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    struct nouveau_heap *heap;
 122    int ret;
 123    unsigned size;
 124
 125    if (!prog->translated) {
 126       prog->translated = nv50_program_translate(prog);
 127       if (!prog->translated)
 128          return FALSE;
 129    } else
 130    if (prog->mem)
 131       return TRUE;
 132
 133    if (prog->type == PIPE_SHADER_FRAGMENT) heap = nv50->screen->fp_code_heap;
 134    else
 135    if (prog->type == PIPE_SHADER_GEOMETRY) heap = nv50->screen->gp_code_heap;
 136    else
 137       heap = nv50->screen->vp_code_heap;
 138
 139    size = align(prog->code_size, 0x100);
 140
 141    ret = nouveau_heap_alloc(heap, size, prog, &prog->mem);
 142    if (ret) {
 143       NOUVEAU_ERR("out of code space for shader type %i\n", prog->type);
 144       return FALSE;
 145    }
 146    prog->code_base = prog->mem->start;
 147
 148    nv50_relocate_program(prog, prog->code_base, 0);
 149
 150    nv50_sifc_linear_u8(&nv50->base, nv50->screen->code,
 151                        (prog->type << NV50_CODE_BO_SIZE_LOG2) + prog->code_base,
 152                        NOUVEAU_BO_VRAM, prog->code_size, prog->code);
 153
 154    BEGIN_NV04(nv50->base.pushbuf, NV50_3D(CODE_CB_FLUSH), 1);
 155    PUSH_DATA (nv50->base.pushbuf, 0);
 156
 157    return TRUE;
 158 }
 159
 160 static INLINE void
 161 nv50_program_update_context_state(struct nv50_context *nv50,
 162                                   struct nv50_program *prog, int stage)
 163 {
 164    const unsigned flags = NOUVEAU_BO_VRAM | NOUVEAU_BO_RDWR;
 165
 166    if (prog && prog->uses_lmem) {
 167       if (!nv50->state.tls_required)
 168          BCTX_REFN_bo(nv50->bufctx_3d, TLS, flags, nv50->screen->tls_bo);
 169       nv50->state.tls_required |= 1 << stage;
 170    } else {
 171       if (nv50->state.tls_required == (1 << stage))
 172          nouveau_bufctx_reset(nv50->bufctx_3d, NV50_BIND_TLS);
 173       nv50->state.tls_required &= ~(1 << stage);
 174    }
 175 }
 176
 177 void
 178 nv50_vertprog_validate(struct nv50_context *nv50)
 179 {
 180    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 181    struct nv50_program *vp = nv50->vertprog;
 182
 183    if (!nv50_program_validate(nv50, vp))
 184          return;
 185    nv50_program_update_context_state(nv50, vp, 0);
 186
 187    BEGIN_NV04(push, NV50_3D(VP_ATTR_EN(0)), 2);
 188    PUSH_DATA (push, vp->vp.attrs[0]);
 189    PUSH_DATA (push, vp->vp.attrs[1]);
 190    BEGIN_NV04(push, NV50_3D(VP_REG_ALLOC_RESULT), 1);
 191    PUSH_DATA (push, vp->max_out);
 192    BEGIN_NV04(push, NV50_3D(VP_REG_ALLOC_TEMP), 1);
 193    PUSH_DATA (push, vp->max_gpr);
 194    BEGIN_NV04(push, NV50_3D(VP_START_ID), 1);
 195    PUSH_DATA (push, vp->code_base);
 196 }
 197
 198 void
 199 nv50_fragprog_validate(struct nv50_context *nv50)
 200 {
 201    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 202    struct nv50_program *fp = nv50->fragprog;
 203
 204    if (!nv50_program_validate(nv50, fp))
 205          return;
 206    nv50_program_update_context_state(nv50, fp, 1);
 207
 208    BEGIN_NV04(push, NV50_3D(FP_REG_ALLOC_TEMP), 1);
 209    PUSH_DATA (push, fp->max_gpr);
 210    BEGIN_NV04(push, NV50_3D(FP_RESULT_COUNT), 1);
 211    PUSH_DATA (push, fp->max_out);
 212    BEGIN_NV04(push, NV50_3D(FP_CONTROL), 1);
 213    PUSH_DATA (push, fp->fp.flags[0]);
 214    BEGIN_NV04(push, NV50_3D(FP_CTRL_UNK196C), 1);
 215    PUSH_DATA (push, fp->fp.flags[1]);
 216    BEGIN_NV04(push, NV50_3D(FP_START_ID), 1);
 217    PUSH_DATA (push, fp->code_base);
 218 }
 219
 220 void
 221 nv50_gmtyprog_validate(struct nv50_context *nv50)
 222 {
 223    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 224    struct nv50_program *gp = nv50->gmtyprog;
 225
 226    if (gp) {
 227       BEGIN_NV04(push, NV50_3D(GP_REG_ALLOC_TEMP), 1);
 228       PUSH_DATA (push, gp->max_gpr);
 229       BEGIN_NV04(push, NV50_3D(GP_REG_ALLOC_RESULT), 1);
 230       PUSH_DATA (push, gp->max_out);
 231       BEGIN_NV04(push, NV50_3D(GP_OUTPUT_PRIMITIVE_TYPE), 1);
 232       PUSH_DATA (push, gp->gp.prim_type);
 233       BEGIN_NV04(push, NV50_3D(GP_VERTEX_OUTPUT_COUNT), 1);
 234       PUSH_DATA (push, gp->gp.vert_count);
 235       BEGIN_NV04(push, NV50_3D(GP_START_ID), 1);
 236       PUSH_DATA (push, gp->code_base);
 237    }
 238    nv50_program_update_context_state(nv50, gp, 2);
 239
 240    /* GP_ENABLE is updated in linkage validation */
 241 }
 242
 243 static void
 244 nv50_sprite_coords_validate(struct nv50_context *nv50)
 245 {
 246    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 247    uint32_t pntc[8], mode;
 248    struct nv50_program *fp = nv50->fragprog;
 249    unsigned i, c;
 250    unsigned m = (nv50->state.interpolant_ctrl >> 8) & 0xff;
 251
 252    if (!nv50->rast->pipe.point_quad_rasterization) {
 253       if (nv50->state.point_sprite) {
 254          BEGIN_NV04(push, NV50_3D(POINT_COORD_REPLACE_MAP(0)), 8);
 255          for (i = 0; i < 8; ++i)
 256             PUSH_DATA(push, 0);
 257
 258          nv50->state.point_sprite = FALSE;
 259       }
 260       return;
 261    } else {
 262       nv50->state.point_sprite = TRUE;
 263    }
 264
 265    memset(pntc, 0, sizeof(pntc));
 266
 267    for (i = 0; i < fp->in_nr; i++) {
 268       unsigned n = util_bitcount(fp->in[i].mask);
 269
 270       if (fp->in[i].sn != TGSI_SEMANTIC_GENERIC) {
 271          m += n;
 272          continue;
 273       }
 274       if (!(nv50->rast->pipe.sprite_coord_enable & (1 << fp->in[i].si))) {
 275          m += n;
 276          continue;
 277       }
 278
 279       for (c = 0; c < 4; ++c) {
 280          if (fp->in[i].mask & (1 << c)) {
 281             pntc[m / 8] |= (c + 1) << ((m % 8) * 4);
 282             ++m;
 283          }
 284       }
 285    }
 286
 287    if (nv50->rast->pipe.sprite_coord_mode == PIPE_SPRITE_COORD_LOWER_LEFT)
 288       mode = 0x00;
 289    else
 290       mode = 0x10;
 291
 292    BEGIN_NV04(push, NV50_3D(POINT_SPRITE_CTRL), 1);
 293    PUSH_DATA (push, mode);
 294
 295    BEGIN_NV04(push, NV50_3D(POINT_COORD_REPLACE_MAP(0)), 8);
 296    PUSH_DATAp(push, pntc, 8);
 297 }
 298
 299 /* Validate state derived from shaders and the rasterizer cso. */
 300 void
 301 nv50_validate_derived_rs(struct nv50_context *nv50)
 302 {
 303    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 304    uint32_t color, psize;
 305
 306    nv50_sprite_coords_validate(nv50);
 307
 308    if (nv50->dirty & NV50_NEW_FRAGPROG)
 309       return;
 310    psize = nv50->state.semantic_psize & ~NV50_3D_SEMANTIC_PTSZ_PTSZ_EN__MASK;
 311    color = nv50->state.semantic_color & ~NV50_3D_SEMANTIC_COLOR_CLMP_EN;
 312
 313    if (nv50->rast->pipe.clamp_vertex_color)
 314       color |= NV50_3D_SEMANTIC_COLOR_CLMP_EN;
 315
 316    if (color != nv50->state.semantic_color) {
 317       nv50->state.semantic_color = color;
 318       BEGIN_NV04(push, NV50_3D(SEMANTIC_COLOR), 1);
 319       PUSH_DATA (push, color);
 320    }
 321
 322    if (nv50->rast->pipe.point_size_per_vertex)
 323       psize |= NV50_3D_SEMANTIC_PTSZ_PTSZ_EN__MASK;
 324
 325    if (psize != nv50->state.semantic_psize) {
 326       nv50->state.semantic_psize = psize;
 327       BEGIN_NV04(push, NV50_3D(SEMANTIC_PTSZ), 1);
 328       PUSH_DATA (push, psize);
 329    }
 330 }
 331
 332 static int
 333 nv50_vec4_map(uint8_t *map, int mid, uint32_t lin[4],
 334               struct nv50_varying *in, struct nv50_varying *out)
 335 {
 336    int c;
 337    uint8_t mv = out->mask, mf = in->mask, oid = out->hw;
 338
 339    for (c = 0; c < 4; ++c) {
 340       if (mf & 1) {
 341          if (in->linear)
 342             lin[mid / 32] |= 1 << (mid % 32);
 343          if (mv & 1)
 344             map[mid] = oid;
 345          else
 346          if (c == 3)
 347             map[mid] |= 1;
 348          ++mid;
 349       }
 350
 351       oid += mv & 1;
 352       mf >>= 1;
 353       mv >>= 1;
 354    }
 355
 356    return mid;
 357 }
 358
 359 void
 360 nv50_fp_linkage_validate(struct nv50_context *nv50)
 361 {
 362    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 363    struct nv50_program *vp = nv50->gmtyprog ? nv50->gmtyprog : nv50->vertprog;
 364    struct nv50_program *fp = nv50->fragprog;
 365    struct nv50_varying dummy;
 366    int i, n, c, m;
 367    uint32_t primid = 0;
 368    uint32_t psiz = 0x000;
 369    uint32_t interp = fp->fp.interp;
 370    uint32_t colors = fp->fp.colors;
 371    uint32_t lin[4];
 372    uint8_t map[64];
 373
 374    memset(lin, 0x00, sizeof(lin));
 375
 376    /* XXX: in buggy-endian mode, is the first element of map (u32)0x000000xx
 377     *  or is it the first byte ?
 378     */
 379    memset(map, nv50->gmtyprog ? 0x80 : 0x40, sizeof(map));
 380
 381    dummy.mask = 0xf; /* map all components of HPOS */
 382    dummy.linear = 0;
 383    m = nv50_vec4_map(map, 0, lin, &dummy, &vp->out[0]);
 384
 385    for (c = 0; c < vp->vp.clpd_nr; ++c)
 386       map[m++] = vp->vp.clpd + c;
 387
 388    colors |= m << 8; /* adjust BFC0 id */
 389
 390    /* if light_twoside is active, FFC0_ID == BFC0_ID is invalid */
 391    if (nv50->rast->pipe.light_twoside) {
 392       for (i = 0; i < 2; ++i)
 393          m = nv50_vec4_map(map, m, lin,
 394                            &fp->in[fp->vp.bfc[i]], &vp->out[vp->vp.bfc[i]]);
 395    }
 396    colors += m - 4; /* adjust FFC0 id */
 397    interp |= m << 8; /* set map id where 'normal' FP inputs start */
 398
 399    dummy.mask = 0x0;
 400    for (i = 0; i < fp->in_nr; ++i) {
 401       for (n = 0; n < vp->out_nr; ++n)
 402          if (vp->out[n].sn == fp->in[i].sn &&
 403              vp->out[n].si == fp->in[i].si)
 404             break;
 405       m = nv50_vec4_map(map, m, lin,
 406                         &fp->in[i], (n < vp->out_nr) ? &vp->out[n] : &dummy);
 407    }
 408
 409    /* PrimitiveID either is replaced by the system value, or
 410     * written by the geometry shader into an output register
 411     */
 412    if (fp->gp.primid < 0x40) {
 413       primid = m;
 414       map[m++] = vp->gp.primid;
 415    }
 416
 417    if (nv50->rast->pipe.point_size_per_vertex) {
 418       psiz = (m << 4) | 1;
 419       map[m++] = vp->vp.psiz;
 420    }
 421
 422    if (nv50->rast->pipe.clamp_vertex_color)
 423       colors |= NV50_3D_SEMANTIC_COLOR_CLMP_EN;
 424
 425    n = (m + 3) / 4;
 426    assert(m <= 64);
 427
 428    if (unlikely(nv50->gmtyprog)) {
 429       BEGIN_NV04(push, NV50_3D(GP_RESULT_MAP_SIZE), 1);
 430       PUSH_DATA (push, m);
 431       BEGIN_NV04(push, NV50_3D(GP_RESULT_MAP(0)), n);
 432       PUSH_DATAp(push, map, n);
 433    } else {
 434       BEGIN_NV04(push, NV50_3D(VP_GP_BUILTIN_ATTR_EN), 1);
 435       PUSH_DATA (push, vp->vp.attrs[2]);
 436
 437       BEGIN_NV04(push, NV50_3D(SEMANTIC_PRIM_ID), 1);
 438       PUSH_DATA (push, primid);
 439
 440       BEGIN_NV04(push, NV50_3D(VP_RESULT_MAP_SIZE), 1);
 441       PUSH_DATA (push, m);
 442       BEGIN_NV04(push, NV50_3D(VP_RESULT_MAP(0)), n);
 443       PUSH_DATAp(push, map, n);
 444    }
 445
 446    BEGIN_NV04(push, NV50_3D(SEMANTIC_COLOR), 4);
 447    PUSH_DATA (push, colors);
 448    PUSH_DATA (push, (vp->vp.clpd_nr << 8) | 4);
 449    PUSH_DATA (push, 0);
 450    PUSH_DATA (push, psiz);
 451
 452    BEGIN_NV04(push, NV50_3D(FP_INTERPOLANT_CTRL), 1);
 453    PUSH_DATA (push, interp);
 454
 455    nv50->state.interpolant_ctrl = interp;
 456
 457    nv50->state.semantic_color = colors;
 458    nv50->state.semantic_psize = psiz;
 459
 460    BEGIN_NV04(push, NV50_3D(NOPERSPECTIVE_BITMAP(0)), 4);
 461    PUSH_DATAp(push, lin, 4);
 462
 463    BEGIN_NV04(push, NV50_3D(GP_ENABLE), 1);
 464    PUSH_DATA (push, nv50->gmtyprog ? 1 : 0);
 465 }
 466
 467 static int
 468 nv50_vp_gp_mapping(uint8_t *map, int m,
 469                    struct nv50_program *vp, struct nv50_program *gp)
 470 {
 471    int i, j, c;
 472
 473    for (i = 0; i < gp->in_nr; ++i) {
 474       uint8_t oid = 0, mv = 0, mg = gp->in[i].mask;
 475
 476       for (j = 0; j < vp->out_nr; ++j) {
 477          if (vp->out[j].sn == gp->in[i].sn &&
 478              vp->out[j].si == gp->in[i].si) {
 479             mv = vp->out[j].mask;
 480             oid = vp->out[j].hw;
 481             break;
 482          }
 483       }
 484
 485       for (c = 0; c < 4; ++c, mv >>= 1, mg >>= 1) {
 486          if (mg & mv & 1)
 487             map[m++] = oid;
 488          else
 489          if (mg & 1)
 490             map[m++] = (c == 3) ? 0x41 : 0x40;
 491          oid += mv & 1;
 492       }
 493    }
 494    return m;
 495 }
 496
 497 void
 498 nv50_gp_linkage_validate(struct nv50_context *nv50)
 499 {
 500    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 501    struct nv50_program *vp = nv50->vertprog;
 502    struct nv50_program *gp = nv50->gmtyprog;
 503    int m = 0;
 504    int n;
 505    uint8_t map[64];
 506
 507    if (!gp)
 508       return;
 509    memset(map, 0, sizeof(map));
 510
 511    m = nv50_vp_gp_mapping(map, m, vp, gp);
 512
 513    n = (m + 3) / 4;
 514
 515    BEGIN_NV04(push, NV50_3D(VP_GP_BUILTIN_ATTR_EN), 1);
 516    PUSH_DATA (push, vp->vp.attrs[2] | gp->vp.attrs[2]);
 517
 518    BEGIN_NV04(push, NV50_3D(VP_RESULT_MAP_SIZE), 1);
 519    PUSH_DATA (push, m);
 520    BEGIN_NV04(push, NV50_3D(VP_RESULT_MAP(0)), n);
 521    PUSH_DATAp(push, map, n);
 522 }