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 #define FALLBACK_DEBUG 0
 241 #if FALLBACK_DEBUG == 1
 242 #define FALLBACK(fmt,args...) do {                                      \
 243         ErrorF("FALLBACK %s:%d> " fmt, __func__, __LINE__, ##args);     \
 244         return FALSE;                                                   \
 245 } while(0)
 246 #else
 247 #define FALLBACK(fmt,args...) do { \
 248         return FALSE;              \
 249 } while(0)
 250 #endif
 251
 252 static void
 253 NV30_SetupBlend(ScrnInfoPtr pScrn, nv_pict_op_t *blend,
 254                 PictFormatShort dest_format, Bool component_alpha)
 255 {
 256         NVPtr pNv = NVPTR(pScrn);
 257         uint32_t sblend, dblend;
 258
 259         sblend = blend->src_card_op;
 260         dblend = blend->dst_card_op;
 261
 262         if (blend->dst_alpha) {
 263                 if (!PICT_FORMAT_A(dest_format)) {
 264                         if (sblend == BF(DST_ALPHA)) {
 265                                 sblend = BF(ONE);
 266                         } else if (sblend == BF(ONE_MINUS_DST_ALPHA)) {
 267                                 sblend = BF(ZERO);
 268                         }
 269                 } else if (dest_format == PICT_a8) {
 270                         if (sblend == BF(DST_ALPHA)) {
 271                                 sblend = BF(DST_COLOR);
 272                         } else if (sblend == BF(ONE_MINUS_DST_ALPHA)) {
 273                                 sblend = BF(ONE_MINUS_DST_COLOR);
 274                         }
 275                 }
 276         }
 277
 278         if (blend->src_alpha && (component_alpha || dest_format == PICT_a8)) {
 279                 if (dblend == BF(SRC_ALPHA)) {
 280                         dblend = BF(SRC_COLOR);
 281                 } else if (dblend == BF(ONE_MINUS_SRC_ALPHA)) {
 282                         dblend = BF(ONE_MINUS_SRC_COLOR);
 283                 }
 284         }
 285
 286         if (sblend == BF(ONE) && dblend == BF(ZERO)) {
 287                 BEGIN_RING(Nv3D, NV34TCL_BLEND_FUNC_ENABLE, 1);
 288                 OUT_RING  (0);
 289         } else {
 290                 BEGIN_RING(Nv3D, NV34TCL_BLEND_FUNC_ENABLE, 3);
 291                 OUT_RING  (1);
 292                 OUT_RING  ((sblend << 16) | sblend);
 293                 OUT_RING  ((dblend << 16) | dblend);
 294         }
 295 }
 296
 297 static Bool
 298 NV30EXATexture(ScrnInfoPtr pScrn, PixmapPtr pPix, PicturePtr pPict, int unit)
 299 {
 300         NVPtr pNv = NVPTR(pScrn);
 301         nv_pict_texture_format_t *fmt;
 302         uint32_t card_filter, card_repeat;
 303         NV30EXA_STATE;
 304
 305         fmt = NV30_GetPictTextureFormat(pPict->format);
 306         if (!fmt)
 307                 return FALSE;
 308
 309         card_repeat = 3; /* repeatNone */
 310
 311         if (pPict->filter == PictFilterBilinear)
 312                 card_filter = 2;
 313         else
 314                 card_filter = 1;
 315
 316         BEGIN_RING(Nv3D, NV34TCL_TX_OFFSET(unit), 8);
 317         OUT_PIXMAPl(pPix, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
 318
 319         OUT_RING  (NV34TCL_TX_FORMAT_DIMS_2D |
 320                         (fmt->card_fmt << NV34TCL_TX_FORMAT_FORMAT_SHIFT) |
 321                         (1 << NV34TCL_TX_FORMAT_MIPMAP_LEVELS_SHIFT) |
 322                         (log2i(pPix->drawable.width)  << NV34TCL_TX_FORMAT_BASE_SIZE_U_SHIFT) |
 323                         (log2i(pPix->drawable.height) << NV34TCL_TX_FORMAT_BASE_SIZE_V_SHIFT) |
 324                         8 |
 325                         NV34TCL_TX_FORMAT_DMA0);
 326
 327         OUT_RING  ((card_repeat << NV34TCL_TX_WRAP_S_SHIFT) |
 328                         (card_repeat << NV34TCL_TX_WRAP_T_SHIFT) |
 329                         (card_repeat << NV34TCL_TX_WRAP_R_SHIFT));
 330         OUT_RING  (NV34TCL_TX_ENABLE_ENABLE);
 331         OUT_RING  ((((uint32_t)exaGetPixmapPitch(pPix)) << NV34TCL_TX_SWIZZLE_RECT_PITCH_SHIFT ) |
 332                         fmt->card_swz);
 333
 334         OUT_RING  ((card_filter << NV34TCL_TX_FILTER_MINIFY_SHIFT) /* min */ |
 335                         (card_filter << NV34TCL_TX_FILTER_MAGNIFY_SHIFT) /* mag */ |
 336                         0x2000 /* engine lock */);
 337         OUT_RING  ((pPix->drawable.width << NV34TCL_TX_NPOT_SIZE_W_SHIFT) | pPix->drawable.height);
 338         OUT_RING  (0); /* border ARGB */
 339
 340         state->unit[unit].width         = (float)pPix->drawable.width;
 341         state->unit[unit].height        = (float)pPix->drawable.height;
 342         state->unit[unit].transform     = pPict->transform;
 343
 344         return TRUE;
 345 }
 346
 347 static Bool
 348 NV30_SetupSurface(ScrnInfoPtr pScrn, PixmapPtr pPix, PicturePtr pPict)
 349 {
 350         NVPtr pNv = NVPTR(pScrn);
 351         nv_pict_surface_format_t *fmt;
 352
 353         fmt = NV30_GetPictSurfaceFormat(pPict->format);
 354         if (!fmt) {
 355                 ErrorF("AIII no format\n");
 356                 return FALSE;
 357         }
 358
 359         uint32_t pitch = (uint32_t)exaGetPixmapPitch(pPix);
 360
 361         BEGIN_RING(Nv3D, NV34TCL_RT_FORMAT, 3);
 362         OUT_RING  (fmt->card_fmt); /* format */
 363         OUT_RING  (pitch << 16 | pitch);
 364         OUT_PIXMAPl(pPix, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
 365
 366         return TRUE;
 367 }
 368
 369 static Bool
 370 NV30EXACheckCompositeTexture(PicturePtr pPict)
 371 {
 372         nv_pict_texture_format_t *fmt;
 373         int w = pPict->pDrawable->width;
 374         int h = pPict->pDrawable->height;
 375
 376         if ((w > 4096) || (h>4096))
 377                 FALLBACK("picture too large, %dx%d\n", w, h);
 378
 379         fmt = NV30_GetPictTextureFormat(pPict->format);
 380         if (!fmt)
 381                 FALLBACK("picture format 0x%08x not supported\n",
 382                                 pPict->format);
 383
 384         if (pPict->filter != PictFilterNearest &&
 385                         pPict->filter != PictFilterBilinear)
 386                 FALLBACK("filter 0x%x not supported\n", pPict->filter);
 387
 388         if (!(w==1 && h==1) && pPict->repeat && pPict->repeatType != RepeatNone)
 389                 FALLBACK("repeat 0x%x not supported (surface %dx%d)\n", pPict->repeatType,w,h);
 390
 391         return TRUE;
 392 }
 393
 394 Bool
 395 NV30EXACheckComposite(int op, PicturePtr psPict,
 396                 PicturePtr pmPict,
 397                 PicturePtr pdPict)
 398 {
 399         nv_pict_surface_format_t *fmt;
 400         nv_pict_op_t *opr;
 401
 402         opr = NV30_GetPictOpRec(op);
 403         if (!opr)
 404                 FALLBACK("unsupported blend op 0x%x\n", op);
 405
 406         fmt = NV30_GetPictSurfaceFormat(pdPict->format);
 407         if (!fmt)
 408                 FALLBACK("dst picture format 0x%08x not supported\n",
 409                                 pdPict->format);
 410
 411         if (!NV30EXACheckCompositeTexture(psPict))
 412                 FALLBACK("src picture\n");
 413         if (pmPict) {
 414                 if (pmPict->componentAlpha &&
 415                                 PICT_FORMAT_RGB(pmPict->format) &&
 416                                 opr->src_alpha && opr->src_card_op != BF(ZERO))
 417                         FALLBACK("mask CA + SA\n");
 418                 if (!NV30EXACheckCompositeTexture(pmPict))
 419                         FALLBACK("mask picture\n");
 420         }
 421
 422         return TRUE;
 423 }
 424
 425 Bool
 426 NV30EXAPrepareComposite(int op, PicturePtr psPict,
 427                 PicturePtr pmPict,
 428                 PicturePtr pdPict,
 429                 PixmapPtr  psPix,
 430                 PixmapPtr  pmPix,
 431                 PixmapPtr  pdPix)
 432 {
 433         ScrnInfoPtr pScrn = xf86Screens[psPix->drawable.pScreen->myNum];
 434         NVPtr pNv = NVPTR(pScrn);
 435         nv_pict_op_t *blend;
 436         int fpid = NV30EXA_FPID_PASS_COL0;
 437         NV30EXA_STATE;
 438
 439         blend = NV30_GetPictOpRec(op);
 440
 441         NV30_SetupBlend(pScrn, blend, pdPict->format,
 442                         (pmPict && pmPict->componentAlpha &&
 443                          PICT_FORMAT_RGB(pmPict->format)));
 444
 445         NV30_SetupSurface(pScrn, pdPix, pdPict);
 446         NV30EXATexture(pScrn, psPix, psPict, 0);
 447
 448 #if 0
 449 #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)
 450         ErrorF("Preparecomposite src(%dx%d)",psPict->pDrawable->width,psPict->pDrawable->height);
 451         printformat((psPict->format));
 452         ErrorF(" dst(%dx%d)",pdPict->pDrawable->width,pdPict->pDrawable->height);
 453         printformat((pdPict->format));
 454         if (pmPict)
 455         {
 456                 ErrorF(" mask(%dx%d)",pmPict->pDrawable->width,pmPict->pDrawable->height);
 457                 printformat((pmPict->format));
 458         }
 459         ErrorF("\n");
 460 #endif
 461
 462         if (pmPict) {
 463                 NV30EXATexture(pScrn, pmPix, pmPict, 1);
 464
 465                 if (pmPict->componentAlpha && PICT_FORMAT_RGB(pmPict->format)) {
 466                         if (blend->src_alpha)
 467                                 fpid = NV30EXA_FPID_COMPOSITE_MASK_SA_CA;
 468                         else
 469                                 fpid = NV30EXA_FPID_COMPOSITE_MASK_CA;
 470                 } else {
 471                         fpid = NV30EXA_FPID_COMPOSITE_MASK;
 472                 }
 473
 474                 state->have_mask = TRUE;
 475         } else {
 476                 fpid = NV30EXA_FPID_PASS_TEX0;
 477
 478                 state->have_mask = FALSE;
 479         }
 480
 481         if (pdPict->format == PICT_a8)
 482                 NV30_LoadFragProg(pScrn, nv40_fp_map_a8[fpid]);
 483         else
 484                 NV30_LoadFragProg(pScrn, nv40_fp_map[fpid]);
 485
 486         BEGIN_RING(Nv3D, 0x23c, 1);
 487         OUT_RING  (pmPict?3:1);
 488
 489         BEGIN_RING(Nv3D, NV34TCL_VERTEX_BEGIN_END, 1);
 490         OUT_RING  (NV34TCL_VERTEX_BEGIN_END_TRIANGLES);
 491
 492         return TRUE;
 493 }
 494
 495 #define xFixedToFloat(v) \
 496         ((float)xFixedToInt((v)) + ((float)xFixedFrac(v) / 65536.0))
 497
 498 static void
 499 NV30EXATransformCoord(PictTransformPtr t, int x, int y, float sx, float sy,
 500                                           float *x_ret, float *y_ret)
 501 {
 502         PictVector v;
 503
 504         if (t) {
 505                 v.vector[0] = IntToxFixed(x);
 506                 v.vector[1] = IntToxFixed(y);
 507                 v.vector[2] = xFixed1;
 508                 PictureTransformPoint(t, &v);
 509                 *x_ret = xFixedToFloat(v.vector[0]);
 510                 *y_ret = xFixedToFloat(v.vector[1]);
 511         } else {
 512                 *x_ret = (float)x;
 513                 *y_ret = (float)y;
 514         }
 515 }
 516
 517 #define CV_OUTm(sx,sy,mx,my,dx,dy) do {                                        \
 518         BEGIN_RING(Nv3D, NV34TCL_VERTEX_ATTR_2F_X(8), 4);        \
 519         OUT_RINGf ((sx)); OUT_RINGf ((sy));                          \
 520         OUT_RINGf ((mx)); OUT_RINGf ((my));                          \
 521         BEGIN_RING(Nv3D, NV34TCL_VERTEX_ATTR_2I(0), 1);          \
 522         OUT_RING  (((dy)<<16)|(dx));                                           \
 523 } while(0)
 524 #define CV_OUT(sx,sy,dx,dy) do {                                               \
 525         BEGIN_RING(Nv3D, NV34TCL_VERTEX_ATTR_2F_X(8), 2);        \
 526         OUT_RINGf ((sx)); OUT_RINGf ((sy));                          \
 527         BEGIN_RING(Nv3D, NV34TCL_VERTEX_ATTR_2I(0), 1);          \
 528         OUT_RING  (((dy)<<16)|(dx));                                           \
 529 } while(0)
 530
 531 void
 532 NV30EXAComposite(PixmapPtr pdPix, int srcX , int srcY,
 533                                   int maskX, int maskY,
 534                                   int dstX , int dstY,
 535                                   int width, int height)
 536 {
 537         ScrnInfoPtr pScrn = xf86Screens[pdPix->drawable.pScreen->myNum];
 538         NVPtr pNv = NVPTR(pScrn);
 539         float sX0, sX1, sX2, sY0, sY1, sY2;
 540         float mX0, mX1, mX2, mY0, mY1, mY2;
 541         NV30EXA_STATE;
 542
 543         /* We're drawing a triangle, we need to scissor it to a quad. */
 544         /* The scissors are here for a good reason, we don't get the full image, but just a part. */
 545         /* Handling the cliprects is done for us already. */
 546         BEGIN_RING(Nv3D, NV34TCL_SCISSOR_HORIZ, 2);
 547         OUT_RING  ((width << 16) | dstX);
 548         OUT_RING  ((height << 16) | dstY);
 549
 550 #if 0
 551         ErrorF("Composite [%dx%d] (%d,%d)IN(%d,%d)OP(%d,%d)\n",width,height,srcX,srcY,maskX,maskY,dstX,dstY);
 552 #endif
 553         NV30EXATransformCoord(state->unit[0].transform,
 554                                 srcX, srcY - height,
 555                                 state->unit[0].width,
 556                                 state->unit[0].height, &sX0, &sY0);
 557         NV30EXATransformCoord(state->unit[0].transform,
 558                                 srcX, srcY + height,
 559                                 state->unit[0].width,
 560                                 state->unit[0].height, &sX1, &sY1);
 561         NV30EXATransformCoord(state->unit[0].transform,
 562                                 srcX + 2*width, srcY + height,
 563                                 state->unit[0].width,
 564                                 state->unit[0].height, &sX2, &sY2);
 565
 566         if (state->have_mask) {
 567                 NV30EXATransformCoord(state->unit[1].transform,
 568                                         maskX, maskY - height,
 569                                         state->unit[1].width,
 570                                         state->unit[1].height, &mX0, &mY0);
 571                 NV30EXATransformCoord(state->unit[1].transform,
 572                                         maskX, maskY + height,
 573                                         state->unit[1].width,
 574                                         state->unit[1].height, &mX1, &mY1);
 575                 NV30EXATransformCoord(state->unit[1].transform,
 576                                         maskX + 2*width, maskY + height,
 577                                         state->unit[1].width,
 578                                         state->unit[1].height, &mX2, &mY2);
 579
 580                 CV_OUTm(sX0 , sY0 , mX0, mY0, dstX                      ,       dstY - height);
 581                 CV_OUTm(sX1 , sY1 , mX1, mY1, dstX                      ,       dstY + height);
 582                 CV_OUTm(sX2 , sY2 , mX2, mY2, dstX + 2*width    ,       dstY + height);
 583         } else {
 584                 CV_OUT(sX0 , sY0 , dstX                 ,       dstY - height);
 585                 CV_OUT(sX1 , sY1 , dstX                 ,       dstY + height);
 586                 CV_OUT(sX2 , sY2 , dstX + 2*width       ,       dstY + height);
 587         }
 588 }
 589
 590 void
 591 NV30EXADoneComposite(PixmapPtr pdPix)
 592 {
 593         ScrnInfoPtr pScrn = xf86Screens[pdPix->drawable.pScreen->myNum];
 594         NVPtr pNv = NVPTR(pScrn);
 595
 596         BEGIN_RING(Nv3D, NV34TCL_VERTEX_BEGIN_END, 1);
 597         OUT_RING  (0);
 598
 599         FIRE_RING();
 600 }
 601
 602 Bool
 603 NVAccelInitNV30TCL(ScrnInfoPtr pScrn)
 604 {
 605         NVPtr pNv = NVPTR(pScrn);
 606         uint32_t class = 0, chipset;
 607         int i;
 608
 609         NV30EXAHackupA8Shaders(pScrn);
 610
 611 #define NV30TCL_CHIPSET_3X_MASK 0x00000003
 612 #define NV35TCL_CHIPSET_3X_MASK 0x000001e0
 613 #define NV34TCL_CHIPSET_3X_MASK 0x00000010
 614
 615         chipset = (nvReadMC(pNv, 0) >> 20) & 0xff;
 616         if ((chipset & 0xf0) != NV_ARCH_30)
 617                 return TRUE;
 618         chipset &= 0xf;
 619
 620         if (NV30TCL_CHIPSET_3X_MASK & (1<<chipset))
 621                 class = NV30TCL;
 622         else if (NV35TCL_CHIPSET_3X_MASK & (1<<chipset))
 623                 class = NV35TCL;
 624         else if (NV34TCL_CHIPSET_3X_MASK & (1<<chipset))
 625                 class = NV34TCL;
 626         else {
 627                 xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
 628                            "NV30EXA: Unknown chipset NV3%1x\n", chipset);
 629                 return FALSE;
 630         }
 631
 632
 633         if (!pNv->Nv3D) {
 634                 if (nouveau_grobj_alloc(pNv->chan, Nv3D, class, &pNv->Nv3D))
 635                         return FALSE;
 636         }
 637
 638         BEGIN_RING(Nv3D, NV34TCL_DMA_TEXTURE0, 3);
 639         OUT_RING  (pNv->chan->vram->handle);
 640         OUT_RING  (pNv->chan->gart->handle);
 641         OUT_RING  (pNv->chan->vram->handle);
 642         BEGIN_RING(Nv3D, NV34TCL_DMA_IN_MEMORY7, 1);
 643         OUT_RING  (pNv->chan->vram->handle);
 644         BEGIN_RING(Nv3D, NV34TCL_DMA_COLOR0, 2);
 645         OUT_RING  (pNv->chan->vram->handle);
 646         OUT_RING  (pNv->chan->vram->handle);
 647         BEGIN_RING(Nv3D, NV34TCL_DMA_IN_MEMORY8, 1);
 648         OUT_RING  (pNv->chan->vram->handle);
 649
 650         for (i=1; i<8; i++) {
 651                 BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_CLIP_HORIZ(i), 2);
 652                 OUT_RING  (0);
 653                 OUT_RING  (0);
 654         }
 655
 656         BEGIN_RING(Nv3D, 0x220, 1);
 657         OUT_RING  (1);
 658
 659         BEGIN_RING(Nv3D, 0x03b0, 1);
 660         OUT_RING  (0x00100000);
 661         BEGIN_RING(Nv3D, 0x1454, 1);
 662         OUT_RING  (0);
 663         BEGIN_RING(Nv3D, 0x1d80, 1);
 664         OUT_RING  (3);
 665         BEGIN_RING(Nv3D, 0x1450, 1);
 666         OUT_RING  (0x00030004);
 667
 668         /* NEW */
 669         BEGIN_RING(Nv3D, 0x1e98, 1);
 670         OUT_RING  (0);
 671         BEGIN_RING(Nv3D, 0x17e0, 3);
 672         OUT_RING  (0);
 673         OUT_RING  (0);
 674         OUT_RING  (0x3f800000);
 675         BEGIN_RING(Nv3D, 0x1f80, 16);
 676         OUT_RING  (0); OUT_RING  (0); OUT_RING  (0); OUT_RING  (0);
 677         OUT_RING  (0); OUT_RING  (0); OUT_RING  (0); OUT_RING  (0);
 678         OUT_RING  (0x0000ffff);
 679         OUT_RING  (0); OUT_RING  (0); OUT_RING  (0); OUT_RING  (0);
 680         OUT_RING  (0); OUT_RING  (0); OUT_RING  (0);
 681
 682         BEGIN_RING(Nv3D, 0x120, 3);
 683         OUT_RING  (0);
 684         OUT_RING  (1);
 685         OUT_RING  (2);
 686
 687         BEGIN_RING(NvImageBlit, 0x120, 3);
 688         OUT_RING  (0);
 689         OUT_RING  (1);
 690         OUT_RING  (2);
 691
 692         BEGIN_RING(Nv3D, 0x1d88, 1);
 693         OUT_RING  (0x00001200);
 694
 695         BEGIN_RING(Nv3D, NV34TCL_RC_ENABLE, 1);
 696         OUT_RING  (0);
 697
 698         /* Attempt to setup a known state.. Probably missing a heap of
 699          * stuff here..
 700          */
 701         BEGIN_RING(Nv3D, NV34TCL_STENCIL_FRONT_ENABLE, 1);
 702         OUT_RING  (0);
 703         BEGIN_RING(Nv3D, NV34TCL_STENCIL_BACK_ENABLE, 1);
 704         OUT_RING  (0);
 705         BEGIN_RING(Nv3D, NV34TCL_ALPHA_FUNC_ENABLE, 1);
 706         OUT_RING  (0);
 707         BEGIN_RING(Nv3D, NV34TCL_DEPTH_WRITE_ENABLE, 2);
 708         OUT_RING  (0); /* wr disable */
 709         OUT_RING  (0); /* test disable */
 710         BEGIN_RING(Nv3D, NV34TCL_COLOR_MASK, 1);
 711         OUT_RING  (0x01010101); /* TR,TR,TR,TR */
 712         BEGIN_RING(Nv3D, NV34TCL_CULL_FACE_ENABLE, 1);
 713         OUT_RING  (0);
 714         BEGIN_RING(Nv3D, NV34TCL_BLEND_FUNC_ENABLE, 5);
 715         OUT_RING  (0);                          /* Blend enable */
 716         OUT_RING  (0);                          /* Blend src */
 717         OUT_RING  (0);                          /* Blend dst */
 718         OUT_RING  (0x00000000);                 /* Blend colour */
 719         OUT_RING  (0x8006);                     /* FUNC_ADD */
 720         BEGIN_RING(Nv3D, NV34TCL_COLOR_LOGIC_OP_ENABLE, 2);
 721         OUT_RING  (0);
 722         OUT_RING  (0x1503 /*GL_COPY*/);
 723         BEGIN_RING(Nv3D, NV34TCL_DITHER_ENABLE, 1);
 724         OUT_RING  (1);
 725         BEGIN_RING(Nv3D, NV34TCL_SHADE_MODEL, 1);
 726         OUT_RING  (0x1d01 /*GL_SMOOTH*/);
 727         BEGIN_RING(Nv3D, NV34TCL_POLYGON_OFFSET_FACTOR,2);
 728         OUT_RINGf (0.0);
 729         OUT_RINGf (0.0);
 730         BEGIN_RING(Nv3D, NV34TCL_POLYGON_MODE_FRONT, 2);
 731         OUT_RING  (0x1b02 /*GL_FILL*/);
 732         OUT_RING  (0x1b02 /*GL_FILL*/);
 733         /* - Disable texture units
 734          * - Set fragprog to MOVR result.color, fragment.color */
 735         for (i=0;i<4;i++) {
 736                 BEGIN_RING(Nv3D, NV34TCL_TX_ENABLE(i), 1);
 737                 OUT_RING  (0);
 738         }
 739         /* Polygon stipple */
 740         BEGIN_RING(Nv3D, NV34TCL_POLYGON_STIPPLE_PATTERN(0), 0x20);
 741         for (i=0;i<0x20;i++)
 742                 OUT_RING  (0xFFFFFFFF);
 743
 744         BEGIN_RING(Nv3D, NV34TCL_DEPTH_RANGE_NEAR, 2);
 745         OUT_RINGf  (0.0);
 746         OUT_RINGf  (1.0);
 747
 748         /* Ok.  If you start X with the nvidia driver, kill it, and then
 749          * start X with nouveau you will get black rendering instead of
 750          * what you'd expect.  This fixes the problem, and it seems that
 751          * it's not needed between nouveau restarts - which suggests that
 752          * the 3D context (wherever it's stored?) survives somehow.
 753          */
 754         //BEGIN_RING(Nv3D, 0x1d60,1);
 755         //OUT_RING  (0x03008000);
 756
 757         int w=4096;
 758         int h=4096;
 759         int pitch=4096*4;
 760         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_HORIZ, 5);
 761         OUT_RING  (w<<16);
 762         OUT_RING  (h<<16);
 763         OUT_RING  (0x148); /* format */
 764         OUT_RING  (pitch << 16 | pitch);
 765         OUT_RING  (0x0);
 766         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_TX_ORIGIN, 1);
 767         OUT_RING  (0);
 768         BEGIN_RING(Nv3D, 0x0a00, 2);
 769         OUT_RING  ((w<<16) | 0);
 770         OUT_RING  ((h<<16) | 0);
 771         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_CLIP_HORIZ(0), 2);
 772         OUT_RING  ((w-1)<<16);
 773         OUT_RING  ((h-1)<<16);
 774         BEGIN_RING(Nv3D, NV34TCL_SCISSOR_HORIZ, 2);
 775         OUT_RING  (w<<16);
 776         OUT_RING  (h<<16);
 777         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_HORIZ, 2);
 778         OUT_RING  (w<<16);
 779         OUT_RING  (h<<16);
 780
 781         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_SCALE0_X, 8);
 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 (1.0);
 788         OUT_RINGf (1.0);
 789         OUT_RINGf (0.0);
 790
 791         BEGIN_RING(Nv3D, NV34TCL_MODELVIEW_MATRIX(0), 16);
 792         OUT_RINGf (1.0);
 793         OUT_RINGf (0.0);
 794         OUT_RINGf (0.0);
 795         OUT_RINGf (0.0);
 796         OUT_RINGf (0.0);
 797         OUT_RINGf (1.0);
 798         OUT_RINGf (0.0);
 799         OUT_RINGf (0.0);
 800         OUT_RINGf (0.0);
 801         OUT_RINGf (0.0);
 802         OUT_RINGf (1.0);
 803         OUT_RINGf (0.0);
 804         OUT_RINGf (0.0);
 805         OUT_RINGf (0.0);
 806         OUT_RINGf (0.0);
 807         OUT_RINGf (1.0);
 808
 809         BEGIN_RING(Nv3D, NV34TCL_PROJECTION_MATRIX(0), 16);
 810         OUT_RINGf (1.0);
 811         OUT_RINGf (0.0);
 812         OUT_RINGf (0.0);
 813         OUT_RINGf (0.0);
 814         OUT_RINGf (0.0);
 815         OUT_RINGf (1.0);
 816         OUT_RINGf (0.0);
 817         OUT_RINGf (0.0);
 818         OUT_RINGf (0.0);
 819         OUT_RINGf (0.0);
 820         OUT_RINGf (1.0);
 821         OUT_RINGf (0.0);
 822         OUT_RINGf (0.0);
 823         OUT_RINGf (0.0);
 824         OUT_RINGf (0.0);
 825         OUT_RINGf (1.0);
 826
 827         BEGIN_RING(Nv3D, NV34TCL_SCISSOR_HORIZ, 2);
 828         OUT_RING  (4096<<16);
 829         OUT_RING  (4096<<16);
 830
 831         return TRUE;
 832 }
 833