fb_depthfunc)
1931                         {
1932                         default:
1933                         case GL_ALWAYS:  for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = true; break;
1934                         case GL_LESS:    for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = depthpixel[x] < d; break;
1935                         case GL_LEQUAL:  for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = depthpixel[x] <= d; break;
1936                         case GL_EQUAL:   for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = depthpixel[x] == d; break;
1937                         case GL_GEQUAL:  for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = depthpixel[x] >= d; break;
1938                         case GL_GREATER: for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = depthpixel[x] > d; break;
1939                         case GL_NEVER:   for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = false; break;
1940                         }
1941                         //colorpixel = dpsoftrast.fb_colorpixels[0] + span->start;
1942                         //for (x = 0;x < span->length;x++)
1943                         //      colorpixel[x] = (depthpixel[x] & 0xFF000000) ? (0x00FF0000) : (depthpixel[x] & 0x00FF0000);
1944                         // if there is no color buffer, skip pixel shader
1945                         startx = 0;
1946                         endx = span->length;
1947                         while (startx < endx && !pixelmask[startx])
1948                                 startx++;
1949                         while (endx > startx && !pixelmask[endx-1])
1950                                 endx--;
1951                         if (startx >= endx)
1952                                 continue; // no pixels to fill
1953                         span->pixelmask = pixelmask;
1954                         span->startx = startx;
1955                         span->endx = endx;
1956                         // run pixel shader if appropriate
1957                         // do this before running depthmask code, to allow the pixelshader
1958                         // to clear pixelmask values for alpha testing
1959                         if (dpsoftrast.fb_colorpixels[0] && dpsoftrast.fb_colormask)
1960                                 DPSOFTRAST_Draw_PixelShaderSpan(span);
1961                         if (dpsoftrast.user.depthmask)
1962                                 for (x = 0, d = depth;x < span->length;x++, d += depthslope)
1963                                         if (pixelmask[x])
1964                                                 depthpixel[x] = d;
1965                 }
1966                 else
1967                 {
1968                         // no depth testing means we're just dealing with color...
1969                         // if there is no color buffer, skip pixel shader
1970                         if (dpsoftrast.fb_colorpixels[0] && dpsoftrast.fb_colormask)
1971                         {
1972                                 memset(pixelmask, 1, span->length);
1973                                 span->pixelmask = pixelmask;
1974                                 span->startx = 0;
1975                                 span->endx = span->length;
1976                                 DPSOFTRAST_Draw_PixelShaderSpan(span);
1977                         }
1978                 }
1979         }
1980 }
1981
1982 void DPSOFTRAST_Draw_ProcessTriangles(int firstvertex, int numvertices, int numtriangles, const int *element3i, const unsigned short *element3s, unsigned char *arraymask)
1983 {
1984         int cullface = dpsoftrast.user.cullface;
1985         int width = dpsoftrast.fb_width;
1986         int height = dpsoftrast.fb_height;
1987         int i;
1988         int j;
1989         int k;
1990         int y;
1991         int mip;
1992         int e[3];
1993         int screenx[4];
1994         int screeny[4];
1995         int screenyless[4];
1996         int numpoints;
1997         int clipflags;
1998         int edge0p;
1999         int edge0n;
2000         int edge1p;
2001         int edge1n;
2002         int extent[4];
2003         int startx;
2004         int endx;
2005         float startxf;
2006         float endxf;
2007         float edge0ylerp;
2008         float edge0yilerp;
2009         float edge1ylerp;
2010         float edge1yilerp;
2011         float edge0xf;
2012         float edge1xf;
2013         float spanilength;
2014         float startxlerp;
2015         float yc;
2016         float w;
2017         float frac;
2018         float ifrac;
2019         float trianglearea2;
2020         float triangleedge[2][4];
2021         float trianglenormal[4];
2022         float clipdist[4];
2023         float clipped[DPSOFTRAST_ARRAY_TOTAL][4][4];
2024         float screen[4][4];
2025         float proj[DPSOFTRAST_ARRAY_TOTAL][4][4];
2026         DPSOFTRAST_State_Draw_Span *span;
2027         DPSOFTRAST_State_Draw_Span *oldspan;
2028         for (i = 0;i < numtriangles;i++)
2029         {
2030                 // generate the 3 edges of this triangle
2031                 // generate spans for the triangle - switch based on left split or right split classification of triangle
2032                 if (element3i)
2033                 {
2034                         e[0] = element3i[i*3+0] - firstvertex;
2035                         e[1] = element3i[i*3+1] - firstvertex;
2036                         e[2] = element3i[i*3+2] - firstvertex;
2037                 }
2038                 else if (element3s)
2039                 {
2040                         e[0] = element3s[i*3+0] - firstvertex;
2041                         e[1] = element3s[i*3+1] - firstvertex;
2042                         e[2] = element3s[i*3+2] - firstvertex;
2043                 }
2044                 else
2045                 {
2046                         e[0] = i*3+0;
2047                         e[1] = i*3+1;
2048                         e[2] = i*3+2;
2049                 }
2050                 triangleedge[0][0] = dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[0]*4+0] - dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[1]*4+0];
2051                 triangleedge[0][1] = dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[0]*4+1] - dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[1]*4+1];
2052                 triangleedge[0][2] = dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[0]*4+2] - dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[1]*4+2];
2053                 triangleedge[1][0] = dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[2]*4+0] - dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[1]*4+0];
2054                 triangleedge[1][1] = dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[2]*4+1] - dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[1]*4+1];
2055                 triangleedge[1][2] = dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[2]*4+2] - dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[1]*4+2];
2056                 trianglenormal[0] = triangleedge[0][1] * triangleedge[1][2] - triangleedge[0][2] * triangleedge[1][1];
2057                 trianglenormal[1] = triangleedge[0][2] * triangleedge[1][0] - triangleedge[0][0] * triangleedge[1][2];
2058                 trianglenormal[2] = triangleedge[0][0] * triangleedge[1][1] - triangleedge[0][1] * triangleedge[1][0];
2059                 trianglearea2 = trianglenormal[0] * trianglenormal[0] + trianglenormal[1] * trianglenormal[1] + trianglenormal[2] * trianglenormal[2];
2060                 // skip degenerate triangles, nothing good can come from them...
2061                 if (trianglearea2 == 0.0f)
2062                         continue;
2063                 // apply current cullface mode (this culls many triangles)
2064                 switch(cullface)
2065                 {
2066                 case GL_BACK:
2067                         if (trianglenormal[2] < 0)
2068                                 continue;
2069                         break;
2070                 case GL_FRONT:
2071                         if (trianglenormal[2] > 0)
2072                                 continue;
2073                         break;
2074                 }
2075                 // calculate distance from nearplane
2076                 clipdist[0] = dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[0]*4+2] + 1.0f;
2077                 clipdist[1] = dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[1]*4+2] + 1.0f;
2078                 clipdist[2] = dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][e[2]*4+2] + 1.0f;
2079                 clipflags = 0;
2080                 if (clipdist[0] < 0.0f)
2081                         clipflags |= 1;
2082                 if (clipdist[1] < 0.0f)
2083                         clipflags |= 2;
2084                 if (clipdist[2] < 0.0f)
2085                         clipflags |= 4;
2086                 // clip triangle if necessary
2087                 switch(clipflags)
2088                 {
2089                 case 0: /*000*/
2090                         // triangle is entirely in front of nearplane
2091
2092                         // macros for clipping vertices
2093 #define CLIPPEDVERTEXLERP(k,p1,p2) \
2094                         frac = clipdist[p1] / (clipdist[p1] - clipdist[p2]);\
2095                         ifrac = 1.0f - frac;\
2096                         for (j = 0;j < DPSOFTRAST_ARRAY_TOTAL;j++)\
2097                         {\
2098                                 if (arraymask[j])\
2099                                 {\
2100                                         clipped[j][k][0] = dpsoftrast.draw.post_array4f[j][e[p1]*4+0]*ifrac+dpsoftrast.draw.post_array4f[j][e[p2]*4+0]*frac;\
2101                                         clipped[j][k][1] = dpsoftrast.draw.post_array4f[j][e[p1]*4+1]*ifrac+dpsoftrast.draw.post_array4f[j][e[p2]*4+1]*frac;\
2102                                         clipped[j][k][2] = dpsoftrast.draw.post_array4f[j][e[p1]*4+2]*ifrac+dpsoftrast.draw.post_array4f[j][e[p2]*4+2]*frac;\
2103                                         clipped[j][k][3] = dpsoftrast.draw.post_array4f[j][e[p1]*4+3]*ifrac+dpsoftrast.draw.post_array4f[j][e[p2]*4+3]*frac;\
2104                                 }\
2105                         }\
2106                         DPSOFTRAST_Draw_ProjectVertices(screen[k], clipped[DPSOFTRAST_ARRAY_POSITION][k], 1)
2107 #define CLIPPEDVERTEXCOPY(k,p1) \
2108                         for (j = 0;j < DPSOFTRAST_ARRAY_TOTAL;j++)\
2109                         {\
2110                                 if (arraymask[j])\
2111                                 {\
2112                                         clipped[j][k][0] = dpsoftrast.draw.post_array4f[j][e[p1]*4+0];\
2113                                         clipped[j][k][1] = dpsoftrast.draw.post_array4f[j][e[p1]*4+1];\
2114                                         clipped[j][k][2] = dpsoftrast.draw.post_array4f[j][e[p1]*4+2];\
2115                                         clipped[j][k][3] = dpsoftrast.draw.post_array4f[j][e[p1]*4+3];\
2116                                 }\
2117                         }\
2118                         screen[k][0] = dpsoftrast.draw.screencoord4f[e[p1]*4+0];\
2119                         screen[k][1] = dpsoftrast.draw.screencoord4f[e[p1]*4+1];\
2120                         screen[k][2] = dpsoftrast.draw.screencoord4f[e[p1]*4+2];\
2121                         screen[k][3] = dpsoftrast.draw.screencoord4f[e[p1]*4+3];
2122
2123                         CLIPPEDVERTEXCOPY(0,0);
2124                         CLIPPEDVERTEXCOPY(1,1);
2125                         CLIPPEDVERTEXCOPY(2,2);
2126                         numpoints = 3;
2127                         break;
2128                 case 1: /*100*/
2129                         CLIPPEDVERTEXLERP(0,0,1);
2130                         CLIPPEDVERTEXCOPY(1,1);
2131                         CLIPPEDVERTEXCOPY(2,2);
2132                         CLIPPEDVERTEXLERP(3,2,0);
2133                         numpoints = 4;
2134                         break;
2135                 case 2: /*010*/
2136                         CLIPPEDVERTEXCOPY(0,0);
2137                         CLIPPEDVERTEXLERP(1,0,1);
2138                         CLIPPEDVERTEXLERP(2,1,2);
2139                         CLIPPEDVERTEXCOPY(3,2);
2140                         numpoints = 4;
2141                         break;
2142                 case 3: /*110*/
2143                         CLIPPEDVERTEXLERP(0,1,2);
2144                         CLIPPEDVERTEXCOPY(1,2);
2145                         CLIPPEDVERTEXLERP(2,2,0);
2146                         numpoints = 3;
2147                         break;
2148                 case 4: /*001*/
2149                         CLIPPEDVERTEXCOPY(0,0);
2150                         CLIPPEDVERTEXCOPY(1,1);
2151                         CLIPPEDVERTEXLERP(2,1,2);
2152                         CLIPPEDVERTEXLERP(3,2,0);
2153                         numpoints = 4;
2154                         break;
2155                 case 5: /*101*/
2156                         CLIPPEDVERTEXLERP(0,0,1);
2157                         CLIPPEDVERTEXCOPY(1,1);
2158                         CLIPPEDVERTEXLERP(2,1,2);
2159                         numpoints = 3;
2160                         break;
2161                 case 6: /*011*/
2162                         CLIPPEDVERTEXCOPY(0,0);
2163                         CLIPPEDVERTEXLERP(1,0,1);
2164                         CLIPPEDVERTEXLERP(2,2,0);
2165                         numpoints = 3;
2166                         break;
2167                 case 7: /*111*/
2168                         // triangle is entirely behind nearplane
2169                         continue;
2170                 }
2171                 // calculate integer y coords for triangle points
2172                 screenx[0] = (int)(screen[0][0]);
2173                 screeny[0] = (int)(screen[0][1]);
2174                 screenx[1] = (int)(screen[1][0]);
2175                 screeny[1] = (int)(screen[1][1]);
2176                 screenx[2] = (int)(screen[2][0]);
2177                 screeny[2] = (int)(screen[2][1]);
2178                 screenx[3] = (int)(screen[3][0]);
2179                 screeny[3] = (int)(screen[3][1]);
2180                 // figure out the extents (bounding box) of the triangle
2181                 extent[0] = screenx[0];
2182                 extent[1] = screeny[0];
2183                 extent[2] = screenx[0];
2184                 extent[3] = screeny[0];
2185                 for (j = 1;j < numpoints;j++)
2186                 {
2187                         if (extent[0] > screenx[j]) extent[0] = screenx[j];
2188                         if (extent[1] > screeny[j]) extent[1] = screeny[j];
2189                         if (extent[2] < screenx[j]) extent[2] = screenx[j];
2190                         if (extent[3] < screeny[j]) extent[3] = screeny[j];
2191                 }
2192                 //extent[0]--;
2193                 //extent[1]--;
2194                 extent[2]++;
2195                 extent[3]++;
2196                 if (extent[0] < 0)
2197                         extent[0] = 0;
2198                 if (extent[1] < 0)
2199                         extent[1] = 0;
2200                 if (extent[2] > width)
2201                         extent[2] = width;
2202                 if (extent[3] > height)
2203                         extent[3] = height;
2204                 // skip offscreen triangles
2205                 if (extent[2] <= extent[0] || extent[3] <= extent[1])
2206                         continue;
2207                 // TODO: adjust texture LOD by texture density
2208                 mip = 0;
2209                 // okay, this triangle is going to produce spans, we'd better project
2210                 // the interpolants now (this is what gives perspective texturing),
2211                 // this consists of simply multiplying all arrays by the W coord
2212                 // (which is basically 1/Z), which will be undone per-pixel
2213                 // (multiplying by Z again) to get the perspective-correct array
2214                 // values
2215                 for (j = 0;j < DPSOFTRAST_ARRAY_TOTAL;j++)
2216                 {
2217                         if (arraymask[j])
2218                         {
2219                                 for (k = 0;k < numpoints;k++)
2220                                 {
2221                                         w = screen[k][3];
2222                                         proj[j][k][0] = clipped[j][k][0] * w;
2223                                         proj[j][k][1] = clipped[j][k][1] * w;
2224                                         proj[j][k][2] = clipped[j][k][2] * w;
2225                                         proj[j][k][3] = clipped[j][k][3] * w;
2226                                 }
2227                         }
2228                 }
2229                 // iterate potential spans
2230                 // TODO: optimize?  if we figured out the edge order beforehand, this
2231                 //       could do loops over the edges in the proper order rather than
2232                 //       selecting them for each span
2233                 // TODO: optimize?  the edges could have data slopes calculated
2234                 // TODO: optimize?  the data slopes could be calculated as a plane
2235                 //       (2D slopes) to avoid any interpolation along edges at all
2236                 for (y = extent[1];y < extent[3];y++)
2237                 {
2238                         // get center of pixel y
2239                         yc = y;
2240                         // do the compares all at once
2241                         screenyless[0] = y <= screeny[0];
2242                         screenyless[1] = y <= screeny[1];
2243                         screenyless[2] = y <= screeny[2];
2244                         screenyless[3] = y <= screeny[3];
2245                         if (numpoints == 4)
2246                         {
2247                                 switch(screenyless[0] + screenyless[1] * 2 + screenyless[2] * 4 + screenyless[3] * 8)
2248                                 {
2249                                 case  0: /*0000*/ continue;
2250                                 case  1: /*1000*/ edge0p = 3;edge0n = 0;edge1p = 0;edge1n = 1;break;
2251                                 case  2: /*0100*/ edge0p = 0;edge0n = 1;edge1p = 1;edge1n = 2;break;
2252                                 case  3: /*1100*/ edge0p = 3;edge0n = 0;edge1p = 1;edge1n = 2;break;
2253                                 case  4: /*0010*/ edge0p = 1;edge0n = 2;edge1p = 2;edge1n = 3;break;
2254                                 case  5: /*1010*/ edge0p = 1;edge0n = 2;edge1p = 2;edge1n = 3;break; // concave - nonsense
2255                                 case  6: /*0110*/ edge0p = 0;edge0n = 1;edge1p = 2;edge1n = 3;break;
2256                                 case  7: /*1110*/ edge0p = 3;edge0n = 0;edge1p = 2;edge1n = 3;break;
2257                                 case  8: /*0001*/ edge0p = 2;edge0n = 3;edge1p = 3;edge1n = 0;break;
2258                                 case  9: /*1001*/ edge0p = 2;edge0n = 3;edge1p = 0;edge1n = 1;break;
2259                                 case 10: /*0101*/ edge0p = 2;edge0n = 3;edge1p = 1;edge1n = 2;break; // concave - nonsense
2260                                 case 11: /*1101*/ edge0p = 2;edge0n = 3;edge1p = 1;edge1n = 2;break;
2261                                 case 12: /*0011*/ edge0p = 1;edge0n = 2;edge1p = 3;edge1n = 0;break;
2262                                 case 13: /*1011*/ edge0p = 1;edge0n = 2;edge1p = 0;edge1n = 1;break;
2263                                 case 14: /*0111*/ edge0p = 0;edge0n = 1;edge1p = 3;edge1n = 0;break;
2264                                 case 15: /*1111*/ continue;
2265                                 }
2266                         }
2267                         else
2268                         {
2269                                 switch(screenyless[0] + screenyless[1] * 2 + screenyless[2] * 4)
2270                                 {
2271                                 case 0: /*000*/ continue;
2272                                 case 1: /*100*/ edge0p = 2;edge0n = 0;edge1p = 0;edge1n = 1;break;
2273                                 case 2: /*010*/ edge0p = 0;edge0n = 1;edge1p = 1;edge1n = 2;break;
2274                                 case 3: /*110*/ edge0p = 2;edge0n = 0;edge1p = 1;edge1n = 2;break;
2275                                 case 4: /*001*/ edge0p = 1;edge0n = 2;edge1p = 2;edge1n = 0;break;
2276                                 case 5: /*101*/ edge0p = 1;edge0n = 2;edge1p = 0;edge1n = 1;break;
2277                                 case 6: /*011*/ edge0p = 0;edge0n = 1;edge1p = 2;edge1n = 0;break;
2278                                 case 7: /*111*/ continue;
2279                                 }
2280                         }
2281 #if 0
2282                 {
2283                         int foundedges = 0;
2284                         int cedge0p = 0;
2285                         int cedge0n = 0;
2286                         int cedge1p = 0;
2287                         int cedge1n = 0;
2288                         for (j = 0, k = numpoints-1;j < numpoints;k = j, j++)
2289                         {
2290                                 if (screenyless[k] && !screenyless[j])
2291                                 {
2292                                         cedge1p = k;
2293                                         cedge1n = j;
2294                                         foundedges |= 1;
2295                                 }
2296                                 else if (screenyless[j] && !screenyless[k])
2297                                 {
2298                                         cedge0p = k;
2299                                         cedge0n = j;
2300                                         foundedges |= 2;
2301                                 }
2302                         }
2303                         if (foundedges != 3)
2304                                 continue;
2305                         if (cedge0p != edge0p || cedge0n != edge0n || cedge1p != edge1p || cedge1n != edge1n)
2306                         {
2307                                 if (numpoints == 4)
2308                                         printf("case %i%i%i%i is broken %i %i %i %i != %i %i %i %i\n", screenyless[0], screenyless[1], screenyless[2], screenyless[3], cedge0p, cedge0n, cedge1p, cedge1n, edge0p, edge0n, edge1p, edge1n);
2309                                 else
2310                                         printf("case %i%i%i is broken %i %i %i %i != %i %i %i %i\n", screenyless[0], screenyless[1], screenyless[2], cedge0p, cedge0n, cedge1p, cedge1n, edge0p, edge0n, edge1p, edge1n);
2311                         }
2312                 }
2313 #endif
2314                         edge0ylerp = (yc - screen[edge0p][1]) / (screen[edge0n][1] - screen[edge0p][1]);
2315                         edge1ylerp = (yc - screen[edge1p][1]) / (screen[edge1n][1] - screen[edge1p][1]);
2316                         if (edge0ylerp < 0 || edge0ylerp > 1 || edge1ylerp < 0 || edge1ylerp > 1)
2317                                 continue;
2318                         edge0yilerp = 1.0f - edge0ylerp;
2319                         edge1yilerp = 1.0f - edge1ylerp;
2320                         edge0xf = screen[edge0p][0] * edge0yilerp + screen[edge0n][0] * edge0ylerp;
2321                         edge1xf = screen[edge1p][0] * edge1yilerp + screen[edge1n][0] * edge1ylerp;
2322                         if (edge0xf < edge1xf)
2323                         {
2324                                 startxf = edge0xf;
2325                                 endxf = edge1xf;
2326                         }
2327                         else
2328                         {
2329                                 startxf = edge1xf;
2330                                 endxf = edge0xf;
2331                         }
2332                         startx = (int)ceil(startxf);
2333                         endx = (int)ceil(endxf);
2334                         if (startx < 0)
2335                                 startx = 0;
2336                         if (endx > width)
2337                                 endx = width;
2338                         if (startx >= endx)
2339                                 continue;
2340                         if (startxf > startx || endxf < endx-1) { printf("%s:%i X wrong (%i to %i is outside %f to %f)\n", __FILE__, __LINE__, startx, endx, startxf, endxf); }
2341                         spanilength = 1.0f / (endxf - startxf);
2342                         startxlerp = startx - startxf;
2343                         span = &dpsoftrast.draw.spanqueue[dpsoftrast.draw.numspans++];
2344                         span->mip = mip;
2345                         span->start = y * width + startx;
2346                         span->length = endx - startx;
2347                         j = DPSOFTRAST_ARRAY_TOTAL;
2348                         if (edge0xf < edge1xf)
2349                         {
2350                                 span->data[0][j][0] = screen[edge0p][0] * edge0yilerp + screen[edge0n][0] * edge0ylerp;
2351                                 span->data[0][j][1] = screen[edge0p][1] * edge0yilerp + screen[edge0n][1] * edge0ylerp;
2352                                 span->data[0][j][2] = screen[edge0p][2] * edge0yilerp + screen[edge0n][2] * edge0ylerp;
2353                                 span->data[0][j][3] = screen[edge0p][3] * edge0yilerp + screen[edge0n][3] * edge0ylerp;
2354                                 span->data[1][j][0] = screen[edge1p][0] * edge1yilerp + screen[edge1n][0] * edge1ylerp;
2355                                 span->data[1][j][1] = screen[edge1p][1] * edge1yilerp + screen[edge1n][1] * edge1ylerp;
2356                                 span->data[1][j][2] = screen[edge1p][2] * edge1yilerp + screen[edge1n][2] * edge1ylerp;
2357                                 span->data[1][j][3] = screen[edge1p][3] * edge1yilerp + screen[edge1n][3] * edge1ylerp;
2358                                 for (j = 0;j < DPSOFTRAST_ARRAY_TOTAL;j++)
2359                                 {
2360                                         if (arraymask[j])
2361                                         {
2362                                                 span->data[0][j][0] = proj[j][edge0p][0] * edge0yilerp + proj[j][edge0n][0] * edge0ylerp;
2363                                                 span->data[0][j][1] = proj[j][edge0p][1] * edge0yilerp + proj[j][edge0n][1] * edge0ylerp;
2364                                                 span->data[0][j][2] = proj[j][edge0p][2] * edge0yilerp + proj[j][edge0n][2] * edge0ylerp;
2365                                                 span->data[0][j][3] = proj[j][edge0p][3] * edge0yilerp + proj[j][edge0n][3] * edge0ylerp;
2366                                                 span->data[1][j][0] = proj[j][edge1p][0] * edge1yilerp + proj[j][edge1n][0] * edge1ylerp;
2367                                                 span->data[1][j][1] = proj[j][edge1p][1] * edge1yilerp + proj[j][edge1n][1] * edge1ylerp;
2368                                                 span->data[1][j][2] = proj[j][edge1p][2] * edge1yilerp + proj[j][edge1n][2] * edge1ylerp;
2369                                                 span->data[1][j][3] = proj[j][edge1p][3] * edge1yilerp + proj[j][edge1n][3] * edge1ylerp;
2370                                         }
2371                                 }
2372                         }
2373                         else
2374                         {
2375                                 span->data[0][j][0] = screen[edge1p][0] * edge1yilerp + screen[edge1n][0] * edge1ylerp;
2376                                 span->data[0][j][1] = screen[edge1p][1] * edge1yilerp + screen[edge1n][1] * edge1ylerp;
2377                                 span->data[0][j][2] = screen[edge1p][2] * edge1yilerp + screen[edge1n][2] * edge1ylerp;
2378                                 span->data[0][j][3] = screen[edge1p][3] * edge1yilerp + screen[edge1n][3] * edge1ylerp;
2379                                 span->data[1][j][0] = screen[edge0p][0] * edge0yilerp + screen[edge0n][0] * edge0ylerp;
2380                                 span->data[1][j][1] = screen[edge0p][1] * edge0yilerp + screen[edge0n][1] * edge0ylerp;
2381                                 span->data[1][j][2] = screen[edge0p][2] * edge0yilerp + screen[edge0n][2] * edge0ylerp;
2382                                 span->data[1][j][3] = screen[edge0p][3] * edge0yilerp + screen[edge0n][3] * edge0ylerp;
2383                                 for (j = 0;j < DPSOFTRAST_ARRAY_TOTAL;j++)
2384                                 {
2385                                         if (arraymask[j])
2386                                         {
2387                                                 span->data[0][j][0] = proj[j][edge1p][0] * edge1yilerp + proj[j][edge1n][0] * edge1ylerp;
2388                                                 span->data[0][j][1] = proj[j][edge1p][1] * edge1yilerp + proj[j][edge1n][1] * edge1ylerp;
2389                                                 span->data[0][j][2] = proj[j][edge1p][2] * edge1yilerp + proj[j][edge1n][2] * edge1ylerp;
2390                                                 span->data[0][j][3] = proj[j][edge1p][3] * edge1yilerp + proj[j][edge1n][3] * edge1ylerp;
2391                                                 span->data[1][j][0] = proj[j][edge0p][0] * edge0yilerp + proj[j][edge0n][0] * edge0ylerp;
2392                                                 span->data[1][j][1] = proj[j][edge0p][1] * edge0yilerp + proj[j][edge0n][1] * edge0ylerp;
2393                                                 span->data[1][j][2] = proj[j][edge0p][2] * edge0yilerp + proj[j][edge0n][2] * edge0ylerp;
2394                                                 span->data[1][j][3] = proj[j][edge0p][3] * edge0yilerp + proj[j][edge0n][3] * edge0ylerp;
2395                                         }
2396                                 }
2397                         }
2398                         // change data[1][n][] to be a data slope
2399                         j = DPSOFTRAST_ARRAY_TOTAL;
2400                         span->data[1][j][0] = (span->data[1][j][0] - span->data[0][j][0]) * spanilength;
2401                         span->data[1][j][1] = (span->data[1][j][1] - span->data[0][j][1]) * spanilength;
2402                         span->data[1][j][2] = (span->data[1][j][2] - span->data[0][j][2]) * spanilength;
2403                         span->data[1][j][3] = (span->data[1][j][3] - span->data[0][j][3]) * spanilength;
2404                         for (j = 0;j < DPSOFTRAST_ARRAY_TOTAL;j++)
2405                         {
2406                                 if (arraymask[j])
2407                                 {
2408                                         span->data[1][j][0] = (span->data[1][j][0] - span->data[0][j][0]) * spanilength;
2409                                         span->data[1][j][1] = (span->data[1][j][1] - span->data[0][j][1]) * spanilength;
2410                                         span->data[1][j][2] = (span->data[1][j][2] - span->data[0][j][2]) * spanilength;
2411                                         span->data[1][j][3] = (span->data[1][j][3] - span->data[0][j][3]) * spanilength;
2412                                 }
2413                         }
2414                         // adjust the data[0][n][] to be correct for the pixel centers
2415                         // this also handles horizontal clipping where a major part of the
2416                         // span may be off the left side of the screen
2417                         j = DPSOFTRAST_ARRAY_TOTAL;
2418                         span->data[0][j][0] += span->data[1][j][0] * startxlerp;
2419                         span->data[0][j][1] += span->data[1][j][1] * startxlerp;
2420                         span->data[0][j][2] += span->data[1][j][2] * startxlerp;
2421                         span->data[0][j][3] += span->data[1][j][3] * startxlerp;
2422                         for (j = 0;j < DPSOFTRAST_ARRAY_TOTAL;j++)
2423                         {
2424                                 if (arraymask[j])
2425                                 {
2426                                         span->data[0][j][0] += span->data[1][j][0] * startxlerp;
2427                                         span->data[0][j][1] += span->data[1][j][1] * startxlerp;
2428                                         span->data[0][j][2] += span->data[1][j][2] * startxlerp;
2429                                         span->data[0][j][3] += span->data[1][j][3] * startxlerp;
2430                                 }
2431                         }
2432                         // to keep the shader routines from needing more than a small
2433                         // buffer for pixel intermediate data, we split long spans...
2434                         while (span->length > DPSOFTRAST_DRAW_MAXSPANLENGTH)
2435                         {
2436                                 span->length = DPSOFTRAST_DRAW_MAXSPANLENGTH;
2437                                 if (dpsoftrast.draw.numspans >= DPSOFTRAST_DRAW_MAXSPANQUEUE)
2438                                 {
2439                                         DPSOFTRAST_Draw_ProcessSpans();
2440                                         dpsoftrast.draw.numspans = 0;
2441                                 }
2442                                 oldspan = span;
2443                                 span = &dpsoftrast.draw.spanqueue[dpsoftrast.draw.numspans++];
2444                                 *span = *oldspan;
2445                                 startx += DPSOFTRAST_DRAW_MAXSPANLENGTH;
2446                                 span->start = y * width + startx;
2447                                 span->length = endx - startx;
2448                                 j = DPSOFTRAST_ARRAY_TOTAL;
2449                                 span->data[0][j][0] += span->data[1][j][0] * DPSOFTRAST_DRAW_MAXSPANLENGTH;
2450                                 span->data[0][j][1] += span->data[1][j][1] * DPSOFTRAST_DRAW_MAXSPANLENGTH;
2451                                 span->data[0][j][2] += span->data[1][j][2] * DPSOFTRAST_DRAW_MAXSPANLENGTH;
2452                                 span->data[0][j][3] += span->data[1][j][3] * DPSOFTRAST_DRAW_MAXSPANLENGTH;
2453                                 for (j = 0;j < DPSOFTRAST_ARRAY_TOTAL;j++)
2454                                 {
2455                                         if (arraymask[j])
2456                                         {
2457                                                 span->data[0][j][0] += span->data[1][j][0] * DPSOFTRAST_DRAW_MAXSPANLENGTH;
2458                                                 span->data[0][j][1] += span->data[1][j][1] * DPSOFTRAST_DRAW_MAXSPANLENGTH;
2459                                                 span->data[0][j][2] += span->data[1][j][2] * DPSOFTRAST_DRAW_MAXSPANLENGTH;
2460                                                 span->data[0][j][3] += span->data[1][j][3] * DPSOFTRAST_DRAW_MAXSPANLENGTH;
2461                                         }
2462                                 }
2463                         }
2464                         // after all that, we have a span suitable for the pixel shader...
2465                         if (dpsoftrast.draw.numspans >= DPSOFTRAST_DRAW_MAXSPANQUEUE)
2466                         {
2467                                 DPSOFTRAST_Draw_ProcessSpans();
2468                                 dpsoftrast.draw.numspans = 0;
2469                         }
2470                 }
2471                 // draw outlines over triangle for debugging
2472         //      for (j = 0, k = numpoints-1;j < numpoints;k = j, j++)
2473         //              DPSOFTRAST_Draw_DebugEdgePoints(screen[k], screen[j]);
2474         }
2475         if (dpsoftrast.draw.numspans)
2476         {
2477                 DPSOFTRAST_Draw_ProcessSpans();
2478                 dpsoftrast.draw.numspans = 0;
2479         }
2480 }
2481
2482 void DPSOFTRAST_Draw_DebugPoints(void)
2483 {
2484         int i;
2485         int x;
2486         int y;
2487         int numvertices = dpsoftrast.draw.numvertices;
2488         int w = dpsoftrast.fb_width;
2489         int bounds[4];
2490         unsigned int *pixels = dpsoftrast.fb_colorpixels[0];
2491         const float *c4f;
2492         bounds[0] = dpsoftrast.fb_viewportscissor[0];
2493         bounds[1] = dpsoftrast.fb_viewportscissor[1];
2494         bounds[2] = dpsoftrast.fb_viewportscissor[0] + dpsoftrast.fb_viewportscissor[2];
2495         bounds[3] = dpsoftrast.fb_viewportscissor[1] + dpsoftrast.fb_viewportscissor[3];
2496         for (i = 0;i < numvertices;i++)
2497         {
2498                 // check nearclip
2499                 //if (dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][i*4+3] != 1.0f)
2500                 //      continue;
2501                 x = (int)(dpsoftrast.draw.screencoord4f[i*4+0]);
2502                 y = (int)(dpsoftrast.draw.screencoord4f[i*4+1]);
2503                 //x = (int)(dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][i*4+0] + 0.5f);
2504                 //y = (int)(dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][i*4+1] + 0.5f);
2505                 //x = (int)((dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][i*4+0] + 1.0f) * dpsoftrast.fb_width * 0.5f + 0.5f);
2506                 //y = (int)((dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION][i*4+1] + 1.0f) * dpsoftrast.fb_height * 0.5f + 0.5f);
2507                 if (x < bounds[0] || y < bounds[1] || x >= bounds[2] || y >= bounds[3])
2508                         continue;
2509                 c4f = dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_COLOR] + i*4;
2510                 pixels[y*w+x] = DPSOFTRAST_BGRA8_FROM_RGBA32F(c4f[0], c4f[1], c4f[2], c4f[3]);
2511         }
2512 }
2513
2514 void DPSOFTRAST_DrawTriangles(int firstvertex, int numvertices, int numtriangles, const int *element3i, const unsigned short *element3s)
2515 {
2516         unsigned char arraymask[DPSOFTRAST_ARRAY_TOTAL];
2517         arraymask[0] = true;
2518         arraymask[1] = dpsoftrast.fb_colorpixels[0] != NULL; // TODO: optimize (decide based on shadermode)
2519         arraymask[2] = dpsoftrast.pointer_texcoordf[0] != NULL;
2520         arraymask[3] = dpsoftrast.pointer_texcoordf[1] != NULL;
2521         arraymask[4] = dpsoftrast.pointer_texcoordf[2] != NULL;
2522         arraymask[5] = dpsoftrast.pointer_texcoordf[3] != NULL;
2523         arraymask[6] = dpsoftrast.pointer_texcoordf[4] != NULL;
2524         arraymask[7] = dpsoftrast.pointer_texcoordf[5] != NULL;
2525         arraymask[8] = dpsoftrast.pointer_texcoordf[6] != NULL;
2526         arraymask[9] = dpsoftrast.pointer_texcoordf[7] != NULL;
2527         DPSOFTRAST_Validate(DPSOFTRAST_VALIDATE_DRAW);
2528         DPSOFTRAST_Draw_LoadVertices(firstvertex, numvertices, true);
2529         DPSOFTRAST_Draw_VertexShader();
2530         DPSOFTRAST_Draw_ProjectVertices(dpsoftrast.draw.screencoord4f, dpsoftrast.draw.post_array4f[DPSOFTRAST_ARRAY_POSITION], numvertices);
2531         DPSOFTRAST_Draw_ProcessTriangles(firstvertex, numvertices, numtriangles, element3i, element3s, arraymask);
2532 }
2533
2534 void DPSOFTRAST_Init(int width, int height, unsigned int *colorpixels, unsigned int *depthpixels)
2535 {
2536         union
2537         {
2538                 int i;
2539                 unsigned char b[4];
2540         }
2541         u;
2542         u.i = 1;
2543         memset(&dpsoftrast, 0, sizeof(dpsoftrast));
2544         dpsoftrast.bigendian = u.b[3];
2545         dpsoftrast.fb_width = width;
2546         dpsoftrast.fb_height = height;
2547         dpsoftrast.fb_depthpixels = depthpixels;
2548         dpsoftrast.fb_colorpixels[0] = colorpixels;
2549         dpsoftrast.fb_colorpixels[1] = NULL;
2550         dpsoftrast.fb_colorpixels[1] = NULL;
2551         dpsoftrast.fb_colorpixels[1] = NULL;
2552         dpsoftrast.texture_firstfree = 1;
2553         dpsoftrast.texture_end = 1;
2554         dpsoftrast.texture_max = 0;
2555         dpsoftrast.user.colormask[0] = 1;
2556         dpsoftrast.user.colormask[1] = 1;
2557         dpsoftrast.user.colormask[2] = 1;
2558         dpsoftrast.user.colormask[3] = 1;
2559         dpsoftrast.user.blendfunc[0] = GL_ONE;
2560         dpsoftrast.user.blendfunc[1] = GL_ZERO;
2561         dpsoftrast.user.depthmask = true;
2562         dpsoftrast.user.depthtest = true;
2563         dpsoftrast.user.depthfunc = GL_LEQUAL;
2564         dpsoftrast.user.scissortest = false;
2565         dpsoftrast.user.cullface = GL_BACK;
2566         dpsoftrast.user.alphatest = false;
2567         dpsoftrast.user.alphafunc = GL_GREATER;
2568         dpsoftrast.user.alphavalue = 0.5f;
2569         dpsoftrast.user.scissor[0] = 0;
2570         dpsoftrast.user.scissor[1] = 0;
2571         dpsoftrast.user.scissor[2] = dpsoftrast.fb_width;
2572         dpsoftrast.user.scissor[3] = dpsoftrast.fb_height;
2573         dpsoftrast.user.viewport[0] = 0;
2574         dpsoftrast.user.viewport[1] = 0;
2575         dpsoftrast.user.viewport[2] = dpsoftrast.fb_width;
2576         dpsoftrast.user.viewport[3] = dpsoftrast.fb_height;
2577         dpsoftrast.user.depthrange[0] = 0;
2578         dpsoftrast.user.depthrange[1] = 1;
2579         dpsoftrast.user.polygonoffset[0] = 0;
2580         dpsoftrast.user.polygonoffset[1] = 0;
2581         dpsoftrast.user.color[0] = 1;
2582         dpsoftrast.user.color[1] = 1;
2583         dpsoftrast.user.color[2] = 1;
2584         dpsoftrast.user.color[3] = 1;
2585         dpsoftrast.validate = -1;
2586         DPSOFTRAST_Validate(-1);
2587         dpsoftrast.validate = 0;
2588 }
2589
2590 void DPSOFTRAST_Shutdown(void)
2591 {
2592         int i;
2593         for (i = 0;i < dpsoftrast.texture_end;i++)
2594                 if (dpsoftrast.texture[i].bytes)
2595                         free(dpsoftrast.texture[i].bytes);
2596         if (dpsoftrast.texture)
2597                 free(dpsoftrast.texture);
2598         memset(&dpsoftrast, 0, sizeof(dpsoftrast));
2599 }