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

   1  /***************************************************************************\
   2 |*                                                                           *|
   3 |*       Copyright 2003 NVIDIA, Corporation.  All rights reserved.           *|
   4 |*                                                                           *|
   5 |*     NOTICE TO USER:   The source code  is copyrighted under  U.S. and     *|
   6 |*     international laws.  Users and possessors of this source code are     *|
   7 |*     hereby granted a nonexclusive,  royalty-free copyright license to     *|
   8 |*     use this code in individual and commercial software.                  *|
   9 |*                                                                           *|
  10 |*     Any use of this source code must include,  in the user documenta-     *|
  11 |*     tion and  internal comments to the code,  notices to the end user     *|
  12 |*     as follows:                                                           *|
  13 |*                                                                           *|
  14 |*       Copyright 2003 NVIDIA, Corporation.  All rights reserved.           *|
  15 |*                                                                           *|
  16 |*     NVIDIA, CORPORATION MAKES NO REPRESENTATION ABOUT THE SUITABILITY     *|
  17 |*     OF  THIS SOURCE  CODE  FOR ANY PURPOSE.  IT IS  PROVIDED  "AS IS"     *|
  18 |*     WITHOUT EXPRESS OR IMPLIED WARRANTY OF ANY KIND.  NVIDIA, CORPOR-     *|
  19 |*     ATION DISCLAIMS ALL WARRANTIES  WITH REGARD  TO THIS SOURCE CODE,     *|
  20 |*     INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY, NONINFRINGE-     *|
  21 |*     MENT,  AND FITNESS  FOR A PARTICULAR PURPOSE.   IN NO EVENT SHALL     *|
  22 |*     NVIDIA, CORPORATION  BE LIABLE FOR ANY SPECIAL,  INDIRECT,  INCI-     *|
  23 |*     DENTAL, OR CONSEQUENTIAL DAMAGES,  OR ANY DAMAGES  WHATSOEVER RE-     *|
  24 |*     SULTING FROM LOSS OF USE,  DATA OR PROFITS,  WHETHER IN AN ACTION     *|
  25 |*     OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION,  ARISING OUT OF     *|
  26 |*     OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOURCE CODE.     *|
  27 |*                                                                           *|
  28 |*     U.S. Government  End  Users.   This source code  is a "commercial     *|
  29 |*     item,"  as that  term is  defined at  48 C.F.R. 2.101 (OCT 1995),     *|
  30 |*     consisting  of "commercial  computer  software"  and  "commercial     *|
  31 |*     computer  software  documentation,"  as such  terms  are  used in     *|
  32 |*     48 C.F.R. 12.212 (SEPT 1995)  and is provided to the U.S. Govern-     *|
  33 |*     ment only as  a commercial end item.   Consistent with  48 C.F.R.     *|
  34 |*     12.212 and  48 C.F.R. 227.7202-1 through  227.7202-4 (JUNE 1995),     *|
  35 |*     all U.S. Government End Users  acquire the source code  with only     *|
  36 |*     those rights set forth herein.                                        *|
  37 |*                                                                           *|
  38  \***************************************************************************/
  39
  40 /*
  41   Exa Modifications (c) Lars Knoll (lars@trolltech.com)
  42  */
  43
  44 #ifdef HAVE_CONFIG_H
  45 #include "config.h"
  46 #endif
  47
  48 #include "nv_include.h"
  49 #include "exa.h"
  50
  51 #include "nv_dma.h"
  52 #include "nv_local.h"
  53
  54 #include <sys/time.h>
  55
  56 const int NVCopyROP[16] =
  57 {
  58    0x00,            /* GXclear */
  59    0x88,            /* GXand */
  60    0x44,            /* GXandReverse */
  61    0xCC,            /* GXcopy */
  62    0x22,            /* GXandInverted */
  63    0xAA,            /* GXnoop */
  64    0x66,            /* GXxor */
  65    0xEE,            /* GXor */
  66    0x11,            /* GXnor */
  67    0x99,            /* GXequiv */
  68    0x55,            /* GXinvert*/
  69    0xDD,            /* GXorReverse */
  70    0x33,            /* GXcopyInverted */
  71    0xBB,            /* GXorInverted */
  72    0x77,            /* GXnand */
  73    0xFF             /* GXset */
  74 };
  75
  76 static void
  77 NVSetPattern(ScrnInfoPtr pScrn, CARD32 clr0, CARD32 clr1,
  78                                 CARD32 pat0, CARD32 pat1)
  79 {
  80         NVPtr pNv = NVPTR(pScrn);
  81
  82         BEGIN_RING(NvImagePattern, NV04_IMAGE_PATTERN_MONOCHROME_COLOR0, 4);
  83         OUT_RING  (clr0);
  84         OUT_RING  (clr1);
  85         OUT_RING  (pat0);
  86         OUT_RING  (pat1);
  87 }
  88
  89 static void
  90 NVSetROP(ScrnInfoPtr pScrn, CARD32 alu, CARD32 planemask)
  91 {
  92         NVPtr pNv = NVPTR(pScrn);
  93         int rop = NVCopyROP[alu] & 0xf0;
  94
  95         if (planemask != ~0) {
  96                 NVSetPattern(pScrn, 0, planemask, ~0, ~0);
  97                 if (pNv->currentRop != (alu + 32)) {
  98                         BEGIN_RING(NvRop, NV03_CONTEXT_ROP_ROP, 1);
  99                         OUT_RING  (rop | 0x0a);
 100                         pNv->currentRop = alu + 32;
 101                 }
 102         } else
 103         if (pNv->currentRop != alu) {
 104                 if(pNv->currentRop >= 16)
 105                         NVSetPattern(pScrn, ~0, ~0, ~0, ~0);
 106                 BEGIN_RING(NvRop, NV03_CONTEXT_ROP_ROP, 1);
 107                 OUT_RING  (rop | (rop >> 4));
 108                 pNv->currentRop = alu;
 109         }
 110 }
 111
 112 static void setM2MFDirection(ScrnInfoPtr pScrn, int dir)
 113 {
 114         NVPtr pNv = NVPTR(pScrn);
 115
 116         if (pNv->M2MFDirection != dir) {
 117
 118                 BEGIN_RING(NvMemFormat,
 119                            NV_MEMORY_TO_MEMORY_FORMAT_DMA_BUFFER_IN, 2);
 120                 OUT_RING  (dir ? NvDmaTT : NvDmaFB);
 121                 OUT_RING  (dir ? NvDmaFB : NvDmaTT);
 122                 pNv->M2MFDirection = dir;
 123         }
 124 }
 125
 126 static CARD32 rectFormat(DrawablePtr pDrawable)
 127 {
 128         switch(pDrawable->bitsPerPixel) {
 129         case 32:
 130         case 24:
 131                 return NV04_GDI_RECTANGLE_TEXT_COLOR_FORMAT_A8R8G8B8;
 132                 break;
 133         case 16:
 134                 return NV04_GDI_RECTANGLE_TEXT_COLOR_FORMAT_A16R5G6B5;
 135                 break;
 136         default:
 137                 return NV04_GDI_RECTANGLE_TEXT_COLOR_FORMAT_A8R8G8B8;
 138                 break;
 139         }
 140 }
 141
 142 /* EXA acceleration hooks */
 143 static void NVExaWaitMarker(ScreenPtr pScreen, int marker)
 144 {
 145         NVSync(xf86Screens[pScreen->myNum]);
 146 }
 147
 148 static Bool NVExaPrepareSolid(PixmapPtr pPixmap,
 149                               int   alu,
 150                               Pixel planemask,
 151                               Pixel fg)
 152 {
 153         ScrnInfoPtr pScrn = xf86Screens[pPixmap->drawable.pScreen->myNum];
 154         NVPtr pNv = NVPTR(pScrn);
 155         int fmt;
 156
 157         planemask |= ~0 << pPixmap->drawable.bitsPerPixel;
 158         if (planemask != ~0 || alu != GXcopy) {
 159                 if (pPixmap->drawable.bitsPerPixel == 32)
 160                         return FALSE;
 161                 BEGIN_RING(NvRectangle, NV04_GDI_RECTANGLE_TEXT_OPERATION, 1);
 162                 OUT_RING  (1); /* ROP_AND */
 163                 NVSetROP(pScrn, alu, planemask);
 164         } else {
 165                 BEGIN_RING(NvRectangle, NV04_GDI_RECTANGLE_TEXT_OPERATION, 1);
 166                 OUT_RING  (3); /* SRCCOPY */
 167         }
 168
 169         if (!NVAccelGetCtxSurf2DFormatFromPixmap(pPixmap, &fmt))
 170                 return FALSE;
 171
 172         /* When SURFACE_FORMAT_A8R8G8B8 is used with GDI_RECTANGLE_TEXT, the
 173          * alpha channel gets forced to 0xFF for some reason.  We're using
 174          * SURFACE_FORMAT_Y32 as a workaround
 175          */
 176         if (fmt == NV04_CONTEXT_SURFACES_2D_FORMAT_A8R8G8B8)
 177                 fmt = NV04_CONTEXT_SURFACES_2D_FORMAT_Y32;
 178
 179         if (!NVAccelSetCtxSurf2D(pPixmap, pPixmap, fmt))
 180                 return FALSE;
 181
 182         BEGIN_RING(NvRectangle, NV04_GDI_RECTANGLE_TEXT_COLOR_FORMAT, 1);
 183         OUT_RING  (rectFormat(&pPixmap->drawable));
 184         BEGIN_RING(NvRectangle, NV04_GDI_RECTANGLE_TEXT_COLOR1_A, 1);
 185         OUT_RING  (fg);
 186
 187         pNv->DMAKickoffCallback = NVDmaKickoffCallback;
 188         return TRUE;
 189 }
 190
 191 static void NVExaSolid (PixmapPtr pPixmap, int x1, int y1, int x2, int y2)
 192 {
 193         ScrnInfoPtr pScrn = xf86Screens[pPixmap->drawable.pScreen->myNum];
 194         NVPtr pNv = NVPTR(pScrn);
 195         int width = x2-x1;
 196         int height = y2-y1;
 197
 198         BEGIN_RING(NvRectangle,
 199                    NV04_GDI_RECTANGLE_TEXT_UNCLIPPED_RECTANGLE_POINT(0), 2);
 200         OUT_RING  ((x1 << 16) | y1);
 201         OUT_RING  ((width << 16) | height);
 202
 203         if((width * height) >= 512)
 204                 FIRE_RING();
 205 }
 206
 207 static void NVExaDoneSolid (PixmapPtr pPixmap)
 208 {
 209 }
 210
 211 static Bool NVExaPrepareCopy(PixmapPtr pSrcPixmap,
 212                              PixmapPtr pDstPixmap,
 213                              int       dx,
 214                              int       dy,
 215                              int       alu,
 216                              Pixel     planemask)
 217 {
 218         ScrnInfoPtr pScrn = xf86Screens[pSrcPixmap->drawable.pScreen->myNum];
 219         NVPtr pNv = NVPTR(pScrn);
 220         int fmt;
 221
 222         if (pSrcPixmap->drawable.bitsPerPixel !=
 223                         pDstPixmap->drawable.bitsPerPixel)
 224                 return FALSE;
 225
 226         planemask |= ~0 << pDstPixmap->drawable.bitsPerPixel;
 227         if (planemask != ~0 || alu != GXcopy) {
 228                 if (pDstPixmap->drawable.bitsPerPixel == 32)
 229                         return FALSE;
 230                 BEGIN_RING(NvImageBlit, NV_IMAGE_BLIT_OPERATION, 1);
 231                 OUT_RING  (1); /* ROP_AND */
 232                 NVSetROP(pScrn, alu, planemask);
 233         } else {
 234                 BEGIN_RING(NvImageBlit, NV_IMAGE_BLIT_OPERATION, 1);
 235                 OUT_RING  (3); /* SRCCOPY */
 236         }
 237
 238         if (!NVAccelGetCtxSurf2DFormatFromPixmap(pDstPixmap, &fmt))
 239                 return FALSE;
 240         if (!NVAccelSetCtxSurf2D(pSrcPixmap, pDstPixmap, fmt))
 241                 return FALSE;
 242
 243         pNv->DMAKickoffCallback = NVDmaKickoffCallback;
 244         return TRUE;
 245 }
 246
 247 static void NVExaCopy(PixmapPtr pDstPixmap,
 248                       int       srcX,
 249                       int       srcY,
 250                       int       dstX,
 251                       int       dstY,
 252                       int       width,
 253                       int       height)
 254 {
 255         ScrnInfoPtr pScrn = xf86Screens[pDstPixmap->drawable.pScreen->myNum];
 256         NVPtr pNv = NVPTR(pScrn);
 257
 258         /* We want to catch people who have this bug, to find a decent fix */
 259 #if 0
 260         /* Now check whether we have the same values for srcY and dstY and
 261            whether the used chipset is buggy. Currently we flag all of G70
 262            cards as buggy, which is probably much to broad. KoalaBR
 263            16 is an abritrary threshold. It should define the maximum number
 264            of lines between dstY and srcY  If the number of lines is below
 265            we guess, that the bug won't trigger...
 266          */
 267         if ( ((abs(srcY - dstY)< 16)||(abs(srcX-dstX)<16)) &&
 268                 ((((pNv->Chipset & 0xfff0) == CHIPSET_G70) ||
 269                  ((pNv->Chipset & 0xfff0) == CHIPSET_G71) ||
 270                  ((pNv->Chipset & 0xfff0) == CHIPSET_G72) ||
 271                  ((pNv->Chipset & 0xfff0) == CHIPSET_G73) ||
 272                  ((pNv->Chipset & 0xfff0) == CHIPSET_C512))) )
 273         {
 274                 int dx=abs(srcX - dstX),dy=abs(srcY - dstY);
 275                 // Ok, let's do it manually unless someone comes up with a better idea
 276                 // 1. If dstY and srcY are really the same, do a copy rowwise
 277                 if (dy<dx) {
 278                         int i,xpos,inc;
 279                         NVDEBUG("ExaCopy: Lines identical:\n");
 280                         if (srcX>=dstX) {
 281                                 xpos=0;
 282                                 inc=1;
 283                         } else {
 284                                 xpos=width-1;
 285                                 inc=-1;
 286                         }
 287                         for (i = 0; i < width; i++) {
 288                                 BEGIN_RING(NvImageBlit,
 289                                            NV_IMAGE_BLIT_POINT_IN, 3);
 290                                 OUT_RING  ((srcY << 16) | (srcX+xpos));
 291                                 OUT_RING  ((dstY << 16) | (dstX+xpos));
 292                                 OUT_RING  ((height  << 16) | 1);
 293                                 xpos+=inc;
 294                         }
 295                 } else {
 296                         // 2. Otherwise we will try a line by line copy in the hope to avoid
 297                         //    the card's bug.
 298                         int i,ypos,inc;
 299                         NVDEBUG("ExaCopy: Lines nearly the same srcY=%d, dstY=%d:\n", srcY, dstY);
 300                         if (srcY>=dstY) {
 301                                 ypos=0;
 302                                 inc=1;
 303                         } else {
 304                                 ypos=height-1;
 305                                 inc=-1;
 306                         }
 307                         for (i = 0; i < height; i++) {
 308                                 BEGIN_RING(NvImageBlit,
 309                                            NV_IMAGE_BLIT_POINT_IN, 3);
 310                                 OUT_RING  (((srcY+ypos) << 16) | srcX);
 311                                 OUT_RING  (((dstY+ypos) << 16) | dstX);
 312                                 OUT_RING  ((1  << 16) | width);
 313                                 ypos+=inc;
 314                         }
 315                 }
 316         } else {
 317                 NVDEBUG("ExaCopy: Using default path\n");
 318                 BEGIN_RING(NvImageBlit, NV_IMAGE_BLIT_POINT_IN, 3);
 319                 OUT_RING  ((srcY << 16) | srcX);
 320                 OUT_RING  ((dstY << 16) | dstX);
 321                 OUT_RING  ((height  << 16) | width);
 322         }
 323 #endif /* 0 */
 324
 325         NVDEBUG("ExaCopy: Using default path\n");
 326         BEGIN_RING(NvImageBlit, NV_IMAGE_BLIT_POINT_IN, 3);
 327         OUT_RING  ((srcY << 16) | srcX);
 328         OUT_RING  ((dstY << 16) | dstX);
 329         OUT_RING  ((height  << 16) | width);
 330
 331         if((width * height) >= 512)
 332                 FIRE_RING();
 333 }
 334
 335 static void NVExaDoneCopy (PixmapPtr pDstPixmap) {}
 336
 337 static inline Bool NVAccelMemcpyRect(char *dst, const char *src, int height,
 338                        int dst_pitch, int src_pitch, int line_len)
 339 {
 340         if ((src_pitch == line_len) && (src_pitch == dst_pitch)) {
 341                 memcpy(dst, src, line_len*height);
 342         } else {
 343                 while (height--) {
 344                         memcpy(dst, src, line_len);
 345                         src += src_pitch;
 346                         dst += dst_pitch;
 347                 }
 348         }
 349
 350         return TRUE;
 351 }
 352
 353 static inline Bool
 354 NVAccelDownloadM2MF(ScrnInfoPtr pScrn, char *dst, uint64_t src_offset,
 355                                      int dst_pitch, int src_pitch,
 356                                      int line_len, int line_count)
 357 {
 358         NVPtr pNv = NVPTR(pScrn);
 359
 360         setM2MFDirection(pScrn, 0);
 361
 362         while (line_count) {
 363                 char *src = pNv->GARTScratch->map;
 364                 int lc, i;
 365
 366                 if (line_count * line_len <= pNv->GARTScratch->size) {
 367                         lc = line_count;
 368                 } else {
 369                         lc = pNv->GARTScratch->size / line_len;
 370                         if (lc > line_count)
 371                                 lc = line_count;
 372                 }
 373
 374                 /* HW limitations */
 375                 if (lc > 2047)
 376                         lc = 2047;
 377
 378                 if (pNv->Architecture >= NV_ARCH_50) {
 379                         BEGIN_RING(NvMemFormat, 0x200, 1);
 380                         OUT_RING  (1);
 381                         BEGIN_RING(NvMemFormat, 0x21c, 1);
 382                         OUT_RING  (1);
 383                         /* probably high-order bits of address */
 384                         BEGIN_RING(NvMemFormat, 0x238, 2);
 385                         OUT_RING  (0);
 386                         OUT_RING  (0);
 387                 }
 388
 389                 BEGIN_RING(NvMemFormat,
 390                            NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
 391                 OUT_RING  ((uint32_t)src_offset);
 392                 OUT_RING  ((uint32_t)pNv->GARTScratch->offset);
 393                 OUT_RING  (src_pitch);
 394                 OUT_RING  (line_len);
 395                 OUT_RING  (line_len);
 396                 OUT_RING  (lc);
 397                 OUT_RING  ((1<<8)|1);
 398                 OUT_RING  (0);
 399
 400                 NVNotifierReset(pScrn, pNv->Notifier0);
 401                 BEGIN_RING(NvMemFormat, NV_MEMORY_TO_MEMORY_FORMAT_NOTIFY, 1);
 402                 OUT_RING  (0);
 403                 BEGIN_RING(NvMemFormat, 0x100, 1);
 404                 OUT_RING  (0);
 405                 FIRE_RING();
 406                 if (!NVNotifierWaitStatus(pScrn, pNv->Notifier0, 0, 2000))
 407                         return FALSE;
 408
 409                 if (dst_pitch == line_len) {
 410                         memcpy(dst, src, dst_pitch * lc);
 411                         dst += dst_pitch * lc;
 412                 } else {
 413                         for (i = 0; i < lc; i++) {
 414                                 memcpy(dst, src, line_len);
 415                                 src += line_len;
 416                                 dst += dst_pitch;
 417                         }
 418                 }
 419
 420                 line_count -= lc;
 421                 src_offset += lc * src_pitch;
 422         }
 423
 424         return TRUE;
 425 }
 426
 427 static inline void *
 428 NVExaPixmapMap(PixmapPtr pPix)
 429 {
 430         ScrnInfoPtr pScrn = xf86Screens[pPix->drawable.pScreen->myNum];
 431         NVPtr pNv = NVPTR(pScrn);
 432         void *map;
 433
 434         map = pNv->FB->map + exaGetPixmapOffset(pPix);
 435         return map;
 436 }
 437
 438 static Bool NVDownloadFromScreen(PixmapPtr pSrc,
 439                                  int x,  int y,
 440                                  int w,  int h,
 441                                  char *dst,  int dst_pitch)
 442 {
 443         ScrnInfoPtr pScrn = xf86Screens[pSrc->drawable.pScreen->myNum];
 444         NVPtr pNv = NVPTR(pScrn);
 445         int src_offset, src_pitch, cpp, offset;
 446         const char *src;
 447
 448         src_offset = NVAccelGetPixmapOffset(pSrc);
 449         src_pitch  = exaGetPixmapPitch(pSrc);
 450         cpp = pSrc->drawable.bitsPerPixel >> 3;
 451         offset = (y * src_pitch) + (x * cpp);
 452
 453         if (pNv->GARTScratch) {
 454                 if (NVAccelDownloadM2MF(pScrn, dst,
 455                                         src_offset + offset,
 456                                         dst_pitch, src_pitch, w * cpp, h))
 457                         return TRUE;
 458         }
 459
 460         src = NVExaPixmapMap(pSrc) + offset;
 461         exaWaitSync(pSrc->drawable.pScreen);
 462         if (NVAccelMemcpyRect(dst, src, h, dst_pitch, src_pitch, w*cpp))
 463                 return TRUE;
 464
 465         return FALSE;
 466 }
 467
 468 static inline Bool
 469 NVAccelUploadIFC(ScrnInfoPtr pScrn, const char *src, int src_pitch,
 470                  PixmapPtr pDst, int fmt, int x, int y, int w, int h, int cpp)
 471 {
 472         NVPtr pNv = NVPTR(pScrn);
 473         int line_len = w * cpp;
 474         int iw, id, ifc_fmt;
 475
 476         if (pNv->Architecture >= NV_ARCH_50)
 477                 return FALSE;
 478
 479         if (h > 1024)
 480                 return FALSE;
 481
 482         switch (cpp) {
 483         case 2: ifc_fmt = 1; break;
 484         case 4: ifc_fmt = 4; break;
 485         default:
 486                 return FALSE;
 487         }
 488
 489         NVAccelSetCtxSurf2D(pDst, pDst, fmt);
 490
 491         /* Pad out input width to cover both COLORA() and COLORB() */
 492         iw  = (line_len + 7) & ~7;
 493         id  = iw / 4; /* line push size */
 494         iw /= cpp;
 495
 496         /* Don't support lines longer than max push size yet.. */
 497         if (id > 1792)
 498                 return FALSE;
 499
 500         BEGIN_RING(NvClipRectangle, NV01_CONTEXT_CLIP_RECTANGLE_POINT, 2);
 501         OUT_RING  (0x0);
 502         OUT_RING  (0x7FFF7FFF);
 503
 504         BEGIN_RING(NvImageFromCpu, NV01_IMAGE_FROM_CPU_OPERATION, 2);
 505         OUT_RING  (NV01_IMAGE_FROM_CPU_OPERATION_SRCCOPY);
 506         OUT_RING  (ifc_fmt);
 507         BEGIN_RING(NvImageFromCpu, NV01_IMAGE_FROM_CPU_POINT, 3);
 508         OUT_RING  ((y << 16) | x); /* dst point */
 509         OUT_RING  ((h << 16) | w); /* width/height out */
 510         OUT_RING  ((h << 16) | iw); /* width/height in */
 511
 512         while (h--) {
 513                 char *dst;
 514                 /* send a line */
 515                 BEGIN_RING(NvImageFromCpu, NV01_IMAGE_FROM_CPU_COLOR(0), id);
 516                 dst = (char *)pNv->dmaBase + (pNv->dmaCurrent << 2);
 517                 memcpy(dst, src, line_len);
 518                 pNv->dmaCurrent += id;
 519
 520                 src += src_pitch;
 521         }
 522
 523         return TRUE;
 524 }
 525
 526 static inline Bool
 527 NVAccelUploadM2MF(ScrnInfoPtr pScrn, uint64_t dst_offset, const char *src,
 528                                      int dst_pitch, int src_pitch,
 529                                      int line_len, int line_count)
 530 {
 531         NVPtr pNv = NVPTR(pScrn);
 532
 533         setM2MFDirection(pScrn, 1);
 534
 535         while (line_count) {
 536                 char *dst = pNv->GARTScratch->map;
 537                 int lc, i;
 538
 539                 /* Determine max amount of data we can DMA at once */
 540                 if (line_count * line_len <= pNv->GARTScratch->size) {
 541                         lc = line_count;
 542                 } else {
 543                         lc = pNv->GARTScratch->size / line_len;
 544                         if (lc > line_count)
 545                                 lc = line_count;
 546                 }
 547
 548                 /* HW limitations */
 549                 if (lc > 2047)
 550                         lc = 2047;
 551
 552                 /* Upload to GART */
 553                 if (src_pitch == line_len) {
 554                         memcpy(dst, src, src_pitch * lc);
 555                         src += src_pitch * lc;
 556                 } else {
 557                         for (i = 0; i < lc; i++) {
 558                                 memcpy(dst, src, line_len);
 559                                 src += src_pitch;
 560                                 dst += line_len;
 561                         }
 562                 }
 563
 564                 if (pNv->Architecture >= NV_ARCH_50) {
 565                         BEGIN_RING(NvMemFormat, 0x200, 1);
 566                         OUT_RING  (1);
 567                         BEGIN_RING(NvMemFormat, 0x21c, 1);
 568                         OUT_RING  (1);
 569                         /* probably high-order bits of address */
 570                         BEGIN_RING(NvMemFormat, 0x238, 2);
 571                         OUT_RING  (0);
 572                         OUT_RING  (0);
 573                 }
 574
 575                 /* DMA to VRAM */
 576                 BEGIN_RING(NvMemFormat,
 577                            NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
 578                 OUT_RING  ((uint32_t)pNv->GARTScratch->offset);
 579                 OUT_RING  ((uint32_t)dst_offset);
 580                 OUT_RING  (line_len);
 581                 OUT_RING  (dst_pitch);
 582                 OUT_RING  (line_len);
 583                 OUT_RING  (lc);
 584                 OUT_RING  ((1<<8)|1);
 585                 OUT_RING  (0);
 586
 587                 NVNotifierReset(pScrn, pNv->Notifier0);
 588                 BEGIN_RING(NvMemFormat, NV_MEMORY_TO_MEMORY_FORMAT_NOTIFY, 1);
 589                 OUT_RING  (0);
 590                 BEGIN_RING(NvMemFormat, 0x100, 1);
 591                 OUT_RING  (0);
 592                 FIRE_RING();
 593                 if (!NVNotifierWaitStatus(pScrn, pNv->Notifier0, 0, 2000))
 594                         return FALSE;
 595
 596                 dst_offset += lc * dst_pitch;
 597                 line_count -= lc;
 598         }
 599
 600         return TRUE;
 601 }
 602
 603 static Bool NVUploadToScreen(PixmapPtr pDst,
 604                              int x, int y, int w, int h,
 605                              char *src, int src_pitch)
 606 {
 607         ScrnInfoPtr pScrn = xf86Screens[pDst->drawable.pScreen->myNum];
 608         NVPtr pNv = NVPTR(pScrn);
 609         int dst_offset, dst_pitch, cpp;
 610         char *dst;
 611
 612         dst_offset = NVAccelGetPixmapOffset(pDst);
 613         dst_pitch  = exaGetPixmapPitch(pDst);
 614         cpp = pDst->drawable.bitsPerPixel >> 3;
 615
 616         /* try hostdata transfer */
 617         if (pNv->Architecture < NV_ARCH_50 && w*h*cpp<16*1024) /* heuristic */
 618         {
 619                 int fmt;
 620
 621                 if (NVAccelGetCtxSurf2DFormatFromPixmap(pDst, &fmt)) {
 622                         if (NVAccelUploadIFC(pScrn, src, src_pitch, pDst, fmt,
 623                                                     x, y, w, h, cpp)) {
 624                                 exaMarkSync(pDst->drawable.pScreen);
 625                                 return TRUE;
 626                         }
 627                 }
 628         }
 629
 630         /* try gart-based transfer */
 631         if (pNv->GARTScratch) {
 632                 dst_offset += (y * dst_pitch) + (x * cpp);
 633                 if (NVAccelUploadM2MF(pScrn, dst_offset, src, dst_pitch,
 634                                         src_pitch, w * cpp, h))
 635                         return TRUE;
 636         }
 637
 638         /* fallback to memcpy-based transfer */
 639         dst = NVExaPixmapMap(pDst) + (y * dst_pitch) + (x * cpp);
 640         exaWaitSync(pDst->drawable.pScreen);
 641         if (NVAccelMemcpyRect(dst, src, h, dst_pitch, src_pitch, w*cpp))
 642                 return TRUE;
 643
 644         return FALSE;
 645 }
 646
 647 Bool
 648 NVExaInit(ScreenPtr pScreen)
 649 {
 650         ScrnInfoPtr pScrn = xf86Screens[pScreen->myNum];
 651         NVPtr pNv = NVPTR(pScrn);
 652
 653         if(!(pNv->EXADriverPtr = (ExaDriverPtr) xnfcalloc(sizeof(ExaDriverRec), 1))) {
 654                 pNv->NoAccel = TRUE;
 655                 return FALSE;
 656         }
 657
 658         pNv->EXADriverPtr->exa_major = EXA_VERSION_MAJOR;
 659         pNv->EXADriverPtr->exa_minor = EXA_VERSION_MINOR;
 660
 661         pNv->EXADriverPtr->memoryBase           = pNv->FB->map;
 662         pNv->EXADriverPtr->offScreenBase        =
 663                 pScrn->virtualX * pScrn->virtualY*(pScrn->bitsPerPixel/8);
 664         pNv->EXADriverPtr->memorySize           = pNv->FB->size;
 665         pNv->EXADriverPtr->pixmapOffsetAlign    = 256;
 666         pNv->EXADriverPtr->pixmapPitchAlign     = 64;
 667         pNv->EXADriverPtr->flags                = EXA_OFFSCREEN_PIXMAPS;
 668         pNv->EXADriverPtr->maxX                 = 32768;
 669         pNv->EXADriverPtr->maxY                 = 32768;
 670
 671         pNv->EXADriverPtr->WaitMarker = NVExaWaitMarker;
 672
 673         /* Install default hooks */
 674         pNv->EXADriverPtr->DownloadFromScreen = NVDownloadFromScreen;
 675         pNv->EXADriverPtr->UploadToScreen = NVUploadToScreen;
 676
 677         if (pNv->Architecture < NV_ARCH_50) {
 678                 pNv->EXADriverPtr->PrepareCopy = NVExaPrepareCopy;
 679                 pNv->EXADriverPtr->Copy = NVExaCopy;
 680                 pNv->EXADriverPtr->DoneCopy = NVExaDoneCopy;
 681
 682                 pNv->EXADriverPtr->PrepareSolid = NVExaPrepareSolid;
 683                 pNv->EXADriverPtr->Solid = NVExaSolid;
 684                 pNv->EXADriverPtr->DoneSolid = NVExaDoneSolid;
 685         } else {
 686                 pNv->EXADriverPtr->PrepareCopy = NV50EXAPrepareCopy;
 687                 pNv->EXADriverPtr->Copy = NV50EXACopy;
 688                 pNv->EXADriverPtr->DoneCopy = NV50EXADoneCopy;
 689
 690                 pNv->EXADriverPtr->PrepareSolid = NV50EXAPrepareSolid;
 691                 pNv->EXADriverPtr->Solid = NV50EXASolid;
 692                 pNv->EXADriverPtr->DoneSolid = NV50EXADoneSolid;
 693         }
 694
 695         switch (pNv->Architecture) {
 696
 697         case NV_ARCH_10:
 698         //case NV_ARCH_20: waiting for confirmation from pq
 699                 pNv->EXADriverPtr->CheckComposite   = NV10CheckComposite;
 700                 pNv->EXADriverPtr->PrepareComposite = NV10PrepareComposite;
 701                 pNv->EXADriverPtr->Composite        = NV10Composite;
 702                 pNv->EXADriverPtr->DoneComposite    = NV10DoneComposite;
 703                 break;
 704
 705 #if (X_BYTE_ORDER == X_LITTLE_ENDIAN) && defined(ENABLE_NV30EXA)
 706         case NV_ARCH_30:
 707                 pNv->EXADriverPtr->CheckComposite   = NV30EXACheckComposite;
 708                 pNv->EXADriverPtr->PrepareComposite = NV30EXAPrepareComposite;
 709                 pNv->EXADriverPtr->Composite        = NV30EXAComposite;
 710                 pNv->EXADriverPtr->DoneComposite    = NV30EXADoneComposite;
 711                 break;
 712 #endif
 713         case NV_ARCH_40:
 714                 pNv->EXADriverPtr->CheckComposite   = NV40EXACheckComposite;
 715                 pNv->EXADriverPtr->PrepareComposite = NV40EXAPrepareComposite;
 716                 pNv->EXADriverPtr->Composite        = NV40EXAComposite;
 717                 pNv->EXADriverPtr->DoneComposite    = NV40EXADoneComposite;
 718                 break;
 719         case NV_ARCH_50:
 720                 break;
 721         default:
 722                 break;
 723         }
 724
 725         return exaDriverInit(pScreen, pNv->EXADriverPtr);
 726 }
 727