git.oblomov.eu Git - nouveau/blob - src/nv40_exa.c

   1 /*
   2  * Copyright 2007 Ben Skeggs
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  18  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  19  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  20  * SOFTWARE.
  21  */
  22
  23 #include "nv_include.h"
  24 #include "nv_shaders.h"
  25
  26 typedef struct nv_pict_surface_format {
  27         int      pict_fmt;
  28         uint32_t card_fmt;
  29 } nv_pict_surface_format_t;
  30
  31 typedef struct nv_pict_texture_format {
  32         int      pict_fmt;
  33         uint32_t card_fmt;
  34         uint32_t card_swz;
  35 } nv_pict_texture_format_t;
  36
  37 typedef struct nv_pict_op {
  38         Bool     src_alpha;
  39         Bool     dst_alpha;
  40         uint32_t src_card_op;
  41         uint32_t dst_card_op;
  42 } nv_pict_op_t;
  43
  44 typedef struct nv40_exa_state {
  45         Bool have_mask;
  46
  47         struct {
  48                 PictTransformPtr transform;
  49                 float width;
  50                 float height;
  51         } unit[2];
  52 } nv40_exa_state_t;
  53 static nv40_exa_state_t exa_state;
  54 #define NV40EXA_STATE nv40_exa_state_t *state = &exa_state
  55
  56 static nv_pict_surface_format_t
  57 NV40SurfaceFormat[] = {
  58         { PICT_a8r8g8b8 , NV40TCL_RT_FORMAT_COLOR_A8R8G8B8 },
  59         { PICT_x8r8g8b8 , NV40TCL_RT_FORMAT_COLOR_X8R8G8B8 },
  60         { PICT_r5g6b5   , NV40TCL_RT_FORMAT_COLOR_R5G6B5   },
  61         { PICT_a8       , NV40TCL_RT_FORMAT_COLOR_B8       },
  62         { -1, ~0 }
  63 };
  64
  65 static nv_pict_surface_format_t *
  66 NV40_GetPictSurfaceFormat(int format)
  67 {
  68         int i = 0;
  69
  70         while (NV40SurfaceFormat[i].pict_fmt != -1) {
  71                 if (NV40SurfaceFormat[i].pict_fmt == format)
  72                         return &NV40SurfaceFormat[i];
  73                 i++;
  74         }
  75
  76         return NULL;
  77 }
  78
  79 enum {
  80         NV40EXA_FPID_PASS_COL0 = 0,
  81         NV40EXA_FPID_PASS_TEX0 = 1,
  82         NV40EXA_FPID_COMPOSITE_MASK = 2,
  83         NV40EXA_FPID_COMPOSITE_MASK_SA_CA = 3,
  84         NV40EXA_FPID_COMPOSITE_MASK_CA = 4,
  85         NV40EXA_FPID_MAX = 5
  86 } NV40EXA_FPID;
  87
  88 static nv_shader_t *nv40_fp_map[NV40EXA_FPID_MAX] = {
  89         &nv30_fp_pass_col0,
  90         &nv30_fp_pass_tex0,
  91         &nv30_fp_composite_mask,
  92         &nv30_fp_composite_mask_sa_ca,
  93         &nv30_fp_composite_mask_ca
  94 };
  95
  96 static nv_shader_t *nv40_fp_map_a8[NV40EXA_FPID_MAX];
  97
  98 static void
  99 NV40EXAHackupA8Shaders(ScrnInfoPtr pScrn)
 100 {
 101         int s;
 102
 103         for (s = 0; s < NV40EXA_FPID_MAX; s++) {
 104                 nv_shader_t *def, *a8;
 105
 106                 def = nv40_fp_map[s];
 107                 a8 = xcalloc(1, sizeof(nv_shader_t));
 108                 a8->card_priv.NV30FP.num_regs = def->card_priv.NV30FP.num_regs;
 109                 a8->size = def->size + 4;
 110                 memcpy(a8->data, def->data, def->size * sizeof(uint32_t));
 111                 nv40_fp_map_a8[s] = a8;
 112
 113                 a8->data[a8->size - 8 + 0] &= ~0x00000081;
 114                 a8->data[a8->size - 4 + 0]  = 0x01401e81;
 115                 a8->data[a8->size - 4 + 1]  = 0x1c9dfe00;
 116                 a8->data[a8->size - 4 + 2]  = 0x0001c800;
 117                 a8->data[a8->size - 4 + 3]  = 0x0001c800;
 118         }
 119 }
 120
 121 #define _(r,tf,ts0x,ts0y,ts0z,ts0w,ts1x,ts1y,ts1z,ts1w)                        \
 122   {                                                                            \
 123   PICT_##r, NV40TCL_TEX_FORMAT_FORMAT_##tf,                                    \
 124   NV40TCL_TEX_SWIZZLE_S0_X_##ts0x | NV40TCL_TEX_SWIZZLE_S0_Y_##ts0y |          \
 125   NV40TCL_TEX_SWIZZLE_S0_Z_##ts0z | NV40TCL_TEX_SWIZZLE_S0_W_##ts0w |          \
 126   NV40TCL_TEX_SWIZZLE_S1_X_##ts1x | NV40TCL_TEX_SWIZZLE_S1_Y_##ts1y |          \
 127   NV40TCL_TEX_SWIZZLE_S1_Z_##ts1z | NV40TCL_TEX_SWIZZLE_S1_W_##ts1w,           \
 128   }
 129 static nv_pict_texture_format_t
 130 NV40TextureFormat[] = {
 131         _(a8r8g8b8, A8R8G8B8,   S1,   S1,   S1,   S1, X, Y, Z, W),
 132         _(x8r8g8b8, A8R8G8B8,   S1,   S1,   S1,  ONE, X, Y, Z, W),
 133         _(x8b8g8r8, A8R8G8B8,   S1,   S1,   S1,  ONE, Z, Y, X, W),
 134         _(a1r5g5b5, A1R5G5B5,   S1,   S1,   S1,   S1, X, Y, Z, W),
 135         _(x1r5g5b5, A1R5G5B5,   S1,   S1,   S1,  ONE, X, Y, Z, W),
 136         _(  r5g6b5,   R5G6B5,   S1,   S1,   S1,   S1, X, Y, Z, W),
 137         _(      a8,       L8, ZERO, ZERO, ZERO,   S1, X, X, X, X),
 138         { -1, ~0, ~0 }
 139 };
 140 #undef _
 141
 142 static nv_pict_texture_format_t *
 143 NV40_GetPictTextureFormat(int format)
 144 {
 145         int i = 0;
 146
 147         while (NV40TextureFormat[i].pict_fmt != -1) {
 148                 if (NV40TextureFormat[i].pict_fmt == format)
 149                         return &NV40TextureFormat[i];
 150                 i++;
 151         }
 152
 153         return NULL;
 154 }
 155
 156 #define SF(bf) (NV40TCL_BLEND_FUNC_SRC_RGB_##bf |                              \
 157                 NV40TCL_BLEND_FUNC_SRC_ALPHA_##bf)
 158 #define DF(bf) (NV40TCL_BLEND_FUNC_DST_RGB_##bf |                              \
 159                 NV40TCL_BLEND_FUNC_DST_ALPHA_##bf)
 160 static nv_pict_op_t
 161 NV40PictOp[] = {
 162 /* Clear       */ { 0, 0, SF(               ZERO), DF(               ZERO) },
 163 /* Src         */ { 0, 0, SF(                ONE), DF(               ZERO) },
 164 /* Dst         */ { 0, 0, SF(               ZERO), DF(                ONE) },
 165 /* Over        */ { 1, 0, SF(                ONE), DF(ONE_MINUS_SRC_ALPHA) },
 166 /* OverReverse */ { 0, 1, SF(ONE_MINUS_DST_ALPHA), DF(                ONE) },
 167 /* In          */ { 0, 1, SF(          DST_ALPHA), DF(               ZERO) },
 168 /* InReverse   */ { 1, 0, SF(               ZERO), DF(          SRC_ALPHA) },
 169 /* Out         */ { 0, 1, SF(ONE_MINUS_DST_ALPHA), DF(               ZERO) },
 170 /* OutReverse  */ { 1, 0, SF(               ZERO), DF(ONE_MINUS_SRC_ALPHA) },
 171 /* Atop        */ { 1, 1, SF(          DST_ALPHA), DF(ONE_MINUS_SRC_ALPHA) },
 172 /* AtopReverse */ { 1, 1, SF(ONE_MINUS_DST_ALPHA), DF(          SRC_ALPHA) },
 173 /* Xor         */ { 1, 1, SF(ONE_MINUS_DST_ALPHA), DF(ONE_MINUS_SRC_ALPHA) },
 174 /* Add         */ { 0, 0, SF(                ONE), DF(                ONE) }
 175 };
 176
 177 static nv_pict_op_t *
 178 NV40_GetPictOpRec(int op)
 179 {
 180         if (op >= PictOpSaturate)
 181                 return NULL;
 182         return &NV40PictOp[op];
 183 }
 184
 185 #if 0
 186 #define FALLBACK(fmt,args...) do {                                      \
 187         ErrorF("FALLBACK %s:%d> " fmt, __func__, __LINE__, ##args);     \
 188         return FALSE;                                                   \
 189 } while(0)
 190 #else
 191 #define FALLBACK(fmt,args...) do { \
 192         return FALSE;              \
 193 } while(0)
 194 #endif
 195
 196 static void
 197 NV40_SetupBlend(ScrnInfoPtr pScrn, nv_pict_op_t *blend,
 198                 PictFormatShort dest_format, Bool component_alpha)
 199 {
 200         NVPtr pNv = NVPTR(pScrn);
 201         uint32_t sblend, dblend;
 202
 203         sblend = blend->src_card_op;
 204         dblend = blend->dst_card_op;
 205
 206         if (blend->dst_alpha) {
 207                 if (!PICT_FORMAT_A(dest_format)) {
 208                         if (sblend == SF(DST_ALPHA)) {
 209                                 sblend = SF(ONE);
 210                         } else if (sblend == SF(ONE_MINUS_DST_ALPHA)) {
 211                                 sblend = SF(ZERO);
 212                         }
 213                 } else if (dest_format == PICT_a8) {
 214                         if (sblend == SF(DST_ALPHA)) {
 215                                 sblend = SF(DST_COLOR);
 216                         } else if (sblend == SF(ONE_MINUS_DST_ALPHA)) {
 217                                 sblend = SF(ONE_MINUS_DST_COLOR);
 218                         }
 219                 }
 220         }
 221
 222         if (blend->src_alpha && (component_alpha || dest_format == PICT_a8)) {
 223                 if (dblend == DF(SRC_ALPHA)) {
 224                         dblend = DF(SRC_COLOR);
 225                 } else if (dblend == DF(ONE_MINUS_SRC_ALPHA)) {
 226                         dblend = DF(ONE_MINUS_SRC_COLOR);
 227                 }
 228         }
 229
 230         if (sblend == SF(ONE) && dblend == DF(ZERO)) {
 231                 BEGIN_RING(Nv3D, NV40TCL_BLEND_ENABLE, 1);
 232                 OUT_RING  (0);
 233         } else {
 234                 BEGIN_RING(Nv3D, NV40TCL_BLEND_ENABLE, 5);
 235                 OUT_RING  (1);
 236                 OUT_RING  (sblend);
 237                 OUT_RING  (dblend);
 238                 OUT_RING  (0x00000000);
 239                 OUT_RING  (NV40TCL_BLEND_EQUATION_ALPHA_FUNC_ADD |
 240                            NV40TCL_BLEND_EQUATION_RGB_FUNC_ADD);
 241         }
 242 }
 243
 244 static Bool
 245 NV40EXATexture(ScrnInfoPtr pScrn, PixmapPtr pPix, PicturePtr pPict, int unit)
 246 {
 247         NVPtr pNv = NVPTR(pScrn);
 248         nv_pict_texture_format_t *fmt;
 249         NV40EXA_STATE;
 250
 251         fmt = NV40_GetPictTextureFormat(pPict->format);
 252         if (!fmt)
 253                 return FALSE;
 254
 255         BEGIN_RING(Nv3D, NV40TCL_TEX_OFFSET(unit), 8);
 256         OUT_PIXMAPl(pPix, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_GART | NOUVEAU_BO_RD);
 257         OUT_PIXMAPd(pPix, fmt->card_fmt | NV40TCL_TEX_FORMAT_LINEAR |
 258                     NV40TCL_TEX_FORMAT_DIMS_2D | NV40TCL_TEX_FORMAT_NO_BORDER |
 259                     (0x8000) | (1 << NV40TCL_TEX_FORMAT_MIPMAP_COUNT_SHIFT),
 260                     NOUVEAU_BO_VRAM | NOUVEAU_BO_GART | NOUVEAU_BO_RD,
 261                     NV40TCL_TEX_FORMAT_DMA0, NV40TCL_TEX_FORMAT_DMA1);
 262         if (pPict->repeat) {
 263                 switch(pPict->repeatType) {
 264                         /* At the moment i do not know if we can support RepeatPad. */
 265                         case RepeatReflect:
 266                                 OUT_RING  (NV40TCL_TEX_WRAP_S_MIRRORED_REPEAT |
 267                                 NV40TCL_TEX_WRAP_T_MIRRORED_REPEAT |
 268                                 NV40TCL_TEX_WRAP_R_MIRRORED_REPEAT);
 269                                 break;
 270                         case RepeatNormal:
 271                         default:
 272                                 OUT_RING  (NV40TCL_TEX_WRAP_S_REPEAT |
 273                                 NV40TCL_TEX_WRAP_T_REPEAT |
 274                                 NV40TCL_TEX_WRAP_R_REPEAT);
 275                                 break;
 276                 }
 277         } else {
 278                 OUT_RING  (NV40TCL_TEX_WRAP_S_CLAMP_TO_BORDER |
 279                                 NV40TCL_TEX_WRAP_T_CLAMP_TO_BORDER |
 280                                 NV40TCL_TEX_WRAP_R_CLAMP_TO_BORDER);
 281         }
 282         OUT_RING  (NV40TCL_TEX_ENABLE_ENABLE);
 283         OUT_RING  (fmt->card_swz);
 284         if (pPict->filter == PictFilterBilinear) {
 285                 OUT_RING  (NV40TCL_TEX_FILTER_MIN_LINEAR |
 286                            NV40TCL_TEX_FILTER_MAG_LINEAR |
 287                            0x3fd6);
 288         } else {
 289                 OUT_RING  (NV40TCL_TEX_FILTER_MIN_NEAREST |
 290                            NV40TCL_TEX_FILTER_MAG_NEAREST |
 291                            0x3fd6);
 292         }
 293         OUT_RING  ((pPix->drawable.width << 16) | pPix->drawable.height);
 294         OUT_RING  (0); /* border ARGB */
 295         BEGIN_RING(Nv3D, NV40TCL_TEX_SIZE1(unit), 1);
 296         OUT_RING  ((1 << NV40TCL_TEX_SIZE1_DEPTH_SHIFT) |
 297                    (uint32_t)exaGetPixmapPitch(pPix));
 298
 299         state->unit[unit].width         = (float)pPix->drawable.width;
 300         state->unit[unit].height        = (float)pPix->drawable.height;
 301         state->unit[unit].transform     = pPict->transform;
 302
 303         return TRUE;
 304 }
 305
 306 static Bool
 307 NV40_SetupSurface(ScrnInfoPtr pScrn, PixmapPtr pPix, PictFormatShort format)
 308 {
 309         NVPtr pNv = NVPTR(pScrn);
 310         nv_pict_surface_format_t *fmt;
 311
 312         fmt = NV40_GetPictSurfaceFormat(format);
 313         if (!fmt) {
 314                 ErrorF("AIII no format\n");
 315                 return FALSE;
 316         }
 317
 318         BEGIN_RING(Nv3D, NV40TCL_RT_FORMAT, 3);
 319         OUT_RING  (NV40TCL_RT_FORMAT_TYPE_LINEAR |
 320                    NV40TCL_RT_FORMAT_ZETA_Z24S8 |
 321                    fmt->card_fmt);
 322         OUT_RING  (exaGetPixmapPitch(pPix));
 323         OUT_PIXMAPl(pPix, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
 324
 325         return TRUE;
 326 }
 327
 328 static Bool
 329 NV40EXACheckCompositeTexture(PicturePtr pPict)
 330 {
 331         nv_pict_texture_format_t *fmt;
 332         int w = pPict->pDrawable->width;
 333         int h = pPict->pDrawable->height;
 334
 335         if ((w > 4096) || (h > 4096))
 336                 FALLBACK("picture too large, %dx%d\n", w, h);
 337
 338         fmt = NV40_GetPictTextureFormat(pPict->format);
 339         if (!fmt)
 340                 FALLBACK("picture format 0x%08x not supported\n",
 341                                 pPict->format);
 342
 343         if (pPict->filter != PictFilterNearest &&
 344             pPict->filter != PictFilterBilinear)
 345                 FALLBACK("filter 0x%x not supported\n", pPict->filter);
 346
 347         if (pPict->repeat && pPict->repeatType == RepeatPad)
 348                 FALLBACK("repeat 0x%x not supported\n", pPict->repeatType);
 349
 350         return TRUE;
 351 }
 352
 353 Bool
 354 NV40EXACheckComposite(int op, PicturePtr psPict,
 355                               PicturePtr pmPict,
 356                               PicturePtr pdPict)
 357 {
 358         nv_pict_surface_format_t *fmt;
 359         nv_pict_op_t *opr;
 360
 361         opr = NV40_GetPictOpRec(op);
 362         if (!opr)
 363                 FALLBACK("unsupported blend op 0x%x\n", op);
 364
 365         fmt = NV40_GetPictSurfaceFormat(pdPict->format);
 366         if (!fmt)
 367                 FALLBACK("dst picture format 0x%08x not supported\n",
 368                                 pdPict->format);
 369
 370         if (!NV40EXACheckCompositeTexture(psPict))
 371                 FALLBACK("src picture\n");
 372         if (pmPict) {
 373                 if (pmPict->componentAlpha &&
 374                     PICT_FORMAT_RGB(pmPict->format) &&
 375                     opr->src_alpha && opr->src_card_op != SF(ZERO))
 376                         FALLBACK("mask CA + SA\n");
 377                 if (!NV40EXACheckCompositeTexture(pmPict))
 378                         FALLBACK("mask picture\n");
 379         }
 380
 381         return TRUE;
 382 }
 383
 384 Bool
 385 NV40EXAPrepareComposite(int op, PicturePtr psPict,
 386                                 PicturePtr pmPict,
 387                                 PicturePtr pdPict,
 388                                 PixmapPtr  psPix,
 389                                 PixmapPtr  pmPix,
 390                                 PixmapPtr  pdPix)
 391 {
 392         ScrnInfoPtr pScrn = xf86Screens[psPix->drawable.pScreen->myNum];
 393         NVPtr pNv = NVPTR(pScrn);
 394         nv_pict_op_t *blend;
 395         int fpid = NV40EXA_FPID_PASS_COL0;
 396         NV40EXA_STATE;
 397
 398         blend = NV40_GetPictOpRec(op);
 399
 400         NV40_SetupBlend(pScrn, blend, pdPict->format,
 401                         (pmPict && pmPict->componentAlpha &&
 402                          PICT_FORMAT_RGB(pmPict->format)));
 403
 404         NV40_SetupSurface(pScrn, pdPix, pdPict->format);
 405         NV40EXATexture(pScrn, psPix, psPict, 0);
 406
 407         NV40_LoadVtxProg(pScrn, &nv40_vp_exa_render);
 408         if (pmPict) {
 409                 NV40EXATexture(pScrn, pmPix, pmPict, 1);
 410
 411                 if (pmPict->componentAlpha && PICT_FORMAT_RGB(pmPict->format)) {
 412                         if (blend->src_alpha)
 413                                 fpid = NV40EXA_FPID_COMPOSITE_MASK_SA_CA;
 414                         else
 415                                 fpid = NV40EXA_FPID_COMPOSITE_MASK_CA;
 416                 } else {
 417                         fpid = NV40EXA_FPID_COMPOSITE_MASK;
 418                 }
 419
 420                 state->have_mask = TRUE;
 421         } else {
 422                 fpid = NV40EXA_FPID_PASS_TEX0;
 423
 424                 state->have_mask = FALSE;
 425         }
 426
 427         if (pdPict->format == PICT_a8)
 428                 NV40_LoadFragProg(pScrn, nv40_fp_map_a8[fpid]);
 429         else
 430                 NV40_LoadFragProg(pScrn, nv40_fp_map[fpid]);
 431
 432         /* Appears to be some kind of cache flush, needed here at least
 433          * sometimes.. funky text rendering otherwise :)
 434          */
 435         BEGIN_RING(Nv3D, NV40TCL_TEX_CACHE_CTL, 1);
 436         OUT_RING  (2);
 437         BEGIN_RING(Nv3D, NV40TCL_TEX_CACHE_CTL, 1);
 438         OUT_RING  (1);
 439
 440         BEGIN_RING(Nv3D, NV40TCL_BEGIN_END, 1);
 441         OUT_RING  (NV40TCL_BEGIN_END_QUADS);
 442
 443         return TRUE;
 444 }
 445
 446 #define xFixedToFloat(v) \
 447         ((float)xFixedToInt((v)) + ((float)xFixedFrac(v) / 65536.0))
 448
 449 static inline void
 450 NV40EXATransformCoord(PictTransformPtr t, int x, int y, float sx, float sy,
 451                                           float *x_ret, float *y_ret)
 452 {
 453         if (t) {
 454                 PictVector v;
 455                 v.vector[0] = IntToxFixed(x);
 456                 v.vector[1] = IntToxFixed(y);
 457                 v.vector[2] = xFixed1;
 458                 PictureTransformPoint(t, &v);
 459                 *x_ret = xFixedToFloat(v.vector[0]) / sx;
 460                 *y_ret = xFixedToFloat(v.vector[1]) / sy;
 461         } else {
 462                 *x_ret = (float)x / sx;
 463                 *y_ret = (float)y / sy;
 464         }
 465 }
 466
 467 #define CV_OUTm(sx,sy,mx,my,dx,dy) do {                                        \
 468         BEGIN_RING(Nv3D, NV40TCL_VTX_ATTR_2F_X(8), 4);                         \
 469         OUT_RINGf ((sx)); OUT_RINGf ((sy));                                    \
 470         OUT_RINGf ((mx)); OUT_RINGf ((my));                                    \
 471         BEGIN_RING(Nv3D, NV40TCL_VTX_ATTR_2I(0), 1);                           \
 472         OUT_RING  (((dy)<<16)|(dx));                                           \
 473 } while(0)
 474 #define CV_OUT(sx,sy,dx,dy) do {                                               \
 475         BEGIN_RING(Nv3D, NV40TCL_VTX_ATTR_2F_X(8), 2);                         \
 476         OUT_RINGf ((sx)); OUT_RINGf ((sy));                                    \
 477         BEGIN_RING(Nv3D, NV40TCL_VTX_ATTR_2I(0), 1);                           \
 478         OUT_RING  (((dy)<<16)|(dx));                                           \
 479 } while(0)
 480
 481 void
 482 NV40EXAComposite(PixmapPtr pdPix, int srcX , int srcY,
 483                                   int maskX, int maskY,
 484                                   int dstX , int dstY,
 485                                   int width, int height)
 486 {
 487         ScrnInfoPtr pScrn = xf86Screens[pdPix->drawable.pScreen->myNum];
 488         NVPtr pNv = NVPTR(pScrn);
 489         float sX0, sX1, sX2, sY0, sY1, sY2, sX3, sY3;
 490         float mX0, mX1, mX2, mY0, mY1, mY2, mX3, mY3;
 491         NV40EXA_STATE;
 492
 493         NV40EXATransformCoord(state->unit[0].transform, srcX, srcY,
 494                               state->unit[0].width,
 495                               state->unit[0].height, &sX0, &sY0);
 496         NV40EXATransformCoord(state->unit[0].transform,
 497                               srcX + width, srcY,
 498                               state->unit[0].width,
 499                               state->unit[0].height, &sX1, &sY1);
 500         NV40EXATransformCoord(state->unit[0].transform,
 501                               srcX + width, srcY + height,
 502                               state->unit[0].width,
 503                               state->unit[0].height, &sX2, &sY2);
 504         NV40EXATransformCoord(state->unit[0].transform,
 505                               srcX, srcY + height,
 506                               state->unit[0].width,
 507                               state->unit[0].height, &sX3, &sY3);
 508
 509         if (state->have_mask) {
 510                 NV40EXATransformCoord(state->unit[1].transform, maskX, maskY,
 511                                       state->unit[1].width,
 512                                       state->unit[1].height, &mX0, &mY0);
 513                 NV40EXATransformCoord(state->unit[1].transform,
 514                                       maskX + width, maskY,
 515                                       state->unit[1].width,
 516                                       state->unit[1].height, &mX1, &mY1);
 517                 NV40EXATransformCoord(state->unit[1].transform,
 518                                       maskX + width, maskY + height,
 519                                       state->unit[1].width,
 520                                       state->unit[1].height, &mX2, &mY2);
 521                 NV40EXATransformCoord(state->unit[1].transform,
 522                                       maskX, maskY + height,
 523                                       state->unit[1].width,
 524                                       state->unit[1].height, &mX3, &mY3);
 525
 526                 CV_OUTm(sX0 , sY0 , mX0, mY0, dstX        ,          dstY);
 527                 CV_OUTm(sX1 , sY1 , mX1, mY1, dstX + width,          dstY);
 528                 CV_OUTm(sX2 , sY2 , mX2, mY2, dstX + width, dstY + height);
 529                 CV_OUTm(sX3 , sY3 , mX3, mY3, dstX        , dstY + height);
 530         } else {
 531                 CV_OUT(sX0 , sY0 , dstX        ,          dstY);
 532                 CV_OUT(sX1 , sY1 , dstX + width,          dstY);
 533                 CV_OUT(sX2 , sY2 , dstX + width, dstY + height);
 534                 CV_OUT(sX3 , sY3 , dstX        , dstY + height);
 535         }
 536
 537         FIRE_RING();
 538 }
 539
 540 void
 541 NV40EXADoneComposite(PixmapPtr pdPix)
 542 {
 543         ScrnInfoPtr pScrn = xf86Screens[pdPix->drawable.pScreen->myNum];
 544         NVPtr pNv = NVPTR(pScrn);
 545
 546         BEGIN_RING(Nv3D, NV40TCL_BEGIN_END, 1);
 547         OUT_RING  (NV40TCL_BEGIN_END_STOP);
 548 }
 549
 550 #define NV40TCL_CHIPSET_4X_MASK 0x00000baf
 551 #define NV44TCL_CHIPSET_4X_MASK 0x00005450
 552 Bool
 553 NVAccelInitNV40TCL(ScrnInfoPtr pScrn)
 554 {
 555         NVPtr pNv = NVPTR(pScrn);
 556         uint32_t class = 0, chipset;
 557         int i;
 558
 559         NV40EXAHackupA8Shaders(pScrn);
 560
 561         chipset = (nvReadMC(pNv, 0) >> 20) & 0xff;
 562         if ((chipset & 0xf0) != NV_ARCH_40)
 563                 return TRUE;
 564         chipset &= 0xf;
 565
 566         if (NV40TCL_CHIPSET_4X_MASK & (1<<chipset))
 567                 class = NV40TCL;
 568         else if (NV44TCL_CHIPSET_4X_MASK & (1<<chipset))
 569                 class = NV44TCL;
 570         else {
 571                 xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
 572                            "NV40EXA: Unknown chipset NV4%1x\n", chipset);
 573                 return FALSE;
 574         }
 575
 576         if (!pNv->Nv3D) {
 577                 if (nouveau_grobj_alloc(pNv->chan, Nv3D, class, &pNv->Nv3D))
 578                         return FALSE;
 579         }
 580
 581         BEGIN_RING(Nv3D, NV40TCL_DMA_NOTIFY, 1);
 582         OUT_RING  (pNv->notify0->handle);
 583         BEGIN_RING(Nv3D, NV40TCL_DMA_TEXTURE0, 2);
 584         OUT_RING  (pNv->chan->vram->handle);
 585         OUT_RING  (pNv->chan->gart->handle);
 586         BEGIN_RING(Nv3D, NV40TCL_DMA_COLOR0, 2);
 587         OUT_RING  (pNv->chan->vram->handle);
 588         OUT_RING  (pNv->chan->vram->handle);
 589
 590         /* voodoo */
 591         BEGIN_RING(Nv3D, 0x1ea4, 3);
 592         OUT_RING  (0x00000010);
 593         OUT_RING  (0x01000100);
 594         OUT_RING  (0xff800006);
 595         BEGIN_RING(Nv3D, 0x1fc4, 1);
 596         OUT_RING  (0x06144321);
 597         BEGIN_RING(Nv3D, 0x1fc8, 2);
 598         OUT_RING  (0xedcba987);
 599         OUT_RING  (0x00000021);
 600         BEGIN_RING(Nv3D, 0x1fd0, 1);
 601         OUT_RING  (0x00171615);
 602         BEGIN_RING(Nv3D, 0x1fd4, 1);
 603         OUT_RING  (0x001b1a19);
 604         BEGIN_RING(Nv3D, 0x1ef8, 1);
 605         OUT_RING  (0x0020ffff);
 606         BEGIN_RING(Nv3D, 0x1d64, 1);
 607         OUT_RING  (0x00d30000);
 608         BEGIN_RING(Nv3D, 0x1e94, 1);
 609         OUT_RING  (0x00000001);
 610
 611         BEGIN_RING(Nv3D, NV40TCL_VIEWPORT_TRANSLATE_X, 8);
 612         OUT_RINGf (0.0);
 613         OUT_RINGf (0.0);
 614         OUT_RINGf (0.0);
 615         OUT_RINGf (0.0);
 616         OUT_RINGf (1.0);
 617         OUT_RINGf (1.0);
 618         OUT_RINGf (1.0);
 619         OUT_RINGf (0.0);
 620
 621         /* default 3D state */
 622         /*XXX: replace with the same state that the DRI emits on startup */
 623         BEGIN_RING(Nv3D, NV40TCL_STENCIL_FRONT_ENABLE, 1);
 624         OUT_RING  (0);
 625         BEGIN_RING(Nv3D, NV40TCL_STENCIL_BACK_ENABLE, 1);
 626         OUT_RING  (0);
 627         BEGIN_RING(Nv3D, NV40TCL_ALPHA_TEST_ENABLE, 1);
 628         OUT_RING  (0);
 629         BEGIN_RING(Nv3D, NV40TCL_DEPTH_WRITE_ENABLE, 2);
 630         OUT_RING  (0);
 631         OUT_RING  (0);
 632         BEGIN_RING(Nv3D, NV40TCL_COLOR_MASK, 1);
 633         OUT_RING  (0x01010101); /* TR,TR,TR,TR */
 634         BEGIN_RING(Nv3D, NV40TCL_CULL_FACE_ENABLE, 1);
 635         OUT_RING  (0);
 636         BEGIN_RING(Nv3D, NV40TCL_BLEND_ENABLE, 1);
 637         OUT_RING  (0);
 638         BEGIN_RING(Nv3D, NV40TCL_COLOR_LOGIC_OP_ENABLE, 2);
 639         OUT_RING  (0);
 640         OUT_RING  (NV40TCL_COLOR_LOGIC_OP_COPY);
 641         BEGIN_RING(Nv3D, NV40TCL_DITHER_ENABLE, 1);
 642         OUT_RING  (0);
 643         BEGIN_RING(Nv3D, NV40TCL_SHADE_MODEL, 1);
 644         OUT_RING  (NV40TCL_SHADE_MODEL_SMOOTH);
 645         BEGIN_RING(Nv3D, NV40TCL_POLYGON_OFFSET_FACTOR,2);
 646         OUT_RINGf (0.0);
 647         OUT_RINGf (0.0);
 648         BEGIN_RING(Nv3D, NV40TCL_POLYGON_MODE_FRONT, 2);
 649         OUT_RING  (NV40TCL_POLYGON_MODE_FRONT_FILL);
 650         OUT_RING  (NV40TCL_POLYGON_MODE_BACK_FILL);
 651         BEGIN_RING(Nv3D, NV40TCL_POLYGON_STIPPLE_PATTERN(0), 0x20);
 652         for (i=0;i<0x20;i++)
 653                 OUT_RING  (0xFFFFFFFF);
 654         for (i=0;i<16;i++) {
 655                 BEGIN_RING(Nv3D, NV40TCL_TEX_ENABLE(i), 1);
 656                 OUT_RING  (0);
 657         }
 658
 659         BEGIN_RING(Nv3D, 0x1d78, 1);
 660         OUT_RING  (0x110);
 661
 662         BEGIN_RING(Nv3D, NV40TCL_RT_ENABLE, 1);
 663         OUT_RING  (NV40TCL_RT_ENABLE_COLOR0);
 664
 665         BEGIN_RING(Nv3D, NV40TCL_RT_HORIZ, 2);
 666         OUT_RING  ((4096 << 16));
 667         OUT_RING  ((4096 << 16));
 668         BEGIN_RING(Nv3D, NV40TCL_SCISSOR_HORIZ, 2);
 669         OUT_RING  ((4096 << 16));
 670         OUT_RING  ((4096 << 16));
 671         BEGIN_RING(Nv3D, NV40TCL_VIEWPORT_HORIZ, 2);
 672         OUT_RING  ((4096 << 16));
 673         OUT_RING  ((4096 << 16));
 674         BEGIN_RING(Nv3D, NV40TCL_VIEWPORT_CLIP_HORIZ(0), 2);
 675         OUT_RING  ((4095 << 16));
 676         OUT_RING  ((4095 << 16));
 677
 678         return TRUE;
 679 }
 680