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 "nv_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_LoadFragProg(ScrnInfoPtr pScrn, nv_shader_t *shader)
 254 {
 255         NVPtr pNv = NVPTR(pScrn);
 256         static struct nouveau_bo *fp_mem = NULL;
 257         static int next_hw_id_offset = 0;
 258
 259         if (!fp_mem) {
 260                 if (nouveau_bo_new(pNv->dev, NOUVEAU_BO_VRAM | NOUVEAU_BO_PIN,
 261                                    0, 0x1000, &fp_mem)) {
 262                         xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
 263                                         "Couldn't alloc fragprog buffer!\n");
 264                         return;
 265                 }
 266
 267                 if (nouveau_bo_map(fp_mem, NOUVEAU_BO_RDWR)) {
 268                         xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
 269                                    "Couldn't map fragprog buffer!\n");
 270                 }
 271         }
 272
 273         if (!shader->hw_id) {
 274                 uint32_t *map = fp_mem->map + next_hw_id_offset;
 275                 int i;
 276
 277                 for (i = 0; i < shader->size; i++) {
 278                         uint32_t data = shader->data[i];
 279 #if (X_BYTE_ORDER != X_LITTLE_ENDIAN)
 280                         data = ((data >> 16) | ((data & 0xffff) << 16));
 281 #endif
 282                         map[i] = data;
 283                 }
 284
 285                 shader->hw_id += next_hw_id_offset;
 286                 next_hw_id_offset += (shader->size * sizeof(uint32_t));
 287                 next_hw_id_offset = (next_hw_id_offset + 63) & ~63;
 288         }
 289
 290         BEGIN_RING(Nv3D, NV34TCL_FP_ACTIVE_PROGRAM, 1);
 291         OUT_RELOC (fp_mem, shader->hw_id, NOUVEAU_BO_VRAM | NOUVEAU_BO_GART |
 292                    NOUVEAU_BO_RD | NOUVEAU_BO_LOW | NOUVEAU_BO_OR,
 293                    1 /*NV30TCL_FP_ADDRESS_DMA0*/, 2 /*NV30TCL_FP_ADDRESS_DMA1*/);
 294
 295         BEGIN_RING(Nv3D, 0x1d60, 1);
 296         OUT_RING  (0); /* USES_KIL (1<<7) == 0 */
 297         BEGIN_RING(Nv3D, 0x1450, 1);
 298         OUT_RING  (shader->card_priv.NV30FP.num_regs << 16| 4);
 299         BEGIN_RING(Nv3D, 0x1d7c, 1);
 300         OUT_RING  (0xffff0000);
 301
 302 }
 303
 304 static void
 305 NV30_SetupBlend(ScrnInfoPtr pScrn, nv_pict_op_t *blend,
 306                 PictFormatShort dest_format, Bool component_alpha)
 307 {
 308         NVPtr pNv = NVPTR(pScrn);
 309         uint32_t sblend, dblend;
 310
 311         sblend = blend->src_card_op;
 312         dblend = blend->dst_card_op;
 313
 314         if (blend->dst_alpha) {
 315                 if (!PICT_FORMAT_A(dest_format)) {
 316                         if (sblend == BF(DST_ALPHA)) {
 317                                 sblend = BF(ONE);
 318                         } else if (sblend == BF(ONE_MINUS_DST_ALPHA)) {
 319                                 sblend = BF(ZERO);
 320                         }
 321                 } else if (dest_format == PICT_a8) {
 322                         if (sblend == BF(DST_ALPHA)) {
 323                                 sblend = BF(DST_COLOR);
 324                         } else if (sblend == BF(ONE_MINUS_DST_ALPHA)) {
 325                                 sblend = BF(ONE_MINUS_DST_COLOR);
 326                         }
 327                 }
 328         }
 329
 330         if (blend->src_alpha && (component_alpha || dest_format == PICT_a8)) {
 331                 if (dblend == BF(SRC_ALPHA)) {
 332                         dblend = BF(SRC_COLOR);
 333                 } else if (dblend == BF(ONE_MINUS_SRC_ALPHA)) {
 334                         dblend = BF(ONE_MINUS_SRC_COLOR);
 335                 }
 336         }
 337
 338         if (sblend == BF(ONE) && dblend == BF(ZERO)) {
 339                 BEGIN_RING(Nv3D, NV34TCL_BLEND_FUNC_ENABLE, 1);
 340                 OUT_RING  (0);
 341         } else {
 342                 BEGIN_RING(Nv3D, NV34TCL_BLEND_FUNC_ENABLE, 3);
 343                 OUT_RING  (1);
 344                 OUT_RING  ((sblend << 16) | sblend);
 345                 OUT_RING  ((dblend << 16) | dblend);
 346         }
 347 }
 348
 349 static Bool
 350 NV30EXATexture(ScrnInfoPtr pScrn, PixmapPtr pPix, PicturePtr pPict, int unit)
 351 {
 352         NVPtr pNv = NVPTR(pScrn);
 353         nv_pict_texture_format_t *fmt;
 354         uint32_t card_filter, card_repeat;
 355         NV30EXA_STATE;
 356
 357         fmt = NV30_GetPictTextureFormat(pPict->format);
 358         if (!fmt)
 359                 return FALSE;
 360
 361         card_repeat = 3; /* repeatNone */
 362
 363         if (pPict->filter == PictFilterBilinear)
 364                 card_filter = 2;
 365         else
 366                 card_filter = 1;
 367
 368         BEGIN_RING(Nv3D, NV34TCL_TX_OFFSET(unit), 8);
 369         OUT_PIXMAPl(pPix, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
 370
 371         OUT_RING  ((2 << 4) /* 2D */ |
 372                         (fmt->card_fmt << 8) |
 373                         (1 << 16) /* 1 mipmap level */ |
 374                         (log2i(pPix->drawable.width)  << 20) |
 375                         (log2i(pPix->drawable.height) << 24) |
 376                         9);
 377
 378         OUT_RING  ((card_repeat <<  0) /* S */ |
 379                         (card_repeat <<  8) /* T */ |
 380                         (card_repeat << 16) /* R */);
 381         OUT_RING  (0x40000000); /* enable */
 382         OUT_RING  ((((uint32_t)exaGetPixmapPitch(pPix))<<16) | fmt->card_swz);
 383
 384         OUT_RING  ((card_filter << 16) /* min */ |
 385                         (card_filter << 24) /* mag */ |
 386                         0x2000 /* engine lock */);
 387         OUT_RING  ((pPix->drawable.width << 16) | pPix->drawable.height);
 388         OUT_RING  (0); /* border ARGB */
 389
 390         state->unit[unit].width         = (float)pPix->drawable.width;
 391         state->unit[unit].height        = (float)pPix->drawable.height;
 392         state->unit[unit].transform     = pPict->transform;
 393
 394         return TRUE;
 395 }
 396
 397 static Bool
 398 NV30_SetupSurface(ScrnInfoPtr pScrn, PixmapPtr pPix, PicturePtr pPict)
 399 {
 400         NVPtr pNv = NVPTR(pScrn);
 401         nv_pict_surface_format_t *fmt;
 402
 403         fmt = NV30_GetPictSurfaceFormat(pPict->format);
 404         if (!fmt) {
 405                 ErrorF("AIII no format\n");
 406                 return FALSE;
 407         }
 408
 409         uint32_t pitch = (uint32_t)exaGetPixmapPitch(pPix);
 410
 411         int x = pPict->pDrawable->x;
 412         int y = pPict->pDrawable->y;
 413         int w = pPict->pDrawable->width;
 414         int h = pPict->pDrawable->height;
 415         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_HORIZ, 5);
 416         OUT_RING  ((w<<16)|x);
 417         OUT_RING  ((h<<16)|y);
 418         OUT_RING  (fmt->card_fmt); /* format */
 419         OUT_RING  (pitch << 16 | pitch);
 420         OUT_PIXMAPl(pPix, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
 421         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_CLIP_HORIZ(0), 2);
 422         OUT_RING  ((w-1+x)<<16);
 423         OUT_RING  ((h-1+y)<<16);
 424         if (pNv->NVArch == 0x30 || pNv->NVArch == 0x31) {
 425                 BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_TX_ORIGIN, 1);
 426                 OUT_RING((y<<16)|x);
 427         }
 428
 429         return TRUE;
 430 }
 431
 432 static Bool
 433 NV30EXACheckCompositeTexture(PicturePtr pPict)
 434 {
 435         nv_pict_texture_format_t *fmt;
 436         int w = pPict->pDrawable->width;
 437         int h = pPict->pDrawable->height;
 438
 439         if ((w > 4096) || (h>4096))
 440                 FALLBACK("picture too large, %dx%d\n", w, h);
 441
 442         fmt = NV30_GetPictTextureFormat(pPict->format);
 443         if (!fmt)
 444                 FALLBACK("picture format 0x%08x not supported\n",
 445                                 pPict->format);
 446
 447         if (pPict->filter != PictFilterNearest &&
 448                         pPict->filter != PictFilterBilinear)
 449                 FALLBACK("filter 0x%x not supported\n", pPict->filter);
 450
 451         if (!(w==1 && h==1) && pPict->repeat && pPict->repeatType != RepeatNone)
 452                 FALLBACK("repeat 0x%x not supported (surface %dx%d)\n", pPict->repeatType,w,h);
 453
 454         return TRUE;
 455 }
 456
 457 Bool
 458 NV30EXACheckComposite(int op, PicturePtr psPict,
 459                 PicturePtr pmPict,
 460                 PicturePtr pdPict)
 461 {
 462         nv_pict_surface_format_t *fmt;
 463         nv_pict_op_t *opr;
 464
 465         opr = NV30_GetPictOpRec(op);
 466         if (!opr)
 467                 FALLBACK("unsupported blend op 0x%x\n", op);
 468
 469         fmt = NV30_GetPictSurfaceFormat(pdPict->format);
 470         if (!fmt)
 471                 FALLBACK("dst picture format 0x%08x not supported\n",
 472                                 pdPict->format);
 473
 474         if (!NV30EXACheckCompositeTexture(psPict))
 475                 FALLBACK("src picture\n");
 476         if (pmPict) {
 477                 if (pmPict->componentAlpha &&
 478                                 PICT_FORMAT_RGB(pmPict->format) &&
 479                                 opr->src_alpha && opr->src_card_op != BF(ZERO))
 480                         FALLBACK("mask CA + SA\n");
 481                 if (!NV30EXACheckCompositeTexture(pmPict))
 482                         FALLBACK("mask picture\n");
 483         }
 484
 485         return TRUE;
 486 }
 487
 488 Bool
 489 NV30EXAPrepareComposite(int op, PicturePtr psPict,
 490                 PicturePtr pmPict,
 491                 PicturePtr pdPict,
 492                 PixmapPtr  psPix,
 493                 PixmapPtr  pmPix,
 494                 PixmapPtr  pdPix)
 495 {
 496         ScrnInfoPtr pScrn = xf86Screens[psPix->drawable.pScreen->myNum];
 497         NVPtr pNv = NVPTR(pScrn);
 498         nv_pict_op_t *blend;
 499         int fpid = NV30EXA_FPID_PASS_COL0;
 500         NV30EXA_STATE;
 501
 502         blend = NV30_GetPictOpRec(op);
 503
 504         NV30_SetupBlend(pScrn, blend, pdPict->format,
 505                         (pmPict && pmPict->componentAlpha &&
 506                          PICT_FORMAT_RGB(pmPict->format)));
 507
 508         NV30_SetupSurface(pScrn, pdPix, pdPict);
 509         NV30EXATexture(pScrn, psPix, psPict, 0);
 510
 511 #if 0
 512 #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)
 513         ErrorF("Preparecomposite src(%dx%d)",psPict->pDrawable->width,psPict->pDrawable->height);
 514         printformat((psPict->format));
 515         ErrorF(" dst(%dx%d)",pdPict->pDrawable->width,pdPict->pDrawable->height);
 516         printformat((pdPict->format));
 517         if (pmPict)
 518         {
 519                 ErrorF(" mask(%dx%d)",pmPict->pDrawable->width,pmPict->pDrawable->height);
 520                 printformat((pmPict->format));
 521         }
 522         ErrorF("\n");
 523 #endif
 524
 525         if (pmPict) {
 526                 NV30EXATexture(pScrn, pmPix, pmPict, 1);
 527
 528                 if (pmPict->componentAlpha && PICT_FORMAT_RGB(pmPict->format)) {
 529                         if (blend->src_alpha)
 530                                 fpid = NV30EXA_FPID_COMPOSITE_MASK_SA_CA;
 531                         else
 532                                 fpid = NV30EXA_FPID_COMPOSITE_MASK_CA;
 533                 } else {
 534                         fpid = NV30EXA_FPID_COMPOSITE_MASK;
 535                 }
 536
 537                 state->have_mask = TRUE;
 538         } else {
 539                 fpid = NV30EXA_FPID_PASS_TEX0;
 540
 541                 state->have_mask = FALSE;
 542         }
 543
 544         if (pdPict->format == PICT_a8)
 545                 NV30_LoadFragProg(pScrn, nv40_fp_map_a8[fpid]);
 546         else
 547                 NV30_LoadFragProg(pScrn, nv40_fp_map[fpid]);
 548
 549         BEGIN_RING(Nv3D, 0x23c, 1);
 550         OUT_RING  (pmPict?3:1);
 551
 552         BEGIN_RING(Nv3D, NV34TCL_VERTEX_BEGIN_END, 1);
 553         OUT_RING  (8); /* GL_QUADS */
 554
 555         return TRUE;
 556 }
 557
 558 #define xFixedToFloat(v) \
 559         ((float)xFixedToInt((v)) + ((float)xFixedFrac(v) / 65536.0))
 560
 561 static void
 562 NV30EXATransformCoord(PictTransformPtr t, int x, int y, float sx, float sy,
 563                                           float *x_ret, float *y_ret)
 564 {
 565         PictVector v;
 566
 567         if (t) {
 568                 v.vector[0] = IntToxFixed(x);
 569                 v.vector[1] = IntToxFixed(y);
 570                 v.vector[2] = xFixed1;
 571                 PictureTransformPoint(t, &v);
 572                 *x_ret = xFixedToFloat(v.vector[0]);
 573                 *y_ret = xFixedToFloat(v.vector[1]);
 574         } else {
 575                 *x_ret = (float)x;
 576                 *y_ret = (float)y;
 577         }
 578 }
 579
 580 #define CV_OUTm(sx,sy,mx,my,dx,dy) do {                                        \
 581         BEGIN_RING(Nv3D, NV34TCL_VERTEX_ATTR_2F_X(8), 4);        \
 582         OUT_RINGf ((sx)); OUT_RINGf ((sy));                          \
 583         OUT_RINGf ((mx)); OUT_RINGf ((my));                          \
 584         BEGIN_RING(Nv3D, NV34TCL_VERTEX_ATTR_2I(0), 1);          \
 585         OUT_RING  (((dy)<<16)|(dx));                                           \
 586 } while(0)
 587 #define CV_OUT(sx,sy,dx,dy) do {                                               \
 588         BEGIN_RING(Nv3D, NV34TCL_VERTEX_ATTR_2F_X(8), 2);        \
 589         OUT_RINGf ((sx)); OUT_RINGf ((sy));                          \
 590         BEGIN_RING(Nv3D, NV34TCL_VERTEX_ATTR_2I(0), 1);          \
 591         OUT_RING  (((dy)<<16)|(dx));                                           \
 592 } while(0)
 593
 594 void
 595 NV30EXAComposite(PixmapPtr pdPix, int srcX , int srcY,
 596                                   int maskX, int maskY,
 597                                   int dstX , int dstY,
 598                                   int width, int height)
 599 {
 600         ScrnInfoPtr pScrn = xf86Screens[pdPix->drawable.pScreen->myNum];
 601         NVPtr pNv = NVPTR(pScrn);
 602         float sX0, sX1, sX2, sY0, sY1, sY2, sX3, sY3;
 603         float mX0, mX1, mX2, mY0, mY1, mY2, mX3, mY3;
 604         NV30EXA_STATE;
 605
 606 #if 0
 607         ErrorF("Composite [%dx%d] (%d,%d)IN(%d,%d)OP(%d,%d)\n",width,height,srcX,srcY,maskX,maskY,dstX,dstY);
 608 #endif
 609         NV30EXATransformCoord(state->unit[0].transform, srcX, srcY,
 610                               state->unit[0].width,
 611                               state->unit[0].height, &sX0, &sY0);
 612         NV30EXATransformCoord(state->unit[0].transform,
 613                               srcX + width, srcY,
 614                               state->unit[0].width,
 615                               state->unit[0].height, &sX1, &sY1);
 616         NV30EXATransformCoord(state->unit[0].transform,
 617                               srcX + width, srcY + height,
 618                               state->unit[0].width,
 619                               state->unit[0].height, &sX2, &sY2);
 620         NV30EXATransformCoord(state->unit[0].transform,
 621                               srcX, srcY + height,
 622                               state->unit[0].width,
 623                               state->unit[0].height, &sX3, &sY3);
 624
 625         if (state->have_mask) {
 626                 NV30EXATransformCoord(state->unit[1].transform, maskX, maskY,
 627                                       state->unit[1].width,
 628                                       state->unit[1].height, &mX0, &mY0);
 629                 NV30EXATransformCoord(state->unit[1].transform,
 630                                       maskX + width, maskY,
 631                                       state->unit[1].width,
 632                                       state->unit[1].height, &mX1, &mY1);
 633                 NV30EXATransformCoord(state->unit[1].transform,
 634                                       maskX + width, maskY + height,
 635                                       state->unit[1].width,
 636                                       state->unit[1].height, &mX2, &mY2);
 637                 NV30EXATransformCoord(state->unit[1].transform,
 638                                       maskX, maskY + height,
 639                                       state->unit[1].width,
 640                                       state->unit[1].height, &mX3, &mY3);
 641
 642                 CV_OUTm(sX0 , sY0 , mX0, mY0, dstX        ,          dstY);
 643                 CV_OUTm(sX1 , sY1 , mX1, mY1, dstX + width,          dstY);
 644                 CV_OUTm(sX2 , sY2 , mX2, mY2, dstX + width, dstY + height);
 645                 CV_OUTm(sX3 , sY3 , mX3, mY3, dstX        , dstY + height);
 646         } else {
 647                 CV_OUT(sX0 , sY0 , dstX        ,          dstY);
 648                 CV_OUT(sX1 , sY1 , dstX + width,          dstY);
 649                 CV_OUT(sX2 , sY2 , dstX + width, dstY + height);
 650                 CV_OUT(sX3 , sY3 , dstX        , dstY + height);
 651         }
 652 }
 653
 654 void
 655 NV30EXADoneComposite(PixmapPtr pdPix)
 656 {
 657         ScrnInfoPtr pScrn = xf86Screens[pdPix->drawable.pScreen->myNum];
 658         NVPtr pNv = NVPTR(pScrn);
 659
 660         BEGIN_RING(Nv3D, NV34TCL_VERTEX_BEGIN_END, 1);
 661         OUT_RING  (0);
 662
 663         FIRE_RING();
 664 }
 665
 666 Bool
 667 NVAccelInitNV30TCL(ScrnInfoPtr pScrn)
 668 {
 669         NVPtr pNv = NVPTR(pScrn);
 670         uint32_t class = 0, chipset;
 671         int i;
 672
 673         NV30EXAHackupA8Shaders(pScrn);
 674
 675 #define NV30TCL_CHIPSET_3X_MASK 0x00000003
 676 #define NV35TCL_CHIPSET_3X_MASK 0x000001e0
 677 #define NV34TCL_CHIPSET_3X_MASK 0x00000010
 678
 679         chipset = (nvReadMC(pNv, 0) >> 20) & 0xff;
 680         if ((chipset & 0xf0) != NV_ARCH_30)
 681                 return TRUE;
 682         chipset &= 0xf;
 683
 684         if (NV30TCL_CHIPSET_3X_MASK & (1<<chipset))
 685                 class = NV30TCL;
 686         else if (NV35TCL_CHIPSET_3X_MASK & (1<<chipset))
 687                 class = NV35TCL;
 688         else if (NV34TCL_CHIPSET_3X_MASK & (1<<chipset))
 689                 class = NV34TCL;
 690         else {
 691                 xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
 692                            "NV30EXA: Unknown chipset NV3%1x\n", chipset);
 693                 return FALSE;
 694         }
 695
 696
 697         if (!pNv->Nv3D) {
 698                 if (nouveau_grobj_alloc(pNv->chan, Nv3D, class, &pNv->Nv3D))
 699                         return FALSE;
 700         }
 701
 702         BEGIN_RING(Nv3D, NV34TCL_DMA_TEXTURE0, 3);
 703         OUT_RING  (pNv->chan->vram->handle);
 704         OUT_RING  (pNv->chan->gart->handle);
 705         OUT_RING  (pNv->chan->vram->handle);
 706         BEGIN_RING(Nv3D, NV34TCL_DMA_IN_MEMORY7, 1);
 707         OUT_RING  (pNv->chan->vram->handle);
 708         BEGIN_RING(Nv3D, NV34TCL_DMA_COLOR0, 2);
 709         OUT_RING  (pNv->chan->vram->handle);
 710         OUT_RING  (pNv->chan->vram->handle);
 711         BEGIN_RING(Nv3D, NV34TCL_DMA_IN_MEMORY8, 1);
 712         OUT_RING  (pNv->chan->vram->handle);
 713
 714         for (i=1; i<8; i++) {
 715                 BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_CLIP_HORIZ(i), 2);
 716                 OUT_RING  (0);
 717                 OUT_RING  (0);
 718         }
 719
 720         BEGIN_RING(Nv3D, 0x220, 1);
 721         OUT_RING  (1);
 722
 723         BEGIN_RING(Nv3D, 0x03b0, 1);
 724         OUT_RING  (0x00100000);
 725         BEGIN_RING(Nv3D, 0x1454, 1);
 726         OUT_RING  (0);
 727         BEGIN_RING(Nv3D, 0x1d80, 1);
 728         OUT_RING  (3);
 729         BEGIN_RING(Nv3D, 0x1450, 1);
 730         OUT_RING  (0x00030004);
 731
 732         /* NEW */
 733         BEGIN_RING(Nv3D, 0x1e98, 1);
 734         OUT_RING  (0);
 735         BEGIN_RING(Nv3D, 0x17e0, 3);
 736         OUT_RING  (0);
 737         OUT_RING  (0);
 738         OUT_RING  (0x3f800000);
 739         BEGIN_RING(Nv3D, 0x1f80, 16);
 740         OUT_RING  (0); OUT_RING  (0); OUT_RING  (0); OUT_RING  (0);
 741         OUT_RING  (0); OUT_RING  (0); OUT_RING  (0); OUT_RING  (0);
 742         OUT_RING  (0x0000ffff);
 743         OUT_RING  (0); OUT_RING  (0); OUT_RING  (0); OUT_RING  (0);
 744         OUT_RING  (0); OUT_RING  (0); OUT_RING  (0);
 745
 746         BEGIN_RING(Nv3D, 0x120, 3);
 747         OUT_RING  (0);
 748         OUT_RING  (1);
 749         OUT_RING  (2);
 750
 751         BEGIN_RING(NvImageBlit, 0x120, 3);
 752         OUT_RING  (0);
 753         OUT_RING  (1);
 754         OUT_RING  (2);
 755
 756         BEGIN_RING(Nv3D, 0x1d88, 1);
 757         OUT_RING  (0x00001200);
 758
 759         BEGIN_RING(Nv3D, NV34TCL_RC_ENABLE, 1);
 760         OUT_RING  (0);
 761
 762         /* Attempt to setup a known state.. Probably missing a heap of
 763          * stuff here..
 764          */
 765         BEGIN_RING(Nv3D, NV34TCL_STENCIL_FRONT_ENABLE, 1);
 766         OUT_RING  (0);
 767         BEGIN_RING(Nv3D, NV34TCL_STENCIL_BACK_ENABLE, 1);
 768         OUT_RING  (0);
 769         BEGIN_RING(Nv3D, NV34TCL_ALPHA_FUNC_ENABLE, 1);
 770         OUT_RING  (0);
 771         BEGIN_RING(Nv3D, NV34TCL_DEPTH_WRITE_ENABLE, 2);
 772         OUT_RING  (0); /* wr disable */
 773         OUT_RING  (0); /* test disable */
 774         BEGIN_RING(Nv3D, NV34TCL_COLOR_MASK, 1);
 775         OUT_RING  (0x01010101); /* TR,TR,TR,TR */
 776         BEGIN_RING(Nv3D, NV34TCL_CULL_FACE_ENABLE, 1);
 777         OUT_RING  (0);
 778         BEGIN_RING(Nv3D, NV34TCL_BLEND_FUNC_ENABLE, 5);
 779         OUT_RING  (0);                          /* Blend enable */
 780         OUT_RING  (0);                          /* Blend src */
 781         OUT_RING  (0);                          /* Blend dst */
 782         OUT_RING  (0x00000000);                 /* Blend colour */
 783         OUT_RING  (0x8006);                     /* FUNC_ADD */
 784         BEGIN_RING(Nv3D, NV34TCL_COLOR_LOGIC_OP_ENABLE, 2);
 785         OUT_RING  (0);
 786         OUT_RING  (0x1503 /*GL_COPY*/);
 787         BEGIN_RING(Nv3D, NV34TCL_DITHER_ENABLE, 1);
 788         OUT_RING  (1);
 789         BEGIN_RING(Nv3D, NV34TCL_SHADE_MODEL, 1);
 790         OUT_RING  (0x1d01 /*GL_SMOOTH*/);
 791         BEGIN_RING(Nv3D, NV34TCL_POLYGON_OFFSET_FACTOR,2);
 792         OUT_RINGf (0.0);
 793         OUT_RINGf (0.0);
 794         BEGIN_RING(Nv3D, NV34TCL_POLYGON_MODE_FRONT, 2);
 795         OUT_RING  (0x1b02 /*GL_FILL*/);
 796         OUT_RING  (0x1b02 /*GL_FILL*/);
 797         /* - Disable texture units
 798          * - Set fragprog to MOVR result.color, fragment.color */
 799         for (i=0;i<4;i++) {
 800                 BEGIN_RING(Nv3D, NV34TCL_TX_ENABLE(i), 1);
 801                 OUT_RING  (0);
 802         }
 803         /* Polygon stipple */
 804         BEGIN_RING(Nv3D, NV34TCL_POLYGON_STIPPLE_PATTERN(0), 0x20);
 805         for (i=0;i<0x20;i++)
 806                 OUT_RING  (0xFFFFFFFF);
 807
 808         BEGIN_RING(Nv3D, NV34TCL_DEPTH_RANGE_NEAR, 2);
 809         OUT_RINGf  (0.0);
 810         OUT_RINGf  (1.0);
 811
 812         /* Ok.  If you start X with the nvidia driver, kill it, and then
 813          * start X with nouveau you will get black rendering instead of
 814          * what you'd expect.  This fixes the problem, and it seems that
 815          * it's not needed between nouveau restarts - which suggests that
 816          * the 3D context (wherever it's stored?) survives somehow.
 817          */
 818         //BEGIN_RING(Nv3D, 0x1d60,1);
 819         //OUT_RING  (0x03008000);
 820
 821         int w=4096;
 822         int h=4096;
 823         int pitch=4096*4;
 824         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_HORIZ, 5);
 825         OUT_RING  (w<<16);
 826         OUT_RING  (h<<16);
 827         OUT_RING  (0x148); /* format */
 828         OUT_RING  (pitch << 16 | pitch);
 829         OUT_RING  (0x0);
 830         BEGIN_RING(Nv3D, 0x0a00, 2);
 831         OUT_RING  ((w<<16) | 0);
 832         OUT_RING  ((h<<16) | 0);
 833         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_CLIP_HORIZ(0), 2);
 834         OUT_RING  ((w-1)<<16);
 835         OUT_RING  ((h-1)<<16);
 836         BEGIN_RING(Nv3D, NV34TCL_SCISSOR_HORIZ, 2);
 837         OUT_RING  (w<<16);
 838         OUT_RING  (h<<16);
 839         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_HORIZ, 2);
 840         OUT_RING  (w<<16);
 841         OUT_RING  (h<<16);
 842
 843         BEGIN_RING(Nv3D, NV34TCL_VIEWPORT_SCALE0_X, 8);
 844         OUT_RINGf (0.0);
 845         OUT_RINGf (0.0);
 846         OUT_RINGf (0.0);
 847         OUT_RINGf (0.0);
 848         OUT_RINGf (1.0);
 849         OUT_RINGf (1.0);
 850         OUT_RINGf (1.0);
 851         OUT_RINGf (0.0);
 852
 853         BEGIN_RING(Nv3D, NV34TCL_MODELVIEW_MATRIX(0), 16);
 854         OUT_RINGf (1.0);
 855         OUT_RINGf (0.0);
 856         OUT_RINGf (0.0);
 857         OUT_RINGf (0.0);
 858         OUT_RINGf (0.0);
 859         OUT_RINGf (1.0);
 860         OUT_RINGf (0.0);
 861         OUT_RINGf (0.0);
 862         OUT_RINGf (0.0);
 863         OUT_RINGf (0.0);
 864         OUT_RINGf (1.0);
 865         OUT_RINGf (0.0);
 866         OUT_RINGf (0.0);
 867         OUT_RINGf (0.0);
 868         OUT_RINGf (0.0);
 869         OUT_RINGf (1.0);
 870
 871         BEGIN_RING(Nv3D, NV34TCL_PROJECTION_MATRIX(0), 16);
 872         OUT_RINGf (1.0);
 873         OUT_RINGf (0.0);
 874         OUT_RINGf (0.0);
 875         OUT_RINGf (0.0);
 876         OUT_RINGf (0.0);
 877         OUT_RINGf (1.0);
 878         OUT_RINGf (0.0);
 879         OUT_RINGf (0.0);
 880         OUT_RINGf (0.0);
 881         OUT_RINGf (0.0);
 882         OUT_RINGf (1.0);
 883         OUT_RINGf (0.0);
 884         OUT_RINGf (0.0);
 885         OUT_RINGf (0.0);
 886         OUT_RINGf (0.0);
 887         OUT_RINGf (1.0);
 888
 889         BEGIN_RING(Nv3D, NV34TCL_SCISSOR_HORIZ, 2);
 890         OUT_RING  (4096<<16);
 891         OUT_RING  (4096<<16);
 892
 893         return TRUE;
 894 }
 895