src/gallium/drivers/nv50/nv50_shader_state.c

   1 /*
   2  * Copyright 2008 Ben Skeggs
   3  * Copyright 2010 Christoph Bumiller
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be included in
  13  * all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  19  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  20  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  */
  23
  24 #include "pipe/p_context.h"
  25 #include "pipe/p_defines.h"
  26 #include "pipe/p_state.h"
  27 #include "util/u_inlines.h"
  28
  29 #include "nv50_context.h"
  30
  31 void
  32 nv50_constbufs_validate(struct nv50_context *nv50)
  33 {
  34    struct nouveau_pushbuf *push = nv50->base.pushbuf;
  35    unsigned s;
  36
  37    for (s = 0; s < 3; ++s) {
  38       struct nv04_resource *res;
  39       int i;
  40       unsigned p, b;
  41
  42       if (s == PIPE_SHADER_FRAGMENT)
  43          p = NV50_3D_SET_PROGRAM_CB_PROGRAM_FRAGMENT;
  44       else
  45       if (s == PIPE_SHADER_GEOMETRY)
  46          p = NV50_3D_SET_PROGRAM_CB_PROGRAM_GEOMETRY;
  47       else
  48          p = NV50_3D_SET_PROGRAM_CB_PROGRAM_VERTEX;
  49
  50       while (nv50->constbuf_dirty[s]) {
  51          struct nouveau_bo *bo;
  52          unsigned start = 0;
  53          unsigned words = 0;
  54
  55          i = ffs(nv50->constbuf_dirty[s]) - 1;
  56          nv50->constbuf_dirty[s] &= ~(1 << i);
  57
  58          res = nv04_resource(nv50->constbuf[s][i]);
  59          if (!res) {
  60             if (i != 0) {
  61                BEGIN_NV04(push, NV50_3D(SET_PROGRAM_CB), 1);
  62                PUSH_DATA (push, (i << 8) | p | 0);
  63             }
  64             continue;
  65          }
  66
  67          if (i == 0) {
  68             b = NV50_CB_PVP + s;
  69
  70             /* always upload GL uniforms through CB DATA */
  71             bo = nv50->screen->uniforms;
  72             words = res->base.width0 / 4;
  73          } else {
  74             b = s * 16 + i;
  75
  76             assert(0);
  77
  78             if (!nouveau_resource_mapped_by_gpu(&res->base)) {
  79                nouveau_buffer_migrate(&nv50->base, res, NOUVEAU_BO_VRAM);
  80
  81                BEGIN_NV04(push, NV50_3D(CODE_CB_FLUSH), 1);
  82                PUSH_DATA (push, 0);
  83             }
  84             BEGIN_NV04(push, NV50_3D(CB_DEF_ADDRESS_HIGH), 3);
  85             PUSH_DATAh(push, res->address);
  86             PUSH_DATA (push, res->address);
  87             PUSH_DATA (push, (b << 16) | (res->base.width0 & 0xffff));
  88             BEGIN_NV04(push, NV50_3D(SET_PROGRAM_CB), 1);
  89             PUSH_DATA (push, (b << 12) | (i << 8) | p | 1);
  90
  91             bo = res->bo;
  92          }
  93
  94          if (bo != nv50->screen->uniforms)
  95             BCTX_REFN(nv50->bufctx_3d, CB(s, i), res, RD);
  96
  97          while (words) {
  98             unsigned nr;
  99
 100             if (!PUSH_SPACE(push, 16))
 101                break;
 102             nr = PUSH_AVAIL(push);
 103             assert(nr >= 16);
 104             nr = MIN2(MIN2(nr - 3, words), NV04_PFIFO_MAX_PACKET_LEN);
 105
 106             BEGIN_NV04(push, NV50_3D(CB_ADDR), 1);
 107             PUSH_DATA (push, (start << 8) | b);
 108             BEGIN_NI04(push, NV50_3D(CB_DATA(0)), nr);
 109             PUSH_DATAp(push, &res->data[start * 4], nr);
 110
 111             start += nr;
 112             words -= nr;
 113          }
 114       }
 115    }
 116 }
 117
 118 static boolean
 119 nv50_program_validate(struct nv50_context *nv50, struct nv50_program *prog)
 120 {
 121    if (!prog->translated) {
 122       prog->translated = nv50_program_translate(
 123          prog, nv50->screen->base.device->chipset);
 124       if (!prog->translated)
 125          return FALSE;
 126    } else
 127    if (prog->mem)
 128       return TRUE;
 129
 130    return nv50_program_upload_code(nv50, prog);
 131 }
 132
 133 static INLINE void
 134 nv50_program_update_context_state(struct nv50_context *nv50,
 135                                   struct nv50_program *prog, int stage)
 136 {
 137    const unsigned flags = NOUVEAU_BO_VRAM | NOUVEAU_BO_RDWR;
 138
 139    if (prog && prog->uses_lmem) {
 140       if (!nv50->state.tls_required)
 141          BCTX_REFN_bo(nv50->bufctx_3d, TLS, flags, nv50->screen->tls_bo);
 142       nv50->state.tls_required |= 1 << stage;
 143    } else {
 144       if (nv50->state.tls_required == (1 << stage))
 145          nouveau_bufctx_reset(nv50->bufctx_3d, NV50_BIND_TLS);
 146       nv50->state.tls_required &= ~(1 << stage);
 147    }
 148 }
 149
 150 void
 151 nv50_vertprog_validate(struct nv50_context *nv50)
 152 {
 153    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 154    struct nv50_program *vp = nv50->vertprog;
 155
 156    if (!nv50_program_validate(nv50, vp))
 157          return;
 158    nv50_program_update_context_state(nv50, vp, 0);
 159
 160    BEGIN_NV04(push, NV50_3D(VP_ATTR_EN(0)), 2);
 161    PUSH_DATA (push, vp->vp.attrs[0]);
 162    PUSH_DATA (push, vp->vp.attrs[1]);
 163    BEGIN_NV04(push, NV50_3D(VP_REG_ALLOC_RESULT), 1);
 164    PUSH_DATA (push, vp->max_out);
 165    BEGIN_NV04(push, NV50_3D(VP_REG_ALLOC_TEMP), 1);
 166    PUSH_DATA (push, vp->max_gpr);
 167    BEGIN_NV04(push, NV50_3D(VP_START_ID), 1);
 168    PUSH_DATA (push, vp->code_base);
 169 }
 170
 171 void
 172 nv50_fragprog_validate(struct nv50_context *nv50)
 173 {
 174    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 175    struct nv50_program *fp = nv50->fragprog;
 176
 177    if (!nv50_program_validate(nv50, fp))
 178          return;
 179    nv50_program_update_context_state(nv50, fp, 1);
 180
 181    BEGIN_NV04(push, NV50_3D(FP_REG_ALLOC_TEMP), 1);
 182    PUSH_DATA (push, fp->max_gpr);
 183    BEGIN_NV04(push, NV50_3D(FP_RESULT_COUNT), 1);
 184    PUSH_DATA (push, fp->max_out);
 185    BEGIN_NV04(push, NV50_3D(FP_CONTROL), 1);
 186    PUSH_DATA (push, fp->fp.flags[0]);
 187    BEGIN_NV04(push, NV50_3D(FP_CTRL_UNK196C), 1);
 188    PUSH_DATA (push, fp->fp.flags[1]);
 189    BEGIN_NV04(push, NV50_3D(FP_START_ID), 1);
 190    PUSH_DATA (push, fp->code_base);
 191 }
 192
 193 void
 194 nv50_gmtyprog_validate(struct nv50_context *nv50)
 195 {
 196    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 197    struct nv50_program *gp = nv50->gmtyprog;
 198
 199    if (gp) {
 200       BEGIN_NV04(push, NV50_3D(GP_REG_ALLOC_TEMP), 1);
 201       PUSH_DATA (push, gp->max_gpr);
 202       BEGIN_NV04(push, NV50_3D(GP_REG_ALLOC_RESULT), 1);
 203       PUSH_DATA (push, gp->max_out);
 204       BEGIN_NV04(push, NV50_3D(GP_OUTPUT_PRIMITIVE_TYPE), 1);
 205       PUSH_DATA (push, gp->gp.prim_type);
 206       BEGIN_NV04(push, NV50_3D(GP_VERTEX_OUTPUT_COUNT), 1);
 207       PUSH_DATA (push, gp->gp.vert_count);
 208       BEGIN_NV04(push, NV50_3D(GP_START_ID), 1);
 209       PUSH_DATA (push, gp->code_base);
 210    }
 211    nv50_program_update_context_state(nv50, gp, 2);
 212
 213    /* GP_ENABLE is updated in linkage validation */
 214 }
 215
 216 static void
 217 nv50_sprite_coords_validate(struct nv50_context *nv50)
 218 {
 219    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 220    uint32_t pntc[8], mode;
 221    struct nv50_program *fp = nv50->fragprog;
 222    unsigned i, c;
 223    unsigned m = (nv50->state.interpolant_ctrl >> 8) & 0xff;
 224
 225    if (!nv50->rast->pipe.point_quad_rasterization) {
 226       if (nv50->state.point_sprite) {
 227          BEGIN_NV04(push, NV50_3D(POINT_COORD_REPLACE_MAP(0)), 8);
 228          for (i = 0; i < 8; ++i)
 229             PUSH_DATA(push, 0);
 230
 231          nv50->state.point_sprite = FALSE;
 232       }
 233       return;
 234    } else {
 235       nv50->state.point_sprite = TRUE;
 236    }
 237
 238    memset(pntc, 0, sizeof(pntc));
 239
 240    for (i = 0; i < fp->in_nr; i++) {
 241       unsigned n = util_bitcount(fp->in[i].mask);
 242
 243       if (fp->in[i].sn != TGSI_SEMANTIC_GENERIC) {
 244          m += n;
 245          continue;
 246       }
 247       if (!(nv50->rast->pipe.sprite_coord_enable & (1 << fp->in[i].si))) {
 248          m += n;
 249          continue;
 250       }
 251
 252       for (c = 0; c < 4; ++c) {
 253          if (fp->in[i].mask & (1 << c)) {
 254             pntc[m / 8] |= (c + 1) << ((m % 8) * 4);
 255             ++m;
 256          }
 257       }
 258    }
 259
 260    if (nv50->rast->pipe.sprite_coord_mode == PIPE_SPRITE_COORD_LOWER_LEFT)
 261       mode = 0x00;
 262    else
 263       mode = 0x10;
 264
 265    BEGIN_NV04(push, NV50_3D(POINT_SPRITE_CTRL), 1);
 266    PUSH_DATA (push, mode);
 267
 268    BEGIN_NV04(push, NV50_3D(POINT_COORD_REPLACE_MAP(0)), 8);
 269    PUSH_DATAp(push, pntc, 8);
 270 }
 271
 272 /* Validate state derived from shaders and the rasterizer cso. */
 273 void
 274 nv50_validate_derived_rs(struct nv50_context *nv50)
 275 {
 276    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 277    uint32_t color, psize;
 278
 279    nv50_sprite_coords_validate(nv50);
 280
 281    if (nv50->dirty & NV50_NEW_FRAGPROG)
 282       return;
 283    psize = nv50->state.semantic_psize & ~NV50_3D_SEMANTIC_PTSZ_PTSZ_EN__MASK;
 284    color = nv50->state.semantic_color & ~NV50_3D_SEMANTIC_COLOR_CLMP_EN;
 285
 286    if (nv50->rast->pipe.clamp_vertex_color)
 287       color |= NV50_3D_SEMANTIC_COLOR_CLMP_EN;
 288
 289    if (color != nv50->state.semantic_color) {
 290       nv50->state.semantic_color = color;
 291       BEGIN_NV04(push, NV50_3D(SEMANTIC_COLOR), 1);
 292       PUSH_DATA (push, color);
 293    }
 294
 295    if (nv50->rast->pipe.point_size_per_vertex)
 296       psize |= NV50_3D_SEMANTIC_PTSZ_PTSZ_EN__MASK;
 297
 298    if (psize != nv50->state.semantic_psize) {
 299       nv50->state.semantic_psize = psize;
 300       BEGIN_NV04(push, NV50_3D(SEMANTIC_PTSZ), 1);
 301       PUSH_DATA (push, psize);
 302    }
 303 }
 304
 305 static int
 306 nv50_vec4_map(uint8_t *map, int mid, uint32_t lin[4],
 307               struct nv50_varying *in, struct nv50_varying *out)
 308 {
 309    int c;
 310    uint8_t mv = out->mask, mf = in->mask, oid = out->hw;
 311
 312    for (c = 0; c < 4; ++c) {
 313       if (mf & 1) {
 314          if (in->linear)
 315             lin[mid / 32] |= 1 << (mid % 32);
 316          if (mv & 1)
 317             map[mid] = oid;
 318          else
 319          if (c == 3)
 320             map[mid] |= 1;
 321          ++mid;
 322       }
 323
 324       oid += mv & 1;
 325       mf >>= 1;
 326       mv >>= 1;
 327    }
 328
 329    return mid;
 330 }
 331
 332 void
 333 nv50_fp_linkage_validate(struct nv50_context *nv50)
 334 {
 335    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 336    struct nv50_program *vp = nv50->gmtyprog ? nv50->gmtyprog : nv50->vertprog;
 337    struct nv50_program *fp = nv50->fragprog;
 338    struct nv50_varying dummy;
 339    int i, n, c, m;
 340    uint32_t primid = 0;
 341    uint32_t psiz = 0x000;
 342    uint32_t interp = fp->fp.interp;
 343    uint32_t colors = fp->fp.colors;
 344    uint32_t lin[4];
 345    uint8_t map[64];
 346
 347    if (!(nv50->dirty & (NV50_NEW_VERTPROG |
 348                         NV50_NEW_FRAGPROG |
 349                         NV50_NEW_GMTYPROG))) {
 350       uint8_t bfc, ffc;
 351       ffc = (nv50->state.semantic_color & NV50_3D_SEMANTIC_COLOR_FFC0_ID__MASK);
 352       bfc = (nv50->state.semantic_color & NV50_3D_SEMANTIC_COLOR_BFC0_ID__MASK)
 353          >> 8;
 354       if (nv50->rast->pipe.light_twoside == ((ffc == bfc) ? 0 : 1))
 355          return;
 356    }
 357
 358    memset(lin, 0x00, sizeof(lin));
 359
 360    /* XXX: in buggy-endian mode, is the first element of map (u32)0x000000xx
 361     *  or is it the first byte ?
 362     */
 363    memset(map, nv50->gmtyprog ? 0x80 : 0x40, sizeof(map));
 364
 365    dummy.mask = 0xf; /* map all components of HPOS */
 366    dummy.linear = 0;
 367    m = nv50_vec4_map(map, 0, lin, &dummy, &vp->out[0]);
 368
 369    for (c = 0; c < vp->vp.clpd_nr; ++c)
 370       map[m++] = vp->vp.clpd[c / 4] + (c % 4);
 371
 372    colors |= m << 8; /* adjust BFC0 id */
 373
 374    dummy.mask = 0x0;
 375
 376    /* if light_twoside is active, FFC0_ID == BFC0_ID is invalid */
 377    if (nv50->rast->pipe.light_twoside) {
 378       for (i = 0; i < 2; ++i) {
 379          n = vp->vp.bfc[i];
 380          if (fp->vp.bfc[i] >= fp->in_nr)
 381             continue;
 382          m = nv50_vec4_map(map, m, lin, &fp->in[fp->vp.bfc[i]],
 383                            (n < vp->out_nr) ? &vp->out[n] : &dummy);
 384       }
 385    }
 386    colors += m - 4; /* adjust FFC0 id */
 387    interp |= m << 8; /* set map id where 'normal' FP inputs start */
 388
 389    for (i = 0; i < fp->in_nr; ++i) {
 390       for (n = 0; n < vp->out_nr; ++n)
 391          if (vp->out[n].sn == fp->in[i].sn &&
 392              vp->out[n].si == fp->in[i].si)
 393             break;
 394       m = nv50_vec4_map(map, m, lin,
 395                         &fp->in[i], (n < vp->out_nr) ? &vp->out[n] : &dummy);
 396    }
 397
 398    /* PrimitiveID either is replaced by the system value, or
 399     * written by the geometry shader into an output register
 400     */
 401    if (fp->gp.primid < 0x80) {
 402       primid = m;
 403       map[m++] = vp->gp.primid;
 404    }
 405
 406    if (nv50->rast->pipe.point_size_per_vertex) {
 407       psiz = (m << 4) | 1;
 408       map[m++] = vp->vp.psiz;
 409    }
 410
 411    if (nv50->rast->pipe.clamp_vertex_color)
 412       colors |= NV50_3D_SEMANTIC_COLOR_CLMP_EN;
 413
 414    n = (m + 3) / 4;
 415    assert(m <= 64);
 416
 417    if (unlikely(nv50->gmtyprog)) {
 418       BEGIN_NV04(push, NV50_3D(GP_RESULT_MAP_SIZE), 1);
 419       PUSH_DATA (push, m);
 420       BEGIN_NV04(push, NV50_3D(GP_RESULT_MAP(0)), n);
 421       PUSH_DATAp(push, map, n);
 422    } else {
 423       BEGIN_NV04(push, NV50_3D(VP_GP_BUILTIN_ATTR_EN), 1);
 424       PUSH_DATA (push, vp->vp.attrs[2]);
 425
 426       BEGIN_NV04(push, NV50_3D(SEMANTIC_PRIM_ID), 1);
 427       PUSH_DATA (push, primid);
 428
 429       BEGIN_NV04(push, NV50_3D(VP_RESULT_MAP_SIZE), 1);
 430       PUSH_DATA (push, m);
 431       BEGIN_NV04(push, NV50_3D(VP_RESULT_MAP(0)), n);
 432       PUSH_DATAp(push, map, n);
 433    }
 434
 435    BEGIN_NV04(push, NV50_3D(SEMANTIC_COLOR), 4);
 436    PUSH_DATA (push, colors);
 437    PUSH_DATA (push, (vp->vp.clpd_nr << 8) | 4);
 438    PUSH_DATA (push, 0);
 439    PUSH_DATA (push, psiz);
 440
 441    BEGIN_NV04(push, NV50_3D(FP_INTERPOLANT_CTRL), 1);
 442    PUSH_DATA (push, interp);
 443
 444    nv50->state.interpolant_ctrl = interp;
 445
 446    nv50->state.semantic_color = colors;
 447    nv50->state.semantic_psize = psiz;
 448
 449    BEGIN_NV04(push, NV50_3D(NOPERSPECTIVE_BITMAP(0)), 4);
 450    PUSH_DATAp(push, lin, 4);
 451
 452    BEGIN_NV04(push, NV50_3D(GP_ENABLE), 1);
 453    PUSH_DATA (push, nv50->gmtyprog ? 1 : 0);
 454 }
 455
 456 static int
 457 nv50_vp_gp_mapping(uint8_t *map, int m,
 458                    struct nv50_program *vp, struct nv50_program *gp)
 459 {
 460    int i, j, c;
 461
 462    for (i = 0; i < gp->in_nr; ++i) {
 463       uint8_t oid = 0, mv = 0, mg = gp->in[i].mask;
 464
 465       for (j = 0; j < vp->out_nr; ++j) {
 466          if (vp->out[j].sn == gp->in[i].sn &&
 467              vp->out[j].si == gp->in[i].si) {
 468             mv = vp->out[j].mask;
 469             oid = vp->out[j].hw;
 470             break;
 471          }
 472       }
 473
 474       for (c = 0; c < 4; ++c, mv >>= 1, mg >>= 1) {
 475          if (mg & mv & 1)
 476             map[m++] = oid;
 477          else
 478          if (mg & 1)
 479             map[m++] = (c == 3) ? 0x41 : 0x40;
 480          oid += mv & 1;
 481       }
 482    }
 483    return m;
 484 }
 485
 486 void
 487 nv50_gp_linkage_validate(struct nv50_context *nv50)
 488 {
 489    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 490    struct nv50_program *vp = nv50->vertprog;
 491    struct nv50_program *gp = nv50->gmtyprog;
 492    int m = 0;
 493    int n;
 494    uint8_t map[64];
 495
 496    if (!gp)
 497       return;
 498    memset(map, 0, sizeof(map));
 499
 500    m = nv50_vp_gp_mapping(map, m, vp, gp);
 501
 502    n = (m + 3) / 4;
 503
 504    BEGIN_NV04(push, NV50_3D(VP_GP_BUILTIN_ATTR_EN), 1);
 505    PUSH_DATA (push, vp->vp.attrs[2] | gp->vp.attrs[2]);
 506
 507    BEGIN_NV04(push, NV50_3D(VP_RESULT_MAP_SIZE), 1);
 508    PUSH_DATA (push, m);
 509    BEGIN_NV04(push, NV50_3D(VP_RESULT_MAP(0)), n);
 510    PUSH_DATAp(push, map, n);
 511 }