src/gallium/drivers/lima/lima_submit.c

   1 /*
   2  * Copyright (C) 2017-2019 Lima Project
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  */
  23
  24 #include <stdlib.h>
  25 #include <string.h>
  26
  27 #include "xf86drm.h"
  28 #include "drm-uapi/lima_drm.h"
  29
  30 #include "util/u_math.h"
  31 #include "util/ralloc.h"
  32 #include "util/u_dynarray.h"
  33 #include "util/os_time.h"
  34 #include "util/hash_table.h"
  35 #include "util/u_upload_mgr.h"
  36 #include "util/u_inlines.h"
  37
  38 #include "lima_screen.h"
  39 #include "lima_context.h"
  40 #include "lima_submit.h"
  41 #include "lima_bo.h"
  42 #include "lima_util.h"
  43 #include "lima_format.h"
  44 #include "lima_resource.h"
  45 #include "lima_texture.h"
  46 #include "lima_fence.h"
  47 #include "lima_gpu.h"
  48
  49 struct lima_submit {
  50    int fd;
  51    struct lima_context *ctx;
  52
  53    struct util_dynarray gem_bos[2];
  54    struct util_dynarray bos[2];
  55 };
  56
  57
  58 #define VOID2U64(x) ((uint64_t)(unsigned long)(x))
  59
  60 static struct lima_submit *
  61 lima_submit_create(struct lima_context *ctx)
  62 {
  63    struct lima_submit *s;
  64
  65    s = rzalloc(ctx, struct lima_submit);
  66    if (!s)
  67       return NULL;
  68
  69    s->fd = lima_screen(ctx->base.screen)->fd;
  70    s->ctx = ctx;
  71
  72    for (int i = 0; i < 2; i++) {
  73       util_dynarray_init(s->gem_bos + i, s);
  74       util_dynarray_init(s->bos + i, s);
  75    }
  76
  77    return s;
  78 }
  79
  80 static void
  81 lima_submit_free(struct lima_submit *submit)
  82 {
  83
  84 }
  85
  86 /*
  87  * Note: this function can only be called in draw code path,
  88  * must not exist in flush code path.
  89  */
  90 struct lima_submit *
  91 lima_submit_get(struct lima_context *ctx)
  92 {
  93    return ctx->submit;
  94 }
  95
  96 bool lima_submit_add_bo(struct lima_submit *submit, int pipe,
  97                         struct lima_bo *bo, uint32_t flags)
  98 {
  99    util_dynarray_foreach(submit->gem_bos + pipe, struct drm_lima_gem_submit_bo, gem_bo) {
 100       if (bo->handle == gem_bo->handle) {
 101          gem_bo->flags |= flags;
 102          return true;
 103       }
 104    }
 105
 106    struct drm_lima_gem_submit_bo *submit_bo =
 107       util_dynarray_grow(submit->gem_bos + pipe, struct drm_lima_gem_submit_bo, 1);
 108    submit_bo->handle = bo->handle;
 109    submit_bo->flags = flags;
 110
 111    struct lima_bo **jbo = util_dynarray_grow(submit->bos + pipe, struct lima_bo *, 1);
 112    *jbo = bo;
 113
 114    /* prevent bo from being freed when submit start */
 115    lima_bo_reference(bo);
 116
 117    return true;
 118 }
 119
 120 static bool
 121 lima_submit_start(struct lima_submit *submit, int pipe, void *frame, uint32_t size)
 122 {
 123    struct lima_context *ctx = submit->ctx;
 124    struct drm_lima_gem_submit req = {
 125       .ctx = ctx->id,
 126       .pipe = pipe,
 127       .nr_bos = submit->gem_bos[pipe].size / sizeof(struct drm_lima_gem_submit_bo),
 128       .bos = VOID2U64(util_dynarray_begin(submit->gem_bos + pipe)),
 129       .frame = VOID2U64(frame),
 130       .frame_size = size,
 131       .out_sync = ctx->out_sync[pipe],
 132    };
 133
 134    if (ctx->in_sync_fd >= 0) {
 135       int err = drmSyncobjImportSyncFile(submit->fd, ctx->in_sync[pipe],
 136                                          ctx->in_sync_fd);
 137       if (err)
 138          return false;
 139
 140       req.in_sync[0] = ctx->in_sync[pipe];
 141       close(ctx->in_sync_fd);
 142       ctx->in_sync_fd = -1;
 143    }
 144
 145    bool ret = drmIoctl(submit->fd, DRM_IOCTL_LIMA_GEM_SUBMIT, &req) == 0;
 146
 147    util_dynarray_foreach(submit->bos + pipe, struct lima_bo *, bo) {
 148       lima_bo_unreference(*bo);
 149    }
 150
 151    util_dynarray_clear(submit->gem_bos + pipe);
 152    util_dynarray_clear(submit->bos + pipe);
 153    return ret;
 154 }
 155
 156 static bool
 157 lima_submit_wait(struct lima_submit *submit, int pipe, uint64_t timeout_ns)
 158 {
 159    int64_t abs_timeout = os_time_get_absolute_timeout(timeout_ns);
 160    if (abs_timeout == OS_TIMEOUT_INFINITE)
 161       abs_timeout = INT64_MAX;
 162
 163    struct lima_context *ctx = submit->ctx;
 164    return !drmSyncobjWait(submit->fd, ctx->out_sync + pipe, 1, abs_timeout, 0, NULL);
 165 }
 166
 167 bool lima_submit_has_bo(struct lima_submit *submit, struct lima_bo *bo, bool all)
 168 {
 169    for (int i = 0; i < 2; i++) {
 170       util_dynarray_foreach(submit->gem_bos + i, struct drm_lima_gem_submit_bo, gem_bo) {
 171          if (bo->handle == gem_bo->handle) {
 172             if (all || gem_bo->flags & LIMA_SUBMIT_BO_WRITE)
 173                return true;
 174             else
 175                break;
 176          }
 177       }
 178    }
 179
 180    return false;
 181 }
 182
 183 void *
 184 lima_submit_create_stream_bo(struct lima_submit *submit, int pipe,
 185                              unsigned size, uint32_t *va)
 186 {
 187    struct lima_context *ctx = submit->ctx;
 188
 189    void *cpu;
 190    unsigned offset;
 191    struct pipe_resource *pres = NULL;
 192    u_upload_alloc(ctx->uploader, 0, size, 0x40, &offset, &pres, &cpu);
 193
 194    struct lima_resource *res = lima_resource(pres);
 195    *va = res->bo->va + offset;
 196
 197    lima_submit_add_bo(submit, pipe, res->bo, LIMA_SUBMIT_BO_READ);
 198
 199    pipe_resource_reference(&pres, NULL);
 200
 201    return cpu;
 202 }
 203
 204 static inline bool
 205 lima_submit_dirty(struct lima_submit *submit)
 206 {
 207    struct lima_context *ctx = submit->ctx;
 208
 209    return !!ctx->resolve;
 210 }
 211
 212 static inline struct lima_damage_region *
 213 lima_submit_get_damage(struct lima_submit *submit)
 214 {
 215    struct lima_context *ctx = submit->ctx;
 216
 217    if (!(ctx->framebuffer.base.nr_cbufs && (ctx->resolve & PIPE_CLEAR_COLOR0)))
 218       return NULL;
 219
 220    struct lima_surface *surf = lima_surface(ctx->framebuffer.base.cbufs[0]);
 221    struct lima_resource *res = lima_resource(surf->base.texture);
 222    return &res->damage;
 223 }
 224
 225 static bool
 226 lima_fb_need_reload(struct lima_submit *submit)
 227 {
 228    struct lima_context *ctx = submit->ctx;
 229
 230    /* Depth buffer is always discarded */
 231    if (!(ctx->framebuffer.base.nr_cbufs && (ctx->resolve & PIPE_CLEAR_COLOR0)))
 232       return false;
 233
 234    struct lima_surface *surf = lima_surface(ctx->framebuffer.base.cbufs[0]);
 235    struct lima_resource *res = lima_resource(surf->base.texture);
 236    if (res->damage.region) {
 237       /* for EGL_KHR_partial_update, when EGL_EXT_buffer_age is enabled,
 238        * we need to reload damage region, otherwise just want to reload
 239        * the region not aligned to tile boundary */
 240       //if (!res->damage.aligned)
 241       //   return true;
 242       return true;
 243    }
 244    else if (surf->reload)
 245          return true;
 246
 247    return false;
 248 }
 249
 250 static void
 251 lima_pack_reload_plbu_cmd(struct lima_submit *submit)
 252 {
 253    #define lima_reload_render_state_offset 0x0000
 254    #define lima_reload_gl_pos_offset       0x0040
 255    #define lima_reload_varying_offset      0x0080
 256    #define lima_reload_tex_desc_offset     0x00c0
 257    #define lima_reload_tex_array_offset    0x0100
 258    #define lima_reload_buffer_size         0x0140
 259
 260    struct lima_context *ctx = submit->ctx;
 261
 262    uint32_t va;
 263    void *cpu = lima_submit_create_stream_bo(
 264       submit, LIMA_PIPE_PP, lima_reload_buffer_size, &va);
 265
 266    struct lima_screen *screen = lima_screen(ctx->base.screen);
 267
 268    uint32_t reload_shader_first_instr_size =
 269       ((uint32_t *)(screen->pp_buffer->map + pp_reload_program_offset))[0] & 0x1f;
 270    uint32_t reload_shader_va = screen->pp_buffer->va + pp_reload_program_offset;
 271
 272    struct lima_render_state reload_render_state = {
 273       .alpha_blend = 0xf03b1ad2,
 274       .depth_test = 0x0000000e,
 275       .depth_range = 0xffff0000,
 276       .stencil_front = 0x00000007,
 277       .stencil_back = 0x00000007,
 278       .multi_sample = 0x0000f007,
 279       .shader_address = reload_shader_va | reload_shader_first_instr_size,
 280       .varying_types = 0x00000001,
 281       .textures_address = va + lima_reload_tex_array_offset,
 282       .aux0 = 0x00004021,
 283       .varyings_address = va + lima_reload_varying_offset,
 284    };
 285    memcpy(cpu + lima_reload_render_state_offset, &reload_render_state,
 286           sizeof(reload_render_state));
 287
 288    struct lima_context_framebuffer *fb = &ctx->framebuffer;
 289    lima_tex_desc *td = cpu + lima_reload_tex_desc_offset;
 290    memset(td, 0, lima_min_tex_desc_size);
 291    lima_texture_desc_set_res(ctx, td, fb->base.cbufs[0]->texture, 0, 0);
 292    td->unnorm_coords = 1;
 293    td->texture_type = LIMA_TEXTURE_TYPE_2D;
 294    td->min_img_filter_nearest = 1;
 295    td->mag_img_filter_nearest = 1;
 296    td->wrap_s_clamp_to_edge = 1;
 297    td->wrap_t_clamp_to_edge = 1;
 298    td->unknown_2_2 = 0x1;
 299
 300    uint32_t *ta = cpu + lima_reload_tex_array_offset;
 301    ta[0] = va + lima_reload_tex_desc_offset;
 302
 303    float reload_gl_pos[] = {
 304       fb->base.width, 0,               0, 1,
 305       0,              0,               0, 1,
 306       0,              fb->base.height, 0, 1,
 307    };
 308    memcpy(cpu + lima_reload_gl_pos_offset, reload_gl_pos,
 309           sizeof(reload_gl_pos));
 310
 311    float reload_varying[] = {
 312       fb->base.width, 0,               0, 0,
 313       0,              fb->base.height, 0, 0,
 314    };
 315    memcpy(cpu + lima_reload_varying_offset, reload_varying,
 316           sizeof(reload_varying));
 317
 318    PLBU_CMD_BEGIN(&ctx->plbu_cmd_head, 20);
 319
 320    PLBU_CMD_VIEWPORT_LEFT(0);
 321    PLBU_CMD_VIEWPORT_RIGHT(fui(fb->base.width));
 322    PLBU_CMD_VIEWPORT_BOTTOM(0);
 323    PLBU_CMD_VIEWPORT_TOP(fui(fb->base.height));
 324
 325    PLBU_CMD_RSW_VERTEX_ARRAY(
 326       va + lima_reload_render_state_offset,
 327       va + lima_reload_gl_pos_offset);
 328
 329    PLBU_CMD_UNKNOWN2();
 330    PLBU_CMD_UNKNOWN1();
 331
 332    PLBU_CMD_INDICES(screen->pp_buffer->va + pp_shared_index_offset);
 333    PLBU_CMD_INDEXED_DEST(va + lima_reload_gl_pos_offset);
 334    PLBU_CMD_DRAW_ELEMENTS(0xf, 0, 3);
 335
 336    PLBU_CMD_END();
 337 }
 338
 339 static void
 340 lima_pack_head_plbu_cmd(struct lima_submit *submit)
 341 {
 342    struct lima_context *ctx = submit->ctx;
 343    struct lima_context_framebuffer *fb = &ctx->framebuffer;
 344
 345    PLBU_CMD_BEGIN(&ctx->plbu_cmd_head, 10);
 346
 347    PLBU_CMD_UNKNOWN2();
 348    PLBU_CMD_BLOCK_STEP(fb->shift_min, fb->shift_h, fb->shift_w);
 349    PLBU_CMD_TILED_DIMENSIONS(fb->tiled_w, fb->tiled_h);
 350    PLBU_CMD_BLOCK_STRIDE(fb->block_w);
 351
 352    PLBU_CMD_ARRAY_ADDRESS(
 353       ctx->plb_gp_stream->va + ctx->plb_index * ctx->plb_gp_size,
 354       fb->block_w * fb->block_h);
 355
 356    PLBU_CMD_END();
 357
 358    if (lima_fb_need_reload(submit))
 359       lima_pack_reload_plbu_cmd(submit);
 360 }
 361
 362 static void
 363 hilbert_rotate(int n, int *x, int *y, int rx, int ry)
 364 {
 365    if (ry == 0) {
 366       if (rx == 1) {
 367          *x = n-1 - *x;
 368          *y = n-1 - *y;
 369       }
 370
 371       /* Swap x and y */
 372       int t  = *x;
 373       *x = *y;
 374       *y = t;
 375    }
 376 }
 377
 378 static void
 379 hilbert_coords(int n, int d, int *x, int *y)
 380 {
 381    int rx, ry, i, t=d;
 382
 383    *x = *y = 0;
 384
 385    for (i = 0; (1 << i) < n; i++) {
 386
 387       rx = 1 & (t / 2);
 388       ry = 1 & (t ^ rx);
 389
 390       hilbert_rotate(1 << i, x, y, rx, ry);
 391
 392       *x += rx << i;
 393       *y += ry << i;
 394
 395       t /= 4;
 396    }
 397 }
 398
 399 static int
 400 lima_get_pp_stream_size(int num_pp, int tiled_w, int tiled_h, uint32_t *off)
 401 {
 402    /* carefully calculate each stream start address:
 403     * 1. overflow: each stream size may be different due to
 404     *    fb->tiled_w * fb->tiled_h can't be divided by num_pp,
 405     *    extra size should be added to the preceeding stream
 406     * 2. alignment: each stream address should be 0x20 aligned
 407     */
 408    int delta = tiled_w * tiled_h / num_pp * 16 + 16;
 409    int remain = tiled_w * tiled_h % num_pp;
 410    int offset = 0;
 411
 412    for (int i = 0; i < num_pp; i++) {
 413       off[i] = offset;
 414
 415       offset += delta;
 416       if (remain) {
 417          offset += 16;
 418          remain--;
 419       }
 420       offset = align(offset, 0x20);
 421    }
 422
 423    return offset;
 424 }
 425
 426 static bool
 427 inside_damage_region(int x, int y, struct lima_damage_region *ds)
 428 {
 429    if (!ds || !ds->region)
 430       return true;
 431
 432    for (int i = 0; i < ds->num_region; i++) {
 433       struct pipe_scissor_state *ss = ds->region + i;
 434       if (x >= ss->minx && x < ss->maxx &&
 435           y >= ss->miny && y < ss->maxy)
 436          return true;
 437    }
 438
 439    return false;
 440 }
 441
 442 static void
 443 lima_generate_pp_stream(struct lima_submit *submit, int off_x, int off_y,
 444                       int tiled_w, int tiled_h)
 445 {
 446    struct lima_context *ctx = submit->ctx;
 447    struct lima_pp_stream_state *ps = &ctx->pp_stream;
 448    struct lima_context_framebuffer *fb = &ctx->framebuffer;
 449    struct lima_damage_region *damage = lima_submit_get_damage(submit);
 450    struct lima_screen *screen = lima_screen(ctx->base.screen);
 451    int i, num_pp = screen->num_pp;
 452
 453    /* use hilbert_coords to generates 1D to 2D relationship.
 454     * 1D for pp stream index and 2D for plb block x/y on framebuffer.
 455     * if multi pp, interleave the 1D index to make each pp's render target
 456     * close enough which should result close workload
 457     */
 458    int max = MAX2(tiled_w, tiled_h);
 459    int index = 0;
 460    uint32_t *stream[4];
 461    int si[4] = {0};
 462    int dim = 0;
 463    int count = 0;
 464
 465    /* Don't update count if we get zero rect. We'll just generate
 466     * PP stream with just terminators in it.
 467     */
 468    if ((tiled_w * tiled_h) != 0) {
 469       dim = util_logbase2_ceil(max);
 470       count = 1 << (dim + dim);
 471    }
 472
 473    for (i = 0; i < num_pp; i++)
 474       stream[i] = ps->map + ps->offset[i];
 475
 476    for (i = 0; i < count; i++) {
 477       int x, y;
 478       hilbert_coords(max, i, &x, &y);
 479       if (x < tiled_w && y < tiled_h) {
 480          x += off_x;
 481          y += off_y;
 482
 483          if (!inside_damage_region(x, y, damage))
 484             continue;
 485
 486          int pp = index % num_pp;
 487          int offset = ((y >> fb->shift_h) * fb->block_w +
 488                        (x >> fb->shift_w)) * LIMA_CTX_PLB_BLK_SIZE;
 489          int plb_va = ctx->plb[ctx->plb_index]->va + offset;
 490
 491          stream[pp][si[pp]++] = 0;
 492          stream[pp][si[pp]++] = 0xB8000000 | x | (y << 8);
 493          stream[pp][si[pp]++] = 0xE0000002 | ((plb_va >> 3) & ~0xE0000003);
 494          stream[pp][si[pp]++] = 0xB0000000;
 495
 496          index++;
 497       }
 498    }
 499
 500    for (i = 0; i < num_pp; i++) {
 501       stream[i][si[i]++] = 0;
 502       stream[i][si[i]++] = 0xBC000000;
 503       stream[i][si[i]++] = 0;
 504       stream[i][si[i]++] = 0;
 505
 506       lima_dump_command_stream_print(
 507          stream[i], si[i] * 4, false, "pp plb stream %d at va %x\n",
 508          i, ps->va + ps->offset[i]);
 509    }
 510 }
 511
 512 static void
 513 lima_update_damage_pp_stream(struct lima_submit *submit)
 514 {
 515    struct lima_context *ctx = submit->ctx;
 516    struct lima_damage_region *ds = lima_submit_get_damage(submit);
 517    struct lima_context_framebuffer *fb = &ctx->framebuffer;
 518    struct pipe_scissor_state bound;
 519
 520    if (ds && ds->region) {
 521       struct pipe_scissor_state *dbound = &ds->bound;
 522       bound.minx = MAX2(dbound->minx, ctx->damage_rect.minx >> 4);
 523       bound.miny = MAX2(dbound->miny, ctx->damage_rect.miny >> 4);
 524       bound.maxx = MIN2(dbound->maxx, (ctx->damage_rect.maxx + 0xf) >> 4);
 525       bound.maxy = MIN2(dbound->maxy, (ctx->damage_rect.maxy + 0xf) >> 4);
 526    } else {
 527       bound.minx = ctx->damage_rect.minx >> 4;
 528       bound.miny = ctx->damage_rect.miny >> 4;
 529       bound.maxx = (ctx->damage_rect.maxx + 0xf) >> 4;
 530       bound.maxy = (ctx->damage_rect.maxy + 0xf) >> 4;
 531    }
 532
 533    /* Clamp to FB size */
 534    bound.minx = MIN2(bound.minx, fb->tiled_w);
 535    bound.miny = MIN2(bound.miny, fb->tiled_h);
 536    bound.maxx = MIN2(bound.maxx, fb->tiled_w);
 537    bound.maxy = MIN2(bound.maxy, fb->tiled_h);
 538
 539    int tiled_w = bound.maxx - bound.minx;
 540    int tiled_h = bound.maxy - bound.miny;
 541
 542    struct lima_screen *screen = lima_screen(ctx->base.screen);
 543    int size = lima_get_pp_stream_size(
 544       screen->num_pp, tiled_w, tiled_h, ctx->pp_stream.offset);
 545
 546    ctx->pp_stream.map = lima_submit_create_stream_bo(
 547       submit, LIMA_PIPE_PP, size, &ctx->pp_stream.va);
 548
 549    lima_generate_pp_stream(submit, bound.minx, bound.miny, tiled_w, tiled_h);
 550 }
 551
 552 static void
 553 lima_update_full_pp_stream(struct lima_submit *submit)
 554 {
 555    struct lima_context *ctx = submit->ctx;
 556    struct lima_context_framebuffer *fb = &ctx->framebuffer;
 557    struct lima_ctx_plb_pp_stream_key key = {
 558       .plb_index = ctx->plb_index,
 559       .tiled_w = fb->tiled_w,
 560       .tiled_h = fb->tiled_h,
 561    };
 562
 563    struct hash_entry *entry =
 564       _mesa_hash_table_search(ctx->plb_pp_stream, &key);
 565    struct lima_ctx_plb_pp_stream *s = entry->data;
 566
 567    if (s->bo) {
 568       ctx->pp_stream.map = lima_bo_map(s->bo);
 569       ctx->pp_stream.va = s->bo->va;
 570       memcpy(ctx->pp_stream.offset, s->offset, sizeof(s->offset));
 571    }
 572    else {
 573       struct lima_screen *screen = lima_screen(ctx->base.screen);
 574       int size = lima_get_pp_stream_size(
 575          screen->num_pp, fb->tiled_w, fb->tiled_h, s->offset);
 576       s->bo = lima_bo_create(screen, size, 0);
 577
 578       ctx->pp_stream.map = lima_bo_map(s->bo);
 579       ctx->pp_stream.va = s->bo->va;
 580       memcpy(ctx->pp_stream.offset, s->offset, sizeof(s->offset));
 581
 582       lima_generate_pp_stream(submit, 0, 0, fb->tiled_w, fb->tiled_h);
 583    }
 584
 585    lima_submit_add_bo(submit, LIMA_PIPE_PP, s->bo, LIMA_SUBMIT_BO_READ);
 586 }
 587
 588 static bool
 589 lima_damage_fullscreen(struct lima_submit *submit)
 590 {
 591    struct lima_context *ctx = submit->ctx;
 592
 593    return ctx->damage_rect.minx == 0 &&
 594           ctx->damage_rect.miny == 0 &&
 595           ctx->damage_rect.maxx == ctx->framebuffer.base.width &&
 596           ctx->damage_rect.maxy == ctx->framebuffer.base.height;
 597 }
 598
 599 static void
 600 lima_update_pp_stream(struct lima_submit *submit)
 601 {
 602    struct lima_context *ctx = submit->ctx;
 603    struct lima_damage_region *damage = lima_submit_get_damage(submit);
 604    if ((damage && damage->region) || !lima_damage_fullscreen(submit))
 605       lima_update_damage_pp_stream(submit);
 606    else if (ctx->plb_pp_stream)
 607       lima_update_full_pp_stream(submit);
 608    else
 609       ctx->pp_stream.map = NULL;
 610 }
 611
 612 static void
 613 lima_update_submit_bo(struct lima_submit *submit)
 614 {
 615    struct lima_context *ctx = submit->ctx;
 616
 617    lima_submit_add_bo(submit, LIMA_PIPE_GP, ctx->plb_gp_stream,
 618                       LIMA_SUBMIT_BO_READ);
 619    lima_submit_add_bo(submit, LIMA_PIPE_GP, ctx->plb[ctx->plb_index],
 620                       LIMA_SUBMIT_BO_WRITE);
 621    lima_submit_add_bo(submit, LIMA_PIPE_GP, ctx->gp_tile_heap[ctx->plb_index],
 622                       LIMA_SUBMIT_BO_WRITE);
 623
 624    lima_dump_command_stream_print(
 625       ctx->plb_gp_stream->map + ctx->plb_index * ctx->plb_gp_size,
 626       ctx->plb_gp_size, false, "gp plb stream at va %x\n",
 627       ctx->plb_gp_stream->va + ctx->plb_index * ctx->plb_gp_size);
 628
 629    lima_submit_add_bo(submit, LIMA_PIPE_PP, ctx->plb[ctx->plb_index],
 630                       LIMA_SUBMIT_BO_READ);
 631    lima_submit_add_bo(submit, LIMA_PIPE_PP, ctx->gp_tile_heap[ctx->plb_index],
 632                       LIMA_SUBMIT_BO_READ);
 633
 634    struct lima_screen *screen = lima_screen(ctx->base.screen);
 635    lima_submit_add_bo(submit, LIMA_PIPE_PP, screen->pp_buffer, LIMA_SUBMIT_BO_READ);
 636 }
 637
 638 static void
 639 lima_finish_plbu_cmd(struct util_dynarray *plbu_cmd_array)
 640 {
 641    int i = 0;
 642    uint32_t *plbu_cmd = util_dynarray_ensure_cap(plbu_cmd_array, plbu_cmd_array->size + 2 * 4);
 643
 644    plbu_cmd[i++] = 0x00000000;
 645    plbu_cmd[i++] = 0x50000000; /* END */
 646
 647    plbu_cmd_array->size += i * 4;
 648 }
 649
 650 static void
 651 lima_pack_wb_zsbuf_reg(struct lima_submit *submit, uint32_t *wb_reg, int wb_idx)
 652 {
 653    struct lima_context *ctx = submit->ctx;
 654    struct lima_context_framebuffer *fb = &ctx->framebuffer;
 655    struct lima_resource *res = lima_resource(fb->base.zsbuf->texture);
 656    int level = fb->base.zsbuf->u.tex.level;
 657    uint32_t format = lima_format_get_pixel(fb->base.zsbuf->format);
 658
 659    struct lima_pp_wb_reg *wb = (void *)wb_reg;
 660    wb[wb_idx].type = 0x01; /* 1 for depth, stencil */
 661    wb[wb_idx].address = res->bo->va + res->levels[level].offset;
 662    wb[wb_idx].pixel_format = format;
 663    if (res->tiled) {
 664       wb[wb_idx].pixel_layout = 0x2;
 665       wb[wb_idx].pitch = fb->tiled_w;
 666    } else {
 667       wb[wb_idx].pixel_layout = 0x0;
 668       wb[wb_idx].pitch = res->levels[level].stride / 8;
 669    }
 670    wb[wb_idx].mrt_bits = 0;
 671 }
 672
 673 static void
 674 lima_pack_wb_cbuf_reg(struct lima_submit *submit, uint32_t *wb_reg, int wb_idx)
 675 {
 676    struct lima_context *ctx = submit->ctx;
 677    struct lima_context_framebuffer *fb = &ctx->framebuffer;
 678    struct lima_resource *res = lima_resource(fb->base.cbufs[0]->texture);
 679    int level = fb->base.cbufs[0]->u.tex.level;
 680    unsigned layer = fb->base.cbufs[0]->u.tex.first_layer;
 681    uint32_t format = lima_format_get_pixel(fb->base.cbufs[0]->format);
 682    bool swap_channels = lima_format_get_swap_rb(fb->base.cbufs[0]->format);
 683
 684    struct lima_pp_wb_reg *wb = (void *)wb_reg;
 685    wb[wb_idx].type = 0x02; /* 2 for color buffer */
 686    wb[wb_idx].address = res->bo->va + res->levels[level].offset + layer * res->levels[level].layer_stride;
 687    wb[wb_idx].pixel_format = format;
 688    if (res->tiled) {
 689       wb[wb_idx].pixel_layout = 0x2;
 690       wb[wb_idx].pitch = fb->tiled_w;
 691    } else {
 692       wb[wb_idx].pixel_layout = 0x0;
 693       wb[wb_idx].pitch = res->levels[level].stride / 8;
 694    }
 695    wb[wb_idx].mrt_bits = swap_channels ? 0x4 : 0x0;
 696 }
 697
 698 static void
 699 lima_pack_pp_frame_reg(struct lima_submit *submit, uint32_t *frame_reg,
 700                        uint32_t *wb_reg)
 701 {
 702    struct lima_context *ctx = submit->ctx;
 703    struct lima_context_framebuffer *fb = &ctx->framebuffer;
 704    struct lima_pp_frame_reg *frame = (void *)frame_reg;
 705    struct lima_screen *screen = lima_screen(ctx->base.screen);
 706    int wb_idx = 0;
 707
 708    frame->render_address = screen->pp_buffer->va + pp_frame_rsw_offset;
 709    frame->flags = 0x02;
 710    frame->clear_value_depth = ctx->clear.depth;
 711    frame->clear_value_stencil = ctx->clear.stencil;
 712    frame->clear_value_color = ctx->clear.color_8pc;
 713    frame->clear_value_color_1 = ctx->clear.color_8pc;
 714    frame->clear_value_color_2 = ctx->clear.color_8pc;
 715    frame->clear_value_color_3 = ctx->clear.color_8pc;
 716    frame->one = 1;
 717
 718    frame->width = fb->base.width - 1;
 719    frame->height = fb->base.height - 1;
 720
 721    /* frame->fragment_stack_address is overwritten per-pp in the kernel
 722     * by the values of pp_frame.fragment_stack_address[i] */
 723
 724    /* These are "stack size" and "stack offset" shifted,
 725     * here they are assumed to be always the same. */
 726    frame->fragment_stack_size = ctx->pp_max_stack_size << 16 | ctx->pp_max_stack_size;
 727
 728    /* related with MSAA and different value when r4p0/r7p0 */
 729    frame->supersampled_height = fb->base.height * 2 - 1;
 730    frame->scale = 0xE0C;
 731
 732    frame->dubya = 0x77;
 733    frame->onscreen = 1;
 734    frame->blocking = (fb->shift_min << 28) | (fb->shift_h << 16) | fb->shift_w;
 735    frame->foureight = 0x8888;
 736
 737    if (fb->base.nr_cbufs && (ctx->resolve & PIPE_CLEAR_COLOR0))
 738       lima_pack_wb_cbuf_reg(submit, wb_reg, wb_idx++);
 739
 740    if (fb->base.zsbuf &&
 741        (ctx->resolve & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)))
 742       lima_pack_wb_zsbuf_reg(submit, wb_reg, wb_idx++);
 743 }
 744
 745 static void
 746 lima_do_submit(struct lima_submit *submit)
 747 {
 748    #define pp_stack_pp_size 0x400
 749
 750    struct lima_context *ctx = submit->ctx;
 751
 752    lima_pack_head_plbu_cmd(submit);
 753    lima_finish_plbu_cmd(&ctx->plbu_cmd_array);
 754
 755    lima_update_submit_bo(submit);
 756
 757    int vs_cmd_size = ctx->vs_cmd_array.size;
 758    uint32_t vs_cmd_va = 0;
 759
 760    if (vs_cmd_size) {
 761       void *vs_cmd = lima_submit_create_stream_bo(
 762          submit, LIMA_PIPE_GP, vs_cmd_size, &vs_cmd_va);
 763       memcpy(vs_cmd, util_dynarray_begin(&ctx->vs_cmd_array), vs_cmd_size);
 764       util_dynarray_clear(&ctx->vs_cmd_array);
 765
 766       lima_dump_command_stream_print(
 767          vs_cmd, vs_cmd_size, false, "flush vs cmd at va %x\n", vs_cmd_va);
 768       lima_dump_vs_command_stream_print(vs_cmd, vs_cmd_size, vs_cmd_va);
 769    }
 770
 771    uint32_t plbu_cmd_va;
 772    int plbu_cmd_size = ctx->plbu_cmd_array.size + ctx->plbu_cmd_head.size;
 773    void *plbu_cmd = lima_submit_create_stream_bo(
 774       submit, LIMA_PIPE_GP, plbu_cmd_size, &plbu_cmd_va);
 775    memcpy(plbu_cmd,
 776           util_dynarray_begin(&ctx->plbu_cmd_head),
 777           ctx->plbu_cmd_head.size);
 778    memcpy(plbu_cmd + ctx->plbu_cmd_head.size,
 779           util_dynarray_begin(&ctx->plbu_cmd_array),
 780           ctx->plbu_cmd_array.size);
 781    util_dynarray_clear(&ctx->plbu_cmd_array);
 782    util_dynarray_clear(&ctx->plbu_cmd_head);
 783
 784    lima_dump_command_stream_print(
 785       plbu_cmd, plbu_cmd_size, false, "flush plbu cmd at va %x\n", plbu_cmd_va);
 786    lima_dump_plbu_command_stream_print(plbu_cmd, plbu_cmd_size, plbu_cmd_va);
 787
 788    struct lima_screen *screen = lima_screen(ctx->base.screen);
 789    struct drm_lima_gp_frame gp_frame;
 790    struct lima_gp_frame_reg *gp_frame_reg = (void *)gp_frame.frame;
 791    gp_frame_reg->vs_cmd_start = vs_cmd_va;
 792    gp_frame_reg->vs_cmd_end = vs_cmd_va + vs_cmd_size;
 793    gp_frame_reg->plbu_cmd_start = plbu_cmd_va;
 794    gp_frame_reg->plbu_cmd_end = plbu_cmd_va + plbu_cmd_size;
 795    gp_frame_reg->tile_heap_start = ctx->gp_tile_heap[ctx->plb_index]->va;
 796    gp_frame_reg->tile_heap_end = ctx->gp_tile_heap[ctx->plb_index]->va + ctx->gp_tile_heap_size;
 797
 798    lima_dump_command_stream_print(
 799       &gp_frame, sizeof(gp_frame), false, "add gp frame\n");
 800
 801    if (!lima_submit_start(submit, LIMA_PIPE_GP, &gp_frame, sizeof(gp_frame)))
 802       fprintf(stderr, "gp submit error\n");
 803
 804    if (lima_dump_command_stream) {
 805       if (lima_submit_wait(submit, LIMA_PIPE_GP, PIPE_TIMEOUT_INFINITE)) {
 806          if (ctx->gp_output) {
 807             float *pos = lima_bo_map(ctx->gp_output);
 808             lima_dump_command_stream_print(
 809                pos, 4 * 4 * 16, true, "gl_pos dump at va %x\n",
 810                ctx->gp_output->va);
 811          }
 812
 813          uint32_t *plb = lima_bo_map(ctx->plb[ctx->plb_index]);
 814          lima_dump_command_stream_print(
 815             plb, LIMA_CTX_PLB_BLK_SIZE, false, "plb dump at va %x\n",
 816             ctx->plb[ctx->plb_index]->va);
 817       }
 818       else {
 819          fprintf(stderr, "gp submit wait error\n");
 820          exit(1);
 821       }
 822    }
 823
 824    uint32_t pp_stack_va = 0;
 825    if (ctx->pp_max_stack_size) {
 826       lima_submit_create_stream_bo(
 827          submit, LIMA_PIPE_PP,
 828          screen->num_pp * ctx->pp_max_stack_size * pp_stack_pp_size,
 829          &pp_stack_va);
 830    }
 831
 832    lima_update_pp_stream(submit);
 833
 834    struct lima_pp_stream_state *ps = &ctx->pp_stream;
 835    if (screen->gpu_type == DRM_LIMA_PARAM_GPU_ID_MALI400) {
 836       struct drm_lima_m400_pp_frame pp_frame = {0};
 837       lima_pack_pp_frame_reg(submit, pp_frame.frame, pp_frame.wb);
 838       pp_frame.num_pp = screen->num_pp;
 839
 840       for (int i = 0; i < screen->num_pp; i++) {
 841          pp_frame.plbu_array_address[i] = ps->va + ps->offset[i];
 842          if (ctx->pp_max_stack_size)
 843             pp_frame.fragment_stack_address[i] = pp_stack_va +
 844                ctx->pp_max_stack_size * pp_stack_pp_size * i;
 845       }
 846
 847       lima_dump_command_stream_print(
 848          &pp_frame, sizeof(pp_frame), false, "add pp frame\n");
 849
 850       if (!lima_submit_start(submit, LIMA_PIPE_PP, &pp_frame, sizeof(pp_frame)))
 851          fprintf(stderr, "pp submit error\n");
 852    }
 853    else {
 854       struct drm_lima_m450_pp_frame pp_frame = {0};
 855       lima_pack_pp_frame_reg(submit, pp_frame.frame, pp_frame.wb);
 856       pp_frame.num_pp = screen->num_pp;
 857
 858       if (ctx->pp_max_stack_size)
 859          for (int i = 0; i < screen->num_pp; i++)
 860             pp_frame.fragment_stack_address[i] = pp_stack_va +
 861                ctx->pp_max_stack_size * pp_stack_pp_size * i;
 862
 863       if (ps->map) {
 864          for (int i = 0; i < screen->num_pp; i++)
 865             pp_frame.plbu_array_address[i] = ps->va + ps->offset[i];
 866       }
 867       else {
 868          pp_frame.use_dlbu = true;
 869
 870          struct lima_context_framebuffer *fb = &ctx->framebuffer;
 871          pp_frame.dlbu_regs[0] = ctx->plb[ctx->plb_index]->va;
 872          pp_frame.dlbu_regs[1] = ((fb->tiled_h - 1) << 16) | (fb->tiled_w - 1);
 873          unsigned s = util_logbase2(LIMA_CTX_PLB_BLK_SIZE) - 7;
 874          pp_frame.dlbu_regs[2] = (s << 28) | (fb->shift_h << 16) | fb->shift_w;
 875          pp_frame.dlbu_regs[3] = ((fb->tiled_h - 1) << 24) | ((fb->tiled_w - 1) << 16);
 876       }
 877
 878       lima_dump_command_stream_print(
 879          &pp_frame, sizeof(pp_frame), false, "add pp frame\n");
 880
 881       if (!lima_submit_start(submit, LIMA_PIPE_PP, &pp_frame, sizeof(pp_frame)))
 882          fprintf(stderr, "pp submit error\n");
 883    }
 884
 885    if (lima_dump_command_stream) {
 886       if (!lima_submit_wait(submit, LIMA_PIPE_PP, PIPE_TIMEOUT_INFINITE)) {
 887          fprintf(stderr, "pp wait error\n");
 888          exit(1);
 889       }
 890    }
 891
 892    ctx->plb_index = (ctx->plb_index + 1) % lima_ctx_num_plb;
 893
 894    if (ctx->framebuffer.base.nr_cbufs && (ctx->resolve & PIPE_CLEAR_COLOR0)) {
 895       /* Set reload flag for next draw. It'll be unset if buffer is cleared */
 896       struct lima_surface *surf = lima_surface(ctx->framebuffer.base.cbufs[0]);
 897       surf->reload = true;
 898    }
 899
 900    ctx->pp_max_stack_size = 0;
 901
 902    ctx->damage_rect.minx = ctx->damage_rect.miny = 0xffff;
 903    ctx->damage_rect.maxx = ctx->damage_rect.maxy = 0;
 904
 905    ctx->resolve = 0;
 906
 907    lima_dump_file_next();
 908 }
 909
 910 void
 911 lima_flush(struct lima_context *ctx)
 912 {
 913    if (!lima_submit_dirty(ctx->submit))
 914       return;
 915
 916    lima_do_submit(ctx->submit);
 917 }
 918
 919 static void
 920 lima_pipe_flush(struct pipe_context *pctx, struct pipe_fence_handle **fence,
 921                 unsigned flags)
 922 {
 923    struct lima_context *ctx = lima_context(pctx);
 924    if (lima_submit_dirty(ctx->submit))
 925       lima_do_submit(ctx->submit);
 926
 927    if (fence) {
 928       int drm_fd = lima_screen(ctx->base.screen)->fd;
 929       int fd;
 930
 931       if (!drmSyncobjExportSyncFile(drm_fd, ctx->out_sync[LIMA_PIPE_PP], &fd))
 932          *fence = lima_fence_create(fd);
 933    }
 934 }
 935
 936 bool lima_submit_init(struct lima_context *ctx)
 937 {
 938    int fd = lima_screen(ctx->base.screen)->fd;
 939
 940    ctx->submit = lima_submit_create(ctx);
 941    if (!ctx->submit)
 942       return false;
 943
 944    ctx->in_sync_fd = -1;
 945
 946    for (int i = 0; i < 2; i++) {
 947       if (drmSyncobjCreate(fd, DRM_SYNCOBJ_CREATE_SIGNALED, ctx->in_sync + i) ||
 948           drmSyncobjCreate(fd, DRM_SYNCOBJ_CREATE_SIGNALED, ctx->out_sync + i))
 949          return false;
 950    }
 951
 952    ctx->base.flush = lima_pipe_flush;
 953
 954    return true;
 955 }
 956
 957 void lima_submit_fini(struct lima_context *ctx)
 958 {
 959    int fd = lima_screen(ctx->base.screen)->fd;
 960
 961    for (int i = 0; i < 2; i++) {
 962       if (ctx->in_sync[i])
 963          drmSyncobjDestroy(fd, ctx->in_sync[i]);
 964       if (ctx->out_sync[i])
 965          drmSyncobjDestroy(fd, ctx->out_sync[i]);
 966    }
 967
 968    if (ctx->in_sync_fd >= 0)
 969       close(ctx->in_sync_fd);
 970
 971    if (ctx->submit)
 972       lima_submit_free(ctx->submit);
 973 }