src/gallium/drivers/nouveau/nvc0/nvc0_transfer.c

   1
   2 #include "util/u_format.h"
   3
   4 #include "nvc0/nvc0_context.h"
   5
   6 struct nvc0_transfer {
   7    struct pipe_transfer base;
   8    struct nv50_m2mf_rect rect[2];
   9    uint32_t nblocksx;
  10    uint16_t nblocksy;
  11    uint16_t nlayers;
  12 };
  13
  14 static void
  15 nvc0_m2mf_transfer_rect(struct nvc0_context *nvc0,
  16                         const struct nv50_m2mf_rect *dst,
  17                         const struct nv50_m2mf_rect *src,
  18                         uint32_t nblocksx, uint32_t nblocksy)
  19 {
  20    struct nouveau_pushbuf *push = nvc0->base.pushbuf;
  21    struct nouveau_bufctx *bctx = nvc0->bufctx;
  22    const int cpp = dst->cpp;
  23    uint32_t src_ofst = src->base;
  24    uint32_t dst_ofst = dst->base;
  25    uint32_t height = nblocksy;
  26    uint32_t sy = src->y;
  27    uint32_t dy = dst->y;
  28    uint32_t exec = (1 << 20);
  29
  30    assert(dst->cpp == src->cpp);
  31
  32    nouveau_bufctx_refn(bctx, 0, src->bo, src->domain | NOUVEAU_BO_RD);
  33    nouveau_bufctx_refn(bctx, 0, dst->bo, dst->domain | NOUVEAU_BO_WR);
  34    nouveau_pushbuf_bufctx(push, bctx);
  35    nouveau_pushbuf_validate(push);
  36
  37    if (nouveau_bo_memtype(src->bo)) {
  38       BEGIN_NVC0(push, NVC0_M2MF(TILING_MODE_IN), 5);
  39       PUSH_DATA (push, src->tile_mode);
  40       PUSH_DATA (push, src->width * cpp);
  41       PUSH_DATA (push, src->height);
  42       PUSH_DATA (push, src->depth);
  43       PUSH_DATA (push, src->z);
  44    } else {
  45       src_ofst += src->y * src->pitch + src->x * cpp;
  46
  47       BEGIN_NVC0(push, NVC0_M2MF(PITCH_IN), 1);
  48       PUSH_DATA (push, src->width * cpp);
  49
  50       exec |= NVC0_M2MF_EXEC_LINEAR_IN;
  51    }
  52
  53    if (nouveau_bo_memtype(dst->bo)) {
  54       BEGIN_NVC0(push, NVC0_M2MF(TILING_MODE_OUT), 5);
  55       PUSH_DATA (push, dst->tile_mode);
  56       PUSH_DATA (push, dst->width * cpp);
  57       PUSH_DATA (push, dst->height);
  58       PUSH_DATA (push, dst->depth);
  59       PUSH_DATA (push, dst->z);
  60    } else {
  61       dst_ofst += dst->y * dst->pitch + dst->x * cpp;
  62
  63       BEGIN_NVC0(push, NVC0_M2MF(PITCH_OUT), 1);
  64       PUSH_DATA (push, dst->width * cpp);
  65
  66       exec |= NVC0_M2MF_EXEC_LINEAR_OUT;
  67    }
  68
  69    while (height) {
  70       int line_count = height > 2047 ? 2047 : height;
  71
  72       BEGIN_NVC0(push, NVC0_M2MF(OFFSET_IN_HIGH), 2);
  73       PUSH_DATAh(push, src->bo->offset + src_ofst);
  74       PUSH_DATA (push, src->bo->offset + src_ofst);
  75
  76       BEGIN_NVC0(push, NVC0_M2MF(OFFSET_OUT_HIGH), 2);
  77       PUSH_DATAh(push, dst->bo->offset + dst_ofst);
  78       PUSH_DATA (push, dst->bo->offset + dst_ofst);
  79
  80       if (!(exec & NVC0_M2MF_EXEC_LINEAR_IN)) {
  81          BEGIN_NVC0(push, NVC0_M2MF(TILING_POSITION_IN_X), 2);
  82          PUSH_DATA (push, src->x * cpp);
  83          PUSH_DATA (push, sy);
  84       } else {
  85          src_ofst += line_count * src->pitch;
  86       }
  87       if (!(exec & NVC0_M2MF_EXEC_LINEAR_OUT)) {
  88          BEGIN_NVC0(push, NVC0_M2MF(TILING_POSITION_OUT_X), 2);
  89          PUSH_DATA (push, dst->x * cpp);
  90          PUSH_DATA (push, dy);
  91       } else {
  92          dst_ofst += line_count * dst->pitch;
  93       }
  94
  95       BEGIN_NVC0(push, NVC0_M2MF(LINE_LENGTH_IN), 2);
  96       PUSH_DATA (push, nblocksx * cpp);
  97       PUSH_DATA (push, line_count);
  98       BEGIN_NVC0(push, NVC0_M2MF(EXEC), 1);
  99       PUSH_DATA (push, exec);
 100
 101       height -= line_count;
 102       sy += line_count;
 103       dy += line_count;
 104    }
 105
 106    nouveau_bufctx_reset(bctx, 0);
 107 }
 108
 109 static void
 110 nve4_m2mf_transfer_rect(struct nvc0_context *nvc0,
 111                         const struct nv50_m2mf_rect *dst,
 112                         const struct nv50_m2mf_rect *src,
 113                         uint32_t nblocksx, uint32_t nblocksy)
 114 {
 115    struct nouveau_pushbuf *push = nvc0->base.pushbuf;
 116    struct nouveau_bufctx *bctx = nvc0->bufctx;
 117    uint32_t exec;
 118    uint32_t src_base = src->base;
 119    uint32_t dst_base = dst->base;
 120    const int cpp = dst->cpp;
 121
 122    assert(dst->cpp == src->cpp);
 123
 124    nouveau_bufctx_refn(bctx, 0, dst->bo, dst->domain | NOUVEAU_BO_WR);
 125    nouveau_bufctx_refn(bctx, 0, src->bo, src->domain | NOUVEAU_BO_RD);
 126    nouveau_pushbuf_bufctx(push, bctx);
 127    nouveau_pushbuf_validate(push);
 128
 129    exec = 0x200 /* 2D_ENABLE */ | 0x6 /* UNK */;
 130
 131    if (!nouveau_bo_memtype(dst->bo)) {
 132       assert(!dst->z);
 133       dst_base += dst->y * dst->pitch + dst->x * cpp;
 134       exec |= 0x100; /* DST_MODE_2D_LINEAR */
 135    }
 136    if (!nouveau_bo_memtype(src->bo)) {
 137       assert(!src->z);
 138       src_base += src->y * src->pitch + src->x * cpp;
 139       exec |= 0x080; /* SRC_MODE_2D_LINEAR */
 140    }
 141
 142    BEGIN_NVC0(push, SUBC_COPY(0x070c), 6);
 143    PUSH_DATA (push, 0x1000 | dst->tile_mode);
 144    PUSH_DATA (push, dst->pitch);
 145    PUSH_DATA (push, dst->height);
 146    PUSH_DATA (push, dst->depth);
 147    PUSH_DATA (push, dst->z);
 148    PUSH_DATA (push, (dst->y << 16) | (dst->x * cpp));
 149
 150    BEGIN_NVC0(push, SUBC_COPY(0x0728), 6);
 151    PUSH_DATA (push, 0x1000 | src->tile_mode);
 152    PUSH_DATA (push, src->pitch);
 153    PUSH_DATA (push, src->height);
 154    PUSH_DATA (push, src->depth);
 155    PUSH_DATA (push, src->z);
 156    PUSH_DATA (push, (src->y << 16) | (src->x * cpp));
 157
 158    BEGIN_NVC0(push, SUBC_COPY(0x0400), 8);
 159    PUSH_DATAh(push, src->bo->offset + src_base);
 160    PUSH_DATA (push, src->bo->offset + src_base);
 161    PUSH_DATAh(push, dst->bo->offset + dst_base);
 162    PUSH_DATA (push, dst->bo->offset + dst_base);
 163    PUSH_DATA (push, src->pitch);
 164    PUSH_DATA (push, dst->pitch);
 165    PUSH_DATA (push, nblocksx * cpp);
 166    PUSH_DATA (push, nblocksy);
 167
 168    BEGIN_NVC0(push, SUBC_COPY(0x0300), 1);
 169    PUSH_DATA (push, exec);
 170
 171    nouveau_bufctx_reset(bctx, 0);
 172 }
 173
 174 void
 175 nvc0_m2mf_push_linear(struct nouveau_context *nv,
 176                       struct nouveau_bo *dst, unsigned offset, unsigned domain,
 177                       unsigned size, const void *data)
 178 {
 179    struct nvc0_context *nvc0 = nvc0_context(&nv->pipe);
 180    struct nouveau_pushbuf *push = nv->pushbuf;
 181    uint32_t *src = (uint32_t *)data;
 182    unsigned count = (size + 3) / 4;
 183
 184    nouveau_bufctx_refn(nvc0->bufctx, 0, dst, domain | NOUVEAU_BO_WR);
 185    nouveau_pushbuf_bufctx(push, nvc0->bufctx);
 186    nouveau_pushbuf_validate(push);
 187
 188    while (count) {
 189       unsigned nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN);
 190
 191       if (!PUSH_SPACE(push, nr + 9))
 192          break;
 193
 194       BEGIN_NVC0(push, NVC0_M2MF(OFFSET_OUT_HIGH), 2);
 195       PUSH_DATAh(push, dst->offset + offset);
 196       PUSH_DATA (push, dst->offset + offset);
 197       BEGIN_NVC0(push, NVC0_M2MF(LINE_LENGTH_IN), 2);
 198       PUSH_DATA (push, MIN2(size, nr * 4));
 199       PUSH_DATA (push, 1);
 200       BEGIN_NVC0(push, NVC0_M2MF(EXEC), 1);
 201       PUSH_DATA (push, 0x100111);
 202
 203       /* must not be interrupted (trap on QUERY fence, 0x50 works however) */
 204       BEGIN_NIC0(push, NVC0_M2MF(DATA), nr);
 205       PUSH_DATAp(push, src, nr);
 206
 207       count -= nr;
 208       src += nr;
 209       offset += nr * 4;
 210       size -= nr * 4;
 211    }
 212
 213    nouveau_bufctx_reset(nvc0->bufctx, 0);
 214 }
 215
 216 void
 217 nve4_p2mf_push_linear(struct nouveau_context *nv,
 218                       struct nouveau_bo *dst, unsigned offset, unsigned domain,
 219                       unsigned size, const void *data)
 220 {
 221    struct nvc0_context *nvc0 = nvc0_context(&nv->pipe);
 222    struct nouveau_pushbuf *push = nv->pushbuf;
 223    uint32_t *src = (uint32_t *)data;
 224    unsigned count = (size + 3) / 4;
 225
 226    nouveau_bufctx_refn(nvc0->bufctx, 0, dst, domain | NOUVEAU_BO_WR);
 227    nouveau_pushbuf_bufctx(push, nvc0->bufctx);
 228    nouveau_pushbuf_validate(push);
 229
 230    while (count) {
 231       unsigned nr = MIN2(count, (NV04_PFIFO_MAX_PACKET_LEN - 1));
 232
 233       if (!PUSH_SPACE(push, nr + 10))
 234          break;
 235
 236       BEGIN_NVC0(push, NVE4_P2MF(UPLOAD_DST_ADDRESS_HIGH), 2);
 237       PUSH_DATAh(push, dst->offset + offset);
 238       PUSH_DATA (push, dst->offset + offset);
 239       BEGIN_NVC0(push, NVE4_P2MF(UPLOAD_LINE_LENGTH_IN), 2);
 240       PUSH_DATA (push, MIN2(size, nr * 4));
 241       PUSH_DATA (push, 1);
 242       /* must not be interrupted (trap on QUERY fence, 0x50 works however) */
 243       BEGIN_1IC0(push, NVE4_P2MF(UPLOAD_EXEC), nr + 1);
 244       PUSH_DATA (push, 0x1001);
 245       PUSH_DATAp(push, src, nr);
 246
 247       count -= nr;
 248       src += nr;
 249       offset += nr * 4;
 250       size -= nr * 4;
 251    }
 252
 253    nouveau_bufctx_reset(nvc0->bufctx, 0);
 254 }
 255
 256 static void
 257 nvc0_m2mf_copy_linear(struct nouveau_context *nv,
 258                       struct nouveau_bo *dst, unsigned dstoff, unsigned dstdom,
 259                       struct nouveau_bo *src, unsigned srcoff, unsigned srcdom,
 260                       unsigned size)
 261 {
 262    struct nouveau_pushbuf *push = nv->pushbuf;
 263    struct nouveau_bufctx *bctx = nvc0_context(&nv->pipe)->bufctx;
 264
 265    nouveau_bufctx_refn(bctx, 0, src, srcdom | NOUVEAU_BO_RD);
 266    nouveau_bufctx_refn(bctx, 0, dst, dstdom | NOUVEAU_BO_WR);
 267    nouveau_pushbuf_bufctx(push, bctx);
 268    nouveau_pushbuf_validate(push);
 269
 270    while (size) {
 271       unsigned bytes = MIN2(size, 1 << 17);
 272
 273       BEGIN_NVC0(push, NVC0_M2MF(OFFSET_OUT_HIGH), 2);
 274       PUSH_DATAh(push, dst->offset + dstoff);
 275       PUSH_DATA (push, dst->offset + dstoff);
 276       BEGIN_NVC0(push, NVC0_M2MF(OFFSET_IN_HIGH), 2);
 277       PUSH_DATAh(push, src->offset + srcoff);
 278       PUSH_DATA (push, src->offset + srcoff);
 279       BEGIN_NVC0(push, NVC0_M2MF(LINE_LENGTH_IN), 2);
 280       PUSH_DATA (push, bytes);
 281       PUSH_DATA (push, 1);
 282       BEGIN_NVC0(push, NVC0_M2MF(EXEC), 1);
 283       PUSH_DATA (push, NVC0_M2MF_EXEC_QUERY_SHORT |
 284                  NVC0_M2MF_EXEC_LINEAR_IN | NVC0_M2MF_EXEC_LINEAR_OUT);
 285
 286       srcoff += bytes;
 287       dstoff += bytes;
 288       size -= bytes;
 289    }
 290
 291    nouveau_bufctx_reset(bctx, 0);
 292 }
 293
 294 static void
 295 nve4_m2mf_copy_linear(struct nouveau_context *nv,
 296                       struct nouveau_bo *dst, unsigned dstoff, unsigned dstdom,
 297                       struct nouveau_bo *src, unsigned srcoff, unsigned srcdom,
 298                       unsigned size)
 299 {
 300    struct nouveau_pushbuf *push = nv->pushbuf;
 301    struct nouveau_bufctx *bctx = nvc0_context(&nv->pipe)->bufctx;
 302
 303    nouveau_bufctx_refn(bctx, 0, src, srcdom | NOUVEAU_BO_RD);
 304    nouveau_bufctx_refn(bctx, 0, dst, dstdom | NOUVEAU_BO_WR);
 305    nouveau_pushbuf_bufctx(push, bctx);
 306    nouveau_pushbuf_validate(push);
 307
 308    BEGIN_NVC0(push, SUBC_COPY(0x0400), 4);
 309    PUSH_DATAh(push, src->offset + srcoff);
 310    PUSH_DATA (push, src->offset + srcoff);
 311    PUSH_DATAh(push, dst->offset + dstoff);
 312    PUSH_DATA (push, dst->offset + dstoff);
 313    BEGIN_NVC0(push, SUBC_COPY(0x0418), 1);
 314    PUSH_DATA (push, size);
 315    BEGIN_NVC0(push, SUBC_COPY(0x0300), 1);
 316    PUSH_DATA (push, 0x186);
 317
 318    nouveau_bufctx_reset(bctx, 0);
 319 }
 320
 321
 322 static inline bool
 323 nvc0_mt_transfer_can_map_directly(struct nv50_miptree *mt)
 324 {
 325    if (mt->base.domain == NOUVEAU_BO_VRAM)
 326       return false;
 327    if (mt->base.base.usage != PIPE_USAGE_STAGING)
 328       return false;
 329    return !nouveau_bo_memtype(mt->base.bo);
 330 }
 331
 332 static inline bool
 333 nvc0_mt_sync(struct nvc0_context *nvc0, struct nv50_miptree *mt, unsigned usage)
 334 {
 335    if (!mt->base.mm) {
 336       uint32_t access = (usage & PIPE_TRANSFER_WRITE) ?
 337          NOUVEAU_BO_WR : NOUVEAU_BO_RD;
 338       return !nouveau_bo_wait(mt->base.bo, access, nvc0->base.client);
 339    }
 340    if (usage & PIPE_TRANSFER_WRITE)
 341       return !mt->base.fence || nouveau_fence_wait(mt->base.fence, &nvc0->base.debug);
 342    return !mt->base.fence_wr || nouveau_fence_wait(mt->base.fence_wr, &nvc0->base.debug);
 343 }
 344
 345 void *
 346 nvc0_miptree_transfer_map(struct pipe_context *pctx,
 347                           struct pipe_resource *res,
 348                           unsigned level,
 349                           unsigned usage,
 350                           const struct pipe_box *box,
 351                           struct pipe_transfer **ptransfer)
 352 {
 353    struct nvc0_context *nvc0 = nvc0_context(pctx);
 354    struct nouveau_device *dev = nvc0->screen->base.device;
 355    struct nv50_miptree *mt = nv50_miptree(res);
 356    struct nvc0_transfer *tx;
 357    uint32_t size;
 358    int ret;
 359    unsigned flags = 0;
 360
 361    if (nvc0_mt_transfer_can_map_directly(mt)) {
 362       ret = !nvc0_mt_sync(nvc0, mt, usage);
 363       if (!ret)
 364          ret = nouveau_bo_map(mt->base.bo, 0, NULL);
 365       if (ret &&
 366           (usage & PIPE_TRANSFER_MAP_DIRECTLY))
 367          return NULL;
 368       if (!ret)
 369          usage |= PIPE_TRANSFER_MAP_DIRECTLY;
 370    } else
 371    if (usage & PIPE_TRANSFER_MAP_DIRECTLY)
 372       return NULL;
 373
 374    tx = CALLOC_STRUCT(nvc0_transfer);
 375    if (!tx)
 376       return NULL;
 377
 378    pipe_resource_reference(&tx->base.resource, res);
 379
 380    tx->base.level = level;
 381    tx->base.usage = usage;
 382    tx->base.box = *box;
 383
 384    if (util_format_is_plain(res->format)) {
 385       tx->nblocksx = box->width << mt->ms_x;
 386       tx->nblocksy = box->height << mt->ms_y;
 387    } else {
 388       tx->nblocksx = util_format_get_nblocksx(res->format, box->width);
 389       tx->nblocksy = util_format_get_nblocksy(res->format, box->height);
 390    }
 391    tx->nlayers = box->depth;
 392
 393    if (usage & PIPE_TRANSFER_MAP_DIRECTLY) {
 394       tx->base.stride = mt->level[level].pitch;
 395       tx->base.layer_stride = mt->layer_stride;
 396       uint32_t offset = box->y * tx->base.stride +
 397          util_format_get_stride(res->format, box->x);
 398       if (!mt->layout_3d)
 399          offset += mt->layer_stride * box->z;
 400       else
 401          offset += nvc0_mt_zslice_offset(mt, level, box->z);
 402       *ptransfer = &tx->base;
 403       return mt->base.bo->map + mt->base.offset + offset;
 404    }
 405
 406    tx->base.stride = tx->nblocksx * util_format_get_blocksize(res->format);
 407    tx->base.layer_stride = tx->nblocksy * tx->base.stride;
 408
 409    nv50_m2mf_rect_setup(&tx->rect[0], res, level, box->x, box->y, box->z);
 410
 411    size = tx->base.layer_stride;
 412
 413    ret = nouveau_bo_new(dev, NOUVEAU_BO_GART | NOUVEAU_BO_MAP, 0,
 414                         size * tx->nlayers, NULL, &tx->rect[1].bo);
 415    if (ret) {
 416       pipe_resource_reference(&tx->base.resource, NULL);
 417       FREE(tx);
 418       return NULL;
 419    }
 420
 421    tx->rect[1].cpp = tx->rect[0].cpp;
 422    tx->rect[1].width = tx->nblocksx;
 423    tx->rect[1].height = tx->nblocksy;
 424    tx->rect[1].depth = 1;
 425    tx->rect[1].pitch = tx->base.stride;
 426    tx->rect[1].domain = NOUVEAU_BO_GART;
 427
 428    if (usage & PIPE_TRANSFER_READ) {
 429       unsigned base = tx->rect[0].base;
 430       unsigned z = tx->rect[0].z;
 431       unsigned i;
 432       for (i = 0; i < tx->nlayers; ++i) {
 433          nvc0->m2mf_copy_rect(nvc0, &tx->rect[1], &tx->rect[0],
 434                               tx->nblocksx, tx->nblocksy);
 435          if (mt->layout_3d)
 436             tx->rect[0].z++;
 437          else
 438             tx->rect[0].base += mt->layer_stride;
 439          tx->rect[1].base += size;
 440       }
 441       tx->rect[0].z = z;
 442       tx->rect[0].base = base;
 443       tx->rect[1].base = 0;
 444    }
 445
 446    if (tx->rect[1].bo->map) {
 447       *ptransfer = &tx->base;
 448       return tx->rect[1].bo->map;
 449    }
 450
 451    if (usage & PIPE_TRANSFER_READ)
 452       flags = NOUVEAU_BO_RD;
 453    if (usage & PIPE_TRANSFER_WRITE)
 454       flags |= NOUVEAU_BO_WR;
 455
 456    ret = nouveau_bo_map(tx->rect[1].bo, flags, nvc0->screen->base.client);
 457    if (ret) {
 458       pipe_resource_reference(&tx->base.resource, NULL);
 459       nouveau_bo_ref(NULL, &tx->rect[1].bo);
 460       FREE(tx);
 461       return NULL;
 462    }
 463
 464    *ptransfer = &tx->base;
 465    return tx->rect[1].bo->map;
 466 }
 467
 468 void
 469 nvc0_miptree_transfer_unmap(struct pipe_context *pctx,
 470                             struct pipe_transfer *transfer)
 471 {
 472    struct nvc0_context *nvc0 = nvc0_context(pctx);
 473    struct nvc0_transfer *tx = (struct nvc0_transfer *)transfer;
 474    struct nv50_miptree *mt = nv50_miptree(tx->base.resource);
 475    unsigned i;
 476
 477    if (tx->base.usage & PIPE_TRANSFER_MAP_DIRECTLY) {
 478       pipe_resource_reference(&transfer->resource, NULL);
 479
 480       FREE(tx);
 481       return;
 482    }
 483
 484    if (tx->base.usage & PIPE_TRANSFER_WRITE) {
 485       for (i = 0; i < tx->nlayers; ++i) {
 486          nvc0->m2mf_copy_rect(nvc0, &tx->rect[0], &tx->rect[1],
 487                               tx->nblocksx, tx->nblocksy);
 488          if (mt->layout_3d)
 489             tx->rect[0].z++;
 490          else
 491             tx->rect[0].base += mt->layer_stride;
 492          tx->rect[1].base += tx->nblocksy * tx->base.stride;
 493       }
 494       NOUVEAU_DRV_STAT(&nvc0->screen->base, tex_transfers_wr, 1);
 495
 496       /* Allow the copies above to finish executing before freeing the source */
 497       nouveau_fence_work(nvc0->screen->base.fence.current,
 498                          nouveau_fence_unref_bo, tx->rect[1].bo);
 499    } else {
 500       nouveau_bo_ref(NULL, &tx->rect[1].bo);
 501    }
 502    if (tx->base.usage & PIPE_TRANSFER_READ)
 503       NOUVEAU_DRV_STAT(&nvc0->screen->base, tex_transfers_rd, 1);
 504
 505    pipe_resource_reference(&transfer->resource, NULL);
 506
 507    FREE(tx);
 508 }
 509
 510 /* This happens rather often with DTD9/st. */
 511 static void
 512 nvc0_cb_push(struct nouveau_context *nv,
 513              struct nv04_resource *res,
 514              unsigned offset, unsigned words, const uint32_t *data)
 515 {
 516    struct nvc0_context *nvc0 = nvc0_context(&nv->pipe);
 517    struct nvc0_constbuf *cb = NULL;
 518    int s;
 519
 520    /* Go through all the constbuf binding points of this buffer and try to
 521     * find one which contains the region to be updated.
 522     */
 523    for (s = 0; s < 6 && !cb; s++) {
 524       uint16_t bindings = res->cb_bindings[s];
 525       while (bindings) {
 526          int i = ffs(bindings) - 1;
 527          uint32_t cb_offset = nvc0->constbuf[s][i].offset;
 528
 529          bindings &= ~(1 << i);
 530          if (cb_offset <= offset &&
 531              cb_offset + nvc0->constbuf[s][i].size >= offset + words * 4) {
 532             cb = &nvc0->constbuf[s][i];
 533             break;
 534          }
 535       }
 536    }
 537
 538    if (cb) {
 539       nvc0_cb_bo_push(nv, res->bo, res->domain,
 540                       res->offset + cb->offset, cb->size,
 541                       offset - cb->offset, words, data);
 542    } else {
 543       nv->push_data(nv, res->bo, res->offset + offset, res->domain,
 544                     words * 4, data);
 545    }
 546 }
 547
 548 void
 549 nvc0_cb_bo_push(struct nouveau_context *nv,
 550                 struct nouveau_bo *bo, unsigned domain,
 551                 unsigned base, unsigned size,
 552                 unsigned offset, unsigned words, const uint32_t *data)
 553 {
 554    struct nouveau_pushbuf *push = nv->pushbuf;
 555
 556    NOUVEAU_DRV_STAT(nv->screen, constbuf_upload_count, 1);
 557    NOUVEAU_DRV_STAT(nv->screen, constbuf_upload_bytes, words * 4);
 558
 559    assert(!(offset & 3));
 560    size = align(size, 0x100);
 561
 562    assert(offset < size);
 563    assert(offset + words * 4 <= size);
 564
 565    BEGIN_NVC0(push, NVC0_3D(CB_SIZE), 3);
 566    PUSH_DATA (push, size);
 567    PUSH_DATAh(push, bo->offset + base);
 568    PUSH_DATA (push, bo->offset + base);
 569
 570    while (words) {
 571       unsigned nr = MIN2(words, NV04_PFIFO_MAX_PACKET_LEN - 1);
 572
 573       PUSH_SPACE(push, nr + 2);
 574       PUSH_REFN (push, bo, NOUVEAU_BO_WR | domain);
 575       BEGIN_1IC0(push, NVC0_3D(CB_POS), nr + 1);
 576       PUSH_DATA (push, offset);
 577       PUSH_DATAp(push, data, nr);
 578
 579       words -= nr;
 580       data += nr;
 581       offset += nr * 4;
 582    }
 583 }
 584
 585 void
 586 nvc0_init_transfer_functions(struct nvc0_context *nvc0)
 587 {
 588    if (nvc0->screen->base.class_3d >= NVE4_3D_CLASS) {
 589       nvc0->m2mf_copy_rect = nve4_m2mf_transfer_rect;
 590       nvc0->base.copy_data = nve4_m2mf_copy_linear;
 591       nvc0->base.push_data = nve4_p2mf_push_linear;
 592    } else {
 593       nvc0->m2mf_copy_rect = nvc0_m2mf_transfer_rect;
 594       nvc0->base.copy_data = nvc0_m2mf_copy_linear;
 595       nvc0->base.push_data = nvc0_m2mf_push_linear;
 596    }
 597    nvc0->base.push_cb = nvc0_cb_push;
 598 }