git.oblomov.eu Git - nouveau/blob - src/nv50_exa.c

   1 /*
   2  * Copyright 2007 NVIDIA, Corporation
   3  * Copyright 2008 Ben Skeggs
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be included in
  13  * all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  19  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  20  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  */
  23
  24 #include "nv_include.h"
  25 #include "nv_rop.h"
  26
  27 #include "nv50_accel.h"
  28 #include "nv50_texture.h"
  29
  30 struct nv50_exa_state {
  31         Bool have_mask;
  32
  33         struct {
  34                 PictTransformPtr transform;
  35                 float width;
  36                 float height;
  37         } unit[2];
  38 };
  39 static struct nv50_exa_state exa_state;
  40
  41 #define NV50EXA_LOCALS(p)                                              \
  42         ScrnInfoPtr pScrn = xf86Screens[(p)->drawable.pScreen->myNum]; \
  43         NVPtr pNv = NVPTR(pScrn);                                      \
  44         struct nouveau_channel *chan = pNv->chan; (void)chan;          \
  45         struct nouveau_grobj *eng2d = pNv->Nv2D; (void)eng2d;          \
  46         struct nouveau_grobj *tesla = pNv->Nv3D; (void)tesla;          \
  47         struct nv50_exa_state *state = &exa_state; (void)state
  48
  49 #define BF(f) (NV50TCL_BLEND_FUNC_SRC_RGB_##f | 0x4000)
  50
  51 struct nv50_blend_op {
  52         unsigned src_alpha;
  53         unsigned dst_alpha;
  54         unsigned src_blend;
  55         unsigned dst_blend;
  56 };
  57
  58 static struct nv50_blend_op
  59 NV50EXABlendOp[] = {
  60 /* Clear       */ { 0, 0, BF(               ZERO), BF(               ZERO) },
  61 /* Src         */ { 0, 0, BF(                ONE), BF(               ZERO) },
  62 /* Dst         */ { 0, 0, BF(               ZERO), BF(                ONE) },
  63 /* Over        */ { 1, 0, BF(                ONE), BF(ONE_MINUS_SRC_ALPHA) },
  64 /* OverReverse */ { 0, 1, BF(ONE_MINUS_DST_ALPHA), BF(                ONE) },
  65 /* In          */ { 0, 1, BF(          DST_ALPHA), BF(               ZERO) },
  66 /* InReverse   */ { 1, 0, BF(               ZERO), BF(          SRC_ALPHA) },
  67 /* Out         */ { 0, 1, BF(ONE_MINUS_DST_ALPHA), BF(               ZERO) },
  68 /* OutReverse  */ { 1, 0, BF(               ZERO), BF(ONE_MINUS_SRC_ALPHA) },
  69 /* Atop        */ { 1, 1, BF(          DST_ALPHA), BF(ONE_MINUS_SRC_ALPHA) },
  70 /* AtopReverse */ { 1, 1, BF(ONE_MINUS_DST_ALPHA), BF(          SRC_ALPHA) },
  71 /* Xor         */ { 1, 1, BF(ONE_MINUS_DST_ALPHA), BF(ONE_MINUS_SRC_ALPHA) },
  72 /* Add         */ { 0, 0, BF(                ONE), BF(                ONE) },
  73 };
  74
  75 static Bool
  76 NV50EXA2DSurfaceFormat(PixmapPtr ppix, uint32_t *fmt)
  77 {
  78         NV50EXA_LOCALS(ppix);
  79
  80         switch (ppix->drawable.depth) {
  81         case 8 : *fmt = NV50_2D_SRC_FORMAT_8BPP; break;
  82         case 15: *fmt = NV50_2D_SRC_FORMAT_15BPP; break;
  83         case 16: *fmt = NV50_2D_SRC_FORMAT_16BPP; break;
  84         case 24: *fmt = NV50_2D_SRC_FORMAT_24BPP; break;
  85         case 32: *fmt = NV50_2D_SRC_FORMAT_32BPP; break;
  86         default:
  87                  xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
  88                             "Unknown surface format for bpp=%d\n",
  89                             ppix->drawable.depth);
  90                  return FALSE;
  91         }
  92
  93         return TRUE;
  94 }
  95
  96 static void NV50EXASetClip(PixmapPtr ppix, int x, int y, int w, int h)
  97 {
  98         NV50EXA_LOCALS(ppix);
  99
 100         BEGIN_RING(chan, eng2d, NV50_2D_CLIP_X, 4);
 101         OUT_RING  (chan, x);
 102         OUT_RING  (chan, y);
 103         OUT_RING  (chan, w);
 104         OUT_RING  (chan, h);
 105 }
 106
 107 static Bool
 108 NV50EXAAcquireSurface2D(PixmapPtr ppix, int is_src)
 109 {
 110         NV50EXA_LOCALS(ppix);
 111         int mthd = is_src ? NV50_2D_SRC_FORMAT : NV50_2D_DST_FORMAT;
 112         uint32_t fmt, bo_flags;
 113         struct nouveau_pixmap *surf = nouveau_pixmap(ppix);
 114
 115         if (!NV50EXA2DSurfaceFormat(ppix, &fmt))
 116                 return FALSE;
 117
 118         bo_flags  = NOUVEAU_BO_VRAM;
 119         bo_flags |= is_src ? NOUVEAU_BO_RD : NOUVEAU_BO_WR;
 120
 121         if (!surf->bo->tiled) {
 122                 BEGIN_RING(chan, eng2d, mthd, 2);
 123                 OUT_RING  (chan, fmt);
 124                 OUT_RING  (chan, 1);
 125                 BEGIN_RING(chan, eng2d, mthd + 0x14, 1);
 126                 OUT_RING  (chan, (uint32_t)exaGetPixmapPitch(ppix));
 127         } else {
 128                 BEGIN_RING(chan, eng2d, mthd, 5);
 129                 OUT_RING  (chan, fmt);
 130                 OUT_RING  (chan, 0);
 131                 OUT_RING  (chan, 0);
 132                 OUT_RING  (chan, 1);
 133                 OUT_RING  (chan, 0);
 134         }
 135
 136         BEGIN_RING(chan, eng2d, mthd + 0x18, 4);
 137         OUT_RING  (chan, ppix->drawable.width);
 138         OUT_RING  (chan, ppix->drawable.height);
 139         OUT_RELOCh(chan, surf->bo, 0, bo_flags);
 140         OUT_RELOCl(chan, surf->bo, 0, bo_flags);
 141
 142         if (is_src == 0)
 143                 NV50EXASetClip(ppix, 0, 0, ppix->drawable.width, ppix->drawable.height);
 144
 145         return TRUE;
 146 }
 147
 148 static void
 149 NV50EXASetPattern(PixmapPtr pdpix, int col0, int col1, int pat0, int pat1)
 150 {
 151         NV50EXA_LOCALS(pdpix);
 152
 153         BEGIN_RING(chan, eng2d, NV50_2D_PATTERN_COLOR(0), 4);
 154         OUT_RING  (chan, col0);
 155         OUT_RING  (chan, col1);
 156         OUT_RING  (chan, pat0);
 157         OUT_RING  (chan, pat1);
 158 }
 159
 160 static void
 161 NV50EXASetROP(PixmapPtr pdpix, int alu, Pixel planemask)
 162 {
 163         NV50EXA_LOCALS(pdpix);
 164         int rop;
 165
 166         if (planemask != ~0)
 167                 rop = NVROP[alu].copy_planemask;
 168         else
 169                 rop = NVROP[alu].copy;
 170
 171         BEGIN_RING(chan, eng2d, NV50_2D_OPERATION, 1);
 172         if (alu == GXcopy && planemask == ~0) {
 173                 OUT_RING  (chan, NV50_2D_OPERATION_SRCCOPY);
 174                 return;
 175         } else {
 176                 OUT_RING  (chan, NV50_2D_OPERATION_SRCCOPY_PREMULT);
 177         }
 178
 179         BEGIN_RING(chan, eng2d, NV50_2D_PATTERN_FORMAT, 2);
 180         switch (pdpix->drawable.depth) {
 181                 case  8: OUT_RING  (chan, 3); break;
 182                 case 15: OUT_RING  (chan, 1); break;
 183                 case 16: OUT_RING  (chan, 0); break;
 184                 case 24:
 185                 case 32:
 186                 default:
 187                          OUT_RING  (chan, 2);
 188                          break;
 189         }
 190         OUT_RING  (chan, 1);
 191
 192         /* There are 16 alu's.
 193          * 0-15: copy
 194          * 16-31: copy_planemask
 195          */
 196
 197         if (planemask != ~0) {
 198                 alu += 16;
 199                 NV50EXASetPattern(pdpix, 0, planemask, ~0, ~0);
 200         } else {
 201                 if (pNv->currentRop > 15)
 202                         NV50EXASetPattern(pdpix, ~0, ~0, ~0, ~0);
 203         }
 204
 205         if (pNv->currentRop != alu) {
 206                 BEGIN_RING(chan, eng2d, NV50_2D_ROP, 1);
 207                 OUT_RING  (chan, rop);
 208                 pNv->currentRop = alu;
 209         }
 210 }
 211
 212 Bool
 213 NV50EXAPrepareSolid(PixmapPtr pdpix, int alu, Pixel planemask, Pixel fg)
 214 {
 215         NV50EXA_LOCALS(pdpix);
 216         uint32_t fmt;
 217
 218         planemask |= ~0 << pScrn->depth;
 219
 220         if (!NV50EXA2DSurfaceFormat(pdpix, &fmt))
 221                 NOUVEAU_FALLBACK("rect format\n");
 222         if (!NV50EXAAcquireSurface2D(pdpix, 0))
 223                 NOUVEAU_FALLBACK("dest pixmap\n");
 224         NV50EXASetROP(pdpix, alu, planemask);
 225
 226         BEGIN_RING(chan, eng2d, 0x580, 3);
 227         OUT_RING  (chan, 4);
 228         OUT_RING  (chan, fmt);
 229         OUT_RING  (chan, fg);
 230
 231         return TRUE;
 232 }
 233
 234 void
 235 NV50EXASolid(PixmapPtr pdpix, int x1, int y1, int x2, int y2)
 236 {
 237         NV50EXA_LOCALS(pdpix);
 238
 239         BEGIN_RING(chan, eng2d, NV50_2D_RECT_X1, 4);
 240         OUT_RING  (chan, x1);
 241         OUT_RING  (chan, y1);
 242         OUT_RING  (chan, x2);
 243         OUT_RING  (chan, y2);
 244
 245         if((x2 - x1) * (y2 - y1) >= 512)
 246                 FIRE_RING (chan);
 247 }
 248
 249 void
 250 NV50EXADoneSolid(PixmapPtr pdpix)
 251 {
 252 }
 253
 254 Bool
 255 NV50EXAPrepareCopy(PixmapPtr pspix, PixmapPtr pdpix, int dx, int dy,
 256                    int alu, Pixel planemask)
 257 {
 258         NV50EXA_LOCALS(pdpix);
 259
 260         planemask |= ~0 << pScrn->depth;
 261
 262         if (!NV50EXAAcquireSurface2D(pspix, 1))
 263                 NOUVEAU_FALLBACK("src pixmap\n");
 264         if (!NV50EXAAcquireSurface2D(pdpix, 0))
 265                 NOUVEAU_FALLBACK("dest pixmap\n");
 266         NV50EXASetROP(pdpix, alu, planemask);
 267
 268         return TRUE;
 269 }
 270
 271 void
 272 NV50EXACopy(PixmapPtr pdpix, int srcX , int srcY,
 273                              int dstX , int dstY,
 274                              int width, int height)
 275 {
 276         NV50EXA_LOCALS(pdpix);
 277
 278         BEGIN_RING(chan, eng2d, 0x0110, 1);
 279         OUT_RING  (chan, 0);
 280         BEGIN_RING(chan, eng2d, 0x088c, 1);
 281         OUT_RING  (chan, 0);
 282         BEGIN_RING(chan, eng2d, NV50_2D_BLIT_DST_X, 12);
 283         OUT_RING  (chan, dstX);
 284         OUT_RING  (chan, dstY);
 285         OUT_RING  (chan, width);
 286         OUT_RING  (chan, height);
 287         OUT_RING  (chan, 0);
 288         OUT_RING  (chan, 1);
 289         OUT_RING  (chan, 0);
 290         OUT_RING  (chan, 1);
 291         OUT_RING  (chan, 0);
 292         OUT_RING  (chan, srcX);
 293         OUT_RING  (chan, 0);
 294         OUT_RING  (chan, srcY);
 295
 296         if(width * height >= 512)
 297                 FIRE_RING (chan);
 298 }
 299
 300 void
 301 NV50EXADoneCopy(PixmapPtr pdpix)
 302 {
 303 }
 304
 305 Bool
 306 NV50EXAUploadSIFC(const char *src, int src_pitch,
 307                   PixmapPtr pdpix, int x, int y, int w, int h, int cpp)
 308 {
 309         NV50EXA_LOCALS(pdpix);
 310         int line_dwords = (w * cpp + 3) / 4;
 311         uint32_t sifc_fmt;
 312
 313         if (!NV50EXA2DSurfaceFormat(pdpix, &sifc_fmt))
 314                 NOUVEAU_FALLBACK("hostdata format\n");
 315         if (!NV50EXAAcquireSurface2D(pdpix, 0))
 316                 NOUVEAU_FALLBACK("dest pixmap\n");
 317
 318         /* If the pitch isn't aligned to a dword, then you can get corruption at the end of a line. */
 319         NV50EXASetClip(pdpix, x, y, w, h);
 320
 321         BEGIN_RING(chan, eng2d, NV50_2D_OPERATION, 1);
 322         OUT_RING  (chan, NV50_2D_OPERATION_SRCCOPY);
 323         BEGIN_RING(chan, eng2d, NV50_2D_SIFC_UNK0800, 2);
 324         OUT_RING  (chan, 0);
 325         OUT_RING  (chan, sifc_fmt);
 326         BEGIN_RING(chan, eng2d, NV50_2D_SIFC_WIDTH, 10);
 327         OUT_RING  (chan, (line_dwords * 4) / cpp);
 328         OUT_RING  (chan, h);
 329         OUT_RING  (chan, 0);
 330         OUT_RING  (chan, 1);
 331         OUT_RING  (chan, 0);
 332         OUT_RING  (chan, 1);
 333         OUT_RING  (chan, 0);
 334         OUT_RING  (chan, x);
 335         OUT_RING  (chan, 0);
 336         OUT_RING  (chan, y);
 337
 338         while (h--) {
 339                 int count = line_dwords;
 340                 const char *p = src;
 341
 342                 while(count) {
 343                         int size = count > 1792 ? 1792 : count;
 344
 345                         BEGIN_RING(chan, eng2d,
 346                                          NV50_2D_SIFC_DATA | 0x40000000, size);
 347                         OUT_RINGp (chan, p, size);
 348
 349                         p += size * cpp;
 350                         count -= size;
 351                 }
 352
 353                 src += src_pitch;
 354         }
 355
 356         return TRUE;
 357 }
 358
 359 static Bool
 360 NV50EXACheckRenderTarget(PicturePtr ppict)
 361 {
 362         if (ppict->pDrawable->width > 8192 ||
 363             ppict->pDrawable->height > 8192)
 364                 NOUVEAU_FALLBACK("render target dimensions exceeded %dx%d\n",
 365                                  ppict->pDrawable->width,
 366                                  ppict->pDrawable->height);
 367
 368         switch (ppict->format) {
 369         case PICT_a8r8g8b8:
 370         case PICT_x8r8g8b8:
 371         case PICT_r5g6b5:
 372         case PICT_a8:
 373                 break;
 374         default:
 375                 NOUVEAU_FALLBACK("picture format 0x%08x\n", ppict->format);
 376         }
 377
 378         return TRUE;
 379 }
 380
 381 static Bool
 382 NV50EXARenderTarget(PixmapPtr ppix, PicturePtr ppict)
 383 {
 384         NV50EXA_LOCALS(ppix);
 385         unsigned format;
 386         struct nouveau_pixmap *rt = nouveau_pixmap(ppix);
 387
 388         /*XXX: Scanout buffer not tiled, someone needs to figure it out */
 389         if (!rt->bo->tiled)
 390                 NOUVEAU_FALLBACK("pixmap is scanout buffer\n");
 391
 392         switch (ppict->format) {
 393         case PICT_a8r8g8b8: format = NV50TCL_RT_FORMAT_32BPP; break;
 394         case PICT_x8r8g8b8: format = NV50TCL_RT_FORMAT_24BPP; break;
 395         case PICT_r5g6b5  : format = NV50TCL_RT_FORMAT_16BPP; break;
 396         case PICT_a8      : format = NV50TCL_RT_FORMAT_8BPP; break;
 397         default:
 398                 NOUVEAU_FALLBACK("invalid picture format\n");
 399         }
 400
 401         BEGIN_RING(chan, tesla, NV50TCL_RT_ADDRESS_HIGH(0), 5);
 402         OUT_RELOCh(chan, rt->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
 403         OUT_RELOCl(chan, rt->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
 404         OUT_RING  (chan, format);
 405         OUT_RING  (chan, 0);
 406         OUT_RING  (chan, 0x00000000);
 407         BEGIN_RING(chan, tesla, NV50TCL_RT_HORIZ(0), 2);
 408         OUT_RING  (chan, ppix->drawable.width);
 409         OUT_RING  (chan, ppix->drawable.height);
 410         BEGIN_RING(chan, tesla, 0x1224, 1);
 411         OUT_RING  (chan, 0x00000001);
 412
 413         return TRUE;
 414 }
 415
 416 static Bool
 417 NV50EXACheckTexture(PicturePtr ppict)
 418 {
 419         if (ppict->pDrawable->width > 8192 ||
 420             ppict->pDrawable->height > 8192)
 421                 NOUVEAU_FALLBACK("texture dimensions exceeded %dx%d\n",
 422                                  ppict->pDrawable->width,
 423                                  ppict->pDrawable->height);
 424
 425         switch (ppict->format) {
 426         case PICT_a8r8g8b8:
 427         case PICT_a8b8g8r8:
 428         case PICT_x8r8g8b8:
 429         case PICT_x8b8g8r8:
 430         case PICT_r5g6b5:
 431         case PICT_a8:
 432                 break;
 433         default:
 434                 NOUVEAU_FALLBACK("picture format 0x%08x\n", ppict->format);
 435         }
 436
 437         switch (ppict->filter) {
 438         case PictFilterNearest:
 439         case PictFilterBilinear:
 440                 break;
 441         default:
 442                 NOUVEAU_FALLBACK("picture filter %d\n", ppict->filter);
 443         }
 444
 445         return TRUE;
 446 }
 447
 448 static Bool
 449 NV50EXATexture(PixmapPtr ppix, PicturePtr ppict, unsigned unit)
 450 {
 451         NV50EXA_LOCALS(ppix);
 452         struct nouveau_pixmap *tex = nouveau_pixmap(ppix);
 453
 454         /*XXX: Scanout buffer not tiled, someone needs to figure it out */
 455         if (!tex->bo->tiled)
 456                 NOUVEAU_FALLBACK("pixmap is scanout buffer\n");
 457
 458         BEGIN_RING(chan, tesla, NV50TCL_CB_ADDR, 1);
 459         OUT_RING  (chan, CB_TIC | ((unit * 8) << NV50TCL_CB_ADDR_ID_SHIFT));
 460         BEGIN_RING(chan, tesla, NV50TCL_CB_DATA(0) | 0x40000000, 8);
 461         switch (ppict->format) {
 462         case PICT_a8r8g8b8:
 463                 OUT_RING  (chan, NV50TIC_0_0_MAPA_C3 | NV50TIC_0_0_TYPEA_UNORM |
 464                          NV50TIC_0_0_MAPR_C0 | NV50TIC_0_0_TYPER_UNORM |
 465                          NV50TIC_0_0_MAPG_C1 | NV50TIC_0_0_TYPEB_UNORM |
 466                          NV50TIC_0_0_MAPB_C2 | NV50TIC_0_0_TYPEG_UNORM |
 467                          NV50TIC_0_0_FMT_8_8_8_8);
 468                 break;
 469         case PICT_a8b8g8r8:
 470                 OUT_RING  (chan, NV50TIC_0_0_MAPA_C3 | NV50TIC_0_0_TYPEA_UNORM |
 471                          NV50TIC_0_0_MAPR_C2 | NV50TIC_0_0_TYPER_UNORM |
 472                          NV50TIC_0_0_MAPG_C1 | NV50TIC_0_0_TYPEB_UNORM |
 473                          NV50TIC_0_0_MAPB_C0 | NV50TIC_0_0_TYPEG_UNORM |
 474                          NV50TIC_0_0_FMT_8_8_8_8);
 475                 break;
 476         case PICT_x8r8g8b8:
 477                 OUT_RING  (chan, NV50TIC_0_0_MAPA_ONE | NV50TIC_0_0_TYPEA_UNORM |
 478                          NV50TIC_0_0_MAPR_C0 | NV50TIC_0_0_TYPER_UNORM |
 479                          NV50TIC_0_0_MAPG_C1 | NV50TIC_0_0_TYPEB_UNORM |
 480                          NV50TIC_0_0_MAPB_C2 | NV50TIC_0_0_TYPEG_UNORM |
 481                          NV50TIC_0_0_FMT_8_8_8_8);
 482                 break;
 483         case PICT_x8b8g8r8:
 484                 OUT_RING  (chan, NV50TIC_0_0_MAPA_ONE | NV50TIC_0_0_TYPEA_UNORM |
 485                          NV50TIC_0_0_MAPR_C2 | NV50TIC_0_0_TYPER_UNORM |
 486                          NV50TIC_0_0_MAPG_C1 | NV50TIC_0_0_TYPEB_UNORM |
 487                          NV50TIC_0_0_MAPB_C0 | NV50TIC_0_0_TYPEG_UNORM |
 488                          NV50TIC_0_0_FMT_8_8_8_8);
 489                 break;
 490         case PICT_r5g6b5:
 491                 OUT_RING  (chan, NV50TIC_0_0_MAPA_ONE | NV50TIC_0_0_TYPEA_UNORM |
 492                          NV50TIC_0_0_MAPR_C0 | NV50TIC_0_0_TYPER_UNORM |
 493                          NV50TIC_0_0_MAPG_C1 | NV50TIC_0_0_TYPEB_UNORM |
 494                          NV50TIC_0_0_MAPB_C2 | NV50TIC_0_0_TYPEG_UNORM |
 495                          NV50TIC_0_0_FMT_5_6_5);
 496                 break;
 497         case PICT_a8:
 498                 OUT_RING  (chan, NV50TIC_0_0_MAPA_C0 | NV50TIC_0_0_TYPEA_UNORM |
 499                          NV50TIC_0_0_MAPR_ZERO | NV50TIC_0_0_TYPER_UNORM |
 500                          NV50TIC_0_0_MAPG_ZERO | NV50TIC_0_0_TYPEB_UNORM |
 501                          NV50TIC_0_0_MAPB_ZERO | NV50TIC_0_0_TYPEG_UNORM |
 502                          NV50TIC_0_0_FMT_8);
 503                 break;
 504         default:
 505                 NOUVEAU_FALLBACK("invalid picture format, this SHOULD NOT HAPPEN. Expect trouble.\n");
 506         }
 507         OUT_RELOCl(chan, tex->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
 508         OUT_RING  (chan, 0xd0005000);
 509         OUT_RING  (chan, 0x00300000);
 510         OUT_RING  (chan, ppix->drawable.width);
 511         OUT_RING  (chan, (1 << NV50TIC_0_5_DEPTH_SHIFT) | ppix->drawable.height);
 512         OUT_RING  (chan, 0x03000000);
 513         OUT_RELOCh(chan, tex->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
 514
 515         BEGIN_RING(chan, tesla, NV50TCL_CB_ADDR, 1);
 516         OUT_RING  (chan, CB_TSC | ((unit * 8) << NV50TCL_CB_ADDR_ID_SHIFT));
 517         BEGIN_RING(chan, tesla, NV50TCL_CB_DATA(0) | 0x40000000, 8);
 518         if (ppict->repeat) {
 519                 switch (ppict->repeatType) {
 520                 case RepeatPad:
 521                         OUT_RING  (chan, NV50TSC_1_0_WRAPS_CLAMP |
 522                                  NV50TSC_1_0_WRAPT_CLAMP |
 523                                  NV50TSC_1_0_WRAPR_CLAMP | 0x00024000);
 524                         break;
 525                 case RepeatReflect:
 526                         OUT_RING  (chan, NV50TSC_1_0_WRAPS_MIRROR_REPEAT |
 527                                  NV50TSC_1_0_WRAPT_MIRROR_REPEAT |
 528                                  NV50TSC_1_0_WRAPR_MIRROR_REPEAT | 0x00024000);
 529                         break;
 530                 case RepeatNormal:
 531                 default:
 532                         OUT_RING  (chan, NV50TSC_1_0_WRAPS_REPEAT |
 533                                  NV50TSC_1_0_WRAPT_REPEAT |
 534                                  NV50TSC_1_0_WRAPR_REPEAT | 0x00024000);
 535                         break;
 536                 }
 537         } else {
 538                 OUT_RING  (chan, NV50TSC_1_0_WRAPS_CLAMP_TO_BORDER |
 539                          NV50TSC_1_0_WRAPT_CLAMP_TO_BORDER |
 540                          NV50TSC_1_0_WRAPR_CLAMP_TO_BORDER | 0x00024000);
 541         }
 542         if (ppict->filter == PictFilterBilinear) {
 543                 OUT_RING  (chan, NV50TSC_1_1_MAGF_LINEAR |
 544                          NV50TSC_1_1_MINF_LINEAR |
 545                          NV50TSC_1_1_MIPF_NONE);
 546         } else {
 547                 OUT_RING  (chan, NV50TSC_1_1_MAGF_NEAREST |
 548                          NV50TSC_1_1_MINF_NEAREST |
 549                          NV50TSC_1_1_MIPF_NONE);
 550         }
 551         OUT_RING  (chan, 0x00000000);
 552         OUT_RING  (chan, 0x00000000);
 553         OUT_RING  (chan, 0x00000000);
 554         OUT_RING  (chan, 0x00000000);
 555         OUT_RING  (chan, 0x00000000);
 556         OUT_RING  (chan, 0x00000000);
 557
 558         state->unit[unit].width = ppix->drawable.width;
 559         state->unit[unit].height = ppix->drawable.height;
 560         state->unit[unit].transform = ppict->transform;
 561         return TRUE;
 562 }
 563
 564 static Bool
 565 NV50EXACheckBlend(int op)
 566 {
 567         if (op > PictOpAdd)
 568                 NOUVEAU_FALLBACK("unsupported blend op %d\n", op);
 569         return TRUE;
 570 }
 571
 572 static void
 573 NV50EXABlend(PixmapPtr ppix, PicturePtr ppict, int op, int component_alpha)
 574 {
 575         NV50EXA_LOCALS(ppix);
 576         struct nv50_blend_op *b = &NV50EXABlendOp[op];
 577         unsigned sblend = b->src_blend;
 578         unsigned dblend = b->dst_blend;
 579
 580         if (b->dst_alpha) {
 581                 if (!PICT_FORMAT_A(ppict->format)) {
 582                         if (sblend == BF(DST_ALPHA))
 583                                 sblend = BF(ONE);
 584                         else
 585                         if (sblend == BF(ONE_MINUS_DST_ALPHA))
 586                                 sblend = BF(ZERO);
 587                 } else
 588                 if (ppict->format == PICT_a8) {
 589                         if (sblend == BF(DST_ALPHA))
 590                                 sblend = BF(DST_COLOR);
 591                         else
 592                         if (sblend == BF(ONE_MINUS_DST_ALPHA))
 593                                 sblend = BF(ONE_MINUS_DST_COLOR);
 594                 }
 595         }
 596
 597         if (b->src_alpha && (component_alpha || ppict->format == PICT_a8)) {
 598                 if (dblend == BF(SRC_ALPHA))
 599                         dblend = BF(SRC_COLOR);
 600                 else
 601                 if (dblend == BF(ONE_MINUS_SRC_ALPHA))
 602                         dblend = BF(ONE_MINUS_SRC_COLOR);
 603         }
 604
 605         if (sblend == BF(ONE) && dblend == BF(ZERO)) {
 606                 BEGIN_RING(chan, tesla, NV50TCL_BLEND_ENABLE(0), 1);
 607                 OUT_RING  (chan, 0);
 608         } else {
 609                 BEGIN_RING(chan, tesla, NV50TCL_BLEND_ENABLE(0), 1);
 610                 OUT_RING  (chan, 1);
 611                 BEGIN_RING(chan, tesla, NV50TCL_BLEND_EQUATION_RGB, 5);
 612                 OUT_RING  (chan, NV50TCL_BLEND_EQUATION_RGB_FUNC_ADD);
 613                 OUT_RING  (chan, sblend);
 614                 OUT_RING  (chan, dblend);
 615                 OUT_RING  (chan, NV50TCL_BLEND_EQUATION_ALPHA_FUNC_ADD);
 616                 OUT_RING  (chan, sblend);
 617                 BEGIN_RING(chan, tesla, NV50TCL_BLEND_FUNC_DST_ALPHA, 1);
 618                 OUT_RING  (chan, dblend);
 619         }
 620 }
 621
 622 Bool
 623 NV50EXACheckComposite(int op,
 624                       PicturePtr pspict, PicturePtr pmpict, PicturePtr pdpict)
 625 {
 626         if (!NV50EXACheckBlend(op))
 627                 NOUVEAU_FALLBACK("blend not supported\n");
 628
 629         if (!NV50EXACheckRenderTarget(pdpict))
 630                 NOUVEAU_FALLBACK("render target invalid\n");
 631
 632         if (!NV50EXACheckTexture(pspict))
 633                 NOUVEAU_FALLBACK("src picture invalid\n");
 634
 635         if (pmpict) {
 636                 if (pmpict->componentAlpha &&
 637                     PICT_FORMAT_RGB(pmpict->format) &&
 638                     NV50EXABlendOp[op].src_alpha &&
 639                     NV50EXABlendOp[op].src_blend != BF(ZERO))
 640                         NOUVEAU_FALLBACK("component-alpha not supported\n");
 641
 642                 if (!NV50EXACheckTexture(pmpict))
 643                         NOUVEAU_FALLBACK("mask picture invalid\n");
 644         }
 645
 646         return TRUE;
 647 }
 648
 649 Bool
 650 NV50EXAPrepareComposite(int op,
 651                         PicturePtr pspict, PicturePtr pmpict, PicturePtr pdpict,
 652                         PixmapPtr pspix, PixmapPtr pmpix, PixmapPtr pdpix)
 653 {
 654         NV50EXA_LOCALS(pspix);
 655
 656         BEGIN_RING(chan, eng2d, 0x0110, 1);
 657         OUT_RING  (chan, 0);
 658
 659         if (!NV50EXARenderTarget(pdpix, pdpict))
 660                 NOUVEAU_FALLBACK("render target invalid\n");
 661
 662         NV50EXABlend(pdpix, pdpict, op, pmpict && pmpict->componentAlpha &&
 663                      PICT_FORMAT_RGB(pmpict->format));
 664
 665         if (pmpict) {
 666                 if (!NV50EXATexture(pspix, pspict, 0))
 667                         NOUVEAU_FALLBACK("src picture invalid\n");
 668                 if (!NV50EXATexture(pmpix, pmpict, 1))
 669                         NOUVEAU_FALLBACK("mask picture invalid\n");
 670                 state->have_mask = TRUE;
 671
 672                 BEGIN_RING(chan, tesla, NV50TCL_FP_START_ID, 1);
 673                 if (pdpict->format == PICT_a8) {
 674                         OUT_RING  (chan, PFP_C_A8);
 675                 } else {
 676                         if (pmpict->componentAlpha &&
 677                             PICT_FORMAT_RGB(pmpict->format)) {
 678                                 if (NV50EXABlendOp[op].src_alpha)
 679                                         OUT_RING  (chan, PFP_CCASA);
 680                                 else
 681                                         OUT_RING  (chan, PFP_CCA);
 682                         } else {
 683                                 OUT_RING  (chan, PFP_C);
 684                         }
 685                 }
 686         } else {
 687                 if (!NV50EXATexture(pspix, pspict, 0))
 688                         NOUVEAU_FALLBACK("src picture invalid\n");
 689                 state->have_mask = FALSE;
 690
 691                 BEGIN_RING(chan, tesla, NV50TCL_FP_START_ID, 1);
 692                 if (pdpict->format == PICT_a8)
 693                         OUT_RING  (chan, PFP_S_A8);
 694                 else
 695                         OUT_RING  (chan, PFP_S);
 696         }
 697
 698         BEGIN_RING(chan, tesla, 0x1334, 1);
 699         OUT_RING  (chan, 0);
 700
 701         BEGIN_RING(chan, tesla, 0x1458, 1);
 702         OUT_RING  (chan, 1);
 703         BEGIN_RING(chan, tesla, 0x1458, 1);
 704         OUT_RING  (chan, 0x203);
 705
 706         BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
 707         OUT_RING  (chan, NV50TCL_VERTEX_BEGIN_QUADS);
 708
 709         return TRUE;
 710 }
 711
 712 #define xFixedToFloat(v) \
 713         ((float)xFixedToInt((v)) + ((float)xFixedFrac(v) / 65536.0))
 714 static inline void
 715 NV50EXATransform(PictTransformPtr t, int x, int y, float sx, float sy,
 716                  float *x_ret, float *y_ret)
 717 {
 718         if (t) {
 719                 PictVector v;
 720
 721                 v.vector[0] = IntToxFixed(x);
 722                 v.vector[1] = IntToxFixed(y);
 723                 v.vector[2] = xFixed1;
 724                 PictureTransformPoint(t, &v);
 725                 *x_ret = xFixedToFloat(v.vector[0]) / sx;
 726                 *y_ret = xFixedToFloat(v.vector[1]) / sy;
 727         } else {
 728                 *x_ret = (float)x / sx;
 729                 *y_ret = (float)y / sy;
 730         }
 731 }
 732
 733 void
 734 NV50EXAComposite(PixmapPtr pdpix, int sx, int sy, int mx, int my,
 735                  int dx, int dy, int w, int h)
 736 {
 737         NV50EXA_LOCALS(pdpix);
 738         float sX0, sX1, sX2, sX3, sY0, sY1, sY2, sY3;
 739         unsigned dX0 = dx, dX1 = dx + w, dY0 = dy, dY1 = dy + h;
 740
 741         NV50EXATransform(state->unit[0].transform, sx, sy,
 742                          state->unit[0].width, state->unit[0].height,
 743                          &sX0, &sY0);
 744         NV50EXATransform(state->unit[0].transform, sx + w, sy,
 745                          state->unit[0].width, state->unit[0].height,
 746                          &sX1, &sY1);
 747         NV50EXATransform(state->unit[0].transform, sx + w, sy + h,
 748                          state->unit[0].width, state->unit[0].height,
 749                          &sX2, &sY2);
 750         NV50EXATransform(state->unit[0].transform, sx, sy + h,
 751                          state->unit[0].width, state->unit[0].height,
 752                          &sX3, &sY3);
 753
 754         if (state->have_mask) {
 755                 float mX0, mX1, mX2, mX3, mY0, mY1, mY2, mY3;
 756
 757                 NV50EXATransform(state->unit[1].transform, mx, my,
 758                                  state->unit[1].width, state->unit[1].height,
 759                                  &mX0, &mY0);
 760                 NV50EXATransform(state->unit[1].transform, mx + w, my,
 761                                  state->unit[1].width, state->unit[1].height,
 762                                  &mX1, &mY1);
 763                 NV50EXATransform(state->unit[1].transform, mx + w, my + h,
 764                                  state->unit[1].width, state->unit[1].height,
 765                                  &mX2, &mY2);
 766                 NV50EXATransform(state->unit[1].transform, mx, my + h,
 767                                  state->unit[1].width, state->unit[1].height,
 768                                  &mX3, &mY3);
 769
 770                 VTX2s(pNv, sX0, sY0, mX0, mY0, dX0, dY0);
 771                 VTX2s(pNv, sX1, sY1, mX1, mY1, dX1, dY0);
 772                 VTX2s(pNv, sX2, sY2, mX2, mY2, dX1, dY1);
 773                 VTX2s(pNv, sX3, sY3, mX3, mY3, dX0, dY1);
 774         } else {
 775                 VTX1s(pNv, sX0, sY0, dX0, dY0);
 776                 VTX1s(pNv, sX1, sY1, dX1, dY0);
 777                 VTX1s(pNv, sX2, sY2, dX1, dY1);
 778                 VTX1s(pNv, sX3, sY3, dX0, dY1);
 779         }
 780 }
 781
 782 void
 783 NV50EXADoneComposite(PixmapPtr pdpix)
 784 {
 785         NV50EXA_LOCALS(pdpix);
 786
 787         BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
 788         OUT_RING  (chan, 0);
 789 }
 790