git.oblomov.eu Git - nouveau/blob - src/nv30_exa.c

   1 /*
   2  * Copyright 2007 Ben Skeggs
   3  * Copyright 2007 Stephane Marchesin
   4  * Copyright 2007 Jeremy Kolb
   5  * Copyright 2007 Patrice Mandin
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the "Software"),
   9  * to deal in the Software without restriction, including without limitation
  10  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  11  * and/or sell copies of the Software, and to permit persons to whom the
  12  * Software is furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice shall be included in
  15  * all copies or substantial portions of the Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  21  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  22  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  */
  25
  26 #include "nv_include.h"
  27 #include "nv30_shaders.h"
  28
  29 typedef struct nv_pict_surface_format {
  30         int      pict_fmt;
  31         uint32_t card_fmt;
  32 } nv_pict_surface_format_t;
  33
  34 typedef struct nv_pict_texture_format {
  35         int      pict_fmt;
  36         uint32_t card_fmt;
  37         uint32_t card_swz;
  38 } nv_pict_texture_format_t;
  39
  40 typedef struct nv_pict_op {
  41         Bool     src_alpha;
  42         Bool     dst_alpha;
  43         uint32_t src_card_op;
  44         uint32_t dst_card_op;
  45 } nv_pict_op_t;
  46
  47 typedef struct nv30_exa_state {
  48         Bool have_mask;
  49
  50         struct {
  51                 PictTransformPtr transform;
  52                 float width;
  53                 float height;
  54         } unit[2];
  55 } nv30_exa_state_t;
  56 static nv30_exa_state_t exa_state;
  57 #define NV30EXA_STATE nv30_exa_state_t *state = &exa_state
  58
  59 static nv_pict_surface_format_t
  60 NV30SurfaceFormat[] = {
  61         { PICT_a8r8g8b8 , 0x148 },
  62         { PICT_a8b8g8r8 , 0x150 },
  63         { PICT_x8r8g8b8 , 0x145 },
  64         { PICT_x8b8g8r8 , 0x14f },
  65         { PICT_r5g6b5   , 0x143 },
  66         { PICT_a8       , 0x149 },
  67 };
  68
  69 static nv_pict_surface_format_t *
  70 NV30_GetPictSurfaceFormat(int format)
  71 {
  72         int i;
  73
  74         for(i=0;i<sizeof(NV30SurfaceFormat)/sizeof(NV30SurfaceFormat[0]);i++)
  75         {
  76                 if (NV30SurfaceFormat[i].pict_fmt == format)
  77                         return &NV30SurfaceFormat[i];
  78         }
  79
  80         return NULL;
  81 }
  82
  83 enum {
  84         NV30EXA_FPID_PASS_COL0 = 0,
  85         NV30EXA_FPID_PASS_TEX0 = 1,
  86         NV30EXA_FPID_COMPOSITE_MASK = 2,
  87         NV30EXA_FPID_COMPOSITE_MASK_SA_CA = 3,
  88         NV30EXA_FPID_COMPOSITE_MASK_CA = 4,
  89         NV30EXA_FPID_MAX = 5
  90 } NV30EXA_FPID;
  91
  92 static nv_shader_t *nv40_fp_map[NV30EXA_FPID_MAX] = {
  93         &nv30_fp_pass_col0,
  94         &nv30_fp_pass_tex0,
  95         &nv30_fp_composite_mask,
  96         &nv30_fp_composite_mask_sa_ca,
  97         &nv30_fp_composite_mask_ca
  98 };
  99
 100 static nv_shader_t *nv40_fp_map_a8[NV30EXA_FPID_MAX];
 101
 102 static void
 103 NV30EXAHackupA8Shaders(ScrnInfoPtr pScrn)
 104 {
 105         int s;
 106
 107         for (s = 0; s < NV30EXA_FPID_MAX; s++) {
 108                 nv_shader_t *def, *a8;
 109
 110                 def = nv40_fp_map[s];
 111                 a8 = xcalloc(1, sizeof(nv_shader_t));
 112                 a8->card_priv.NV30FP.num_regs = def->card_priv.NV30FP.num_regs;
 113                 a8->size = def->size + 4;
 114                 memcpy(a8->data, def->data, def->size * sizeof(uint32_t));
 115                 nv40_fp_map_a8[s] = a8;
 116
 117                 a8->data[a8->size - 8 + 0] &= ~0x00000081;
 118                 a8->data[a8->size - 4 + 0]  = 0x01401e81;
 119                 a8->data[a8->size - 4 + 1]  = 0x1c9dfe00;
 120                 a8->data[a8->size - 4 + 2]  = 0x0001c800;
 121                 a8->data[a8->size - 4 + 3]  = 0x0001c800;
 122         }
 123 }
 124
 125 /* should be in nouveau_reg.h at some point.. */
 126 #define NV34TCL_TX_SWIZZLE_UNIT_S0_X_SHIFT      14
 127 #define NV34TCL_TX_SWIZZLE_UNIT_S0_X_ZERO        0
 128 #define NV34TCL_TX_SWIZZLE_UNIT_S0_X_ONE         1
 129 #define NV34TCL_TX_SWIZZLE_UNIT_S0_X_S1          2
 130 #define NV34TCL_TX_SWIZZLE_UNIT_S0_Y_SHIFT      12
 131 #define NV34TCL_TX_SWIZZLE_UNIT_S0_Z_SHIFT      10
 132 #define NV34TCL_TX_SWIZZLE_UNIT_S0_W_SHIFT       8
 133 #define NV34TCL_TX_SWIZZLE_UNIT_S1_X_SHIFT       6
 134 #define NV34TCL_TX_SWIZZLE_UNIT_S1_X_X           3
 135 #define NV34TCL_TX_SWIZZLE_UNIT_S1_X_Y           2
 136 #define NV34TCL_TX_SWIZZLE_UNIT_S1_X_Z           1
 137 #define NV34TCL_TX_SWIZZLE_UNIT_S1_X_W           0
 138 #define NV34TCL_TX_SWIZZLE_UNIT_S1_Y_SHIFT       4
 139 #define NV34TCL_TX_SWIZZLE_UNIT_S1_Z_SHIFT       2
 140 #define NV34TCL_TX_SWIZZLE_UNIT_S1_W_SHIFT       0
 141
 142 #define _(r,tf,ts0x,ts0y,ts0z,ts0w,ts1x,ts1y,ts1z,ts1w)                       \
 143   {                                                                           \
 144   PICT_##r,                                                                   \
 145   (tf),                                                                       \
 146   (NV34TCL_TX_SWIZZLE_UNIT_S0_X_##ts0x << NV34TCL_TX_SWIZZLE_UNIT_S0_X_SHIFT)|\
 147   (NV34TCL_TX_SWIZZLE_UNIT_S0_X_##ts0y << NV34TCL_TX_SWIZZLE_UNIT_S0_Y_SHIFT)|\
 148   (NV34TCL_TX_SWIZZLE_UNIT_S0_X_##ts0z << NV34TCL_TX_SWIZZLE_UNIT_S0_Z_SHIFT)|\
 149   (NV34TCL_TX_SWIZZLE_UNIT_S0_X_##ts0w << NV34TCL_TX_SWIZZLE_UNIT_S0_W_SHIFT)|\
 150   (NV34TCL_TX_SWIZZLE_UNIT_S1_X_##ts1x << NV34TCL_TX_SWIZZLE_UNIT_S1_X_SHIFT)|\
 151   (NV34TCL_TX_SWIZZLE_UNIT_S1_X_##ts1y << NV34TCL_TX_SWIZZLE_UNIT_S1_Y_SHIFT)|\
 152   (NV34TCL_TX_SWIZZLE_UNIT_S1_X_##ts1z << NV34TCL_TX_SWIZZLE_UNIT_S1_Z_SHIFT)|\
 153   (NV34TCL_TX_SWIZZLE_UNIT_S1_X_##ts1w << NV34TCL_TX_SWIZZLE_UNIT_S1_W_SHIFT)\
 154   }
 155
 156 static nv_pict_texture_format_t
 157 NV30TextureFormat[] = {
 158         _(a8r8g8b8, 0x12,   S1,   S1,   S1,   S1, X, Y, Z, W),
 159         _(a8b8g8r8, 0x12,   S1,   S1,   S1,   S1, Z, Y, X, W),
 160         _(x8r8g8b8, 0x12,   S1,   S1,   S1,  ONE, X, Y, Z, W),
 161         _(x8b8g8r8, 0x12,   S1,   S1,   S1,  ONE, Z, Y, X, W),
 162         _(a1r5g5b5, 0x10,   S1,   S1,   S1,   S1, X, Y, Z, W),
 163         _(x1r5g5b5, 0x10,   S1,   S1,   S1,  ONE, X, Y, Z, W),
 164         _(x4r4g4b4, 0x1d,   S1,   S1,   S1,  ONE, X, Y, Z, W),
 165         _(a4r4g4b4, 0x1d,   S1,   S1,   S1,   S1, X, Y, Z, W),
 166         _(      a8, 0x1b, ZERO, ZERO, ZERO,   S1, X, X, X, X),
 167 };
 168
 169
 170 static nv_pict_texture_format_t *
 171 NV30_GetPictTextureFormat(int format)
 172 {
 173         int i;
 174
 175         for(i=0;i<sizeof(NV30TextureFormat)/sizeof(NV30TextureFormat[0]);i++)
 176         {
 177                 if (NV30TextureFormat[i].pict_fmt == format)
 178                         return &NV30TextureFormat[i];
 179         }
 180
 181         return NULL;
 182 }
 183
 184 #define NV34TCL_BF_ZERO                                     0x0000
 185 #define NV34TCL_BF_ONE                                      0x0001
 186 #define NV34TCL_BF_SRC_COLOR                                0x0300
 187 #define NV34TCL_BF_ONE_MINUS_SRC_COLOR                      0x0301
 188 #define NV34TCL_BF_SRC_ALPHA                                0x0302
 189 #define NV34TCL_BF_ONE_MINUS_SRC_ALPHA                      0x0303
 190 #define NV34TCL_BF_DST_ALPHA                                0x0304
 191 #define NV34TCL_BF_ONE_MINUS_DST_ALPHA                      0x0305
 192 #define NV34TCL_BF_DST_COLOR                                0x0306
 193 #define NV34TCL_BF_ONE_MINUS_DST_COLOR                      0x0307
 194 #define NV34TCL_BF_ALPHA_SATURATE                           0x0308
 195 #define BF(bf) NV34TCL_BF_##bf
 196
 197 static nv_pict_op_t
 198 NV30PictOp[] = {
 199 /* Clear       */ { 0, 0, BF(               ZERO), BF(               ZERO) },
 200 /* Src         */ { 0, 0, BF(                ONE), BF(               ZERO) },
 201 /* Dst         */ { 0, 0, BF(               ZERO), BF(                ONE) },
 202 /* Over        */ { 1, 0, BF(                ONE), BF(ONE_MINUS_SRC_ALPHA) },
 203 /* OverReverse */ { 0, 1, BF(ONE_MINUS_DST_ALPHA), BF(                ONE) },
 204 /* In          */ { 0, 1, BF(          DST_ALPHA), BF(               ZERO) },
 205 /* InReverse   */ { 1, 0, BF(               ZERO), BF(          SRC_ALPHA) },
 206 /* Out         */ { 0, 1, BF(ONE_MINUS_DST_ALPHA), BF(               ZERO) },
 207 /* OutReverse  */ { 1, 0, BF(               ZERO), BF(ONE_MINUS_SRC_ALPHA) },
 208 /* Atop        */ { 1, 1, BF(          DST_ALPHA), BF(ONE_MINUS_SRC_ALPHA) },
 209 /* AtopReverse */ { 1, 1, BF(ONE_MINUS_DST_ALPHA), BF(          SRC_ALPHA) },
 210 /* Xor         */ { 1, 1, BF(ONE_MINUS_DST_ALPHA), BF(ONE_MINUS_SRC_ALPHA) },
 211 /* Add         */ { 0, 0, BF(                ONE), BF(                ONE) }
 212 };
 213
 214 static nv_pict_op_t *
 215 NV30_GetPictOpRec(int op)
 216 {
 217         if (op >= PictOpSaturate)
 218                 return NULL;
 219 #if 0
 220         switch(op)
 221         {
 222                 case 0:ErrorF("Op Clear\n");break;
 223                 case 1:ErrorF("Op Src\n");break;
 224                 case 2:ErrorF("Op Dst\n");break;
 225                 case 3:ErrorF("Op Over\n");break;
 226                 case 4:ErrorF("Op OverReverse\n");break;
 227                 case 5:ErrorF("Op In\n");break;
 228                 case 6:ErrorF("Op InReverse\n");break;
 229                 case 7:ErrorF("Op Out\n");break;
 230                 case 8:ErrorF("Op OutReverse\n");break;
 231                 case 9:ErrorF("Op Atop\n");break;
 232                 case 10:ErrorF("Op AtopReverse\n");break;
 233                 case 11:ErrorF("Op Xor\n");break;
 234                 case 12:ErrorF("Op Add\n");break;
 235         }
 236 #endif
 237         return &NV30PictOp[op];
 238 }
 239
 240 static void
 241 NV30_SetupBlend(ScrnInfoPtr pScrn, nv_pict_op_t *blend,
 242                 PictFormatShort dest_format, Bool component_alpha)
 243 {
 244         NVPtr pNv = NVPTR(pScrn);
 245         uint32_t sblend, dblend;
 246
 247         sblend = blend->src_card_op;
 248         dblend = blend->dst_card_op;
 249
 250         if (blend->dst_alpha) {
 251                 if (!PICT_FORMAT_A(dest_format)) {
 252                         if (sblend == BF(DST_ALPHA)) {
 253                                 sblend = BF(ONE);
 254                         } else if (sblend == BF(ONE_MINUS_DST_ALPHA)) {
 255                                 sblend = BF(ZERO);
 256                         }
 257                 } else if (dest_format == PICT_a8) {
 258                         if (sblend == BF(DST_ALPHA)) {
 259                                 sblend = BF(DST_COLOR);
 260                         } else if (sblend == BF(ONE_MINUS_DST_ALPHA)) {
 261                                 sblend = BF(ONE_MINUS_DST_COLOR);
 262                         }
 263                 }
 264         }
 265
 266         if (blend->src_alpha && (component_alpha || dest_format == PICT_a8)) {
 267                 if (dblend == BF(SRC_ALPHA)) {
 268                         dblend = BF(SRC_COLOR);
 269                 } else if (dblend == BF(ONE_MINUS_SRC_ALPHA)) {
 270                         dblend = BF(ONE_MINUS_SRC_COLOR);
 271                 }
 272         }
 273
 274         if (sblend == BF(ONE) && dblend == BF(ZERO)) {
 275                 BEGIN_RING(Nv3D, NV34TCL_BLEND_FUNC_ENABLE, 1);
 276                 OUT_RING  (0);
 277         } else {
 278                 BEGIN_RING(Nv3D, NV34TCL_BLEND_FUNC_ENABLE, 3);
 279                 OUT_RING  (1);
 280                 OUT_RING  ((sblend << 16) | sblend);
 281                 OUT_RING  ((dblend << 16) | dblend);
 282         }
 283 }
 284
 285 static Bool
 286 NV30EXATexture(ScrnInfoPtr pScrn, PixmapPtr pPix, PicturePtr pPict, int unit)
 287 {
 288         NVPtr pNv = NVPTR(pScrn);
 289         nv_pict_texture_format_t *fmt;
 290         uint32_t card_filter, card_repeat;
 291         NV30EXA_STATE;
 292
 293         fmt = NV30_GetPictTextureFormat(pPict->format);
 294         if (!fmt)
 295                 return FALSE;
 296
 297         card_repeat = 3; /* repeatNone */
 298
 299         if (pPict->filter == PictFilterBilinear)
 300                 card_filter = 2;
 301         else
 302                 card_filter = 1;
 303
 304         BEGIN_RING(Nv3D, NV34TCL_TX_OFFSET(unit), 8);
 305         OUT_PIXMAPl(pPix, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
 306
 307         OUT_RING  (NV34TCL_TX_FORMAT_DIMS_2D |
 308                         (fmt->card_fmt << NV34TCL_TX_FORMAT_FORMAT_SHIFT) |
 309                         (1 << NV34TCL_TX_FORMAT_MIPMAP_LEVELS_SHIFT) |
 310                         (log2i(pPix->drawable.width)  << NV34TCL_TX_FORMAT_BASE_SIZE_U_SHIFT) |
 311                         (log2i(pPix->drawable.height) << NV34TCL_TX_FORMAT_BASE_SIZE_V_SHIFT) |
 312                         8 |
 313                         NV34TCL_TX_FORMAT_DMA0);
 314
 315         OUT_RING  ((card_repeat << NV34TCL_TX_WRAP_S_SHIFT) |
 316                         (card_repeat << NV34TCL_TX_WRAP_T_SHIFT) |
 317                         (card_repeat << NV34TCL_TX_WRAP_R_SHIFT));
 318         OUT_RING  (NV34TCL_TX_ENABLE_ENABLE);
 319         OUT_RING  ((((uint32_t)exaGetPixmapPitch(pPix)) << NV34TCL_TX_SWIZZLE_RECT_PITCH_SHIFT ) |
 320                         fmt->card_swz);
 321
 322         OUT_RING  ((card_filter << NV34TCL_TX_FILTER_MINIFY_SHIFT) /* min */ |
 323                         (card_filter << NV34TCL_TX_FILTER_MAGNIFY_SHIFT) /* mag */ |
 324                         0x2000 /* engine lock */);
 325         OUT_RING  ((pPix->drawable.width << NV34TCL_TX_NPOT_SIZE_W_SHIFT) | pPix->drawable.height);
 326         OUT_RING  (0); /* border ARGB */
 327
 328         state->unit[unit].width         = (float)pPix->drawable.width;
 329         state->unit[unit].height        = (float)pPix->drawable.height;
 330         state->unit[unit].transform     = pPict->transform;
 331
 332         return TRUE;
 333 }
 334
 335 static Bool
 336 NV30_SetupSurface(ScrnInfoPtr pScrn, PixmapPtr pPix, PicturePtr pPict)
 337 {
 338         NVPtr pNv = NVPTR(pScrn);
 339         nv_pict_surface_format_t *fmt;
 340
 341         fmt = NV30_GetPictSurfaceFormat(pPict->format);
 342         if (!fmt) {
 343                 ErrorF("AIII no format\n");
 344                 return FALSE;
 345         }
 346
 347         uint32_t pitch = (uint32_t)exaGetPixmapPitch(pPix);
 348
 349         BEGIN_RING(Nv3D, NV34TCL_RT_FORMAT, 3);
 350         OUT_RING  (fmt->card_fmt); /* format */
 351         OUT_RING  (pitch << 16 | pitch);
 352         OUT_PIXMAPl(pPix, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
 353
 354         return TRUE;
 355 }
 356
 357 static Bool
 358 NV30EXACheckCompositeTexture(PicturePtr pPict)
 359 {
 360         nv_pict_texture_format_t *fmt;
 361         int w = pPict->pDrawable->width;
 362         int h = pPict->pDrawable->height;
 363
 364         if ((w > 4096) || (h>4096))
 365                 NOUVEAU_FALLBACK("picture too large, %dx%d\n", w, h);
 366
 367         fmt = NV30_GetPictTextureFormat(pPict->format);
 368         if (!fmt)
 369                 NOUVEAU_FALLBACK("picture format 0x%08x not supported\n",
 370                                 pPict->format);
 371
 372         if (pPict->filter != PictFilterNearest &&
 373                         pPict->filter != PictFilterBilinear)
 374                 NOUVEAU_FALLBACK("filter 0x%x not supported\n", pPict->filter);
 375
 376         if (!(w==1 && h==1) && pPict->repeat && pPict->repeatType != RepeatNone)
 377                 NOUVEAU_FALLBACK("repeat 0x%x not supported (surface %dx%d)\n",
 378                                  pPict->repeatType,w,h);
 379
 380         return TRUE;
 381 }
 382
 383 Bool
 384 NV30EXACheckComposite(int op, PicturePtr psPict,
 385                 PicturePtr pmPict,
 386                 PicturePtr pdPict)
 387 {
 388         nv_pict_surface_format_t *fmt;
 389         nv_pict_op_t *opr;
 390
 391         opr = NV30_GetPictOpRec(op);
 392         if (!opr)
 393                 NOUVEAU_FALLBACK("unsupported blend op 0x%x\n", op);
 394
 395         fmt = NV30_GetPictSurfaceFormat(pdPict->format);
 396         if (!fmt)
 397                 NOUVEAU_FALLBACK("dst picture format 0x%08x not supported\n",
 398                                 pdPict->format);
 399
 400         if (!NV30EXACheckCompositeTexture(psPict))
 401                 NOUVEAU_FALLBACK("src picture\n");
 402         if (pmPict) {
 403                 if (pmPict->componentAlpha &&
 404                                 PICT_FORMAT_RGB(pmPict->format) &&
 405                                 opr->src_alpha && opr->src_card_op != BF(ZERO))
 406                         NOUVEAU_FALLBACK("mask CA + SA\n");
 407                 if (!NV30EXACheckCompositeTexture(pmPict))
 408                         NOUVEAU_FALLBACK("mask picture\n");
 409         }
 410
 411         return TRUE;
 412 }
 413
 414 Bool
 415 NV30EXAPrepareComposite(int op, PicturePtr psPict,
 416                 PicturePtr pmPict,
 417                 PicturePtr pdPict,
 418                 PixmapPtr  psPix,
 419                 PixmapPtr  pmPix,
 420                 PixmapPtr  pdPix)
 421 {
 422         ScrnInfoPtr pScrn = xf86Screens[psPix->drawable.pScreen->myNum];
 423         NVPtr pNv = NVPTR(pScrn);
 424         nv_pict_op_t *blend;
 425         int fpid = NV30EXA_FPID_PASS_COL0;
 426         NV30EXA_STATE;
 427
 428         blend = NV30_GetPictOpRec(op);
 429
 430         NV30_SetupBlend(pScrn, blend, pdPict->format,
 431                         (pmPict && pmPict->componentAlpha &&
 432                          PICT_FORMAT_RGB(pmPict->format)));
 433
 434         NV30_SetupSurface(pScrn, pdPix, pdPict);
 435         NV30EXATexture(pScrn, psPix, psPict, 0);
 436
 437 #if 0
 438 #define printformat(f) ErrorF("(%xh %s %dbpp A%dR%dG%dB%d)",f,(f>>16)&0xf==2?"ARGB":"ABGR",(f>>24),(f&0xf000)>>12,(f&0xf00)>>8,(f&0xf0)>>4,f&0xf)
 439         ErrorF("Preparecomposite src(%dx%d)",psPict->pDrawable->width,psPict->pDrawable->height);
 440         printformat((psPict->format));
 441         ErrorF(" dst(%dx%d)",pdPict->pDrawable->width,pdPict->pDrawable->height);
 442         printformat((pdPict->format));
 443         if (pmPict)
 444         {
 445                 ErrorF(" mask(%dx%d)",pmPict->pDrawable->width,pmPict->pDrawable->height);
 446                 printformat((pmPict->format));
 447         }
 448         ErrorF("\n");
 449 #endif
 450
 451         if (pmPict) {
 452                 NV30EXATexture(pScrn, pmPix, pmPict, 1);
 453
 454                 if (pmPict->componentAlpha && PICT_FORMAT_RGB(pmPict->format)) {
 455                         if (blend->src_alpha)
 456                                 fpid = NV30EXA_FPID_COMPOSITE_MASK_SA_CA;
 457                         else
 458                                 fpid = NV30EXA_FPID_COMPOSITE_MASK_CA;
 459                 } else {
 460                         fpid = NV30EXA_FPID_COMPOSITE_MASK;
 461                 }
 462
 463                 state->have_mask = TRUE;
 464         } else {
 465                 fpid = NV30EXA_FPID_PASS_TEX0;
 466
 467                 state->have_mask = FALSE;
 468         }
 469
 470         if (pdPict->format == PICT_a8)
 471                 NV30_LoadFragProg(pScrn, nv40_fp_map_a8[fpid]);
 472         else
 473                 NV30_LoadFragProg(pScrn, nv40_fp_map[fpid]);
 474
 475         BEGIN_RING(Nv3D, 0x23c, 1);
 476         OUT_RING  (pmPict?3:1);
 477
 478         BEGIN_RING(Nv3D, NV34TCL_VERTEX_BEGIN_END, 1);
 479         OUT_RING  (NV34TCL_VERTEX_BEGIN_END_TRIANGLES);
 480
 481         return TRUE;
 482 }
 483
 484 #define xFixedToFloat(v) \
 485         ((float)xFixedToInt((v)) + ((float)xFixedFrac(v) / 65536.0))
 486
 487 static void
 488 NV30EXATransformCoord(PictTransformPtr t, int x, int y, float sx, float sy,
 489                                           float *x_ret, float *y_ret)
 490 {
 491         PictVector v;
 492
 493         if (t) {
 494                 v.vector[0] = IntToxFixed(x);
 495                 v.vector[1] = IntToxFixed(y);
 496                 v.vector[2] = xFixed1;
 497                 PictureTransformPoint(t, &v);
 498                 *x_ret = xFixedToFloat(v.vector[0]);
 499                 *y_ret = xFixedToFloat(v.vector[1]);
 500         } else {
 501                 *x_ret = (float)x;
 502                 *y_ret = (float)y;
 503         }
 504 }
 505
 506 #define CV_OUTm(sx,sy,mx,my,dx,dy) do {                                        \
 507         BEGIN_RING(Nv3D, NV34TCL_VTX_ATTR_2F_X(8), 4);                         \
 508         OUT_RINGf ((sx)); OUT_RINGf ((sy));                                    \
 509         OUT_RINGf ((mx)); OUT_RINGf ((my));                                    \
 510         BEGIN_RING(Nv3D, NV34TCL_VTX_ATTR_2I(0), 1);                           \
 511         OUT_RING  (((dy)<<16)|(dx));                                           \
 512 } while(0)
 513 #define CV_OUT(sx,sy,dx,dy) do {                                               \
 514         BEGIN_RING(Nv3D, NV34TCL_VTX_ATTR_2F_X(8), 2);                         \
 515         OUT_RINGf ((sx)); OUT_RINGf ((sy));                                    \
 516         BEGIN_RING(Nv3D, NV34TCL_VTX_ATTR_2I(0), 1);                           \
 517         OUT_RING  (((dy)<<16)|(dx));                                           \
 518 } while(0)
 519
 520 void
 521 NV30EXAComposite(PixmapPtr pdPix, int srcX , int srcY,
 522                                   int maskX, int maskY,
 523                                   int dstX , int dstY,
 524                                   int width, int height)
 525 {
 526         ScrnInfoPtr pScrn = xf86Screens[pdPix->drawable.pScreen->myNum];
 527         NVPtr pNv = NVPTR(pScrn);
 528         float sX0, sX1, sX2, sY0, sY1, sY2;
 529         float mX0, mX1, mX2, mY0, mY1, mY2;
 530         NV30EXA_STATE;
 531
 532         /* We're drawing a triangle, we need to scissor it to a quad. */
 533         /* The scissors are here for a good reason, we don't get the full image, but just a part. */
 534         /* Handling the cliprects is done for us already. */
 535         BEGIN_RING(Nv3D, NV34TCL_SCISSOR_HORIZ, 2);
 536         OUT_RING  ((width << 16) | dstX);
 537         OUT_RING  ((height << 16) | dstY);
 538
 539 #if 0
 540         ErrorF("Composite [%dx%d] (%d,%d)IN(%d,%d)OP(%d,%d)\n",width,height,srcX,srcY,maskX,maskY,dstX,dstY);
 541 #endif
 542         NV30EXATransformCoord(state->unit[0].transform,
 543                                 srcX, srcY - height,
 544                                 state->unit[0].width,
 545                                 state->unit[0].height, &sX0, &sY0);
 546         NV30EXATransformCoord(state->unit[0].transform,
 547                                 srcX, srcY + height,
 548                                 state->unit[0].width,
 549                                 state->unit[0].height, &sX1, &sY1);
 550         NV30EXATransformCoord(state->unit[0].transform,
 551                                 srcX + 2*width, srcY + height,
 552                                 state->unit[0].width,
 553                                 state->unit[0].height, &sX2, &sY2);
 554
 555         if (state->have_mask) {
 556                 NV30EXATransformCoord(state->unit[1].transform,
 557                                         maskX, maskY - height,
 558                                         state->unit[1].width,
 559                                         state->unit[1].height, &mX0, &mY0);
 560                 NV30EXATransformCoord(state->unit[1].transform,
 561                                         maskX, maskY + height,
 562                                         state->unit[1].width,
 563                                         state->unit[1].height, &mX1, &mY1);
 564                 NV30EXATransformCoord(state->unit[1].transform,
 565                                         maskX + 2*width, maskY + height,
 566                                         state->unit[1].width,
 567                                         state->unit[1].height, &mX2, &mY2);
 568
 569                 CV_OUTm(sX0 , sY0 , mX0, mY0, dstX                      ,       dstY - height);
 570                 CV_OUTm(sX1 , sY1 , mX1, mY1, dstX                      ,       dstY + height);
 571                 CV_OUTm(sX2 , sY2 , mX2, mY2, dstX + 2*width    ,       dstY + height);
 572         } else {
 573                 CV_OUT(sX0 , sY0 , dstX                 ,       dstY - height);
 574                 CV_OUT(sX1 , sY1 , dstX                 ,       dstY + height);
 575                 CV_OUT(sX2 , sY2 , dstX + 2*width       ,       dstY + height);
 576         }
 577 }
 578
 579 void
 580 NV30EXADoneComposite(PixmapPtr pdPix)
 581 {
 582         ScrnInfoPtr pScrn = xf86Screens[pdPix->drawable.pScreen->myNum];
 583         NVPtr pNv = NVPTR(pScrn);
 584
 585         BEGIN_RING(Nv3D, NV34TCL_VERTEX_BEGIN_END, 1);
 586         OUT_RING  (0);
 587
 588         FIRE_RING();
 589 }
 590
 591 Bool
 592 NVAccelInitNV30TCL(ScrnInfoPtr pScrn)
 593 {
 594         NVPtr pNv = NVPTR(pScrn);
 595         uint32_t class = 0, chipset;
 596         int i;
 597
 598         NV30EXAHackupA8Shaders(pScrn);
 599
 600 #define NV30TCL_CHIPSET_3X_MASK 0x00000003
 601 #define NV35TCL_CHIPSET_3X_MASK 0x000001e0
 602 #define NV34TCL_CHIPSET_3X_MASK 0x00000010
 603
 604         chipset = (nvReadMC(pNv, NV_PMC_BOOT_0) >> 20) & 0xff;
 605         if ((chipset & 0xf0) != NV_ARCH_30)
 606                 return TRUE;
 607         chipset &= 0xf;
 608
 609         if (NV30TCL_CHIPSET_3X_MASK & (1<<chipset))
 610                 class = NV30TCL;
 611         else if (NV35TCL_CHIPSET_3X_MASK & (1<<chipset))
 612                 class = NV35TCL;
 613         else if (NV34TCL_CHIPSET_3X_MASK & (1<<chipset))
 614                 class = NV34TCL;
 615         else {
 616                 xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
 617                            "NV30EXA: Unknown chipset NV3%1x\n", chipset);
 618                 return FALSE;
 619         }
 620
 621
 622         if (!pNv->Nv3D) {
 623                 if (nouveau_grobj_alloc(pNv->chan, Nv3D, class, &pNv->Nv3D))
 624                         return FALSE;
 625         }
 626
 627         BEGIN_RING(Nv3D, NV34TCL_DMA_TEXTURE0, 3);
 628         OUT_RING  (pNv->chan->vram->handle);
 629         OUT_RING  (pNv->chan->gart->handle);
 630         OUT_RING  (pNv->chan->vram->handle);
 631         BEGIN_RING(Nv3D, NV34TCL_DMA_IN_MEMORY7, 1);
 632         OUT_RING  (pNv->chan->vram->handle);
 633         BEGIN_RING(Nv3D, NV34TCL_DMA_COLOR0, 2);
 634         OUT_RING  (pNv->chan->vram->handle);
 635         OUT_RING  (pNv->chan->vram->handle);
 636         BEGIN_RING(Nv3D, NV34TCL_DMA_IN_MEMORY8, 1);
 637         OUT_RING  (pNv->chan->vram->handle);
 638
 639         for (i=1; i<8; i++) {
 640                 BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_CLIP_HORIZ(i), 2);
 641                 OUT_RING  (0);
 642                 OUT_RING  (0);
 643         }
 644
 645         BEGIN_RING(Nv3D, 0x220, 1);
 646         OUT_RING  (1);
 647
 648         BEGIN_RING(Nv3D, 0x03b0, 1);
 649         OUT_RING  (0x00100000);
 650         BEGIN_RING(Nv3D, 0x1454, 1);
 651         OUT_RING  (0);
 652         BEGIN_RING(Nv3D, 0x1d80, 1);
 653         OUT_RING  (3);
 654         BEGIN_RING(Nv3D, 0x1450, 1);
 655         OUT_RING  (0x00030004);
 656
 657         /* NEW */
 658         BEGIN_RING(Nv3D, 0x1e98, 1);
 659         OUT_RING  (0);
 660         BEGIN_RING(Nv3D, 0x17e0, 3);
 661         OUT_RING  (0);
 662         OUT_RING  (0);
 663         OUT_RING  (0x3f800000);
 664         BEGIN_RING(Nv3D, 0x1f80, 16);
 665         OUT_RING  (0); OUT_RING  (0); OUT_RING  (0); OUT_RING  (0);
 666         OUT_RING  (0); OUT_RING  (0); OUT_RING  (0); OUT_RING  (0);
 667         OUT_RING  (0x0000ffff);
 668         OUT_RING  (0); OUT_RING  (0); OUT_RING  (0); OUT_RING  (0);
 669         OUT_RING  (0); OUT_RING  (0); OUT_RING  (0);
 670
 671         BEGIN_RING(Nv3D, 0x120, 3);
 672         OUT_RING  (0);
 673         OUT_RING  (1);
 674         OUT_RING  (2);
 675
 676         BEGIN_RING(NvImageBlit, 0x120, 3);
 677         OUT_RING  (0);
 678         OUT_RING  (1);
 679         OUT_RING  (2);
 680
 681         BEGIN_RING(Nv3D, 0x1d88, 1);
 682         OUT_RING  (0x00001200);
 683
 684         BEGIN_RING(Nv3D, NV34TCL_RC_ENABLE, 1);
 685         OUT_RING  (0);
 686
 687         /* Attempt to setup a known state.. Probably missing a heap of
 688          * stuff here..
 689          */
 690         BEGIN_RING(Nv3D, NV34TCL_STENCIL_FRONT_ENABLE, 1);
 691         OUT_RING  (0);
 692         BEGIN_RING(Nv3D, NV34TCL_STENCIL_BACK_ENABLE, 1);
 693         OUT_RING  (0);
 694         BEGIN_RING(Nv3D, NV34TCL_ALPHA_FUNC_ENABLE, 1);
 695         OUT_RING  (0);
 696         BEGIN_RING(Nv3D, NV34TCL_DEPTH_WRITE_ENABLE, 2);
 697         OUT_RING  (0); /* wr disable */
 698         OUT_RING  (0); /* test disable */
 699         BEGIN_RING(Nv3D, NV34TCL_COLOR_MASK, 1);
 700         OUT_RING  (0x01010101); /* TR,TR,TR,TR */
 701         BEGIN_RING(Nv3D, NV34TCL_CULL_FACE_ENABLE, 1);
 702         OUT_RING  (0);
 703         BEGIN_RING(Nv3D, NV34TCL_BLEND_FUNC_ENABLE, 5);
 704         OUT_RING  (0);                          /* Blend enable */
 705         OUT_RING  (0);                          /* Blend src */
 706         OUT_RING  (0);                          /* Blend dst */
 707         OUT_RING  (0x00000000);                 /* Blend colour */
 708         OUT_RING  (0x8006);                     /* FUNC_ADD */
 709         BEGIN_RING(Nv3D, NV34TCL_COLOR_LOGIC_OP_ENABLE, 2);
 710         OUT_RING  (0);
 711         OUT_RING  (0x1503 /*GL_COPY*/);
 712         BEGIN_RING(Nv3D, NV34TCL_DITHER_ENABLE, 1);
 713         OUT_RING  (1);
 714         BEGIN_RING(Nv3D, NV34TCL_SHADE_MODEL, 1);
 715         OUT_RING  (0x1d01 /*GL_SMOOTH*/);
 716         BEGIN_RING(Nv3D, NV34TCL_POLYGON_OFFSET_FACTOR,2);
 717         OUT_RINGf (0.0);
 718         OUT_RINGf (0.0);
 719         BEGIN_RING(Nv3D, NV34TCL_POLYGON_MODE_FRONT, 2);
 720         OUT_RING  (0x1b02 /*GL_FILL*/);
 721         OUT_RING  (0x1b02 /*GL_FILL*/);
 722         /* - Disable texture units
 723          * - Set fragprog to MOVR result.color, fragment.color */
 724         for (i=0;i<4;i++) {
 725                 BEGIN_RING(Nv3D, NV34TCL_TX_ENABLE(i), 1);
 726                 OUT_RING  (0);
 727         }
 728         /* Polygon stipple */
 729         BEGIN_RING(Nv3D, NV34TCL_POLYGON_STIPPLE_PATTERN(0), 0x20);
 730         for (i=0;i<0x20;i++)
 731                 OUT_RING  (0xFFFFFFFF);
 732
 733         BEGIN_RING(Nv3D, NV34TCL_DEPTH_RANGE_NEAR, 2);
 734         OUT_RINGf  (0.0);
 735         OUT_RINGf  (1.0);
 736
 737         /* Ok.  If you start X with the nvidia driver, kill it, and then
 738          * start X with nouveau you will get black rendering instead of
 739          * what you'd expect.  This fixes the problem, and it seems that
 740          * it's not needed between nouveau restarts - which suggests that
 741          * the 3D context (wherever it's stored?) survives somehow.
 742          */
 743         //BEGIN_RING(Nv3D, 0x1d60,1);
 744         //OUT_RING  (0x03008000);
 745
 746         int w=4096;
 747         int h=4096;
 748         int pitch=4096*4;
 749         BEGIN_RING(Nv3D, NV34TCL_RT_HORIZ, 5);
 750         OUT_RING  (w<<16);
 751         OUT_RING  (h<<16);
 752         OUT_RING  (0x148); /* format */
 753         OUT_RING  (pitch << 16 | pitch);
 754         OUT_RING  (0x0);
 755         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_TX_ORIGIN, 1);
 756         OUT_RING  (0);
 757         BEGIN_RING(Nv3D, 0x0a00, 2);
 758         OUT_RING  ((w<<16) | 0);
 759         OUT_RING  ((h<<16) | 0);
 760         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_CLIP_HORIZ(0), 2);
 761         OUT_RING  ((w-1)<<16);
 762         OUT_RING  ((h-1)<<16);
 763         BEGIN_RING(Nv3D, NV34TCL_SCISSOR_HORIZ, 2);
 764         OUT_RING  (w<<16);
 765         OUT_RING  (h<<16);
 766         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_HORIZ, 2);
 767         OUT_RING  (w<<16);
 768         OUT_RING  (h<<16);
 769
 770         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_TRANSLATE_X, 8);
 771         OUT_RINGf (0.0);
 772         OUT_RINGf (0.0);
 773         OUT_RINGf (0.0);
 774         OUT_RINGf (0.0);
 775         OUT_RINGf (1.0);
 776         OUT_RINGf (1.0);
 777         OUT_RINGf (1.0);
 778         OUT_RINGf (0.0);
 779
 780         BEGIN_RING(Nv3D, NV34TCL_MODELVIEW_MATRIX(0), 16);
 781         OUT_RINGf (1.0);
 782         OUT_RINGf (0.0);
 783         OUT_RINGf (0.0);
 784         OUT_RINGf (0.0);
 785         OUT_RINGf (0.0);
 786         OUT_RINGf (1.0);
 787         OUT_RINGf (0.0);
 788         OUT_RINGf (0.0);
 789         OUT_RINGf (0.0);
 790         OUT_RINGf (0.0);
 791         OUT_RINGf (1.0);
 792         OUT_RINGf (0.0);
 793         OUT_RINGf (0.0);
 794         OUT_RINGf (0.0);
 795         OUT_RINGf (0.0);
 796         OUT_RINGf (1.0);
 797
 798         BEGIN_RING(Nv3D, NV34TCL_PROJECTION_MATRIX(0), 16);
 799         OUT_RINGf (1.0);
 800         OUT_RINGf (0.0);
 801         OUT_RINGf (0.0);
 802         OUT_RINGf (0.0);
 803         OUT_RINGf (0.0);
 804         OUT_RINGf (1.0);
 805         OUT_RINGf (0.0);
 806         OUT_RINGf (0.0);
 807         OUT_RINGf (0.0);
 808         OUT_RINGf (0.0);
 809         OUT_RINGf (1.0);
 810         OUT_RINGf (0.0);
 811         OUT_RINGf (0.0);
 812         OUT_RINGf (0.0);
 813         OUT_RINGf (0.0);
 814         OUT_RINGf (1.0);
 815
 816         BEGIN_RING(Nv3D, NV34TCL_SCISSOR_HORIZ, 2);
 817         OUT_RING  (4096<<16);
 818         OUT_RING  (4096<<16);
 819
 820         return TRUE;
 821 }
 822