src/freedreno/vulkan/tu_cmd_buffer.c

   1 /*
   2  * Copyright © 2016 Red Hat.
   3  * Copyright © 2016 Bas Nieuwenhuizen
   4  *
   5  * based in part on anv driver which is:
   6  * Copyright © 2015 Intel Corporation
   7  *
   8  * Permission is hereby granted, free of charge, to any person obtaining a
   9  * copy of this software and associated documentation files (the "Software"),
  10  * to deal in the Software without restriction, including without limitation
  11  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  12  * and/or sell copies of the Software, and to permit persons to whom the
  13  * Software is furnished to do so, subject to the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the next
  16  * paragraph) shall be included in all copies or substantial portions of the
  17  * Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  21  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  22  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  23  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  24  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  25  * DEALINGS IN THE SOFTWARE.
  26  */
  27
  28 #include "tu_private.h"
  29
  30 #include "registers/adreno_pm4.xml.h"
  31 #include "registers/adreno_common.xml.h"
  32
  33 #include "vk_format.h"
  34
  35 #include "tu_cs.h"
  36 #include "tu_blit.h"
  37
  38 #define OVERFLOW_FLAG_REG REG_A6XX_CP_SCRATCH_REG(0)
  39
  40 void
  41 tu_bo_list_init(struct tu_bo_list *list)
  42 {
  43    list->count = list->capacity = 0;
  44    list->bo_infos = NULL;
  45 }
  46
  47 void
  48 tu_bo_list_destroy(struct tu_bo_list *list)
  49 {
  50    free(list->bo_infos);
  51 }
  52
  53 void
  54 tu_bo_list_reset(struct tu_bo_list *list)
  55 {
  56    list->count = 0;
  57 }
  58
  59 /**
  60  * \a flags consists of MSM_SUBMIT_BO_FLAGS.
  61  */
  62 static uint32_t
  63 tu_bo_list_add_info(struct tu_bo_list *list,
  64                     const struct drm_msm_gem_submit_bo *bo_info)
  65 {
  66    assert(bo_info->handle != 0);
  67
  68    for (uint32_t i = 0; i < list->count; ++i) {
  69       if (list->bo_infos[i].handle == bo_info->handle) {
  70          assert(list->bo_infos[i].presumed == bo_info->presumed);
  71          list->bo_infos[i].flags |= bo_info->flags;
  72          return i;
  73       }
  74    }
  75
  76    /* grow list->bo_infos if needed */
  77    if (list->count == list->capacity) {
  78       uint32_t new_capacity = MAX2(2 * list->count, 16);
  79       struct drm_msm_gem_submit_bo *new_bo_infos = realloc(
  80          list->bo_infos, new_capacity * sizeof(struct drm_msm_gem_submit_bo));
  81       if (!new_bo_infos)
  82          return TU_BO_LIST_FAILED;
  83       list->bo_infos = new_bo_infos;
  84       list->capacity = new_capacity;
  85    }
  86
  87    list->bo_infos[list->count] = *bo_info;
  88    return list->count++;
  89 }
  90
  91 uint32_t
  92 tu_bo_list_add(struct tu_bo_list *list,
  93                const struct tu_bo *bo,
  94                uint32_t flags)
  95 {
  96    return tu_bo_list_add_info(list, &(struct drm_msm_gem_submit_bo) {
  97                                        .flags = flags,
  98                                        .handle = bo->gem_handle,
  99                                        .presumed = bo->iova,
 100                                     });
 101 }
 102
 103 VkResult
 104 tu_bo_list_merge(struct tu_bo_list *list, const struct tu_bo_list *other)
 105 {
 106    for (uint32_t i = 0; i < other->count; i++) {
 107       if (tu_bo_list_add_info(list, other->bo_infos + i) == TU_BO_LIST_FAILED)
 108          return VK_ERROR_OUT_OF_HOST_MEMORY;
 109    }
 110
 111    return VK_SUCCESS;
 112 }
 113
 114 static bool
 115 is_linear_mipmapped(const struct tu_image_view *iview)
 116 {
 117    return iview->image->layout.tile_mode == TILE6_LINEAR &&
 118           iview->base_mip != iview->image->level_count - 1;
 119 }
 120
 121 static bool
 122 force_sysmem(const struct tu_cmd_buffer *cmd,
 123              const struct VkRect2D *render_area)
 124 {
 125    const struct tu_framebuffer *fb = cmd->state.framebuffer;
 126    const struct tu_physical_device *device = cmd->device->physical_device;
 127    bool has_linear_mipmapped_store = false;
 128    const struct tu_render_pass *pass = cmd->state.pass;
 129
 130    /* Iterate over all the places we call tu6_emit_store_attachment() */
 131    for (unsigned i = 0; i < pass->subpass_count; i++) {
 132       const struct tu_subpass *subpass = &pass->subpasses[i];
 133       if (subpass->resolve_attachments) {
 134          for (unsigned i = 0; i < subpass->color_count; i++) {
 135             uint32_t a = subpass->resolve_attachments[i].attachment;
 136             if (a != VK_ATTACHMENT_UNUSED &&
 137                 cmd->state.pass->attachments[a].store_op == VK_ATTACHMENT_STORE_OP_STORE) {
 138                const struct tu_image_view *iview = fb->attachments[a].attachment;
 139                if (is_linear_mipmapped(iview)) {
 140                   has_linear_mipmapped_store = true;
 141                   break;
 142                }
 143             }
 144          }
 145       }
 146    }
 147
 148    for (unsigned i = 0; i < pass->attachment_count; i++) {
 149       if (pass->attachments[i].gmem_offset >= 0 &&
 150           cmd->state.pass->attachments[i].store_op == VK_ATTACHMENT_STORE_OP_STORE) {
 151          const struct tu_image_view *iview = fb->attachments[i].attachment;
 152          if (is_linear_mipmapped(iview)) {
 153             has_linear_mipmapped_store = true;
 154             break;
 155          }
 156       }
 157    }
 158
 159    /* Linear textures cannot have any padding between mipmap levels and their
 160     * height isn't padded, while at the same time the GMEM->MEM resolve does
 161     * not have per-pixel granularity, so if the image height isn't aligned to
 162     * the resolve granularity and the render area is tall enough, we may wind
 163     * up writing past the bottom of the image into the next miplevel or even
 164     * past the end of the image. For the last miplevel, the layout code should
 165     * insert enough padding so that the overdraw writes to the padding.  To
 166     * work around this, we force-enable sysmem rendering.
 167     */
 168    const uint32_t y2 = render_area->offset.y + render_area->extent.height;
 169    const uint32_t aligned_y2 = ALIGN_POT(y2, device->tile_align_h);
 170
 171    return has_linear_mipmapped_store && aligned_y2 > fb->height;
 172 }
 173
 174 static void
 175 tu_tiling_config_update_tile_layout(struct tu_tiling_config *tiling,
 176                                     const struct tu_device *dev,
 177                                     uint32_t pixels)
 178 {
 179    const uint32_t tile_align_w = dev->physical_device->tile_align_w;
 180    const uint32_t tile_align_h = dev->physical_device->tile_align_h;
 181    const uint32_t max_tile_width = 1024; /* A6xx */
 182
 183    /* note: don't offset the tiling config by render_area.offset,
 184     * because binning pass can't deal with it
 185     * this means we might end up with more tiles than necessary,
 186     * but load/store/etc are still scissored to the render_area
 187     */
 188    tiling->tile0.offset = (VkOffset2D) {};
 189
 190    const uint32_t ra_width =
 191       tiling->render_area.extent.width +
 192       (tiling->render_area.offset.x - tiling->tile0.offset.x);
 193    const uint32_t ra_height =
 194       tiling->render_area.extent.height +
 195       (tiling->render_area.offset.y - tiling->tile0.offset.y);
 196
 197    /* start from 1 tile */
 198    tiling->tile_count = (VkExtent2D) {
 199       .width = 1,
 200       .height = 1,
 201    };
 202    tiling->tile0.extent = (VkExtent2D) {
 203       .width = align(ra_width, tile_align_w),
 204       .height = align(ra_height, tile_align_h),
 205    };
 206
 207    if (unlikely(dev->physical_device->instance->debug_flags & TU_DEBUG_FORCEBIN)) {
 208       /* start with 2x2 tiles */
 209       tiling->tile_count.width = 2;
 210       tiling->tile_count.height = 2;
 211       tiling->tile0.extent.width = align(DIV_ROUND_UP(ra_width, 2), tile_align_w);
 212       tiling->tile0.extent.height = align(DIV_ROUND_UP(ra_height, 2), tile_align_h);
 213    }
 214
 215    /* do not exceed max tile width */
 216    while (tiling->tile0.extent.width > max_tile_width) {
 217       tiling->tile_count.width++;
 218       tiling->tile0.extent.width =
 219          align(DIV_ROUND_UP(ra_width, tiling->tile_count.width), tile_align_w);
 220    }
 221
 222    /* will force to sysmem, don't bother trying to have a valid tile config
 223     * TODO: just skip all GMEM stuff when sysmem is forced?
 224     */
 225    if (!pixels)
 226       return;
 227
 228    /* do not exceed gmem size */
 229    while (tiling->tile0.extent.width * tiling->tile0.extent.height > pixels) {
 230       if (tiling->tile0.extent.width > MAX2(tile_align_w, tiling->tile0.extent.height)) {
 231          tiling->tile_count.width++;
 232          tiling->tile0.extent.width =
 233             align(DIV_ROUND_UP(ra_width, tiling->tile_count.width), tile_align_w);
 234       } else {
 235          /* if this assert fails then layout is impossible.. */
 236          assert(tiling->tile0.extent.height > tile_align_h);
 237          tiling->tile_count.height++;
 238          tiling->tile0.extent.height =
 239             align(DIV_ROUND_UP(ra_height, tiling->tile_count.height), tile_align_h);
 240       }
 241    }
 242 }
 243
 244 static void
 245 tu_tiling_config_update_pipe_layout(struct tu_tiling_config *tiling,
 246                                     const struct tu_device *dev)
 247 {
 248    const uint32_t max_pipe_count = 32; /* A6xx */
 249
 250    /* start from 1 tile per pipe */
 251    tiling->pipe0 = (VkExtent2D) {
 252       .width = 1,
 253       .height = 1,
 254    };
 255    tiling->pipe_count = tiling->tile_count;
 256
 257    while (tiling->pipe_count.width * tiling->pipe_count.height > max_pipe_count) {
 258       if (tiling->pipe0.width < tiling->pipe0.height) {
 259          tiling->pipe0.width += 1;
 260          tiling->pipe_count.width =
 261             DIV_ROUND_UP(tiling->tile_count.width, tiling->pipe0.width);
 262       } else {
 263          tiling->pipe0.height += 1;
 264          tiling->pipe_count.height =
 265             DIV_ROUND_UP(tiling->tile_count.height, tiling->pipe0.height);
 266       }
 267    }
 268 }
 269
 270 static void
 271 tu_tiling_config_update_pipes(struct tu_tiling_config *tiling,
 272                               const struct tu_device *dev)
 273 {
 274    const uint32_t max_pipe_count = 32; /* A6xx */
 275    const uint32_t used_pipe_count =
 276       tiling->pipe_count.width * tiling->pipe_count.height;
 277    const VkExtent2D last_pipe = {
 278       .width = (tiling->tile_count.width - 1) % tiling->pipe0.width + 1,
 279       .height = (tiling->tile_count.height - 1) % tiling->pipe0.height + 1,
 280    };
 281
 282    assert(used_pipe_count <= max_pipe_count);
 283    assert(max_pipe_count <= ARRAY_SIZE(tiling->pipe_config));
 284
 285    for (uint32_t y = 0; y < tiling->pipe_count.height; y++) {
 286       for (uint32_t x = 0; x < tiling->pipe_count.width; x++) {
 287          const uint32_t pipe_x = tiling->pipe0.width * x;
 288          const uint32_t pipe_y = tiling->pipe0.height * y;
 289          const uint32_t pipe_w = (x == tiling->pipe_count.width - 1)
 290                                     ? last_pipe.width
 291                                     : tiling->pipe0.width;
 292          const uint32_t pipe_h = (y == tiling->pipe_count.height - 1)
 293                                     ? last_pipe.height
 294                                     : tiling->pipe0.height;
 295          const uint32_t n = tiling->pipe_count.width * y + x;
 296
 297          tiling->pipe_config[n] = A6XX_VSC_PIPE_CONFIG_REG_X(pipe_x) |
 298                                   A6XX_VSC_PIPE_CONFIG_REG_Y(pipe_y) |
 299                                   A6XX_VSC_PIPE_CONFIG_REG_W(pipe_w) |
 300                                   A6XX_VSC_PIPE_CONFIG_REG_H(pipe_h);
 301          tiling->pipe_sizes[n] = CP_SET_BIN_DATA5_0_VSC_SIZE(pipe_w * pipe_h);
 302       }
 303    }
 304
 305    memset(tiling->pipe_config + used_pipe_count, 0,
 306           sizeof(uint32_t) * (max_pipe_count - used_pipe_count));
 307 }
 308
 309 static void
 310 tu_tiling_config_get_tile(const struct tu_tiling_config *tiling,
 311                           const struct tu_device *dev,
 312                           uint32_t tx,
 313                           uint32_t ty,
 314                           struct tu_tile *tile)
 315 {
 316    /* find the pipe and the slot for tile (tx, ty) */
 317    const uint32_t px = tx / tiling->pipe0.width;
 318    const uint32_t py = ty / tiling->pipe0.height;
 319    const uint32_t sx = tx - tiling->pipe0.width * px;
 320    const uint32_t sy = ty - tiling->pipe0.height * py;
 321    /* last pipe has different width */
 322    const uint32_t pipe_width =
 323       MIN2(tiling->pipe0.width,
 324            tiling->tile_count.width - px * tiling->pipe0.width);
 325
 326    assert(tx < tiling->tile_count.width && ty < tiling->tile_count.height);
 327    assert(px < tiling->pipe_count.width && py < tiling->pipe_count.height);
 328    assert(sx < tiling->pipe0.width && sy < tiling->pipe0.height);
 329
 330    /* convert to 1D indices */
 331    tile->pipe = tiling->pipe_count.width * py + px;
 332    tile->slot = pipe_width * sy + sx;
 333
 334    /* get the blit area for the tile */
 335    tile->begin = (VkOffset2D) {
 336       .x = tiling->tile0.offset.x + tiling->tile0.extent.width * tx,
 337       .y = tiling->tile0.offset.y + tiling->tile0.extent.height * ty,
 338    };
 339    tile->end.x =
 340       (tx == tiling->tile_count.width - 1)
 341          ? tiling->render_area.offset.x + tiling->render_area.extent.width
 342          : tile->begin.x + tiling->tile0.extent.width;
 343    tile->end.y =
 344       (ty == tiling->tile_count.height - 1)
 345          ? tiling->render_area.offset.y + tiling->render_area.extent.height
 346          : tile->begin.y + tiling->tile0.extent.height;
 347 }
 348
 349 enum a3xx_msaa_samples
 350 tu_msaa_samples(uint32_t samples)
 351 {
 352    switch (samples) {
 353    case 1:
 354       return MSAA_ONE;
 355    case 2:
 356       return MSAA_TWO;
 357    case 4:
 358       return MSAA_FOUR;
 359    case 8:
 360       return MSAA_EIGHT;
 361    default:
 362       assert(!"invalid sample count");
 363       return MSAA_ONE;
 364    }
 365 }
 366
 367 static enum a4xx_index_size
 368 tu6_index_size(VkIndexType type)
 369 {
 370    switch (type) {
 371    case VK_INDEX_TYPE_UINT16:
 372       return INDEX4_SIZE_16_BIT;
 373    case VK_INDEX_TYPE_UINT32:
 374       return INDEX4_SIZE_32_BIT;
 375    default:
 376       unreachable("invalid VkIndexType");
 377       return INDEX4_SIZE_8_BIT;
 378    }
 379 }
 380
 381 unsigned
 382 tu6_emit_event_write(struct tu_cmd_buffer *cmd,
 383                      struct tu_cs *cs,
 384                      enum vgt_event_type event,
 385                      bool need_seqno)
 386 {
 387    unsigned seqno = 0;
 388
 389    tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, need_seqno ? 4 : 1);
 390    tu_cs_emit(cs, CP_EVENT_WRITE_0_EVENT(event));
 391    if (need_seqno) {
 392       tu_cs_emit_qw(cs, cmd->scratch_bo.iova);
 393       seqno = ++cmd->scratch_seqno;
 394       tu_cs_emit(cs, seqno);
 395    }
 396
 397    return seqno;
 398 }
 399
 400 static void
 401 tu6_emit_cache_flush(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 402 {
 403    tu6_emit_event_write(cmd, cs, 0x31, false);
 404 }
 405
 406 static void
 407 tu6_emit_lrz_flush(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 408 {
 409    tu6_emit_event_write(cmd, cs, LRZ_FLUSH, false);
 410 }
 411
 412 static void
 413 tu6_emit_wfi(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 414 {
 415    if (cmd->wait_for_idle) {
 416       tu_cs_emit_wfi(cs);
 417       cmd->wait_for_idle = false;
 418    }
 419 }
 420
 421 #define tu_image_view_ubwc_pitches(iview)                                \
 422    .pitch = tu_image_ubwc_pitch(iview->image, iview->base_mip),          \
 423    .array_pitch = tu_image_ubwc_size(iview->image, iview->base_mip) >> 2
 424
 425 static void
 426 tu6_emit_zs(struct tu_cmd_buffer *cmd,
 427             const struct tu_subpass *subpass,
 428             struct tu_cs *cs)
 429 {
 430    const struct tu_framebuffer *fb = cmd->state.framebuffer;
 431
 432    const uint32_t a = subpass->depth_stencil_attachment.attachment;
 433    if (a == VK_ATTACHMENT_UNUSED) {
 434       tu_cs_emit_regs(cs,
 435                       A6XX_RB_DEPTH_BUFFER_INFO(.depth_format = DEPTH6_NONE),
 436                       A6XX_RB_DEPTH_BUFFER_PITCH(0),
 437                       A6XX_RB_DEPTH_BUFFER_ARRAY_PITCH(0),
 438                       A6XX_RB_DEPTH_BUFFER_BASE(0),
 439                       A6XX_RB_DEPTH_BUFFER_BASE_GMEM(0));
 440
 441       tu_cs_emit_regs(cs,
 442                       A6XX_GRAS_SU_DEPTH_BUFFER_INFO(.depth_format = DEPTH6_NONE));
 443
 444       tu_cs_emit_regs(cs,
 445                       A6XX_GRAS_LRZ_BUFFER_BASE(0),
 446                       A6XX_GRAS_LRZ_BUFFER_PITCH(0),
 447                       A6XX_GRAS_LRZ_FAST_CLEAR_BUFFER_BASE(0));
 448
 449       tu_cs_emit_regs(cs, A6XX_RB_STENCIL_INFO(0));
 450
 451       return;
 452    }
 453
 454    const struct tu_image_view *iview = fb->attachments[a].attachment;
 455    enum a6xx_depth_format fmt = tu6_pipe2depth(iview->vk_format);
 456
 457    tu_cs_emit_regs(cs,
 458                    A6XX_RB_DEPTH_BUFFER_INFO(.depth_format = fmt),
 459                    A6XX_RB_DEPTH_BUFFER_PITCH(tu_image_stride(iview->image, iview->base_mip)),
 460                    A6XX_RB_DEPTH_BUFFER_ARRAY_PITCH(iview->image->layout.layer_size),
 461                    A6XX_RB_DEPTH_BUFFER_BASE(tu_image_view_base_ref(iview)),
 462                    A6XX_RB_DEPTH_BUFFER_BASE_GMEM(cmd->state.pass->attachments[a].gmem_offset));
 463
 464    tu_cs_emit_regs(cs,
 465                    A6XX_GRAS_SU_DEPTH_BUFFER_INFO(.depth_format = fmt));
 466
 467    tu_cs_emit_regs(cs,
 468                    A6XX_RB_DEPTH_FLAG_BUFFER_BASE(tu_image_view_ubwc_base_ref(iview)),
 469                    A6XX_RB_DEPTH_FLAG_BUFFER_PITCH(tu_image_view_ubwc_pitches(iview)));
 470
 471    tu_cs_emit_regs(cs,
 472                    A6XX_GRAS_LRZ_BUFFER_BASE(0),
 473                    A6XX_GRAS_LRZ_BUFFER_PITCH(0),
 474                    A6XX_GRAS_LRZ_FAST_CLEAR_BUFFER_BASE(0));
 475
 476    tu_cs_emit_regs(cs,
 477                    A6XX_RB_STENCIL_INFO(0));
 478
 479    /* enable zs? */
 480 }
 481
 482 static void
 483 tu6_emit_mrt(struct tu_cmd_buffer *cmd,
 484              const struct tu_subpass *subpass,
 485              struct tu_cs *cs)
 486 {
 487    const struct tu_framebuffer *fb = cmd->state.framebuffer;
 488    unsigned char mrt_comp[MAX_RTS] = { 0 };
 489    unsigned srgb_cntl = 0;
 490
 491    for (uint32_t i = 0; i < subpass->color_count; ++i) {
 492       uint32_t a = subpass->color_attachments[i].attachment;
 493       if (a == VK_ATTACHMENT_UNUSED)
 494          continue;
 495
 496       const struct tu_image_view *iview = fb->attachments[a].attachment;
 497       const enum a6xx_tile_mode tile_mode =
 498          tu6_get_image_tile_mode(iview->image, iview->base_mip);
 499
 500       mrt_comp[i] = 0xf;
 501
 502       if (vk_format_is_srgb(iview->vk_format))
 503          srgb_cntl |= (1 << i);
 504
 505       const struct tu_native_format format =
 506          tu6_format_color(iview->vk_format, iview->image->layout.tile_mode);
 507
 508       tu_cs_emit_regs(cs,
 509                       A6XX_RB_MRT_BUF_INFO(i,
 510                                            .color_tile_mode = tile_mode,
 511                                            .color_format = format.fmt,
 512                                            .color_swap = format.swap),
 513                       A6XX_RB_MRT_PITCH(i, tu_image_stride(iview->image, iview->base_mip)),
 514                       A6XX_RB_MRT_ARRAY_PITCH(i, iview->image->layout.layer_size),
 515                       A6XX_RB_MRT_BASE(i, tu_image_view_base_ref(iview)),
 516                       A6XX_RB_MRT_BASE_GMEM(i, cmd->state.pass->attachments[a].gmem_offset));
 517
 518       tu_cs_emit_regs(cs,
 519                       A6XX_SP_FS_MRT_REG(i,
 520                                          .color_format = format.fmt,
 521                                          .color_sint = vk_format_is_sint(iview->vk_format),
 522                                          .color_uint = vk_format_is_uint(iview->vk_format)));
 523
 524       tu_cs_emit_regs(cs,
 525                       A6XX_RB_MRT_FLAG_BUFFER_ADDR(i, tu_image_view_ubwc_base_ref(iview)),
 526                       A6XX_RB_MRT_FLAG_BUFFER_PITCH(i, tu_image_view_ubwc_pitches(iview)));
 527    }
 528
 529    tu_cs_emit_regs(cs,
 530                    A6XX_RB_SRGB_CNTL(.dword = srgb_cntl));
 531
 532    tu_cs_emit_regs(cs,
 533                    A6XX_SP_SRGB_CNTL(.dword = srgb_cntl));
 534
 535    tu_cs_emit_regs(cs,
 536                    A6XX_RB_RENDER_COMPONENTS(
 537                       .rt0 = mrt_comp[0],
 538                       .rt1 = mrt_comp[1],
 539                       .rt2 = mrt_comp[2],
 540                       .rt3 = mrt_comp[3],
 541                       .rt4 = mrt_comp[4],
 542                       .rt5 = mrt_comp[5],
 543                       .rt6 = mrt_comp[6],
 544                       .rt7 = mrt_comp[7]));
 545
 546    tu_cs_emit_regs(cs,
 547                    A6XX_SP_FS_RENDER_COMPONENTS(
 548                       .rt0 = mrt_comp[0],
 549                       .rt1 = mrt_comp[1],
 550                       .rt2 = mrt_comp[2],
 551                       .rt3 = mrt_comp[3],
 552                       .rt4 = mrt_comp[4],
 553                       .rt5 = mrt_comp[5],
 554                       .rt6 = mrt_comp[6],
 555                       .rt7 = mrt_comp[7]));
 556 }
 557
 558 static void
 559 tu6_emit_msaa(struct tu_cmd_buffer *cmd,
 560               const struct tu_subpass *subpass,
 561               struct tu_cs *cs)
 562 {
 563    const enum a3xx_msaa_samples samples = tu_msaa_samples(subpass->samples);
 564    bool msaa_disable = samples == MSAA_ONE;
 565
 566    tu_cs_emit_regs(cs,
 567                    A6XX_SP_TP_RAS_MSAA_CNTL(samples),
 568                    A6XX_SP_TP_DEST_MSAA_CNTL(.samples = samples,
 569                                              .msaa_disable = msaa_disable));
 570
 571    tu_cs_emit_regs(cs,
 572                    A6XX_GRAS_RAS_MSAA_CNTL(samples),
 573                    A6XX_GRAS_DEST_MSAA_CNTL(.samples = samples,
 574                                             .msaa_disable = msaa_disable));
 575
 576    tu_cs_emit_regs(cs,
 577                    A6XX_RB_RAS_MSAA_CNTL(samples),
 578                    A6XX_RB_DEST_MSAA_CNTL(.samples = samples,
 579                                           .msaa_disable = msaa_disable));
 580
 581    tu_cs_emit_regs(cs,
 582                    A6XX_RB_MSAA_CNTL(samples));
 583 }
 584
 585 static void
 586 tu6_emit_bin_size(struct tu_cs *cs,
 587                   uint32_t bin_w, uint32_t bin_h, uint32_t flags)
 588 {
 589    tu_cs_emit_regs(cs,
 590                    A6XX_GRAS_BIN_CONTROL(.binw = bin_w,
 591                                          .binh = bin_h,
 592                                          .dword = flags));
 593
 594    tu_cs_emit_regs(cs,
 595                    A6XX_RB_BIN_CONTROL(.binw = bin_w,
 596                                        .binh = bin_h,
 597                                        .dword = flags));
 598
 599    /* no flag for RB_BIN_CONTROL2... */
 600    tu_cs_emit_regs(cs,
 601                    A6XX_RB_BIN_CONTROL2(.binw = bin_w,
 602                                         .binh = bin_h));
 603 }
 604
 605 static void
 606 tu6_emit_render_cntl(struct tu_cmd_buffer *cmd,
 607                      const struct tu_subpass *subpass,
 608                      struct tu_cs *cs,
 609                      bool binning)
 610 {
 611    const struct tu_framebuffer *fb = cmd->state.framebuffer;
 612    uint32_t cntl = 0;
 613    cntl |= A6XX_RB_RENDER_CNTL_UNK4;
 614    if (binning) {
 615       cntl |= A6XX_RB_RENDER_CNTL_BINNING;
 616    } else {
 617       uint32_t mrts_ubwc_enable = 0;
 618       for (uint32_t i = 0; i < subpass->color_count; ++i) {
 619          uint32_t a = subpass->color_attachments[i].attachment;
 620          if (a == VK_ATTACHMENT_UNUSED)
 621             continue;
 622
 623          const struct tu_image_view *iview = fb->attachments[a].attachment;
 624          if (iview->image->layout.ubwc_layer_size != 0)
 625             mrts_ubwc_enable |= 1 << i;
 626       }
 627
 628       cntl |= A6XX_RB_RENDER_CNTL_FLAG_MRTS(mrts_ubwc_enable);
 629
 630       const uint32_t a = subpass->depth_stencil_attachment.attachment;
 631       if (a != VK_ATTACHMENT_UNUSED) {
 632          const struct tu_image_view *iview = fb->attachments[a].attachment;
 633          if (iview->image->layout.ubwc_layer_size != 0)
 634             cntl |= A6XX_RB_RENDER_CNTL_FLAG_DEPTH;
 635       }
 636
 637       /* In the !binning case, we need to set RB_RENDER_CNTL in the draw_cs
 638        * in order to set it correctly for the different subpasses. However,
 639        * that means the packets we're emitting also happen during binning. So
 640        * we need to guard the write on !BINNING at CP execution time.
 641        */
 642       tu_cs_reserve(cs, 3 + 4);
 643       tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
 644       tu_cs_emit(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
 645                      CP_COND_REG_EXEC_0_GMEM | CP_COND_REG_EXEC_0_SYSMEM);
 646       tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(4));
 647    }
 648
 649    tu_cs_emit_pkt7(cs, CP_REG_WRITE, 3);
 650    tu_cs_emit(cs, CP_REG_WRITE_0_TRACKER(TRACK_RENDER_CNTL));
 651    tu_cs_emit(cs, REG_A6XX_RB_RENDER_CNTL);
 652    tu_cs_emit(cs, cntl);
 653 }
 654
 655 static void
 656 tu6_emit_blit_scissor(struct tu_cmd_buffer *cmd, struct tu_cs *cs, bool align)
 657 {
 658    const VkRect2D *render_area = &cmd->state.tiling_config.render_area;
 659    uint32_t x1 = render_area->offset.x;
 660    uint32_t y1 = render_area->offset.y;
 661    uint32_t x2 = x1 + render_area->extent.width - 1;
 662    uint32_t y2 = y1 + render_area->extent.height - 1;
 663
 664    /* TODO: alignment requirement seems to be less than tile_align_w/h */
 665    if (align) {
 666       x1 = x1 & ~cmd->device->physical_device->tile_align_w;
 667       y1 = y1 & ~cmd->device->physical_device->tile_align_h;
 668       x2 = ALIGN_POT(x2 + 1, cmd->device->physical_device->tile_align_w) - 1;
 669       y2 = ALIGN_POT(y2 + 1, cmd->device->physical_device->tile_align_h) - 1;
 670    }
 671
 672    tu_cs_emit_regs(cs,
 673                    A6XX_RB_BLIT_SCISSOR_TL(.x = x1, .y = y1),
 674                    A6XX_RB_BLIT_SCISSOR_BR(.x = x2, .y = y2));
 675 }
 676
 677 static void
 678 tu6_emit_blit_info(struct tu_cmd_buffer *cmd,
 679                    struct tu_cs *cs,
 680                    const struct tu_image_view *iview,
 681                    uint32_t gmem_offset,
 682                    bool resolve)
 683 {
 684    tu_cs_emit_regs(cs,
 685                    A6XX_RB_BLIT_INFO(.unk0 = !resolve, .gmem = !resolve));
 686
 687    const struct tu_native_format format =
 688       tu6_format_color(iview->vk_format, iview->image->layout.tile_mode);
 689
 690    enum a6xx_tile_mode tile_mode =
 691       tu6_get_image_tile_mode(iview->image, iview->base_mip);
 692    tu_cs_emit_regs(cs,
 693                    A6XX_RB_BLIT_DST_INFO(
 694                       .tile_mode = tile_mode,
 695                       .samples = tu_msaa_samples(iview->image->samples),
 696                       .color_format = format.fmt,
 697                       .color_swap = format.swap,
 698                       .flags = iview->image->layout.ubwc_layer_size != 0),
 699                    A6XX_RB_BLIT_DST(tu_image_view_base_ref(iview)),
 700                    A6XX_RB_BLIT_DST_PITCH(tu_image_stride(iview->image, iview->base_mip)),
 701                    A6XX_RB_BLIT_DST_ARRAY_PITCH(iview->image->layout.layer_size));
 702
 703    if (iview->image->layout.ubwc_layer_size) {
 704       tu_cs_emit_regs(cs,
 705                       A6XX_RB_BLIT_FLAG_DST(tu_image_view_ubwc_base_ref(iview)),
 706                       A6XX_RB_BLIT_FLAG_DST_PITCH(tu_image_view_ubwc_pitches(iview)));
 707    }
 708
 709    tu_cs_emit_regs(cs,
 710                    A6XX_RB_BLIT_BASE_GMEM(gmem_offset));
 711 }
 712
 713 static void
 714 tu6_emit_blit(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 715 {
 716    tu6_emit_event_write(cmd, cs, BLIT, false);
 717 }
 718
 719 static void
 720 tu6_emit_window_scissor(struct tu_cmd_buffer *cmd,
 721                         struct tu_cs *cs,
 722                         uint32_t x1,
 723                         uint32_t y1,
 724                         uint32_t x2,
 725                         uint32_t y2)
 726 {
 727    tu_cs_emit_regs(cs,
 728                    A6XX_GRAS_SC_WINDOW_SCISSOR_TL(.x = x1, .y = y1),
 729                    A6XX_GRAS_SC_WINDOW_SCISSOR_BR(.x = x2, .y = y2));
 730
 731    tu_cs_emit_regs(cs,
 732                    A6XX_GRAS_RESOLVE_CNTL_1(.x = x1, .y = y1),
 733                    A6XX_GRAS_RESOLVE_CNTL_2(.x = x2, .y = y2));
 734 }
 735
 736 static void
 737 tu6_emit_window_offset(struct tu_cmd_buffer *cmd,
 738                        struct tu_cs *cs,
 739                        uint32_t x1,
 740                        uint32_t y1)
 741 {
 742    tu_cs_emit_regs(cs,
 743                    A6XX_RB_WINDOW_OFFSET(.x = x1, .y = y1));
 744
 745    tu_cs_emit_regs(cs,
 746                    A6XX_RB_WINDOW_OFFSET2(.x = x1, .y = y1));
 747
 748    tu_cs_emit_regs(cs,
 749                    A6XX_SP_WINDOW_OFFSET(.x = x1, .y = y1));
 750
 751    tu_cs_emit_regs(cs,
 752                    A6XX_SP_TP_WINDOW_OFFSET(.x = x1, .y = y1));
 753 }
 754
 755 static bool
 756 use_hw_binning(struct tu_cmd_buffer *cmd)
 757 {
 758    const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
 759
 760    if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_NOBIN))
 761       return false;
 762
 763    if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_FORCEBIN))
 764       return true;
 765
 766    return (tiling->tile_count.width * tiling->tile_count.height) > 2;
 767 }
 768
 769 static bool
 770 use_sysmem_rendering(struct tu_cmd_buffer *cmd)
 771 {
 772    if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_SYSMEM))
 773       return true;
 774
 775    /* can't fit attachments into gmem */
 776    if (!cmd->state.pass->gmem_pixels)
 777       return true;
 778
 779    return cmd->state.tiling_config.force_sysmem;
 780 }
 781
 782 static void
 783 tu6_emit_tile_select(struct tu_cmd_buffer *cmd,
 784                      struct tu_cs *cs,
 785                      const struct tu_tile *tile)
 786 {
 787    tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
 788    tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_YIELD));
 789
 790    tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
 791    tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_GMEM));
 792
 793    const uint32_t x1 = tile->begin.x;
 794    const uint32_t y1 = tile->begin.y;
 795    const uint32_t x2 = tile->end.x - 1;
 796    const uint32_t y2 = tile->end.y - 1;
 797    tu6_emit_window_scissor(cmd, cs, x1, y1, x2, y2);
 798    tu6_emit_window_offset(cmd, cs, x1, y1);
 799
 800    tu_cs_emit_regs(cs,
 801                    A6XX_VPC_SO_OVERRIDE(.so_disable = false));
 802
 803    if (use_hw_binning(cmd)) {
 804       tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
 805
 806       tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
 807       tu_cs_emit(cs, 0x0);
 808
 809       tu_cs_emit_pkt7(cs, CP_REG_TEST, 1);
 810       tu_cs_emit(cs, A6XX_CP_REG_TEST_0_REG(OVERFLOW_FLAG_REG) |
 811                      A6XX_CP_REG_TEST_0_BIT(0) |
 812                      A6XX_CP_REG_TEST_0_WAIT_FOR_ME);
 813
 814       tu_cs_reserve(cs, 3 + 11);
 815       tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
 816       tu_cs_emit(cs, CP_COND_REG_EXEC_0_MODE(PRED_TEST));
 817       tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(11));
 818
 819       /* if (no overflow) */ {
 820          tu_cs_emit_pkt7(cs, CP_SET_BIN_DATA5, 7);
 821          tu_cs_emit(cs, cmd->state.tiling_config.pipe_sizes[tile->pipe] |
 822                         CP_SET_BIN_DATA5_0_VSC_N(tile->slot));
 823          tu_cs_emit_qw(cs, cmd->vsc_data.iova + tile->pipe * cmd->vsc_data_pitch);
 824          tu_cs_emit_qw(cs, cmd->vsc_data.iova + (tile->pipe * 4) + (32 * cmd->vsc_data_pitch));
 825          tu_cs_emit_qw(cs, cmd->vsc_data2.iova + (tile->pipe * cmd->vsc_data2_pitch));
 826
 827          tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
 828          tu_cs_emit(cs, 0x0);
 829
 830          /* use a NOP packet to skip over the 'else' side: */
 831          tu_cs_emit_pkt7(cs, CP_NOP, 2);
 832       } /* else */ {
 833          tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
 834          tu_cs_emit(cs, 0x1);
 835       }
 836
 837       tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
 838       tu_cs_emit(cs, 0x0);
 839
 840       tu_cs_emit_regs(cs,
 841                       A6XX_RB_UNKNOWN_8804(0));
 842
 843       tu_cs_emit_regs(cs,
 844                       A6XX_SP_TP_UNKNOWN_B304(0));
 845
 846       tu_cs_emit_regs(cs,
 847                       A6XX_GRAS_UNKNOWN_80A4(0));
 848    } else {
 849       tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
 850       tu_cs_emit(cs, 0x1);
 851
 852       tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
 853       tu_cs_emit(cs, 0x0);
 854    }
 855 }
 856
 857 static void
 858 tu6_emit_load_attachment(struct tu_cmd_buffer *cmd, struct tu_cs *cs, uint32_t a)
 859 {
 860    const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
 861    const struct tu_framebuffer *fb = cmd->state.framebuffer;
 862    const struct tu_image_view *iview = fb->attachments[a].attachment;
 863    const struct tu_render_pass_attachment *attachment =
 864       &cmd->state.pass->attachments[a];
 865
 866    if (attachment->gmem_offset < 0)
 867       return;
 868
 869    const uint32_t x1 = tiling->render_area.offset.x;
 870    const uint32_t y1 = tiling->render_area.offset.y;
 871    const uint32_t x2 = x1 + tiling->render_area.extent.width;
 872    const uint32_t y2 = y1 + tiling->render_area.extent.height;
 873    const uint32_t tile_x2 =
 874       tiling->tile0.offset.x + tiling->tile0.extent.width * tiling->tile_count.width;
 875    const uint32_t tile_y2 =
 876       tiling->tile0.offset.y + tiling->tile0.extent.height * tiling->tile_count.height;
 877    bool need_load =
 878       x1 != tiling->tile0.offset.x || x2 != MIN2(fb->width, tile_x2) ||
 879       y1 != tiling->tile0.offset.y || y2 != MIN2(fb->height, tile_y2);
 880
 881    if (need_load)
 882       tu_finishme("improve handling of unaligned render area");
 883
 884    if (attachment->load_op == VK_ATTACHMENT_LOAD_OP_LOAD)
 885       need_load = true;
 886
 887    if (vk_format_has_stencil(iview->vk_format) &&
 888        attachment->stencil_load_op == VK_ATTACHMENT_LOAD_OP_LOAD)
 889       need_load = true;
 890
 891    if (need_load) {
 892       tu6_emit_blit_info(cmd, cs, iview, attachment->gmem_offset, false);
 893       tu6_emit_blit(cmd, cs);
 894    }
 895 }
 896
 897 static void
 898 tu6_emit_clear_attachment(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
 899                           uint32_t a,
 900                           const VkRenderPassBeginInfo *info)
 901 {
 902    const struct tu_framebuffer *fb = cmd->state.framebuffer;
 903    const struct tu_image_view *iview = fb->attachments[a].attachment;
 904    const struct tu_render_pass_attachment *attachment =
 905       &cmd->state.pass->attachments[a];
 906    unsigned clear_mask = 0;
 907
 908    /* note: this means it isn't used by any subpass and shouldn't be cleared anyway */
 909    if (attachment->gmem_offset < 0)
 910       return;
 911
 912    if (attachment->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR)
 913       clear_mask = 0xf;
 914
 915    if (vk_format_has_stencil(iview->vk_format)) {
 916       clear_mask &= 0x1;
 917       if (attachment->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR)
 918          clear_mask |= 0x2;
 919    }
 920    if (!clear_mask)
 921       return;
 922
 923    tu_clear_gmem_attachment(cmd, cs, a, clear_mask,
 924                             &info->pClearValues[a]);
 925 }
 926
 927 static void
 928 tu6_emit_predicated_blit(struct tu_cmd_buffer *cmd,
 929                          struct tu_cs *cs,
 930                          uint32_t a,
 931                          uint32_t gmem_a,
 932                          bool resolve)
 933 {
 934    tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_GMEM);
 935
 936    tu6_emit_blit_info(cmd, cs,
 937                       cmd->state.framebuffer->attachments[a].attachment,
 938                       cmd->state.pass->attachments[gmem_a].gmem_offset, resolve);
 939    tu6_emit_blit(cmd, cs);
 940
 941    tu_cond_exec_end(cs);
 942 }
 943
 944 static void
 945 tu6_emit_sysmem_resolve(struct tu_cmd_buffer *cmd,
 946                         struct tu_cs *cs,
 947                         uint32_t a,
 948                         uint32_t gmem_a)
 949 {
 950    const struct tu_framebuffer *fb = cmd->state.framebuffer;
 951    const struct tu_image_view *dst = fb->attachments[a].attachment;
 952    const struct tu_image_view *src = fb->attachments[gmem_a].attachment;
 953
 954    tu_blit(cmd, cs, &(struct tu_blit) {
 955       .dst = sysmem_attachment_surf(dst, dst->base_layer,
 956                                     &cmd->state.tiling_config.render_area),
 957       .src = sysmem_attachment_surf(src, src->base_layer,
 958                                     &cmd->state.tiling_config.render_area),
 959       .layers = fb->layers,
 960    });
 961 }
 962
 963
 964 /* Emit a MSAA resolve operation, with both gmem and sysmem paths. */
 965 static void tu6_emit_resolve(struct tu_cmd_buffer *cmd,
 966                              struct tu_cs *cs,
 967                              uint32_t a,
 968                              uint32_t gmem_a)
 969 {
 970    if (cmd->state.pass->attachments[a].store_op == VK_ATTACHMENT_STORE_OP_DONT_CARE)
 971       return;
 972
 973    tu6_emit_predicated_blit(cmd, cs, a, gmem_a, true);
 974
 975    tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_SYSMEM);
 976    tu6_emit_sysmem_resolve(cmd, cs, a, gmem_a);
 977    tu_cond_exec_end(cs);
 978 }
 979
 980 static void
 981 tu6_emit_store_attachment(struct tu_cmd_buffer *cmd,
 982                           struct tu_cs *cs,
 983                           uint32_t a,
 984                           uint32_t gmem_a)
 985 {
 986    if (cmd->state.pass->attachments[a].store_op == VK_ATTACHMENT_STORE_OP_DONT_CARE)
 987       return;
 988
 989    tu6_emit_blit_info(cmd, cs,
 990                       cmd->state.framebuffer->attachments[a].attachment,
 991                       cmd->state.pass->attachments[gmem_a].gmem_offset, true);
 992    tu6_emit_blit(cmd, cs);
 993 }
 994
 995 static void
 996 tu6_emit_tile_store(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 997 {
 998    const struct tu_render_pass *pass = cmd->state.pass;
 999    const struct tu_subpass *subpass = &pass->subpasses[pass->subpass_count-1];
1000
1001    tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3);
1002    tu_cs_emit(cs, CP_SET_DRAW_STATE__0_COUNT(0) |
1003                      CP_SET_DRAW_STATE__0_DISABLE_ALL_GROUPS |
1004                      CP_SET_DRAW_STATE__0_GROUP_ID(0));
1005    tu_cs_emit(cs, CP_SET_DRAW_STATE__1_ADDR_LO(0));
1006    tu_cs_emit(cs, CP_SET_DRAW_STATE__2_ADDR_HI(0));
1007
1008    tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
1009    tu_cs_emit(cs, 0x0);
1010
1011    tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
1012    tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_RESOLVE));
1013
1014    tu6_emit_blit_scissor(cmd, cs, true);
1015
1016    for (uint32_t a = 0; a < pass->attachment_count; ++a) {
1017       if (pass->attachments[a].gmem_offset >= 0)
1018          tu6_emit_store_attachment(cmd, cs, a, a);
1019    }
1020
1021    if (subpass->resolve_attachments) {
1022       for (unsigned i = 0; i < subpass->color_count; i++) {
1023          uint32_t a = subpass->resolve_attachments[i].attachment;
1024          if (a != VK_ATTACHMENT_UNUSED)
1025             tu6_emit_store_attachment(cmd, cs, a,
1026                                       subpass->color_attachments[i].attachment);
1027       }
1028    }
1029 }
1030
1031 static void
1032 tu6_emit_restart_index(struct tu_cs *cs, uint32_t restart_index)
1033 {
1034    tu_cs_emit_regs(cs,
1035                    A6XX_PC_RESTART_INDEX(restart_index));
1036 }
1037
1038 static void
1039 tu6_init_hw(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
1040 {
1041    tu6_emit_cache_flush(cmd, cs);
1042
1043    tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UPDATE_CNTL, 0xfffff);
1044
1045    tu_cs_emit_write_reg(cs, REG_A6XX_RB_CCU_CNTL, 0x10000000);
1046    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8E04, 0x00100000);
1047    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE04, 0x8);
1048    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE00, 0);
1049    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE0F, 0x3f);
1050    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_B605, 0x44);
1051    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_B600, 0x100000);
1052    tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UNKNOWN_BE00, 0x80);
1053    tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UNKNOWN_BE01, 0);
1054
1055    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9600, 0);
1056    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_8600, 0x880);
1057    tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UNKNOWN_BE04, 0);
1058    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE03, 0x00000410);
1059    tu_cs_emit_write_reg(cs, REG_A6XX_SP_IBO_COUNT, 0);
1060    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_B182, 0);
1061    tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UNKNOWN_BB11, 0);
1062    tu_cs_emit_write_reg(cs, REG_A6XX_UCHE_UNKNOWN_0E12, 0x3200000);
1063    tu_cs_emit_write_reg(cs, REG_A6XX_UCHE_CLIENT_PF, 4);
1064    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8E01, 0x0);
1065    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_A982, 0);
1066    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_A9A8, 0);
1067    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AB00, 0x5);
1068    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_GS_SIV_CNTL, 0x0000ffff);
1069
1070    tu_cs_emit_write_reg(cs, REG_A6XX_VFD_ADD_OFFSET, A6XX_VFD_ADD_OFFSET_VERTEX);
1071    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8811, 0x00000010);
1072    tu_cs_emit_write_reg(cs, REG_A6XX_PC_MODE_CNTL, 0x1f);
1073
1074    tu_cs_emit_write_reg(cs, REG_A6XX_RB_SRGB_CNTL, 0);
1075
1076    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_8110, 0);
1077
1078    tu_cs_emit_write_reg(cs, REG_A6XX_RB_RENDER_CONTROL0, 0x401);
1079    tu_cs_emit_write_reg(cs, REG_A6XX_RB_RENDER_CONTROL1, 0);
1080    tu_cs_emit_write_reg(cs, REG_A6XX_RB_FS_OUTPUT_CNTL0, 0);
1081    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8818, 0);
1082    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8819, 0);
1083    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_881A, 0);
1084    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_881B, 0);
1085    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_881C, 0);
1086    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_881D, 0);
1087    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_881E, 0);
1088    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_88F0, 0);
1089
1090    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9101, 0xffff00);
1091    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9107, 0);
1092
1093    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9236, 1);
1094    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9300, 0);
1095
1096    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_SO_OVERRIDE,
1097                         A6XX_VPC_SO_OVERRIDE_SO_DISABLE);
1098
1099    tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9801, 0);
1100    tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9806, 0);
1101    tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9980, 0);
1102    tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9990, 0);
1103
1104    tu_cs_emit_write_reg(cs, REG_A6XX_PC_PRIMITIVE_CNTL_6, 0);
1105    tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9B07, 0);
1106
1107    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_A81B, 0);
1108
1109    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_B183, 0);
1110
1111    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_8099, 0);
1112    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_809B, 0);
1113    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A0, 2);
1114    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80AF, 0);
1115    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9210, 0);
1116    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9211, 0);
1117    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9602, 0);
1118    tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9981, 0x3);
1119    tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9E72, 0);
1120    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9108, 0x3);
1121    tu_cs_emit_write_reg(cs, REG_A6XX_SP_TP_UNKNOWN_B304, 0);
1122    tu_cs_emit_write_reg(cs, REG_A6XX_SP_TP_UNKNOWN_B309, 0x000000a2);
1123    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8804, 0);
1124    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A4, 0);
1125    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A5, 0);
1126    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A6, 0);
1127    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8805, 0);
1128    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8806, 0);
1129    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8878, 0);
1130    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8879, 0);
1131    tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_CONTROL_5_REG, 0xfc);
1132
1133    tu_cs_emit_write_reg(cs, REG_A6XX_VFD_MODE_CNTL, 0x00000000);
1134
1135    tu_cs_emit_write_reg(cs, REG_A6XX_VFD_UNKNOWN_A008, 0);
1136
1137    tu_cs_emit_write_reg(cs, REG_A6XX_PC_MODE_CNTL, 0x0000001f);
1138
1139    /* we don't use this yet.. probably best to disable.. */
1140    tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3);
1141    tu_cs_emit(cs, CP_SET_DRAW_STATE__0_COUNT(0) |
1142                      CP_SET_DRAW_STATE__0_DISABLE_ALL_GROUPS |
1143                      CP_SET_DRAW_STATE__0_GROUP_ID(0));
1144    tu_cs_emit(cs, CP_SET_DRAW_STATE__1_ADDR_LO(0));
1145    tu_cs_emit(cs, CP_SET_DRAW_STATE__2_ADDR_HI(0));
1146
1147    /* Set not to use streamout by default, */
1148    tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 4);
1149    tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
1150    tu_cs_emit(cs, 0);
1151    tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
1152    tu_cs_emit(cs, 0);
1153
1154    tu_cs_emit_regs(cs,
1155                    A6XX_SP_HS_CTRL_REG0(0));
1156
1157    tu_cs_emit_regs(cs,
1158                    A6XX_SP_GS_CTRL_REG0(0));
1159
1160    tu_cs_emit_regs(cs,
1161                    A6XX_GRAS_LRZ_CNTL(0));
1162
1163    tu_cs_emit_regs(cs,
1164                    A6XX_RB_LRZ_CNTL(0));
1165
1166    tu_cs_sanity_check(cs);
1167 }
1168
1169 static void
1170 tu6_cache_flush(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
1171 {
1172    unsigned seqno;
1173
1174    seqno = tu6_emit_event_write(cmd, cs, CACHE_FLUSH_AND_INV_EVENT, true);
1175
1176    tu_cs_emit_pkt7(cs, CP_WAIT_REG_MEM, 6);
1177    tu_cs_emit(cs, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ) |
1178                   CP_WAIT_REG_MEM_0_POLL_MEMORY);
1179    tu_cs_emit_qw(cs, cmd->scratch_bo.iova);
1180    tu_cs_emit(cs, CP_WAIT_REG_MEM_3_REF(seqno));
1181    tu_cs_emit(cs, CP_WAIT_REG_MEM_4_MASK(~0));
1182    tu_cs_emit(cs, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(16));
1183
1184    seqno = tu6_emit_event_write(cmd, cs, CACHE_FLUSH_TS, true);
1185
1186    tu_cs_emit_pkt7(cs, CP_WAIT_MEM_GTE, 4);
1187    tu_cs_emit(cs, CP_WAIT_MEM_GTE_0_RESERVED(0));
1188    tu_cs_emit_qw(cs, cmd->scratch_bo.iova);
1189    tu_cs_emit(cs, CP_WAIT_MEM_GTE_3_REF(seqno));
1190 }
1191
1192 static void
1193 update_vsc_pipe(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
1194 {
1195    const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
1196
1197    tu_cs_emit_regs(cs,
1198                    A6XX_VSC_BIN_SIZE(.width = tiling->tile0.extent.width,
1199                                      .height = tiling->tile0.extent.height),
1200                    A6XX_VSC_SIZE_ADDRESS(.bo = &cmd->vsc_data,
1201                                          .bo_offset = 32 * cmd->vsc_data_pitch));
1202
1203    tu_cs_emit_regs(cs,
1204                    A6XX_VSC_BIN_COUNT(.nx = tiling->tile_count.width,
1205                                       .ny = tiling->tile_count.height));
1206
1207    tu_cs_emit_pkt4(cs, REG_A6XX_VSC_PIPE_CONFIG_REG(0), 32);
1208    for (unsigned i = 0; i < 32; i++)
1209       tu_cs_emit(cs, tiling->pipe_config[i]);
1210
1211    tu_cs_emit_regs(cs,
1212                    A6XX_VSC_PIPE_DATA2_ADDRESS(.bo = &cmd->vsc_data2),
1213                    A6XX_VSC_PIPE_DATA2_PITCH(cmd->vsc_data2_pitch),
1214                    A6XX_VSC_PIPE_DATA2_ARRAY_PITCH(cmd->vsc_data2.size));
1215
1216    tu_cs_emit_regs(cs,
1217                    A6XX_VSC_PIPE_DATA_ADDRESS(.bo = &cmd->vsc_data),
1218                    A6XX_VSC_PIPE_DATA_PITCH(cmd->vsc_data_pitch),
1219                    A6XX_VSC_PIPE_DATA_ARRAY_PITCH(cmd->vsc_data.size));
1220 }
1221
1222 static void
1223 emit_vsc_overflow_test(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
1224 {
1225    const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
1226    const uint32_t used_pipe_count =
1227       tiling->pipe_count.width * tiling->pipe_count.height;
1228
1229    /* Clear vsc_scratch: */
1230    tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 3);
1231    tu_cs_emit_qw(cs, cmd->scratch_bo.iova + VSC_SCRATCH);
1232    tu_cs_emit(cs, 0x0);
1233
1234    /* Check for overflow, write vsc_scratch if detected: */
1235    for (int i = 0; i < used_pipe_count; i++) {
1236       tu_cs_emit_pkt7(cs, CP_COND_WRITE5, 8);
1237       tu_cs_emit(cs, CP_COND_WRITE5_0_FUNCTION(WRITE_GE) |
1238             CP_COND_WRITE5_0_WRITE_MEMORY);
1239       tu_cs_emit(cs, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_SIZE_REG(i)));
1240       tu_cs_emit(cs, CP_COND_WRITE5_2_POLL_ADDR_HI(0));
1241       tu_cs_emit(cs, CP_COND_WRITE5_3_REF(cmd->vsc_data_pitch));
1242       tu_cs_emit(cs, CP_COND_WRITE5_4_MASK(~0));
1243       tu_cs_emit_qw(cs, cmd->scratch_bo.iova + VSC_SCRATCH);
1244       tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(1 + cmd->vsc_data_pitch));
1245
1246       tu_cs_emit_pkt7(cs, CP_COND_WRITE5, 8);
1247       tu_cs_emit(cs, CP_COND_WRITE5_0_FUNCTION(WRITE_GE) |
1248             CP_COND_WRITE5_0_WRITE_MEMORY);
1249       tu_cs_emit(cs, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_SIZE2_REG(i)));
1250       tu_cs_emit(cs, CP_COND_WRITE5_2_POLL_ADDR_HI(0));
1251       tu_cs_emit(cs, CP_COND_WRITE5_3_REF(cmd->vsc_data2_pitch));
1252       tu_cs_emit(cs, CP_COND_WRITE5_4_MASK(~0));
1253       tu_cs_emit_qw(cs, cmd->scratch_bo.iova + VSC_SCRATCH);
1254       tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(3 + cmd->vsc_data2_pitch));
1255    }
1256
1257    tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
1258
1259    tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
1260
1261    tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
1262    tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(OVERFLOW_FLAG_REG) |
1263          CP_MEM_TO_REG_0_CNT(1 - 1));
1264    tu_cs_emit_qw(cs, cmd->scratch_bo.iova + VSC_SCRATCH);
1265
1266    /*
1267     * This is a bit awkward, we really want a way to invert the
1268     * CP_REG_TEST/CP_COND_REG_EXEC logic, so that we can conditionally
1269     * execute cmds to use hwbinning when a bit is *not* set.  This
1270     * dance is to invert OVERFLOW_FLAG_REG
1271     *
1272     * A CP_NOP packet is used to skip executing the 'else' clause
1273     * if (b0 set)..
1274     */
1275
1276    /* b0 will be set if VSC_DATA or VSC_DATA2 overflow: */
1277    tu_cs_emit_pkt7(cs, CP_REG_TEST, 1);
1278    tu_cs_emit(cs, A6XX_CP_REG_TEST_0_REG(OVERFLOW_FLAG_REG) |
1279          A6XX_CP_REG_TEST_0_BIT(0) |
1280          A6XX_CP_REG_TEST_0_WAIT_FOR_ME);
1281
1282    tu_cs_reserve(cs, 3 + 7);
1283    tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
1284    tu_cs_emit(cs, CP_COND_REG_EXEC_0_MODE(PRED_TEST));
1285    tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(7));
1286
1287    /* if (b0 set) */ {
1288       /*
1289        * On overflow, mirror the value to control->vsc_overflow
1290        * which CPU is checking to detect overflow (see
1291        * check_vsc_overflow())
1292        */
1293       tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
1294       tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(OVERFLOW_FLAG_REG) |
1295             CP_REG_TO_MEM_0_CNT(0));
1296       tu_cs_emit_qw(cs, cmd->scratch_bo.iova + VSC_OVERFLOW);
1297
1298       tu_cs_emit_pkt4(cs, OVERFLOW_FLAG_REG, 1);
1299       tu_cs_emit(cs, 0x0);
1300
1301       tu_cs_emit_pkt7(cs, CP_NOP, 2);  /* skip 'else' when 'if' is taken */
1302    } /* else */ {
1303       tu_cs_emit_pkt4(cs, OVERFLOW_FLAG_REG, 1);
1304       tu_cs_emit(cs, 0x1);
1305    }
1306 }
1307
1308 static void
1309 tu6_emit_binning_pass(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
1310 {
1311    struct tu_physical_device *phys_dev = cmd->device->physical_device;
1312    const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
1313
1314    uint32_t x1 = tiling->tile0.offset.x;
1315    uint32_t y1 = tiling->tile0.offset.y;
1316    uint32_t x2 = tiling->render_area.offset.x + tiling->render_area.extent.width - 1;
1317    uint32_t y2 = tiling->render_area.offset.y + tiling->render_area.extent.height - 1;
1318
1319    tu6_emit_window_scissor(cmd, cs, x1, y1, x2, y2);
1320
1321    tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
1322    tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_BINNING));
1323
1324    tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
1325    tu_cs_emit(cs, 0x1);
1326
1327    tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
1328    tu_cs_emit(cs, 0x1);
1329
1330    tu_cs_emit_wfi(cs);
1331
1332    tu_cs_emit_regs(cs,
1333                    A6XX_VFD_MODE_CNTL(.binning_pass = true));
1334
1335    update_vsc_pipe(cmd, cs);
1336
1337    tu_cs_emit_regs(cs,
1338                    A6XX_PC_UNKNOWN_9805(.unknown = phys_dev->magic.PC_UNKNOWN_9805));
1339
1340    tu_cs_emit_regs(cs,
1341                    A6XX_SP_UNKNOWN_A0F8(.unknown = phys_dev->magic.SP_UNKNOWN_A0F8));
1342
1343    tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 1);
1344    tu_cs_emit(cs, UNK_2C);
1345
1346    tu_cs_emit_regs(cs,
1347                    A6XX_RB_WINDOW_OFFSET(.x = 0, .y = 0));
1348
1349    tu_cs_emit_regs(cs,
1350                    A6XX_SP_TP_WINDOW_OFFSET(.x = 0, .y = 0));
1351
1352    /* emit IB to binning drawcmds: */
1353    tu_cs_emit_call(cs, &cmd->draw_cs);
1354
1355    tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3);
1356    tu_cs_emit(cs, CP_SET_DRAW_STATE__0_COUNT(0) |
1357                   CP_SET_DRAW_STATE__0_DISABLE_ALL_GROUPS |
1358                   CP_SET_DRAW_STATE__0_GROUP_ID(0));
1359    tu_cs_emit(cs, CP_SET_DRAW_STATE__1_ADDR_LO(0));
1360    tu_cs_emit(cs, CP_SET_DRAW_STATE__2_ADDR_HI(0));
1361
1362    tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 1);
1363    tu_cs_emit(cs, UNK_2D);
1364
1365    tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE, false);
1366    tu6_cache_flush(cmd, cs);
1367
1368    tu_cs_emit_wfi(cs);
1369
1370    tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
1371
1372    emit_vsc_overflow_test(cmd, cs);
1373
1374    tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
1375    tu_cs_emit(cs, 0x0);
1376
1377    tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
1378    tu_cs_emit(cs, 0x0);
1379
1380    tu_cs_emit_wfi(cs);
1381
1382    tu_cs_emit_regs(cs,
1383                    A6XX_RB_CCU_CNTL(.unknown = phys_dev->magic.RB_CCU_CNTL_gmem));
1384
1385    cmd->wait_for_idle = false;
1386 }
1387
1388 static void
1389 tu_emit_sysmem_clear_attachment(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
1390                                 uint32_t a,
1391                                 const VkRenderPassBeginInfo *info)
1392 {
1393    const struct tu_framebuffer *fb = cmd->state.framebuffer;
1394    const struct tu_image_view *iview = fb->attachments[a].attachment;
1395    const struct tu_render_pass_attachment *attachment =
1396       &cmd->state.pass->attachments[a];
1397    unsigned clear_mask = 0;
1398
1399    /* note: this means it isn't used by any subpass and shouldn't be cleared anyway */
1400    if (attachment->gmem_offset < 0)
1401       return;
1402
1403    if (attachment->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
1404       clear_mask = 0xf;
1405    }
1406
1407    if (vk_format_has_stencil(iview->vk_format)) {
1408       clear_mask &= 0x1;
1409       if (attachment->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR)
1410          clear_mask |= 0x2;
1411       if (clear_mask != 0x3)
1412          tu_finishme("depth/stencil only load op");
1413    }
1414
1415    if (!clear_mask)
1416       return;
1417
1418    tu_clear_sysmem_attachment(cmd, cs, a,
1419                               &info->pClearValues[a], &(struct VkClearRect) {
1420       .rect = info->renderArea,
1421       .baseArrayLayer = iview->base_layer,
1422       .layerCount = iview->layer_count,
1423    });
1424 }
1425
1426 static void
1427 tu_emit_load_clear(struct tu_cmd_buffer *cmd,
1428                    const VkRenderPassBeginInfo *info)
1429 {
1430    struct tu_cs *cs = &cmd->draw_cs;
1431
1432    tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_GMEM);
1433
1434    tu6_emit_blit_scissor(cmd, cs, true);
1435
1436    for (uint32_t i = 0; i < cmd->state.pass->attachment_count; ++i)
1437       tu6_emit_load_attachment(cmd, cs, i);
1438
1439    tu6_emit_blit_scissor(cmd, cs, false);
1440
1441    for (uint32_t i = 0; i < cmd->state.pass->attachment_count; ++i)
1442       tu6_emit_clear_attachment(cmd, cs, i, info);
1443
1444    tu_cond_exec_end(cs);
1445
1446    /* invalidate because reading input attachments will cache GMEM and
1447     * the cache isn''t updated when GMEM is written
1448     * TODO: is there a no-cache bit for textures?
1449     */
1450    if (cmd->state.subpass->input_count)
1451       tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE, false);
1452
1453    tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_SYSMEM);
1454
1455    for (uint32_t i = 0; i < cmd->state.pass->attachment_count; ++i)
1456       tu_emit_sysmem_clear_attachment(cmd, cs, i, info);
1457
1458    tu_cond_exec_end(cs);
1459 }
1460
1461 static void
1462 tu6_sysmem_render_begin(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
1463                         const struct VkRect2D *renderArea)
1464 {
1465    const struct tu_framebuffer *fb = cmd->state.framebuffer;
1466
1467    assert(fb->width > 0 && fb->height > 0);
1468    tu6_emit_window_scissor(cmd, cs, 0, 0, fb->width - 1, fb->height - 1);
1469    tu6_emit_window_offset(cmd, cs, 0, 0);
1470
1471    tu6_emit_bin_size(cs, 0, 0, 0xc00000); /* 0xc00000 = BYPASS? */
1472
1473    tu6_emit_lrz_flush(cmd, cs);
1474
1475    tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
1476    tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_BYPASS));
1477
1478    tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
1479    tu_cs_emit(cs, 0x0);
1480
1481    tu6_emit_event_write(cmd, cs, PC_CCU_INVALIDATE_COLOR, false);
1482    tu6_emit_event_write(cmd, cs, PC_CCU_INVALIDATE_DEPTH, false);
1483    tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE, false);
1484
1485    tu6_emit_wfi(cmd, cs);
1486    tu_cs_emit_regs(cs,
1487                    A6XX_RB_CCU_CNTL(0x10000000));
1488
1489    /* enable stream-out, with sysmem there is only one pass: */
1490    tu_cs_emit_regs(cs,
1491                    A6XX_VPC_SO_OVERRIDE(.so_disable = false));
1492
1493    tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
1494    tu_cs_emit(cs, 0x1);
1495
1496    tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
1497    tu_cs_emit(cs, 0x0);
1498
1499    tu_cs_sanity_check(cs);
1500 }
1501
1502 static void
1503 tu6_sysmem_render_end(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
1504 {
1505    /* Do any resolves of the last subpass. These are handled in the
1506     * tile_store_ib in the gmem path.
1507     */
1508
1509    const struct tu_subpass *subpass = cmd->state.subpass;
1510    if (subpass->resolve_attachments) {
1511       for (unsigned i = 0; i < subpass->color_count; i++) {
1512          uint32_t a = subpass->resolve_attachments[i].attachment;
1513          if (a != VK_ATTACHMENT_UNUSED)
1514             tu6_emit_sysmem_resolve(cmd, cs, a,
1515                                     subpass->color_attachments[i].attachment);
1516       }
1517    }
1518
1519    tu_cs_emit_call(cs, &cmd->draw_epilogue_cs);
1520
1521    tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
1522    tu_cs_emit(cs, 0x0);
1523
1524    tu6_emit_lrz_flush(cmd, cs);
1525
1526    tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_COLOR_TS, true);
1527    tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_DEPTH_TS, true);
1528
1529    tu_cs_sanity_check(cs);
1530 }
1531
1532
1533 static void
1534 tu6_tile_render_begin(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
1535 {
1536    struct tu_physical_device *phys_dev = cmd->device->physical_device;
1537
1538    tu6_emit_lrz_flush(cmd, cs);
1539
1540    /* lrz clear? */
1541
1542    tu6_emit_cache_flush(cmd, cs);
1543
1544    tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
1545    tu_cs_emit(cs, 0x0);
1546
1547    /* 0x10000000 for BYPASS.. 0x7c13c080 for GMEM: */
1548    tu6_emit_wfi(cmd, cs);
1549    tu_cs_emit_regs(cs,
1550                    A6XX_RB_CCU_CNTL(phys_dev->magic.RB_CCU_CNTL_gmem));
1551
1552    const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
1553    if (use_hw_binning(cmd)) {
1554       /* enable stream-out during binning pass: */
1555       tu_cs_emit_regs(cs, A6XX_VPC_SO_OVERRIDE(.so_disable=false));
1556
1557       tu6_emit_bin_size(cs,
1558                         tiling->tile0.extent.width,
1559                         tiling->tile0.extent.height,
1560                         A6XX_RB_BIN_CONTROL_BINNING_PASS | 0x6000000);
1561
1562       tu6_emit_render_cntl(cmd, cmd->state.subpass, cs, true);
1563
1564       tu6_emit_binning_pass(cmd, cs);
1565
1566       /* and disable stream-out for draw pass: */
1567       tu_cs_emit_regs(cs, A6XX_VPC_SO_OVERRIDE(.so_disable=true));
1568
1569       tu6_emit_bin_size(cs,
1570                         tiling->tile0.extent.width,
1571                         tiling->tile0.extent.height,
1572                         A6XX_RB_BIN_CONTROL_USE_VIZ | 0x6000000);
1573
1574       tu_cs_emit_regs(cs,
1575                       A6XX_VFD_MODE_CNTL(0));
1576
1577       tu_cs_emit_regs(cs, A6XX_PC_UNKNOWN_9805(.unknown = phys_dev->magic.PC_UNKNOWN_9805));
1578
1579       tu_cs_emit_regs(cs, A6XX_SP_UNKNOWN_A0F8(.unknown = phys_dev->magic.SP_UNKNOWN_A0F8));
1580
1581       tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
1582       tu_cs_emit(cs, 0x1);
1583    } else {
1584       /* no binning pass, so enable stream-out for draw pass:: */
1585       tu_cs_emit_regs(cs, A6XX_VPC_SO_OVERRIDE(.so_disable=false));
1586
1587       tu6_emit_bin_size(cs,
1588                         tiling->tile0.extent.width,
1589                         tiling->tile0.extent.height,
1590                         0x6000000);
1591    }
1592
1593    tu_cs_sanity_check(cs);
1594 }
1595
1596 static void
1597 tu6_render_tile(struct tu_cmd_buffer *cmd,
1598                 struct tu_cs *cs,
1599                 const struct tu_tile *tile)
1600 {
1601    tu6_emit_tile_select(cmd, cs, tile);
1602
1603    tu_cs_emit_call(cs, &cmd->draw_cs);
1604    cmd->wait_for_idle = true;
1605
1606    if (use_hw_binning(cmd)) {
1607       tu_cs_emit_pkt7(cs, CP_REG_TEST, 1);
1608       tu_cs_emit(cs, A6XX_CP_REG_TEST_0_REG(OVERFLOW_FLAG_REG) |
1609                      A6XX_CP_REG_TEST_0_BIT(0) |
1610                      A6XX_CP_REG_TEST_0_WAIT_FOR_ME);
1611
1612       tu_cs_reserve(cs, 3 + 2);
1613       tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
1614       tu_cs_emit(cs, CP_COND_REG_EXEC_0_MODE(PRED_TEST));
1615       tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(2));
1616
1617       /* if (no overflow) */ {
1618          tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
1619          tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_ENDVIS));
1620       }
1621    }
1622
1623    tu_cs_emit_ib(cs, &cmd->state.tile_store_ib);
1624
1625    tu_cs_sanity_check(cs);
1626 }
1627
1628 static void
1629 tu6_tile_render_end(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
1630 {
1631    tu_cs_emit_call(cs, &cmd->draw_epilogue_cs);
1632
1633    tu_cs_emit_regs(cs,
1634                    A6XX_GRAS_LRZ_CNTL(0));
1635
1636    tu6_emit_lrz_flush(cmd, cs);
1637
1638    tu6_emit_event_write(cmd, cs, CACHE_FLUSH_TS, true);
1639
1640    tu_cs_sanity_check(cs);
1641 }
1642
1643 static void
1644 tu_cmd_render_tiles(struct tu_cmd_buffer *cmd)
1645 {
1646    const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
1647
1648    tu6_tile_render_begin(cmd, &cmd->cs);
1649
1650    for (uint32_t y = 0; y < tiling->tile_count.height; y++) {
1651       for (uint32_t x = 0; x < tiling->tile_count.width; x++) {
1652          struct tu_tile tile;
1653          tu_tiling_config_get_tile(tiling, cmd->device, x, y, &tile);
1654          tu6_render_tile(cmd, &cmd->cs, &tile);
1655       }
1656    }
1657
1658    tu6_tile_render_end(cmd, &cmd->cs);
1659 }
1660
1661 static void
1662 tu_cmd_render_sysmem(struct tu_cmd_buffer *cmd)
1663 {
1664    const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
1665
1666    tu6_sysmem_render_begin(cmd, &cmd->cs, &tiling->render_area);
1667
1668    tu_cs_emit_call(&cmd->cs, &cmd->draw_cs);
1669    cmd->wait_for_idle = true;
1670
1671    tu6_sysmem_render_end(cmd, &cmd->cs);
1672 }
1673
1674 static void
1675 tu_cmd_prepare_tile_store_ib(struct tu_cmd_buffer *cmd)
1676 {
1677    const uint32_t tile_store_space = 32 + 23 * cmd->state.pass->attachment_count;
1678    struct tu_cs sub_cs;
1679
1680    VkResult result =
1681       tu_cs_begin_sub_stream(&cmd->sub_cs, tile_store_space, &sub_cs);
1682    if (result != VK_SUCCESS) {
1683       cmd->record_result = result;
1684       return;
1685    }
1686
1687    /* emit to tile-store sub_cs */
1688    tu6_emit_tile_store(cmd, &sub_cs);
1689
1690    cmd->state.tile_store_ib = tu_cs_end_sub_stream(&cmd->sub_cs, &sub_cs);
1691 }
1692
1693 static void
1694 tu_cmd_update_tiling_config(struct tu_cmd_buffer *cmd,
1695                             const VkRect2D *render_area)
1696 {
1697    const struct tu_device *dev = cmd->device;
1698    struct tu_tiling_config *tiling = &cmd->state.tiling_config;
1699
1700    tiling->render_area = *render_area;
1701    tiling->force_sysmem = force_sysmem(cmd, render_area);
1702
1703    tu_tiling_config_update_tile_layout(tiling, dev, cmd->state.pass->gmem_pixels);
1704    tu_tiling_config_update_pipe_layout(tiling, dev);
1705    tu_tiling_config_update_pipes(tiling, dev);
1706 }
1707
1708 const struct tu_dynamic_state default_dynamic_state = {
1709    .viewport =
1710      {
1711        .count = 0,
1712      },
1713    .scissor =
1714      {
1715        .count = 0,
1716      },
1717    .line_width = 1.0f,
1718    .depth_bias =
1719      {
1720        .bias = 0.0f,
1721        .clamp = 0.0f,
1722        .slope = 0.0f,
1723      },
1724    .blend_constants = { 0.0f, 0.0f, 0.0f, 0.0f },
1725    .depth_bounds =
1726      {
1727        .min = 0.0f,
1728        .max = 1.0f,
1729      },
1730    .stencil_compare_mask =
1731      {
1732        .front = ~0u,
1733        .back = ~0u,
1734      },
1735    .stencil_write_mask =
1736      {
1737        .front = ~0u,
1738        .back = ~0u,
1739      },
1740    .stencil_reference =
1741      {
1742        .front = 0u,
1743        .back = 0u,
1744      },
1745 };
1746
1747 static void UNUSED /* FINISHME */
1748 tu_bind_dynamic_state(struct tu_cmd_buffer *cmd_buffer,
1749                       const struct tu_dynamic_state *src)
1750 {
1751    struct tu_dynamic_state *dest = &cmd_buffer->state.dynamic;
1752    uint32_t copy_mask = src->mask;
1753    uint32_t dest_mask = 0;
1754
1755    tu_use_args(cmd_buffer); /* FINISHME */
1756
1757    /* Make sure to copy the number of viewports/scissors because they can
1758     * only be specified at pipeline creation time.
1759     */
1760    dest->viewport.count = src->viewport.count;
1761    dest->scissor.count = src->scissor.count;
1762    dest->discard_rectangle.count = src->discard_rectangle.count;
1763
1764    if (copy_mask & TU_DYNAMIC_VIEWPORT) {
1765       if (memcmp(&dest->viewport.viewports, &src->viewport.viewports,
1766                  src->viewport.count * sizeof(VkViewport))) {
1767          typed_memcpy(dest->viewport.viewports, src->viewport.viewports,
1768                       src->viewport.count);
1769          dest_mask |= TU_DYNAMIC_VIEWPORT;
1770       }
1771    }
1772
1773    if (copy_mask & TU_DYNAMIC_SCISSOR) {
1774       if (memcmp(&dest->scissor.scissors, &src->scissor.scissors,
1775                  src->scissor.count * sizeof(VkRect2D))) {
1776          typed_memcpy(dest->scissor.scissors, src->scissor.scissors,
1777                       src->scissor.count);
1778          dest_mask |= TU_DYNAMIC_SCISSOR;
1779       }
1780    }
1781
1782    if (copy_mask & TU_DYNAMIC_LINE_WIDTH) {
1783       if (dest->line_width != src->line_width) {
1784          dest->line_width = src->line_width;
1785          dest_mask |= TU_DYNAMIC_LINE_WIDTH;
1786       }
1787    }
1788
1789    if (copy_mask & TU_DYNAMIC_DEPTH_BIAS) {
1790       if (memcmp(&dest->depth_bias, &src->depth_bias,
1791                  sizeof(src->depth_bias))) {
1792          dest->depth_bias = src->depth_bias;
1793          dest_mask |= TU_DYNAMIC_DEPTH_BIAS;
1794       }
1795    }
1796
1797    if (copy_mask & TU_DYNAMIC_BLEND_CONSTANTS) {
1798       if (memcmp(&dest->blend_constants, &src->blend_constants,
1799                  sizeof(src->blend_constants))) {
1800          typed_memcpy(dest->blend_constants, src->blend_constants, 4);
1801          dest_mask |= TU_DYNAMIC_BLEND_CONSTANTS;
1802       }
1803    }
1804
1805    if (copy_mask & TU_DYNAMIC_DEPTH_BOUNDS) {
1806       if (memcmp(&dest->depth_bounds, &src->depth_bounds,
1807                  sizeof(src->depth_bounds))) {
1808          dest->depth_bounds = src->depth_bounds;
1809          dest_mask |= TU_DYNAMIC_DEPTH_BOUNDS;
1810       }
1811    }
1812
1813    if (copy_mask & TU_DYNAMIC_STENCIL_COMPARE_MASK) {
1814       if (memcmp(&dest->stencil_compare_mask, &src->stencil_compare_mask,
1815                  sizeof(src->stencil_compare_mask))) {
1816          dest->stencil_compare_mask = src->stencil_compare_mask;
1817          dest_mask |= TU_DYNAMIC_STENCIL_COMPARE_MASK;
1818       }
1819    }
1820
1821    if (copy_mask & TU_DYNAMIC_STENCIL_WRITE_MASK) {
1822       if (memcmp(&dest->stencil_write_mask, &src->stencil_write_mask,
1823                  sizeof(src->stencil_write_mask))) {
1824          dest->stencil_write_mask = src->stencil_write_mask;
1825          dest_mask |= TU_DYNAMIC_STENCIL_WRITE_MASK;
1826       }
1827    }
1828
1829    if (copy_mask & TU_DYNAMIC_STENCIL_REFERENCE) {
1830       if (memcmp(&dest->stencil_reference, &src->stencil_reference,
1831                  sizeof(src->stencil_reference))) {
1832          dest->stencil_reference = src->stencil_reference;
1833          dest_mask |= TU_DYNAMIC_STENCIL_REFERENCE;
1834       }
1835    }
1836
1837    if (copy_mask & TU_DYNAMIC_DISCARD_RECTANGLE) {
1838       if (memcmp(&dest->discard_rectangle.rectangles,
1839                  &src->discard_rectangle.rectangles,
1840                  src->discard_rectangle.count * sizeof(VkRect2D))) {
1841          typed_memcpy(dest->discard_rectangle.rectangles,
1842                       src->discard_rectangle.rectangles,
1843                       src->discard_rectangle.count);
1844          dest_mask |= TU_DYNAMIC_DISCARD_RECTANGLE;
1845       }
1846    }
1847 }
1848
1849 static VkResult
1850 tu_create_cmd_buffer(struct tu_device *device,
1851                      struct tu_cmd_pool *pool,
1852                      VkCommandBufferLevel level,
1853                      VkCommandBuffer *pCommandBuffer)
1854 {
1855    struct tu_cmd_buffer *cmd_buffer;
1856    cmd_buffer = vk_zalloc(&pool->alloc, sizeof(*cmd_buffer), 8,
1857                           VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1858    if (cmd_buffer == NULL)
1859       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1860
1861    cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1862    cmd_buffer->device = device;
1863    cmd_buffer->pool = pool;
1864    cmd_buffer->level = level;
1865
1866    if (pool) {
1867       list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
1868       cmd_buffer->queue_family_index = pool->queue_family_index;
1869
1870    } else {
1871       /* Init the pool_link so we can safely call list_del when we destroy
1872        * the command buffer
1873        */
1874       list_inithead(&cmd_buffer->pool_link);
1875       cmd_buffer->queue_family_index = TU_QUEUE_GENERAL;
1876    }
1877
1878    tu_bo_list_init(&cmd_buffer->bo_list);
1879    tu_cs_init(&cmd_buffer->cs, device, TU_CS_MODE_GROW, 4096);
1880    tu_cs_init(&cmd_buffer->draw_cs, device, TU_CS_MODE_GROW, 4096);
1881    tu_cs_init(&cmd_buffer->draw_epilogue_cs, device, TU_CS_MODE_GROW, 4096);
1882    tu_cs_init(&cmd_buffer->sub_cs, device, TU_CS_MODE_SUB_STREAM, 2048);
1883
1884    *pCommandBuffer = tu_cmd_buffer_to_handle(cmd_buffer);
1885
1886    list_inithead(&cmd_buffer->upload.list);
1887
1888    VkResult result = tu_bo_init_new(device, &cmd_buffer->scratch_bo, 0x1000);
1889    if (result != VK_SUCCESS)
1890       goto fail_scratch_bo;
1891
1892    /* TODO: resize on overflow */
1893    cmd_buffer->vsc_data_pitch = device->vsc_data_pitch;
1894    cmd_buffer->vsc_data2_pitch = device->vsc_data2_pitch;
1895    cmd_buffer->vsc_data = device->vsc_data;
1896    cmd_buffer->vsc_data2 = device->vsc_data2;
1897
1898    return VK_SUCCESS;
1899
1900 fail_scratch_bo:
1901    list_del(&cmd_buffer->pool_link);
1902    return result;
1903 }
1904
1905 static void
1906 tu_cmd_buffer_destroy(struct tu_cmd_buffer *cmd_buffer)
1907 {
1908    tu_bo_finish(cmd_buffer->device, &cmd_buffer->scratch_bo);
1909
1910    list_del(&cmd_buffer->pool_link);
1911
1912    for (unsigned i = 0; i < VK_PIPELINE_BIND_POINT_RANGE_SIZE; i++)
1913       free(cmd_buffer->descriptors[i].push_set.set.mapped_ptr);
1914
1915    tu_cs_finish(&cmd_buffer->cs);
1916    tu_cs_finish(&cmd_buffer->draw_cs);
1917    tu_cs_finish(&cmd_buffer->draw_epilogue_cs);
1918    tu_cs_finish(&cmd_buffer->sub_cs);
1919
1920    tu_bo_list_destroy(&cmd_buffer->bo_list);
1921    vk_free(&cmd_buffer->pool->alloc, cmd_buffer);
1922 }
1923
1924 static VkResult
1925 tu_reset_cmd_buffer(struct tu_cmd_buffer *cmd_buffer)
1926 {
1927    cmd_buffer->wait_for_idle = true;
1928
1929    cmd_buffer->record_result = VK_SUCCESS;
1930
1931    tu_bo_list_reset(&cmd_buffer->bo_list);
1932    tu_cs_reset(&cmd_buffer->cs);
1933    tu_cs_reset(&cmd_buffer->draw_cs);
1934    tu_cs_reset(&cmd_buffer->draw_epilogue_cs);
1935    tu_cs_reset(&cmd_buffer->sub_cs);
1936
1937    for (unsigned i = 0; i < VK_PIPELINE_BIND_POINT_RANGE_SIZE; i++) {
1938       cmd_buffer->descriptors[i].valid = 0;
1939       cmd_buffer->descriptors[i].push_dirty = false;
1940    }
1941
1942    cmd_buffer->status = TU_CMD_BUFFER_STATUS_INITIAL;
1943
1944    return cmd_buffer->record_result;
1945 }
1946
1947 VkResult
1948 tu_AllocateCommandBuffers(VkDevice _device,
1949                           const VkCommandBufferAllocateInfo *pAllocateInfo,
1950                           VkCommandBuffer *pCommandBuffers)
1951 {
1952    TU_FROM_HANDLE(tu_device, device, _device);
1953    TU_FROM_HANDLE(tu_cmd_pool, pool, pAllocateInfo->commandPool);
1954
1955    VkResult result = VK_SUCCESS;
1956    uint32_t i;
1957
1958    for (i = 0; i < pAllocateInfo->commandBufferCount; i++) {
1959
1960       if (!list_is_empty(&pool->free_cmd_buffers)) {
1961          struct tu_cmd_buffer *cmd_buffer = list_first_entry(
1962             &pool->free_cmd_buffers, struct tu_cmd_buffer, pool_link);
1963
1964          list_del(&cmd_buffer->pool_link);
1965          list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
1966
1967          result = tu_reset_cmd_buffer(cmd_buffer);
1968          cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1969          cmd_buffer->level = pAllocateInfo->level;
1970
1971          pCommandBuffers[i] = tu_cmd_buffer_to_handle(cmd_buffer);
1972       } else {
1973          result = tu_create_cmd_buffer(device, pool, pAllocateInfo->level,
1974                                        &pCommandBuffers[i]);
1975       }
1976       if (result != VK_SUCCESS)
1977          break;
1978    }
1979
1980    if (result != VK_SUCCESS) {
1981       tu_FreeCommandBuffers(_device, pAllocateInfo->commandPool, i,
1982                             pCommandBuffers);
1983
1984       /* From the Vulkan 1.0.66 spec:
1985        *
1986        * "vkAllocateCommandBuffers can be used to create multiple
1987        *  command buffers. If the creation of any of those command
1988        *  buffers fails, the implementation must destroy all
1989        *  successfully created command buffer objects from this
1990        *  command, set all entries of the pCommandBuffers array to
1991        *  NULL and return the error."
1992        */
1993       memset(pCommandBuffers, 0,
1994              sizeof(*pCommandBuffers) * pAllocateInfo->commandBufferCount);
1995    }
1996
1997    return result;
1998 }
1999
2000 void
2001 tu_FreeCommandBuffers(VkDevice device,
2002                       VkCommandPool commandPool,
2003                       uint32_t commandBufferCount,
2004                       const VkCommandBuffer *pCommandBuffers)
2005 {
2006    for (uint32_t i = 0; i < commandBufferCount; i++) {
2007       TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, pCommandBuffers[i]);
2008
2009       if (cmd_buffer) {
2010          if (cmd_buffer->pool) {
2011             list_del(&cmd_buffer->pool_link);
2012             list_addtail(&cmd_buffer->pool_link,
2013                          &cmd_buffer->pool->free_cmd_buffers);
2014          } else
2015             tu_cmd_buffer_destroy(cmd_buffer);
2016       }
2017    }
2018 }
2019
2020 VkResult
2021 tu_ResetCommandBuffer(VkCommandBuffer commandBuffer,
2022                       VkCommandBufferResetFlags flags)
2023 {
2024    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
2025    return tu_reset_cmd_buffer(cmd_buffer);
2026 }
2027
2028 VkResult
2029 tu_BeginCommandBuffer(VkCommandBuffer commandBuffer,
2030                       const VkCommandBufferBeginInfo *pBeginInfo)
2031 {
2032    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
2033    VkResult result = VK_SUCCESS;
2034
2035    if (cmd_buffer->status != TU_CMD_BUFFER_STATUS_INITIAL) {
2036       /* If the command buffer has already been resetted with
2037        * vkResetCommandBuffer, no need to do it again.
2038        */
2039       result = tu_reset_cmd_buffer(cmd_buffer);
2040       if (result != VK_SUCCESS)
2041          return result;
2042    }
2043
2044    memset(&cmd_buffer->state, 0, sizeof(cmd_buffer->state));
2045    cmd_buffer->usage_flags = pBeginInfo->flags;
2046
2047    tu_cs_begin(&cmd_buffer->cs);
2048    tu_cs_begin(&cmd_buffer->draw_cs);
2049    tu_cs_begin(&cmd_buffer->draw_epilogue_cs);
2050
2051    cmd_buffer->scratch_seqno = 0;
2052
2053    /* setup initial configuration into command buffer */
2054    if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
2055       switch (cmd_buffer->queue_family_index) {
2056       case TU_QUEUE_GENERAL:
2057          tu6_init_hw(cmd_buffer, &cmd_buffer->cs);
2058          break;
2059       default:
2060          break;
2061       }
2062    } else if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY &&
2063               (pBeginInfo->flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT)) {
2064       assert(pBeginInfo->pInheritanceInfo);
2065       cmd_buffer->state.pass = tu_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
2066       cmd_buffer->state.subpass = &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
2067    }
2068
2069    cmd_buffer->status = TU_CMD_BUFFER_STATUS_RECORDING;
2070
2071    return VK_SUCCESS;
2072 }
2073
2074 void
2075 tu_CmdBindVertexBuffers(VkCommandBuffer commandBuffer,
2076                         uint32_t firstBinding,
2077                         uint32_t bindingCount,
2078                         const VkBuffer *pBuffers,
2079                         const VkDeviceSize *pOffsets)
2080 {
2081    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2082
2083    assert(firstBinding + bindingCount <= MAX_VBS);
2084
2085    for (uint32_t i = 0; i < bindingCount; i++) {
2086       cmd->state.vb.buffers[firstBinding + i] =
2087          tu_buffer_from_handle(pBuffers[i]);
2088       cmd->state.vb.offsets[firstBinding + i] = pOffsets[i];
2089    }
2090
2091    /* VB states depend on VkPipelineVertexInputStateCreateInfo */
2092    cmd->state.dirty |= TU_CMD_DIRTY_VERTEX_BUFFERS;
2093 }
2094
2095 void
2096 tu_CmdBindIndexBuffer(VkCommandBuffer commandBuffer,
2097                       VkBuffer buffer,
2098                       VkDeviceSize offset,
2099                       VkIndexType indexType)
2100 {
2101    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2102    TU_FROM_HANDLE(tu_buffer, buf, buffer);
2103
2104    /* initialize/update the restart index */
2105    if (!cmd->state.index_buffer || cmd->state.index_type != indexType) {
2106       struct tu_cs *draw_cs = &cmd->draw_cs;
2107
2108       tu6_emit_restart_index(
2109          draw_cs, indexType == VK_INDEX_TYPE_UINT32 ? 0xffffffff : 0xffff);
2110
2111       tu_cs_sanity_check(draw_cs);
2112    }
2113
2114    /* track the BO */
2115    if (cmd->state.index_buffer != buf)
2116       tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
2117
2118    cmd->state.index_buffer = buf;
2119    cmd->state.index_offset = offset;
2120    cmd->state.index_type = indexType;
2121 }
2122
2123 void
2124 tu_CmdBindDescriptorSets(VkCommandBuffer commandBuffer,
2125                          VkPipelineBindPoint pipelineBindPoint,
2126                          VkPipelineLayout _layout,
2127                          uint32_t firstSet,
2128                          uint32_t descriptorSetCount,
2129                          const VkDescriptorSet *pDescriptorSets,
2130                          uint32_t dynamicOffsetCount,
2131                          const uint32_t *pDynamicOffsets)
2132 {
2133    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
2134    TU_FROM_HANDLE(tu_pipeline_layout, layout, _layout);
2135    unsigned dyn_idx = 0;
2136
2137    struct tu_descriptor_state *descriptors_state =
2138       tu_get_descriptors_state(cmd_buffer, pipelineBindPoint);
2139
2140    for (unsigned i = 0; i < descriptorSetCount; ++i) {
2141       unsigned idx = i + firstSet;
2142       TU_FROM_HANDLE(tu_descriptor_set, set, pDescriptorSets[i]);
2143
2144       descriptors_state->sets[idx] = set;
2145       descriptors_state->valid |= (1u << idx);
2146
2147       for(unsigned j = 0; j < set->layout->dynamic_offset_count; ++j, ++dyn_idx) {
2148          unsigned idx = j + layout->set[i + firstSet].dynamic_offset_start;
2149          assert(dyn_idx < dynamicOffsetCount);
2150
2151          descriptors_state->dynamic_buffers[idx] =
2152          set->dynamic_descriptors[j].va + pDynamicOffsets[dyn_idx];
2153       }
2154    }
2155
2156    cmd_buffer->state.dirty |= TU_CMD_DIRTY_DESCRIPTOR_SETS;
2157 }
2158
2159 void tu_CmdBindTransformFeedbackBuffersEXT(VkCommandBuffer commandBuffer,
2160                                            uint32_t firstBinding,
2161                                            uint32_t bindingCount,
2162                                            const VkBuffer *pBuffers,
2163                                            const VkDeviceSize *pOffsets,
2164                                            const VkDeviceSize *pSizes)
2165 {
2166    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2167    assert(firstBinding + bindingCount <= IR3_MAX_SO_BUFFERS);
2168
2169    for (uint32_t i = 0; i < bindingCount; i++) {
2170       uint32_t idx = firstBinding + i;
2171       TU_FROM_HANDLE(tu_buffer, buf, pBuffers[i]);
2172
2173       if (pOffsets[i] != 0)
2174          cmd->state.streamout_reset |= 1 << idx;
2175
2176       cmd->state.streamout_buf.buffers[idx] = buf;
2177       cmd->state.streamout_buf.offsets[idx] = pOffsets[i];
2178       cmd->state.streamout_buf.sizes[idx] = pSizes[i];
2179
2180       cmd->state.streamout_enabled |= 1 << idx;
2181    }
2182
2183    cmd->state.dirty |= TU_CMD_DIRTY_STREAMOUT_BUFFERS;
2184 }
2185
2186 void tu_CmdBeginTransformFeedbackEXT(VkCommandBuffer commandBuffer,
2187                                        uint32_t firstCounterBuffer,
2188                                        uint32_t counterBufferCount,
2189                                        const VkBuffer *pCounterBuffers,
2190                                        const VkDeviceSize *pCounterBufferOffsets)
2191 {
2192    assert(firstCounterBuffer + counterBufferCount <= IR3_MAX_SO_BUFFERS);
2193    /* TODO do something with counter buffer? */
2194 }
2195
2196 void tu_CmdEndTransformFeedbackEXT(VkCommandBuffer commandBuffer,
2197                                        uint32_t firstCounterBuffer,
2198                                        uint32_t counterBufferCount,
2199                                        const VkBuffer *pCounterBuffers,
2200                                        const VkDeviceSize *pCounterBufferOffsets)
2201 {
2202    assert(firstCounterBuffer + counterBufferCount <= IR3_MAX_SO_BUFFERS);
2203    /* TODO do something with counter buffer? */
2204
2205    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2206    cmd->state.streamout_enabled = 0;
2207 }
2208
2209 void
2210 tu_CmdPushConstants(VkCommandBuffer commandBuffer,
2211                     VkPipelineLayout layout,
2212                     VkShaderStageFlags stageFlags,
2213                     uint32_t offset,
2214                     uint32_t size,
2215                     const void *pValues)
2216 {
2217    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2218    memcpy((void*) cmd->push_constants + offset, pValues, size);
2219    cmd->state.dirty |= TU_CMD_DIRTY_PUSH_CONSTANTS;
2220 }
2221
2222 VkResult
2223 tu_EndCommandBuffer(VkCommandBuffer commandBuffer)
2224 {
2225    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
2226
2227    if (cmd_buffer->scratch_seqno) {
2228       tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->scratch_bo,
2229                      MSM_SUBMIT_BO_WRITE);
2230    }
2231
2232    if (cmd_buffer->use_vsc_data) {
2233       tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->vsc_data,
2234                      MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
2235       tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->vsc_data2,
2236                      MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
2237    }
2238
2239    for (uint32_t i = 0; i < cmd_buffer->draw_cs.bo_count; i++) {
2240       tu_bo_list_add(&cmd_buffer->bo_list, cmd_buffer->draw_cs.bos[i],
2241                      MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
2242    }
2243
2244    for (uint32_t i = 0; i < cmd_buffer->draw_epilogue_cs.bo_count; i++) {
2245       tu_bo_list_add(&cmd_buffer->bo_list, cmd_buffer->draw_epilogue_cs.bos[i],
2246                      MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
2247    }
2248
2249    for (uint32_t i = 0; i < cmd_buffer->sub_cs.bo_count; i++) {
2250       tu_bo_list_add(&cmd_buffer->bo_list, cmd_buffer->sub_cs.bos[i],
2251                      MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
2252    }
2253
2254    tu_cs_end(&cmd_buffer->cs);
2255    tu_cs_end(&cmd_buffer->draw_cs);
2256    tu_cs_end(&cmd_buffer->draw_epilogue_cs);
2257
2258    cmd_buffer->status = TU_CMD_BUFFER_STATUS_EXECUTABLE;
2259
2260    return cmd_buffer->record_result;
2261 }
2262
2263 void
2264 tu_CmdBindPipeline(VkCommandBuffer commandBuffer,
2265                    VkPipelineBindPoint pipelineBindPoint,
2266                    VkPipeline _pipeline)
2267 {
2268    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2269    TU_FROM_HANDLE(tu_pipeline, pipeline, _pipeline);
2270
2271    switch (pipelineBindPoint) {
2272    case VK_PIPELINE_BIND_POINT_GRAPHICS:
2273       cmd->state.pipeline = pipeline;
2274       cmd->state.dirty |= TU_CMD_DIRTY_PIPELINE;
2275       break;
2276    case VK_PIPELINE_BIND_POINT_COMPUTE:
2277       cmd->state.compute_pipeline = pipeline;
2278       cmd->state.dirty |= TU_CMD_DIRTY_COMPUTE_PIPELINE;
2279       break;
2280    default:
2281       unreachable("unrecognized pipeline bind point");
2282       break;
2283    }
2284
2285    tu_bo_list_add(&cmd->bo_list, &pipeline->program.binary_bo,
2286                   MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
2287    for (uint32_t i = 0; i < pipeline->cs.bo_count; i++) {
2288       tu_bo_list_add(&cmd->bo_list, pipeline->cs.bos[i],
2289                      MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
2290    }
2291 }
2292
2293 void
2294 tu_CmdSetViewport(VkCommandBuffer commandBuffer,
2295                   uint32_t firstViewport,
2296                   uint32_t viewportCount,
2297                   const VkViewport *pViewports)
2298 {
2299    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2300    struct tu_cs *draw_cs = &cmd->draw_cs;
2301
2302    assert(firstViewport == 0 && viewportCount == 1);
2303    tu6_emit_viewport(draw_cs, pViewports);
2304
2305    tu_cs_sanity_check(draw_cs);
2306 }
2307
2308 void
2309 tu_CmdSetScissor(VkCommandBuffer commandBuffer,
2310                  uint32_t firstScissor,
2311                  uint32_t scissorCount,
2312                  const VkRect2D *pScissors)
2313 {
2314    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2315    struct tu_cs *draw_cs = &cmd->draw_cs;
2316
2317    assert(firstScissor == 0 && scissorCount == 1);
2318    tu6_emit_scissor(draw_cs, pScissors);
2319
2320    tu_cs_sanity_check(draw_cs);
2321 }
2322
2323 void
2324 tu_CmdSetLineWidth(VkCommandBuffer commandBuffer, float lineWidth)
2325 {
2326    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2327
2328    cmd->state.dynamic.line_width = lineWidth;
2329
2330    /* line width depends on VkPipelineRasterizationStateCreateInfo */
2331    cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
2332 }
2333
2334 void
2335 tu_CmdSetDepthBias(VkCommandBuffer commandBuffer,
2336                    float depthBiasConstantFactor,
2337                    float depthBiasClamp,
2338                    float depthBiasSlopeFactor)
2339 {
2340    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2341    struct tu_cs *draw_cs = &cmd->draw_cs;
2342
2343    tu6_emit_depth_bias(draw_cs, depthBiasConstantFactor, depthBiasClamp,
2344                        depthBiasSlopeFactor);
2345
2346    tu_cs_sanity_check(draw_cs);
2347 }
2348
2349 void
2350 tu_CmdSetBlendConstants(VkCommandBuffer commandBuffer,
2351                         const float blendConstants[4])
2352 {
2353    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2354    struct tu_cs *draw_cs = &cmd->draw_cs;
2355
2356    tu6_emit_blend_constants(draw_cs, blendConstants);
2357
2358    tu_cs_sanity_check(draw_cs);
2359 }
2360
2361 void
2362 tu_CmdSetDepthBounds(VkCommandBuffer commandBuffer,
2363                      float minDepthBounds,
2364                      float maxDepthBounds)
2365 {
2366 }
2367
2368 void
2369 tu_CmdSetStencilCompareMask(VkCommandBuffer commandBuffer,
2370                             VkStencilFaceFlags faceMask,
2371                             uint32_t compareMask)
2372 {
2373    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2374
2375    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
2376       cmd->state.dynamic.stencil_compare_mask.front = compareMask;
2377    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
2378       cmd->state.dynamic.stencil_compare_mask.back = compareMask;
2379
2380    /* the front/back compare masks must be updated together */
2381    cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
2382 }
2383
2384 void
2385 tu_CmdSetStencilWriteMask(VkCommandBuffer commandBuffer,
2386                           VkStencilFaceFlags faceMask,
2387                           uint32_t writeMask)
2388 {
2389    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2390
2391    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
2392       cmd->state.dynamic.stencil_write_mask.front = writeMask;
2393    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
2394       cmd->state.dynamic.stencil_write_mask.back = writeMask;
2395
2396    /* the front/back write masks must be updated together */
2397    cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
2398 }
2399
2400 void
2401 tu_CmdSetStencilReference(VkCommandBuffer commandBuffer,
2402                           VkStencilFaceFlags faceMask,
2403                           uint32_t reference)
2404 {
2405    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2406
2407    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
2408       cmd->state.dynamic.stencil_reference.front = reference;
2409    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
2410       cmd->state.dynamic.stencil_reference.back = reference;
2411
2412    /* the front/back references must be updated together */
2413    cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
2414 }
2415
2416 void
2417 tu_CmdExecuteCommands(VkCommandBuffer commandBuffer,
2418                       uint32_t commandBufferCount,
2419                       const VkCommandBuffer *pCmdBuffers)
2420 {
2421    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2422    VkResult result;
2423
2424    assert(commandBufferCount > 0);
2425
2426    for (uint32_t i = 0; i < commandBufferCount; i++) {
2427       TU_FROM_HANDLE(tu_cmd_buffer, secondary, pCmdBuffers[i]);
2428
2429       result = tu_bo_list_merge(&cmd->bo_list, &secondary->bo_list);
2430       if (result != VK_SUCCESS) {
2431          cmd->record_result = result;
2432          break;
2433       }
2434
2435       if (secondary->usage_flags &
2436           VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
2437          assert(tu_cs_is_empty(&secondary->cs));
2438
2439          result = tu_cs_add_entries(&cmd->draw_cs, &secondary->draw_cs);
2440          if (result != VK_SUCCESS) {
2441             cmd->record_result = result;
2442             break;
2443          }
2444
2445          result = tu_cs_add_entries(&cmd->draw_epilogue_cs,
2446                &secondary->draw_epilogue_cs);
2447          if (result != VK_SUCCESS) {
2448             cmd->record_result = result;
2449             break;
2450          }
2451       } else {
2452          assert(tu_cs_is_empty(&secondary->draw_cs));
2453          assert(tu_cs_is_empty(&secondary->draw_epilogue_cs));
2454
2455          for (uint32_t j = 0; j < secondary->cs.bo_count; j++) {
2456             tu_bo_list_add(&cmd->bo_list, secondary->cs.bos[j],
2457                            MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
2458          }
2459
2460          tu_cs_emit_call(&cmd->cs, &secondary->cs);
2461       }
2462    }
2463    cmd->state.dirty = ~0u; /* TODO: set dirty only what needs to be */
2464 }
2465
2466 VkResult
2467 tu_CreateCommandPool(VkDevice _device,
2468                      const VkCommandPoolCreateInfo *pCreateInfo,
2469                      const VkAllocationCallbacks *pAllocator,
2470                      VkCommandPool *pCmdPool)
2471 {
2472    TU_FROM_HANDLE(tu_device, device, _device);
2473    struct tu_cmd_pool *pool;
2474
2475    pool = vk_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
2476                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2477    if (pool == NULL)
2478       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2479
2480    if (pAllocator)
2481       pool->alloc = *pAllocator;
2482    else
2483       pool->alloc = device->alloc;
2484
2485    list_inithead(&pool->cmd_buffers);
2486    list_inithead(&pool->free_cmd_buffers);
2487
2488    pool->queue_family_index = pCreateInfo->queueFamilyIndex;
2489
2490    *pCmdPool = tu_cmd_pool_to_handle(pool);
2491
2492    return VK_SUCCESS;
2493 }
2494
2495 void
2496 tu_DestroyCommandPool(VkDevice _device,
2497                       VkCommandPool commandPool,
2498                       const VkAllocationCallbacks *pAllocator)
2499 {
2500    TU_FROM_HANDLE(tu_device, device, _device);
2501    TU_FROM_HANDLE(tu_cmd_pool, pool, commandPool);
2502
2503    if (!pool)
2504       return;
2505
2506    list_for_each_entry_safe(struct tu_cmd_buffer, cmd_buffer,
2507                             &pool->cmd_buffers, pool_link)
2508    {
2509       tu_cmd_buffer_destroy(cmd_buffer);
2510    }
2511
2512    list_for_each_entry_safe(struct tu_cmd_buffer, cmd_buffer,
2513                             &pool->free_cmd_buffers, pool_link)
2514    {
2515       tu_cmd_buffer_destroy(cmd_buffer);
2516    }
2517
2518    vk_free2(&device->alloc, pAllocator, pool);
2519 }
2520
2521 VkResult
2522 tu_ResetCommandPool(VkDevice device,
2523                     VkCommandPool commandPool,
2524                     VkCommandPoolResetFlags flags)
2525 {
2526    TU_FROM_HANDLE(tu_cmd_pool, pool, commandPool);
2527    VkResult result;
2528
2529    list_for_each_entry(struct tu_cmd_buffer, cmd_buffer, &pool->cmd_buffers,
2530                        pool_link)
2531    {
2532       result = tu_reset_cmd_buffer(cmd_buffer);
2533       if (result != VK_SUCCESS)
2534          return result;
2535    }
2536
2537    return VK_SUCCESS;
2538 }
2539
2540 void
2541 tu_TrimCommandPool(VkDevice device,
2542                    VkCommandPool commandPool,
2543                    VkCommandPoolTrimFlags flags)
2544 {
2545    TU_FROM_HANDLE(tu_cmd_pool, pool, commandPool);
2546
2547    if (!pool)
2548       return;
2549
2550    list_for_each_entry_safe(struct tu_cmd_buffer, cmd_buffer,
2551                             &pool->free_cmd_buffers, pool_link)
2552    {
2553       tu_cmd_buffer_destroy(cmd_buffer);
2554    }
2555 }
2556
2557 void
2558 tu_CmdBeginRenderPass(VkCommandBuffer commandBuffer,
2559                       const VkRenderPassBeginInfo *pRenderPassBegin,
2560                       VkSubpassContents contents)
2561 {
2562    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2563    TU_FROM_HANDLE(tu_render_pass, pass, pRenderPassBegin->renderPass);
2564    TU_FROM_HANDLE(tu_framebuffer, fb, pRenderPassBegin->framebuffer);
2565
2566    cmd->state.pass = pass;
2567    cmd->state.subpass = pass->subpasses;
2568    cmd->state.framebuffer = fb;
2569
2570    tu_cmd_update_tiling_config(cmd, &pRenderPassBegin->renderArea);
2571    tu_cmd_prepare_tile_store_ib(cmd);
2572
2573    tu_emit_load_clear(cmd, pRenderPassBegin);
2574
2575    tu6_emit_zs(cmd, cmd->state.subpass, &cmd->draw_cs);
2576    tu6_emit_mrt(cmd, cmd->state.subpass, &cmd->draw_cs);
2577    tu6_emit_msaa(cmd, cmd->state.subpass, &cmd->draw_cs);
2578    tu6_emit_render_cntl(cmd, cmd->state.subpass, &cmd->draw_cs, false);
2579
2580    /* note: use_hw_binning only checks tiling config */
2581    if (use_hw_binning(cmd))
2582       cmd->use_vsc_data = true;
2583
2584    for (uint32_t i = 0; i < fb->attachment_count; ++i) {
2585       const struct tu_image_view *iview = fb->attachments[i].attachment;
2586       tu_bo_list_add(&cmd->bo_list, iview->image->bo,
2587                      MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
2588    }
2589 }
2590
2591 void
2592 tu_CmdBeginRenderPass2(VkCommandBuffer commandBuffer,
2593                        const VkRenderPassBeginInfo *pRenderPassBeginInfo,
2594                        const VkSubpassBeginInfoKHR *pSubpassBeginInfo)
2595 {
2596    tu_CmdBeginRenderPass(commandBuffer, pRenderPassBeginInfo,
2597                          pSubpassBeginInfo->contents);
2598 }
2599
2600 void
2601 tu_CmdNextSubpass(VkCommandBuffer commandBuffer, VkSubpassContents contents)
2602 {
2603    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
2604    const struct tu_render_pass *pass = cmd->state.pass;
2605    struct tu_cs *cs = &cmd->draw_cs;
2606
2607    const struct tu_subpass *subpass = cmd->state.subpass++;
2608    /* TODO:
2609     * if msaa samples change between subpasses,
2610     * attachment store is broken for some attachments
2611     */
2612    if (subpass->resolve_attachments) {
2613       tu6_emit_blit_scissor(cmd, cs, true);
2614       for (unsigned i = 0; i < subpass->color_count; i++) {
2615          uint32_t a = subpass->resolve_attachments[i].attachment;
2616          if (a != VK_ATTACHMENT_UNUSED) {
2617             tu6_emit_resolve(cmd, cs, a,
2618                              subpass->color_attachments[i].attachment);
2619          }
2620       }
2621    }
2622
2623    /* invalidate because reading input attachments will cache GMEM and
2624     * the cache isn''t updated when GMEM is written
2625     * TODO: is there a no-cache bit for textures?
2626     */
2627    if (cmd->state.subpass->input_count)
2628       tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE, false);
2629
2630    /* emit mrt/zs/msaa/ubwc state for the subpass that is starting */
2631    tu6_emit_zs(cmd, cmd->state.subpass, cs);
2632    tu6_emit_mrt(cmd, cmd->state.subpass, cs);
2633    tu6_emit_msaa(cmd, cmd->state.subpass, cs);
2634    tu6_emit_render_cntl(cmd, cmd->state.subpass, cs, false);
2635
2636    /* Emit flushes so that input attachments will read the correct value. This
2637     * is for sysmem only, although it shouldn't do much harm on gmem.
2638     */
2639    tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_COLOR_TS, true);
2640    tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_DEPTH_TS, true);
2641
2642    /* TODO:
2643     * since we don't know how to do GMEM->GMEM resolve,
2644     * resolve attachments are resolved to memory then loaded to GMEM again if needed
2645     */
2646    if (subpass->resolve_attachments) {
2647       for (unsigned i = 0; i < subpass->color_count; i++) {
2648          uint32_t a = subpass->resolve_attachments[i].attachment;
2649          if (a != VK_ATTACHMENT_UNUSED && pass->attachments[a].gmem_offset >= 0) {
2650                tu_finishme("missing GMEM->GMEM resolve, performance will suffer\n");
2651                tu6_emit_predicated_blit(cmd, cs, a, a, false);
2652          }
2653       }
2654    }
2655 }
2656
2657 void
2658 tu_CmdNextSubpass2(VkCommandBuffer commandBuffer,
2659                    const VkSubpassBeginInfoKHR *pSubpassBeginInfo,
2660                    const VkSubpassEndInfoKHR *pSubpassEndInfo)
2661 {
2662    tu_CmdNextSubpass(commandBuffer, pSubpassBeginInfo->contents);
2663 }
2664
2665 struct tu_draw_info
2666 {
2667    /**
2668     * Number of vertices.
2669     */
2670    uint32_t count;
2671
2672    /**
2673     * Index of the first vertex.
2674     */
2675    int32_t vertex_offset;
2676
2677    /**
2678     * First instance id.
2679     */
2680    uint32_t first_instance;
2681
2682    /**
2683     * Number of instances.
2684     */
2685    uint32_t instance_count;
2686
2687    /**
2688     * First index (indexed draws only).
2689     */
2690    uint32_t first_index;
2691
2692    /**
2693     * Whether it's an indexed draw.
2694     */
2695    bool indexed;
2696
2697    /**
2698     * Indirect draw parameters resource.
2699     */
2700    struct tu_buffer *indirect;
2701    uint64_t indirect_offset;
2702    uint32_t stride;
2703
2704    /**
2705     * Draw count parameters resource.
2706     */
2707    struct tu_buffer *count_buffer;
2708    uint64_t count_buffer_offset;
2709
2710    /**
2711     * Stream output parameters resource.
2712     */
2713    struct tu_buffer *streamout_buffer;
2714    uint64_t streamout_buffer_offset;
2715 };
2716
2717 #define ENABLE_ALL (CP_SET_DRAW_STATE__0_BINNING | CP_SET_DRAW_STATE__0_GMEM | CP_SET_DRAW_STATE__0_SYSMEM)
2718 #define ENABLE_DRAW (CP_SET_DRAW_STATE__0_GMEM | CP_SET_DRAW_STATE__0_SYSMEM)
2719
2720 enum tu_draw_state_group_id
2721 {
2722    TU_DRAW_STATE_PROGRAM,
2723    TU_DRAW_STATE_PROGRAM_BINNING,
2724    TU_DRAW_STATE_VI,
2725    TU_DRAW_STATE_VI_BINNING,
2726    TU_DRAW_STATE_VP,
2727    TU_DRAW_STATE_RAST,
2728    TU_DRAW_STATE_DS,
2729    TU_DRAW_STATE_BLEND,
2730    TU_DRAW_STATE_VS_CONST,
2731    TU_DRAW_STATE_FS_CONST,
2732    TU_DRAW_STATE_VS_TEX,
2733    TU_DRAW_STATE_FS_TEX_SYSMEM,
2734    TU_DRAW_STATE_FS_TEX_GMEM,
2735    TU_DRAW_STATE_FS_IBO,
2736    TU_DRAW_STATE_VS_PARAMS,
2737
2738    TU_DRAW_STATE_COUNT,
2739 };
2740
2741 struct tu_draw_state_group
2742 {
2743    enum tu_draw_state_group_id id;
2744    uint32_t enable_mask;
2745    struct tu_cs_entry ib;
2746 };
2747
2748 const static struct tu_sampler*
2749 sampler_ptr(struct tu_descriptor_state *descriptors_state,
2750             const struct tu_descriptor_map *map, unsigned i,
2751             unsigned array_index)
2752 {
2753    assert(descriptors_state->valid & (1 << map->set[i]));
2754
2755    struct tu_descriptor_set *set = descriptors_state->sets[map->set[i]];
2756    assert(map->binding[i] < set->layout->binding_count);
2757
2758    const struct tu_descriptor_set_binding_layout *layout =
2759       &set->layout->binding[map->binding[i]];
2760
2761    if (layout->immutable_samplers_offset) {
2762       const struct tu_sampler *immutable_samplers =
2763          tu_immutable_samplers(set->layout, layout);
2764
2765       return &immutable_samplers[array_index];
2766    }
2767
2768    switch (layout->type) {
2769    case VK_DESCRIPTOR_TYPE_SAMPLER:
2770       return (struct tu_sampler*) &set->mapped_ptr[layout->offset / 4];
2771    case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
2772       return (struct tu_sampler*) &set->mapped_ptr[layout->offset / 4 + A6XX_TEX_CONST_DWORDS +
2773                                                    array_index *
2774                                                    (A6XX_TEX_CONST_DWORDS +
2775                                                     sizeof(struct tu_sampler) / 4)];
2776    default:
2777       unreachable("unimplemented descriptor type");
2778       break;
2779    }
2780 }
2781
2782 static void
2783 write_tex_const(struct tu_cmd_buffer *cmd,
2784                 uint32_t *dst,
2785                 struct tu_descriptor_state *descriptors_state,
2786                 const struct tu_descriptor_map *map,
2787                 unsigned i, unsigned array_index, bool is_sysmem)
2788 {
2789    assert(descriptors_state->valid & (1 << map->set[i]));
2790
2791    struct tu_descriptor_set *set = descriptors_state->sets[map->set[i]];
2792    assert(map->binding[i] < set->layout->binding_count);
2793
2794    const struct tu_descriptor_set_binding_layout *layout =
2795       &set->layout->binding[map->binding[i]];
2796
2797    switch (layout->type) {
2798    case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
2799    case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
2800    case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
2801    case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
2802       memcpy(dst, &set->mapped_ptr[layout->offset / 4 +
2803                                    array_index * A6XX_TEX_CONST_DWORDS],
2804              A6XX_TEX_CONST_DWORDS * 4);
2805       break;
2806    case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
2807       memcpy(dst, &set->mapped_ptr[layout->offset / 4 +
2808                                    array_index *
2809                                    (A6XX_TEX_CONST_DWORDS +
2810                                     sizeof(struct tu_sampler) / 4)],
2811              A6XX_TEX_CONST_DWORDS * 4);
2812       break;
2813    default:
2814       unreachable("unimplemented descriptor type");
2815       break;
2816    }
2817
2818    if (layout->type == VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT && !is_sysmem) {
2819       const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
2820       uint32_t a = cmd->state.subpass->input_attachments[map->value[i] +
2821                                                          array_index].attachment;
2822       const struct tu_render_pass_attachment *att = &cmd->state.pass->attachments[a];
2823
2824       assert(att->gmem_offset >= 0);
2825
2826       dst[0] &= ~(A6XX_TEX_CONST_0_SWAP__MASK | A6XX_TEX_CONST_0_TILE_MODE__MASK);
2827       dst[0] |= A6XX_TEX_CONST_0_TILE_MODE(TILE6_2);
2828       dst[2] &= ~(A6XX_TEX_CONST_2_TYPE__MASK | A6XX_TEX_CONST_2_PITCH__MASK);
2829       dst[2] |=
2830          A6XX_TEX_CONST_2_TYPE(A6XX_TEX_2D) |
2831          A6XX_TEX_CONST_2_PITCH(tiling->tile0.extent.width * att->cpp);
2832       dst[3] = 0;
2833       dst[4] = cmd->device->physical_device->gmem_base + att->gmem_offset;
2834       dst[5] = A6XX_TEX_CONST_5_DEPTH(1);
2835       for (unsigned i = 6; i < A6XX_TEX_CONST_DWORDS; i++)
2836          dst[i] = 0;
2837
2838       if (cmd->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY)
2839          tu_finishme("patch input attachment pitch for secondary cmd buffer");
2840    }
2841 }
2842
2843 static void
2844 write_image_ibo(struct tu_cmd_buffer *cmd,
2845                 uint32_t *dst,
2846                 struct tu_descriptor_state *descriptors_state,
2847                 const struct tu_descriptor_map *map,
2848                 unsigned i, unsigned array_index)
2849 {
2850    assert(descriptors_state->valid & (1 << map->set[i]));
2851
2852    struct tu_descriptor_set *set = descriptors_state->sets[map->set[i]];
2853    assert(map->binding[i] < set->layout->binding_count);
2854
2855    const struct tu_descriptor_set_binding_layout *layout =
2856       &set->layout->binding[map->binding[i]];
2857
2858    assert(layout->type == VK_DESCRIPTOR_TYPE_STORAGE_IMAGE);
2859
2860    memcpy(dst, &set->mapped_ptr[layout->offset / 4 +
2861                                 (array_index * 2 + 1) * A6XX_TEX_CONST_DWORDS],
2862           A6XX_TEX_CONST_DWORDS * 4);
2863 }
2864
2865 static uint64_t
2866 buffer_ptr(struct tu_descriptor_state *descriptors_state,
2867            const struct tu_descriptor_map *map,
2868            unsigned i, unsigned array_index)
2869 {
2870    assert(descriptors_state->valid & (1 << map->set[i]));
2871
2872    struct tu_descriptor_set *set = descriptors_state->sets[map->set[i]];
2873    assert(map->binding[i] < set->layout->binding_count);
2874
2875    const struct tu_descriptor_set_binding_layout *layout =
2876       &set->layout->binding[map->binding[i]];
2877
2878    switch (layout->type) {
2879    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
2880    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
2881       return descriptors_state->dynamic_buffers[layout->dynamic_offset_offset +
2882                                                 array_index];
2883    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
2884    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
2885       return (uint64_t) set->mapped_ptr[layout->offset / 4 + array_index * 2 + 1] << 32 |
2886                         set->mapped_ptr[layout->offset / 4 + array_index * 2];
2887    default:
2888       unreachable("unimplemented descriptor type");
2889       break;
2890    }
2891 }
2892
2893 static inline uint32_t
2894 tu6_stage2opcode(gl_shader_stage type)
2895 {
2896    switch (type) {
2897    case MESA_SHADER_VERTEX:
2898    case MESA_SHADER_TESS_CTRL:
2899    case MESA_SHADER_TESS_EVAL:
2900    case MESA_SHADER_GEOMETRY:
2901       return CP_LOAD_STATE6_GEOM;
2902    case MESA_SHADER_FRAGMENT:
2903    case MESA_SHADER_COMPUTE:
2904    case MESA_SHADER_KERNEL:
2905       return CP_LOAD_STATE6_FRAG;
2906    default:
2907       unreachable("bad shader type");
2908    }
2909 }
2910
2911 static inline enum a6xx_state_block
2912 tu6_stage2shadersb(gl_shader_stage type)
2913 {
2914    switch (type) {
2915    case MESA_SHADER_VERTEX:
2916       return SB6_VS_SHADER;
2917    case MESA_SHADER_FRAGMENT:
2918       return SB6_FS_SHADER;
2919    case MESA_SHADER_COMPUTE:
2920    case MESA_SHADER_KERNEL:
2921       return SB6_CS_SHADER;
2922    default:
2923       unreachable("bad shader type");
2924       return ~0;
2925    }
2926 }
2927
2928 static void
2929 tu6_emit_user_consts(struct tu_cs *cs, const struct tu_pipeline *pipeline,
2930                      struct tu_descriptor_state *descriptors_state,
2931                      gl_shader_stage type,
2932                      uint32_t *push_constants)
2933 {
2934    const struct tu_program_descriptor_linkage *link =
2935       &pipeline->program.link[type];
2936    const struct ir3_ubo_analysis_state *state = &link->ubo_state;
2937
2938    for (uint32_t i = 0; i < ARRAY_SIZE(state->range); i++) {
2939       if (state->range[i].start < state->range[i].end) {
2940          uint32_t size = state->range[i].end - state->range[i].start;
2941          uint32_t offset = state->range[i].start;
2942
2943          /* and even if the start of the const buffer is before
2944           * first_immediate, the end may not be:
2945           */
2946          size = MIN2(size, (16 * link->constlen) - state->range[i].offset);
2947
2948          if (size == 0)
2949             continue;
2950
2951          /* things should be aligned to vec4: */
2952          debug_assert((state->range[i].offset % 16) == 0);
2953          debug_assert((size % 16) == 0);
2954          debug_assert((offset % 16) == 0);
2955
2956          if (i == 0) {
2957             /* push constants */
2958             tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3 + (size / 4));
2959             tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(state->range[i].offset / 16) |
2960                   CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
2961                   CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
2962                   CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
2963                   CP_LOAD_STATE6_0_NUM_UNIT(size / 16));
2964             tu_cs_emit(cs, 0);
2965             tu_cs_emit(cs, 0);
2966             for (unsigned i = 0; i < size / 4; i++)
2967                tu_cs_emit(cs, push_constants[i + offset / 4]);
2968             continue;
2969          }
2970
2971          /* Look through the UBO map to find our UBO index, and get the VA for
2972           * that UBO.
2973           */
2974          uint64_t va = 0;
2975          uint32_t ubo_idx = i - 1;
2976          uint32_t ubo_map_base = 0;
2977          for (int j = 0; j < link->ubo_map.num; j++) {
2978             if (ubo_idx >= ubo_map_base &&
2979                 ubo_idx < ubo_map_base + link->ubo_map.array_size[j]) {
2980                va = buffer_ptr(descriptors_state, &link->ubo_map, j,
2981                                ubo_idx - ubo_map_base);
2982                break;
2983             }
2984             ubo_map_base += link->ubo_map.array_size[j];
2985          }
2986          assert(va);
2987
2988          tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3);
2989          tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(state->range[i].offset / 16) |
2990                CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
2991                CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
2992                CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
2993                CP_LOAD_STATE6_0_NUM_UNIT(size / 16));
2994          tu_cs_emit_qw(cs, va + offset);
2995       }
2996    }
2997 }
2998
2999 static void
3000 tu6_emit_ubos(struct tu_cs *cs, const struct tu_pipeline *pipeline,
3001               struct tu_descriptor_state *descriptors_state,
3002               gl_shader_stage type)
3003 {
3004    const struct tu_program_descriptor_linkage *link =
3005       &pipeline->program.link[type];
3006
3007    uint32_t num = MIN2(link->ubo_map.num_desc, link->const_state.num_ubos);
3008    uint32_t anum = align(num, 2);
3009
3010    if (!num)
3011       return;
3012
3013    tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3 + (2 * anum));
3014    tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(link->const_state.offsets.ubo) |
3015          CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
3016          CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
3017          CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
3018          CP_LOAD_STATE6_0_NUM_UNIT(anum/2));
3019    tu_cs_emit(cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
3020    tu_cs_emit(cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
3021
3022    unsigned emitted = 0;
3023    for (unsigned i = 0; emitted < num && i < link->ubo_map.num; i++) {
3024       for (unsigned j = 0; emitted < num && j < link->ubo_map.array_size[i]; j++) {
3025          tu_cs_emit_qw(cs, buffer_ptr(descriptors_state, &link->ubo_map, i, j));
3026          emitted++;
3027       }
3028    }
3029
3030    for (; emitted < anum; emitted++) {
3031       tu_cs_emit(cs, 0xffffffff);
3032       tu_cs_emit(cs, 0xffffffff);
3033    }
3034 }
3035
3036 static struct tu_cs_entry
3037 tu6_emit_consts(struct tu_cmd_buffer *cmd,
3038                 const struct tu_pipeline *pipeline,
3039                 struct tu_descriptor_state *descriptors_state,
3040                 gl_shader_stage type)
3041 {
3042    struct tu_cs cs;
3043    tu_cs_begin_sub_stream(&cmd->sub_cs, 512, &cs); /* TODO: maximum size? */
3044
3045    tu6_emit_user_consts(&cs, pipeline, descriptors_state, type, cmd->push_constants);
3046    tu6_emit_ubos(&cs, pipeline, descriptors_state, type);
3047
3048    return tu_cs_end_sub_stream(&cmd->sub_cs, &cs);
3049 }
3050
3051 static VkResult
3052 tu6_emit_vs_params(struct tu_cmd_buffer *cmd,
3053                    const struct tu_draw_info *draw,
3054                    struct tu_cs_entry *entry)
3055 {
3056    /* TODO: fill out more than just base instance */
3057    const struct tu_program_descriptor_linkage *link =
3058       &cmd->state.pipeline->program.link[MESA_SHADER_VERTEX];
3059    const struct ir3_const_state *const_state = &link->const_state;
3060    struct tu_cs cs;
3061
3062    if (const_state->offsets.driver_param >= link->constlen) {
3063       *entry = (struct tu_cs_entry) {};
3064       return VK_SUCCESS;
3065    }
3066
3067    VkResult result = tu_cs_begin_sub_stream(&cmd->sub_cs, 8, &cs);
3068    if (result != VK_SUCCESS)
3069       return result;
3070
3071    tu_cs_emit_pkt7(&cs, CP_LOAD_STATE6_GEOM, 3 + 4);
3072    tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(const_state->offsets.driver_param) |
3073          CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
3074          CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
3075          CP_LOAD_STATE6_0_STATE_BLOCK(SB6_VS_SHADER) |
3076          CP_LOAD_STATE6_0_NUM_UNIT(1));
3077    tu_cs_emit(&cs, 0);
3078    tu_cs_emit(&cs, 0);
3079
3080    STATIC_ASSERT(IR3_DP_INSTID_BASE == 2);
3081
3082    tu_cs_emit(&cs, 0);
3083    tu_cs_emit(&cs, 0);
3084    tu_cs_emit(&cs, draw->first_instance);
3085    tu_cs_emit(&cs, 0);
3086
3087    *entry = tu_cs_end_sub_stream(&cmd->sub_cs, &cs);
3088    return VK_SUCCESS;
3089 }
3090
3091 static VkResult
3092 tu6_emit_textures(struct tu_cmd_buffer *cmd,
3093                   const struct tu_pipeline *pipeline,
3094                   struct tu_descriptor_state *descriptors_state,
3095                   gl_shader_stage type,
3096                   struct tu_cs_entry *entry,
3097                   bool *needs_border,
3098                   bool is_sysmem)
3099 {
3100    struct tu_cs *draw_state = &cmd->sub_cs;
3101    const struct tu_program_descriptor_linkage *link =
3102       &pipeline->program.link[type];
3103    VkResult result;
3104
3105    if (link->texture_map.num_desc == 0 && link->sampler_map.num_desc == 0) {
3106       *entry = (struct tu_cs_entry) {};
3107       return VK_SUCCESS;
3108    }
3109
3110    /* allocate and fill texture state */
3111    struct ts_cs_memory tex_const;
3112    result = tu_cs_alloc(draw_state, link->texture_map.num_desc,
3113                         A6XX_TEX_CONST_DWORDS, &tex_const);
3114    if (result != VK_SUCCESS)
3115       return result;
3116
3117    int tex_index = 0;
3118    for (unsigned i = 0; i < link->texture_map.num; i++) {
3119       for (int j = 0; j < link->texture_map.array_size[i]; j++) {
3120          write_tex_const(cmd,
3121                          &tex_const.map[A6XX_TEX_CONST_DWORDS * tex_index++],
3122                          descriptors_state, &link->texture_map, i, j,
3123                          is_sysmem);
3124       }
3125    }
3126
3127    /* allocate and fill sampler state */
3128    struct ts_cs_memory tex_samp = { 0 };
3129    if (link->sampler_map.num_desc) {
3130       result = tu_cs_alloc(draw_state, link->sampler_map.num_desc,
3131                            A6XX_TEX_SAMP_DWORDS, &tex_samp);
3132       if (result != VK_SUCCESS)
3133          return result;
3134
3135       int sampler_index = 0;
3136       for (unsigned i = 0; i < link->sampler_map.num; i++) {
3137          for (int j = 0; j < link->sampler_map.array_size[i]; j++) {
3138             const struct tu_sampler *sampler = sampler_ptr(descriptors_state,
3139                                                            &link->sampler_map,
3140                                                            i, j);
3141             memcpy(&tex_samp.map[A6XX_TEX_SAMP_DWORDS * sampler_index++],
3142                    sampler->state, sizeof(sampler->state));
3143             *needs_border |= sampler->needs_border;
3144          }
3145       }
3146    }
3147
3148    unsigned tex_samp_reg, tex_const_reg, tex_count_reg;
3149    enum a6xx_state_block sb;
3150
3151    switch (type) {
3152    case MESA_SHADER_VERTEX:
3153       sb = SB6_VS_TEX;
3154       tex_samp_reg = REG_A6XX_SP_VS_TEX_SAMP_LO;
3155       tex_const_reg = REG_A6XX_SP_VS_TEX_CONST_LO;
3156       tex_count_reg = REG_A6XX_SP_VS_TEX_COUNT;
3157       break;
3158    case MESA_SHADER_FRAGMENT:
3159       sb = SB6_FS_TEX;
3160       tex_samp_reg = REG_A6XX_SP_FS_TEX_SAMP_LO;
3161       tex_const_reg = REG_A6XX_SP_FS_TEX_CONST_LO;
3162       tex_count_reg = REG_A6XX_SP_FS_TEX_COUNT;
3163       break;
3164    case MESA_SHADER_COMPUTE:
3165       sb = SB6_CS_TEX;
3166       tex_samp_reg = REG_A6XX_SP_CS_TEX_SAMP_LO;
3167       tex_const_reg = REG_A6XX_SP_CS_TEX_CONST_LO;
3168       tex_count_reg = REG_A6XX_SP_CS_TEX_COUNT;
3169       break;
3170    default:
3171       unreachable("bad state block");
3172    }
3173
3174    struct tu_cs cs;
3175    result = tu_cs_begin_sub_stream(draw_state, 16, &cs);
3176    if (result != VK_SUCCESS)
3177       return result;
3178
3179    if (link->sampler_map.num_desc) {
3180       /* output sampler state: */
3181       tu_cs_emit_pkt7(&cs, tu6_stage2opcode(type), 3);
3182       tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(0) |
3183                  CP_LOAD_STATE6_0_STATE_TYPE(ST6_SHADER) |
3184                  CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
3185                  CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
3186                  CP_LOAD_STATE6_0_NUM_UNIT(link->sampler_map.num_desc));
3187       tu_cs_emit_qw(&cs, tex_samp.iova); /* SRC_ADDR_LO/HI */
3188
3189       tu_cs_emit_pkt4(&cs, tex_samp_reg, 2);
3190       tu_cs_emit_qw(&cs, tex_samp.iova); /* SRC_ADDR_LO/HI */
3191    }
3192
3193    /* emit texture state: */
3194    tu_cs_emit_pkt7(&cs, tu6_stage2opcode(type), 3);
3195    tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(0) |
3196       CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
3197       CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
3198       CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
3199       CP_LOAD_STATE6_0_NUM_UNIT(link->texture_map.num_desc));
3200    tu_cs_emit_qw(&cs, tex_const.iova); /* SRC_ADDR_LO/HI */
3201
3202    tu_cs_emit_pkt4(&cs, tex_const_reg, 2);
3203    tu_cs_emit_qw(&cs, tex_const.iova); /* SRC_ADDR_LO/HI */
3204
3205    tu_cs_emit_pkt4(&cs, tex_count_reg, 1);
3206    tu_cs_emit(&cs, link->texture_map.num_desc);
3207
3208    *entry = tu_cs_end_sub_stream(draw_state, &cs);
3209    return VK_SUCCESS;
3210 }
3211
3212 static VkResult
3213 tu6_emit_ibo(struct tu_cmd_buffer *cmd,
3214              const struct tu_pipeline *pipeline,
3215              struct tu_descriptor_state *descriptors_state,
3216              gl_shader_stage type,
3217              struct tu_cs_entry *entry)
3218 {
3219    struct tu_cs *draw_state = &cmd->sub_cs;
3220    const struct tu_program_descriptor_linkage *link =
3221       &pipeline->program.link[type];
3222    VkResult result;
3223
3224    unsigned num_desc = link->ssbo_map.num_desc + link->image_map.num_desc;
3225
3226    if (num_desc == 0) {
3227       *entry = (struct tu_cs_entry) {};
3228       return VK_SUCCESS;
3229    }
3230
3231    struct ts_cs_memory ibo_const;
3232    result = tu_cs_alloc(draw_state, num_desc,
3233                         A6XX_TEX_CONST_DWORDS, &ibo_const);
3234    if (result != VK_SUCCESS)
3235       return result;
3236
3237    int ssbo_index = 0;
3238    for (unsigned i = 0; i < link->ssbo_map.num; i++) {
3239       for (int j = 0; j < link->ssbo_map.array_size[i]; j++) {
3240          uint32_t *dst = &ibo_const.map[A6XX_TEX_CONST_DWORDS * ssbo_index];
3241
3242          uint64_t va = buffer_ptr(descriptors_state, &link->ssbo_map, i, j);
3243          /* We don't expose robustBufferAccess, so leave the size unlimited. */
3244          uint32_t sz = MAX_STORAGE_BUFFER_RANGE / 4;
3245
3246          dst[0] = A6XX_IBO_0_FMT(FMT6_32_UINT);
3247          dst[1] = A6XX_IBO_1_WIDTH(sz & MASK(15)) |
3248                   A6XX_IBO_1_HEIGHT(sz >> 15);
3249          dst[2] = A6XX_IBO_2_UNK4 |
3250                   A6XX_IBO_2_UNK31 |
3251                   A6XX_IBO_2_TYPE(A6XX_TEX_1D);
3252          dst[3] = 0;
3253          dst[4] = va;
3254          dst[5] = va >> 32;
3255          for (int i = 6; i < A6XX_TEX_CONST_DWORDS; i++)
3256             dst[i] = 0;
3257
3258          ssbo_index++;
3259       }
3260    }
3261
3262    for (unsigned i = 0; i < link->image_map.num; i++) {
3263       for (int j = 0; j < link->image_map.array_size[i]; j++) {
3264          uint32_t *dst = &ibo_const.map[A6XX_TEX_CONST_DWORDS * ssbo_index];
3265
3266          write_image_ibo(cmd, dst,
3267                          descriptors_state, &link->image_map, i, j);
3268
3269          ssbo_index++;
3270       }
3271    }
3272
3273    assert(ssbo_index == num_desc);
3274
3275    struct tu_cs cs;
3276    result = tu_cs_begin_sub_stream(draw_state, 7, &cs);
3277    if (result != VK_SUCCESS)
3278       return result;
3279
3280    uint32_t opcode, ibo_addr_reg;
3281    enum a6xx_state_block sb;
3282    enum a6xx_state_type st;
3283
3284    switch (type) {
3285    case MESA_SHADER_FRAGMENT:
3286       opcode = CP_LOAD_STATE6;
3287       st = ST6_SHADER;
3288       sb = SB6_IBO;
3289       ibo_addr_reg = REG_A6XX_SP_IBO_LO;
3290       break;
3291    case MESA_SHADER_COMPUTE:
3292       opcode = CP_LOAD_STATE6_FRAG;
3293       st = ST6_IBO;
3294       sb = SB6_CS_SHADER;
3295       ibo_addr_reg = REG_A6XX_SP_CS_IBO_LO;
3296       break;
3297    default:
3298       unreachable("unsupported stage for ibos");
3299    }
3300
3301    /* emit texture state: */
3302    tu_cs_emit_pkt7(&cs, opcode, 3);
3303    tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(0) |
3304               CP_LOAD_STATE6_0_STATE_TYPE(st) |
3305               CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
3306               CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
3307               CP_LOAD_STATE6_0_NUM_UNIT(num_desc));
3308    tu_cs_emit_qw(&cs, ibo_const.iova); /* SRC_ADDR_LO/HI */
3309
3310    tu_cs_emit_pkt4(&cs, ibo_addr_reg, 2);
3311    tu_cs_emit_qw(&cs, ibo_const.iova); /* SRC_ADDR_LO/HI */
3312
3313    *entry = tu_cs_end_sub_stream(draw_state, &cs);
3314    return VK_SUCCESS;
3315 }
3316
3317 struct PACKED bcolor_entry {
3318    uint32_t fp32[4];
3319    uint16_t ui16[4];
3320    int16_t  si16[4];
3321    uint16_t fp16[4];
3322    uint16_t rgb565;
3323    uint16_t rgb5a1;
3324    uint16_t rgba4;
3325    uint8_t __pad0[2];
3326    uint8_t  ui8[4];
3327    int8_t   si8[4];
3328    uint32_t rgb10a2;
3329    uint32_t z24; /* also s8? */
3330    uint16_t srgb[4];      /* appears to duplicate fp16[], but clamped, used for srgb */
3331    uint8_t  __pad1[56];
3332 } border_color[] = {
3333    [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = {},
3334    [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = {},
3335    [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = {
3336       .fp32[3] = 0x3f800000,
3337       .ui16[3] = 0xffff,
3338       .si16[3] = 0x7fff,
3339       .fp16[3] = 0x3c00,
3340       .rgb5a1 = 0x8000,
3341       .rgba4 = 0xf000,
3342       .ui8[3] = 0xff,
3343       .si8[3] = 0x7f,
3344       .rgb10a2 = 0xc0000000,
3345       .srgb[3] = 0x3c00,
3346    },
3347    [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = {
3348       .fp32[3] = 1,
3349       .fp16[3] = 1,
3350    },
3351    [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = {
3352       .fp32[0 ... 3] = 0x3f800000,
3353       .ui16[0 ... 3] = 0xffff,
3354       .si16[0 ... 3] = 0x7fff,
3355       .fp16[0 ... 3] = 0x3c00,
3356       .rgb565 = 0xffff,
3357       .rgb5a1 = 0xffff,
3358       .rgba4 = 0xffff,
3359       .ui8[0 ... 3] = 0xff,
3360       .si8[0 ... 3] = 0x7f,
3361       .rgb10a2 = 0xffffffff,
3362       .z24 = 0xffffff,
3363       .srgb[0 ... 3] = 0x3c00,
3364    },
3365    [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = {
3366       .fp32[0 ... 3] = 1,
3367       .fp16[0 ... 3] = 1,
3368    },
3369 };
3370
3371 static VkResult
3372 tu6_emit_border_color(struct tu_cmd_buffer *cmd,
3373                       struct tu_cs *cs)
3374 {
3375    STATIC_ASSERT(sizeof(struct bcolor_entry) == 128);
3376
3377    const struct tu_pipeline *pipeline = cmd->state.pipeline;
3378    struct tu_descriptor_state *descriptors_state =
3379       &cmd->descriptors[VK_PIPELINE_BIND_POINT_GRAPHICS];
3380    const struct tu_descriptor_map *vs_sampler =
3381       &pipeline->program.link[MESA_SHADER_VERTEX].sampler_map;
3382    const struct tu_descriptor_map *fs_sampler =
3383       &pipeline->program.link[MESA_SHADER_FRAGMENT].sampler_map;
3384    struct ts_cs_memory ptr;
3385
3386    VkResult result = tu_cs_alloc(&cmd->sub_cs,
3387                                  vs_sampler->num_desc + fs_sampler->num_desc,
3388                                  128 / 4,
3389                                  &ptr);
3390    if (result != VK_SUCCESS)
3391       return result;
3392
3393    for (unsigned i = 0; i < vs_sampler->num; i++) {
3394       for (unsigned j = 0; j < vs_sampler->array_size[i]; j++) {
3395          const struct tu_sampler *sampler = sampler_ptr(descriptors_state,
3396                                                         vs_sampler, i, j);
3397          memcpy(ptr.map, &border_color[sampler->border], 128);
3398          ptr.map += 128 / 4;
3399       }
3400    }
3401
3402    for (unsigned i = 0; i < fs_sampler->num; i++) {
3403       for (unsigned j = 0; j < fs_sampler->array_size[i]; j++) {
3404          const struct tu_sampler *sampler = sampler_ptr(descriptors_state,
3405                                                         fs_sampler, i, j);
3406          memcpy(ptr.map, &border_color[sampler->border], 128);
3407          ptr.map += 128 / 4;
3408       }
3409    }
3410
3411    tu_cs_emit_pkt4(cs, REG_A6XX_SP_TP_BORDER_COLOR_BASE_ADDR_LO, 2);
3412    tu_cs_emit_qw(cs, ptr.iova);
3413    return VK_SUCCESS;
3414 }
3415
3416 static void
3417 tu6_emit_streamout(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
3418 {
3419    struct tu_streamout_state *tf = &cmd->state.pipeline->streamout;
3420
3421    for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
3422       struct tu_buffer *buf = cmd->state.streamout_buf.buffers[i];
3423       if (!buf)
3424          continue;
3425
3426       uint32_t offset;
3427       offset = cmd->state.streamout_buf.offsets[i];
3428
3429       tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_BASE(i, .bo = buf->bo,
3430                                                      .bo_offset = buf->bo_offset));
3431       tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_SIZE(i, buf->size));
3432
3433       if (cmd->state.streamout_reset & (1 << i)) {
3434          offset *= tf->stride[i];
3435
3436          tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_OFFSET(i, offset));
3437          cmd->state.streamout_reset &= ~(1  << i);
3438       } else {
3439          tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
3440          tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(i)) |
3441                         CP_MEM_TO_REG_0_SHIFT_BY_2 | CP_MEM_TO_REG_0_UNK31 |
3442                         CP_MEM_TO_REG_0_CNT(0));
3443          tu_cs_emit_qw(cs, cmd->scratch_bo.iova + VSC_FLUSH * (i + 1));
3444       }
3445
3446       tu_cs_emit_regs(cs, A6XX_VPC_SO_FLUSH_BASE(i, .bo = &cmd->scratch_bo,
3447                                                     .bo_offset = VSC_FLUSH * (i + 1)));
3448    }
3449
3450    if (cmd->state.streamout_enabled) {
3451       tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 12 + (2 * tf->prog_count));
3452       tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
3453       tu_cs_emit(cs, tf->vpc_so_buf_cntl);
3454       tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(0));
3455       tu_cs_emit(cs, tf->ncomp[0]);
3456       tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(1));
3457       tu_cs_emit(cs, tf->ncomp[1]);
3458       tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(2));
3459       tu_cs_emit(cs, tf->ncomp[2]);
3460       tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(3));
3461       tu_cs_emit(cs, tf->ncomp[3]);
3462       tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
3463       tu_cs_emit(cs, A6XX_VPC_SO_CNTL_ENABLE);
3464       for (unsigned i = 0; i < tf->prog_count; i++) {
3465          tu_cs_emit(cs, REG_A6XX_VPC_SO_PROG);
3466          tu_cs_emit(cs, tf->prog[i]);
3467       }
3468    } else {
3469       tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 4);
3470       tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
3471       tu_cs_emit(cs, 0);
3472       tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
3473       tu_cs_emit(cs, 0);
3474    }
3475 }
3476
3477 static VkResult
3478 tu6_bind_draw_states(struct tu_cmd_buffer *cmd,
3479                      struct tu_cs *cs,
3480                      const struct tu_draw_info *draw)
3481 {
3482    const struct tu_pipeline *pipeline = cmd->state.pipeline;
3483    const struct tu_dynamic_state *dynamic = &cmd->state.dynamic;
3484    struct tu_draw_state_group draw_state_groups[TU_DRAW_STATE_COUNT];
3485    uint32_t draw_state_group_count = 0;
3486    VkResult result;
3487
3488    struct tu_descriptor_state *descriptors_state =
3489       &cmd->descriptors[VK_PIPELINE_BIND_POINT_GRAPHICS];
3490
3491    /* TODO lrz */
3492
3493    tu_cs_emit_regs(cs,
3494                    A6XX_PC_PRIMITIVE_CNTL_0(.primitive_restart =
3495                                             pipeline->ia.primitive_restart && draw->indexed));
3496
3497    if (cmd->state.dirty &
3498           (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_DYNAMIC_LINE_WIDTH) &&
3499        (pipeline->dynamic_state.mask & TU_DYNAMIC_LINE_WIDTH)) {
3500       tu6_emit_gras_su_cntl(cs, pipeline->rast.gras_su_cntl,
3501                             dynamic->line_width);
3502    }
3503
3504    if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK) &&
3505        (pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_COMPARE_MASK)) {
3506       tu6_emit_stencil_compare_mask(cs, dynamic->stencil_compare_mask.front,
3507                                     dynamic->stencil_compare_mask.back);
3508    }
3509
3510    if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK) &&
3511        (pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_WRITE_MASK)) {
3512       tu6_emit_stencil_write_mask(cs, dynamic->stencil_write_mask.front,
3513                                   dynamic->stencil_write_mask.back);
3514    }
3515
3516    if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE) &&
3517        (pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_REFERENCE)) {
3518       tu6_emit_stencil_reference(cs, dynamic->stencil_reference.front,
3519                                  dynamic->stencil_reference.back);
3520    }
3521
3522    if (cmd->state.dirty &
3523        (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_VERTEX_BUFFERS)) {
3524       for (uint32_t i = 0; i < pipeline->vi.count; i++) {
3525          const uint32_t binding = pipeline->vi.bindings[i];
3526          const uint32_t stride = pipeline->vi.strides[i];
3527          const struct tu_buffer *buf = cmd->state.vb.buffers[binding];
3528          const VkDeviceSize offset = buf->bo_offset +
3529                                      cmd->state.vb.offsets[binding] +
3530                                      pipeline->vi.offsets[i];
3531          const VkDeviceSize size =
3532             offset < buf->bo->size ? buf->bo->size - offset : 0;
3533
3534          tu_cs_emit_regs(cs,
3535                          A6XX_VFD_FETCH_BASE(i, .bo = buf->bo, .bo_offset = offset),
3536                          A6XX_VFD_FETCH_SIZE(i, size),
3537                          A6XX_VFD_FETCH_STRIDE(i, stride));
3538       }
3539    }
3540
3541    if (cmd->state.dirty & TU_CMD_DIRTY_PIPELINE) {
3542       draw_state_groups[draw_state_group_count++] =
3543          (struct tu_draw_state_group) {
3544             .id = TU_DRAW_STATE_PROGRAM,
3545             .enable_mask = ENABLE_DRAW,
3546             .ib = pipeline->program.state_ib,
3547          };
3548       draw_state_groups[draw_state_group_count++] =
3549          (struct tu_draw_state_group) {
3550             .id = TU_DRAW_STATE_PROGRAM_BINNING,
3551             .enable_mask = CP_SET_DRAW_STATE__0_BINNING,
3552             .ib = pipeline->program.binning_state_ib,
3553          };
3554       draw_state_groups[draw_state_group_count++] =
3555          (struct tu_draw_state_group) {
3556             .id = TU_DRAW_STATE_VI,
3557             .enable_mask = ENABLE_DRAW,
3558             .ib = pipeline->vi.state_ib,
3559          };
3560       draw_state_groups[draw_state_group_count++] =
3561          (struct tu_draw_state_group) {
3562             .id = TU_DRAW_STATE_VI_BINNING,
3563             .enable_mask = CP_SET_DRAW_STATE__0_BINNING,
3564             .ib = pipeline->vi.binning_state_ib,
3565          };
3566       draw_state_groups[draw_state_group_count++] =
3567          (struct tu_draw_state_group) {
3568             .id = TU_DRAW_STATE_VP,
3569             .enable_mask = ENABLE_ALL,
3570             .ib = pipeline->vp.state_ib,
3571          };
3572       draw_state_groups[draw_state_group_count++] =
3573          (struct tu_draw_state_group) {
3574             .id = TU_DRAW_STATE_RAST,
3575             .enable_mask = ENABLE_ALL,
3576             .ib = pipeline->rast.state_ib,
3577          };
3578       draw_state_groups[draw_state_group_count++] =
3579          (struct tu_draw_state_group) {
3580             .id = TU_DRAW_STATE_DS,
3581             .enable_mask = ENABLE_ALL,
3582             .ib = pipeline->ds.state_ib,
3583          };
3584       draw_state_groups[draw_state_group_count++] =
3585          (struct tu_draw_state_group) {
3586             .id = TU_DRAW_STATE_BLEND,
3587             .enable_mask = ENABLE_ALL,
3588             .ib = pipeline->blend.state_ib,
3589          };
3590    }
3591
3592    if (cmd->state.dirty &
3593          (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_DESCRIPTOR_SETS | TU_CMD_DIRTY_PUSH_CONSTANTS)) {
3594       draw_state_groups[draw_state_group_count++] =
3595          (struct tu_draw_state_group) {
3596             .id = TU_DRAW_STATE_VS_CONST,
3597             .enable_mask = ENABLE_ALL,
3598             .ib = tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_VERTEX)
3599          };
3600       draw_state_groups[draw_state_group_count++] =
3601          (struct tu_draw_state_group) {
3602             .id = TU_DRAW_STATE_FS_CONST,
3603             .enable_mask = ENABLE_DRAW,
3604             .ib = tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_FRAGMENT)
3605          };
3606    }
3607
3608    if (cmd->state.dirty & TU_CMD_DIRTY_STREAMOUT_BUFFERS)
3609       tu6_emit_streamout(cmd, cs);
3610
3611    if (cmd->state.dirty &
3612          (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_DESCRIPTOR_SETS)) {
3613       bool needs_border = false;
3614       struct tu_cs_entry vs_tex, fs_tex_sysmem, fs_tex_gmem, fs_ibo;
3615
3616       result = tu6_emit_textures(cmd, pipeline, descriptors_state,
3617                                  MESA_SHADER_VERTEX, &vs_tex, &needs_border,
3618                                  false);
3619       if (result != VK_SUCCESS)
3620          return result;
3621
3622       /* TODO: we could emit just one texture descriptor draw state when there
3623        * are no input attachments, which is the most common case. We could
3624        * also split out the sampler state, which doesn't change even for input
3625        * attachments.
3626        */
3627       result = tu6_emit_textures(cmd, pipeline, descriptors_state,
3628                                  MESA_SHADER_FRAGMENT, &fs_tex_sysmem,
3629                                  &needs_border, true);
3630       if (result != VK_SUCCESS)
3631          return result;
3632
3633       result = tu6_emit_textures(cmd, pipeline, descriptors_state,
3634                                  MESA_SHADER_FRAGMENT, &fs_tex_gmem,
3635                                  &needs_border, false);
3636       if (result != VK_SUCCESS)
3637          return result;
3638
3639       result = tu6_emit_ibo(cmd, pipeline, descriptors_state,
3640                             MESA_SHADER_FRAGMENT, &fs_ibo);
3641       if (result != VK_SUCCESS)
3642          return result;
3643
3644       draw_state_groups[draw_state_group_count++] =
3645          (struct tu_draw_state_group) {
3646             .id = TU_DRAW_STATE_VS_TEX,
3647             .enable_mask = ENABLE_ALL,
3648             .ib = vs_tex,
3649          };
3650       draw_state_groups[draw_state_group_count++] =
3651          (struct tu_draw_state_group) {
3652             .id = TU_DRAW_STATE_FS_TEX_GMEM,
3653             .enable_mask = CP_SET_DRAW_STATE__0_GMEM,
3654             .ib = fs_tex_gmem,
3655          };
3656       draw_state_groups[draw_state_group_count++] =
3657          (struct tu_draw_state_group) {
3658             .id = TU_DRAW_STATE_FS_TEX_SYSMEM,
3659             .enable_mask = CP_SET_DRAW_STATE__0_SYSMEM,
3660             .ib = fs_tex_sysmem,
3661          };
3662       draw_state_groups[draw_state_group_count++] =
3663          (struct tu_draw_state_group) {
3664             .id = TU_DRAW_STATE_FS_IBO,
3665             .enable_mask = ENABLE_DRAW,
3666             .ib = fs_ibo,
3667          };
3668
3669       if (needs_border) {
3670          result = tu6_emit_border_color(cmd, cs);
3671          if (result != VK_SUCCESS)
3672             return result;
3673       }
3674    }
3675
3676    struct tu_cs_entry vs_params;
3677    result = tu6_emit_vs_params(cmd, draw, &vs_params);
3678    if (result != VK_SUCCESS)
3679       return result;
3680
3681    draw_state_groups[draw_state_group_count++] =
3682       (struct tu_draw_state_group) {
3683          .id = TU_DRAW_STATE_VS_PARAMS,
3684          .enable_mask = ENABLE_ALL,
3685          .ib = vs_params,
3686       };
3687
3688    tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * draw_state_group_count);
3689    for (uint32_t i = 0; i < draw_state_group_count; i++) {
3690       const struct tu_draw_state_group *group = &draw_state_groups[i];
3691       debug_assert((group->enable_mask & ~ENABLE_ALL) == 0);
3692       uint32_t cp_set_draw_state =
3693          CP_SET_DRAW_STATE__0_COUNT(group->ib.size / 4) |
3694          group->enable_mask |
3695          CP_SET_DRAW_STATE__0_GROUP_ID(group->id);
3696       uint64_t iova;
3697       if (group->ib.size) {
3698          iova = group->ib.bo->iova + group->ib.offset;
3699       } else {
3700          cp_set_draw_state |= CP_SET_DRAW_STATE__0_DISABLE;
3701          iova = 0;
3702       }
3703
3704       tu_cs_emit(cs, cp_set_draw_state);
3705       tu_cs_emit_qw(cs, iova);
3706    }
3707
3708    tu_cs_sanity_check(cs);
3709
3710    /* track BOs */
3711    if (cmd->state.dirty & TU_CMD_DIRTY_VERTEX_BUFFERS) {
3712       for (uint32_t i = 0; i < MAX_VBS; i++) {
3713          const struct tu_buffer *buf = cmd->state.vb.buffers[i];
3714          if (buf)
3715             tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
3716       }
3717    }
3718    if (cmd->state.dirty & TU_CMD_DIRTY_DESCRIPTOR_SETS) {
3719       unsigned i;
3720       for_each_bit(i, descriptors_state->valid) {
3721          struct tu_descriptor_set *set = descriptors_state->sets[i];
3722          for (unsigned j = 0; j < set->layout->buffer_count; ++j)
3723             if (set->descriptors[j]) {
3724                tu_bo_list_add(&cmd->bo_list, set->descriptors[j],
3725                               MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
3726             }
3727       }
3728    }
3729    if (cmd->state.dirty & TU_CMD_DIRTY_STREAMOUT_BUFFERS) {
3730       for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
3731          const struct tu_buffer *buf = cmd->state.streamout_buf.buffers[i];
3732          if (buf) {
3733             tu_bo_list_add(&cmd->bo_list, buf->bo,
3734                               MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
3735          }
3736       }
3737    }
3738
3739    /* Fragment shader state overwrites compute shader state, so flag the
3740     * compute pipeline for re-emit.
3741     */
3742    cmd->state.dirty = TU_CMD_DIRTY_COMPUTE_PIPELINE;
3743    return VK_SUCCESS;
3744 }
3745
3746 static void
3747 tu6_emit_draw_indirect(struct tu_cmd_buffer *cmd,
3748                      struct tu_cs *cs,
3749                      const struct tu_draw_info *draw)
3750 {
3751    const enum pc_di_primtype primtype = cmd->state.pipeline->ia.primtype;
3752
3753    tu_cs_emit_regs(cs,
3754                    A6XX_VFD_INDEX_OFFSET(draw->vertex_offset),
3755                    A6XX_VFD_INSTANCE_START_OFFSET(draw->first_instance));
3756
3757    if (draw->indexed) {
3758       const enum a4xx_index_size index_size =
3759          tu6_index_size(cmd->state.index_type);
3760       const uint32_t index_bytes =
3761          (cmd->state.index_type == VK_INDEX_TYPE_UINT32) ? 4 : 2;
3762       const struct tu_buffer *index_buf = cmd->state.index_buffer;
3763       unsigned max_indicies =
3764          (index_buf->size - cmd->state.index_offset) / index_bytes;
3765
3766       const uint32_t cp_draw_indx =
3767          CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
3768          CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_DMA) |
3769          CP_DRAW_INDX_OFFSET_0_INDEX_SIZE(index_size) |
3770          CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) | 0x2000;
3771
3772       tu_cs_emit_pkt7(cs, CP_DRAW_INDX_INDIRECT, 6);
3773       tu_cs_emit(cs, cp_draw_indx);
3774       tu_cs_emit_qw(cs, index_buf->bo->iova + cmd->state.index_offset);
3775       tu_cs_emit(cs, A5XX_CP_DRAW_INDX_INDIRECT_3_MAX_INDICES(max_indicies));
3776       tu_cs_emit_qw(cs, draw->indirect->bo->iova + draw->indirect_offset);
3777    } else {
3778       const uint32_t cp_draw_indx =
3779          CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
3780          CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_AUTO_INDEX) |
3781          CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) | 0x2000;
3782
3783       tu_cs_emit_pkt7(cs, CP_DRAW_INDIRECT, 3);
3784       tu_cs_emit(cs, cp_draw_indx);
3785       tu_cs_emit_qw(cs, draw->indirect->bo->iova + draw->indirect_offset);
3786    }
3787
3788    tu_bo_list_add(&cmd->bo_list, draw->indirect->bo, MSM_SUBMIT_BO_READ);
3789 }
3790
3791 static void
3792 tu6_emit_draw_direct(struct tu_cmd_buffer *cmd,
3793                      struct tu_cs *cs,
3794                      const struct tu_draw_info *draw)
3795 {
3796
3797    const enum pc_di_primtype primtype = cmd->state.pipeline->ia.primtype;
3798
3799    tu_cs_emit_regs(cs,
3800                    A6XX_VFD_INDEX_OFFSET(draw->vertex_offset),
3801                    A6XX_VFD_INSTANCE_START_OFFSET(draw->first_instance));
3802
3803    /* TODO hw binning */
3804    if (draw->indexed) {
3805       const enum a4xx_index_size index_size =
3806          tu6_index_size(cmd->state.index_type);
3807       const uint32_t index_bytes =
3808          (cmd->state.index_type == VK_INDEX_TYPE_UINT32) ? 4 : 2;
3809       const struct tu_buffer *buf = cmd->state.index_buffer;
3810       const VkDeviceSize offset = buf->bo_offset + cmd->state.index_offset +
3811                                   index_bytes * draw->first_index;
3812       const uint32_t size = index_bytes * draw->count;
3813
3814       const uint32_t cp_draw_indx =
3815          CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
3816          CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_DMA) |
3817          CP_DRAW_INDX_OFFSET_0_INDEX_SIZE(index_size) |
3818          CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) | 0x2000;
3819
3820       tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 7);
3821       tu_cs_emit(cs, cp_draw_indx);
3822       tu_cs_emit(cs, draw->instance_count);
3823       tu_cs_emit(cs, draw->count);
3824       tu_cs_emit(cs, 0x0); /* XXX */
3825       tu_cs_emit_qw(cs, buf->bo->iova + offset);
3826       tu_cs_emit(cs, size);
3827    } else {
3828       const uint32_t cp_draw_indx =
3829          CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
3830          CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_AUTO_INDEX) |
3831          CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) | 0x2000;
3832
3833       tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 3);
3834       tu_cs_emit(cs, cp_draw_indx);
3835       tu_cs_emit(cs, draw->instance_count);
3836       tu_cs_emit(cs, draw->count);
3837    }
3838 }
3839
3840 static void
3841 tu_draw(struct tu_cmd_buffer *cmd, const struct tu_draw_info *draw)
3842 {
3843    struct tu_cs *cs = &cmd->draw_cs;
3844    VkResult result;
3845
3846    result = tu6_bind_draw_states(cmd, cs, draw);
3847    if (result != VK_SUCCESS) {
3848       cmd->record_result = result;
3849       return;
3850    }
3851
3852    if (draw->indirect)
3853       tu6_emit_draw_indirect(cmd, cs, draw);
3854    else
3855       tu6_emit_draw_direct(cmd, cs, draw);
3856
3857    if (cmd->state.streamout_enabled) {
3858       for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
3859          if (cmd->state.streamout_enabled & (1 << i))
3860             tu6_emit_event_write(cmd, cs, FLUSH_SO_0 + i, false);
3861       }
3862    }
3863
3864    cmd->wait_for_idle = true;
3865
3866    tu_cs_sanity_check(cs);
3867 }
3868
3869 void
3870 tu_CmdDraw(VkCommandBuffer commandBuffer,
3871            uint32_t vertexCount,
3872            uint32_t instanceCount,
3873            uint32_t firstVertex,
3874            uint32_t firstInstance)
3875 {
3876    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
3877    struct tu_draw_info info = {};
3878
3879    info.count = vertexCount;
3880    info.instance_count = instanceCount;
3881    info.first_instance = firstInstance;
3882    info.vertex_offset = firstVertex;
3883
3884    tu_draw(cmd_buffer, &info);
3885 }
3886
3887 void
3888 tu_CmdDrawIndexed(VkCommandBuffer commandBuffer,
3889                   uint32_t indexCount,
3890                   uint32_t instanceCount,
3891                   uint32_t firstIndex,
3892                   int32_t vertexOffset,
3893                   uint32_t firstInstance)
3894 {
3895    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
3896    struct tu_draw_info info = {};
3897
3898    info.indexed = true;
3899    info.count = indexCount;
3900    info.instance_count = instanceCount;
3901    info.first_index = firstIndex;
3902    info.vertex_offset = vertexOffset;
3903    info.first_instance = firstInstance;
3904
3905    tu_draw(cmd_buffer, &info);
3906 }
3907
3908 void
3909 tu_CmdDrawIndirect(VkCommandBuffer commandBuffer,
3910                    VkBuffer _buffer,
3911                    VkDeviceSize offset,
3912                    uint32_t drawCount,
3913                    uint32_t stride)
3914 {
3915    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
3916    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
3917    struct tu_draw_info info = {};
3918
3919    info.count = drawCount;
3920    info.indirect = buffer;
3921    info.indirect_offset = offset;
3922    info.stride = stride;
3923
3924    tu_draw(cmd_buffer, &info);
3925 }
3926
3927 void
3928 tu_CmdDrawIndexedIndirect(VkCommandBuffer commandBuffer,
3929                           VkBuffer _buffer,
3930                           VkDeviceSize offset,
3931                           uint32_t drawCount,
3932                           uint32_t stride)
3933 {
3934    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
3935    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
3936    struct tu_draw_info info = {};
3937
3938    info.indexed = true;
3939    info.count = drawCount;
3940    info.indirect = buffer;
3941    info.indirect_offset = offset;
3942    info.stride = stride;
3943
3944    tu_draw(cmd_buffer, &info);
3945 }
3946
3947 void tu_CmdDrawIndirectByteCountEXT(VkCommandBuffer commandBuffer,
3948                                     uint32_t instanceCount,
3949                                     uint32_t firstInstance,
3950                                     VkBuffer _counterBuffer,
3951                                     VkDeviceSize counterBufferOffset,
3952                                     uint32_t counterOffset,
3953                                     uint32_t vertexStride)
3954 {
3955    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
3956    TU_FROM_HANDLE(tu_buffer, buffer, _counterBuffer);
3957
3958    struct tu_draw_info info = {};
3959
3960    info.instance_count = instanceCount;
3961    info.first_instance = firstInstance;
3962    info.streamout_buffer = buffer;
3963    info.streamout_buffer_offset = counterBufferOffset;
3964    info.stride = vertexStride;
3965
3966    tu_draw(cmd_buffer, &info);
3967 }
3968
3969 struct tu_dispatch_info
3970 {
3971    /**
3972     * Determine the layout of the grid (in block units) to be used.
3973     */
3974    uint32_t blocks[3];
3975
3976    /**
3977     * A starting offset for the grid. If unaligned is set, the offset
3978     * must still be aligned.
3979     */
3980    uint32_t offsets[3];
3981    /**
3982     * Whether it's an unaligned compute dispatch.
3983     */
3984    bool unaligned;
3985
3986    /**
3987     * Indirect compute parameters resource.
3988     */
3989    struct tu_buffer *indirect;
3990    uint64_t indirect_offset;
3991 };
3992
3993 static void
3994 tu_emit_compute_driver_params(struct tu_cs *cs, struct tu_pipeline *pipeline,
3995                               const struct tu_dispatch_info *info)
3996 {
3997    gl_shader_stage type = MESA_SHADER_COMPUTE;
3998    const struct tu_program_descriptor_linkage *link =
3999       &pipeline->program.link[type];
4000    const struct ir3_const_state *const_state = &link->const_state;
4001    uint32_t offset = const_state->offsets.driver_param;
4002
4003    if (link->constlen <= offset)
4004       return;
4005
4006    if (!info->indirect) {
4007       uint32_t driver_params[IR3_DP_CS_COUNT] = {
4008          [IR3_DP_NUM_WORK_GROUPS_X] = info->blocks[0],
4009          [IR3_DP_NUM_WORK_GROUPS_Y] = info->blocks[1],
4010          [IR3_DP_NUM_WORK_GROUPS_Z] = info->blocks[2],
4011          [IR3_DP_LOCAL_GROUP_SIZE_X] = pipeline->compute.local_size[0],
4012          [IR3_DP_LOCAL_GROUP_SIZE_Y] = pipeline->compute.local_size[1],
4013          [IR3_DP_LOCAL_GROUP_SIZE_Z] = pipeline->compute.local_size[2],
4014       };
4015
4016       uint32_t num_consts = MIN2(const_state->num_driver_params,
4017                                  (link->constlen - offset) * 4);
4018       /* push constants */
4019       tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3 + num_consts);
4020       tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(offset) |
4021                  CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
4022                  CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
4023                  CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
4024                  CP_LOAD_STATE6_0_NUM_UNIT(num_consts / 4));
4025       tu_cs_emit(cs, 0);
4026       tu_cs_emit(cs, 0);
4027       uint32_t i;
4028       for (i = 0; i < num_consts; i++)
4029          tu_cs_emit(cs, driver_params[i]);
4030    } else {
4031       tu_finishme("Indirect driver params");
4032    }
4033 }
4034
4035 static void
4036 tu_dispatch(struct tu_cmd_buffer *cmd,
4037             const struct tu_dispatch_info *info)
4038 {
4039    struct tu_cs *cs = &cmd->cs;
4040    struct tu_pipeline *pipeline = cmd->state.compute_pipeline;
4041    struct tu_descriptor_state *descriptors_state =
4042       &cmd->descriptors[VK_PIPELINE_BIND_POINT_COMPUTE];
4043    VkResult result;
4044
4045    if (cmd->state.dirty & TU_CMD_DIRTY_COMPUTE_PIPELINE)
4046       tu_cs_emit_ib(cs, &pipeline->program.state_ib);
4047
4048    struct tu_cs_entry ib;
4049
4050    ib = tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_COMPUTE);
4051    if (ib.size)
4052       tu_cs_emit_ib(cs, &ib);
4053
4054    tu_emit_compute_driver_params(cs, pipeline, info);
4055
4056    bool needs_border;
4057    result = tu6_emit_textures(cmd, pipeline, descriptors_state,
4058                               MESA_SHADER_COMPUTE, &ib, &needs_border, false);
4059    if (result != VK_SUCCESS) {
4060       cmd->record_result = result;
4061       return;
4062    }
4063
4064    if (ib.size)
4065       tu_cs_emit_ib(cs, &ib);
4066
4067    if (needs_border)
4068       tu_finishme("compute border color");
4069
4070    result = tu6_emit_ibo(cmd, pipeline, descriptors_state, MESA_SHADER_COMPUTE, &ib);
4071    if (result != VK_SUCCESS) {
4072       cmd->record_result = result;
4073       return;
4074    }
4075
4076    if (ib.size)
4077       tu_cs_emit_ib(cs, &ib);
4078
4079    /* track BOs */
4080    if (cmd->state.dirty & TU_CMD_DIRTY_DESCRIPTOR_SETS) {
4081       unsigned i;
4082       for_each_bit(i, descriptors_state->valid) {
4083          struct tu_descriptor_set *set = descriptors_state->sets[i];
4084          for (unsigned j = 0; j < set->layout->buffer_count; ++j)
4085             if (set->descriptors[j]) {
4086                tu_bo_list_add(&cmd->bo_list, set->descriptors[j],
4087                               MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
4088             }
4089       }
4090    }
4091
4092    /* Compute shader state overwrites fragment shader state, so we flag the
4093     * graphics pipeline for re-emit.
4094     */
4095    cmd->state.dirty = TU_CMD_DIRTY_PIPELINE;
4096
4097    tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
4098    tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_COMPUTE));
4099
4100    const uint32_t *local_size = pipeline->compute.local_size;
4101    const uint32_t *num_groups = info->blocks;
4102    tu_cs_emit_regs(cs,
4103                    A6XX_HLSQ_CS_NDRANGE_0(.kerneldim = 3,
4104                                           .localsizex = local_size[0] - 1,
4105                                           .localsizey = local_size[1] - 1,
4106                                           .localsizez = local_size[2] - 1),
4107                    A6XX_HLSQ_CS_NDRANGE_1(.globalsize_x = local_size[0] * num_groups[0]),
4108                    A6XX_HLSQ_CS_NDRANGE_2(.globaloff_x = 0),
4109                    A6XX_HLSQ_CS_NDRANGE_3(.globalsize_y = local_size[1] * num_groups[1]),
4110                    A6XX_HLSQ_CS_NDRANGE_4(.globaloff_y = 0),
4111                    A6XX_HLSQ_CS_NDRANGE_5(.globalsize_z = local_size[2] * num_groups[2]),
4112                    A6XX_HLSQ_CS_NDRANGE_6(.globaloff_z = 0));
4113
4114    tu_cs_emit_regs(cs,
4115                    A6XX_HLSQ_CS_KERNEL_GROUP_X(1),
4116                    A6XX_HLSQ_CS_KERNEL_GROUP_Y(1),
4117                    A6XX_HLSQ_CS_KERNEL_GROUP_Z(1));
4118
4119    if (info->indirect) {
4120       uint64_t iova = tu_buffer_iova(info->indirect) + info->indirect_offset;
4121
4122       tu_bo_list_add(&cmd->bo_list, info->indirect->bo,
4123                      MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
4124
4125       tu_cs_emit_pkt7(cs, CP_EXEC_CS_INDIRECT, 4);
4126       tu_cs_emit(cs, 0x00000000);
4127       tu_cs_emit_qw(cs, iova);
4128       tu_cs_emit(cs,
4129                  A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEX(local_size[0] - 1) |
4130                  A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEY(local_size[1] - 1) |
4131                  A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEZ(local_size[2] - 1));
4132    } else {
4133       tu_cs_emit_pkt7(cs, CP_EXEC_CS, 4);
4134       tu_cs_emit(cs, 0x00000000);
4135       tu_cs_emit(cs, CP_EXEC_CS_1_NGROUPS_X(info->blocks[0]));
4136       tu_cs_emit(cs, CP_EXEC_CS_2_NGROUPS_Y(info->blocks[1]));
4137       tu_cs_emit(cs, CP_EXEC_CS_3_NGROUPS_Z(info->blocks[2]));
4138    }
4139
4140    tu_cs_emit_wfi(cs);
4141
4142    tu6_emit_cache_flush(cmd, cs);
4143 }
4144
4145 void
4146 tu_CmdDispatchBase(VkCommandBuffer commandBuffer,
4147                    uint32_t base_x,
4148                    uint32_t base_y,
4149                    uint32_t base_z,
4150                    uint32_t x,
4151                    uint32_t y,
4152                    uint32_t z)
4153 {
4154    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
4155    struct tu_dispatch_info info = {};
4156
4157    info.blocks[0] = x;
4158    info.blocks[1] = y;
4159    info.blocks[2] = z;
4160
4161    info.offsets[0] = base_x;
4162    info.offsets[1] = base_y;
4163    info.offsets[2] = base_z;
4164    tu_dispatch(cmd_buffer, &info);
4165 }
4166
4167 void
4168 tu_CmdDispatch(VkCommandBuffer commandBuffer,
4169                uint32_t x,
4170                uint32_t y,
4171                uint32_t z)
4172 {
4173    tu_CmdDispatchBase(commandBuffer, 0, 0, 0, x, y, z);
4174 }
4175
4176 void
4177 tu_CmdDispatchIndirect(VkCommandBuffer commandBuffer,
4178                        VkBuffer _buffer,
4179                        VkDeviceSize offset)
4180 {
4181    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
4182    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
4183    struct tu_dispatch_info info = {};
4184
4185    info.indirect = buffer;
4186    info.indirect_offset = offset;
4187
4188    tu_dispatch(cmd_buffer, &info);
4189 }
4190
4191 void
4192 tu_CmdEndRenderPass(VkCommandBuffer commandBuffer)
4193 {
4194    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
4195
4196    tu_cs_end(&cmd_buffer->draw_cs);
4197    tu_cs_end(&cmd_buffer->draw_epilogue_cs);
4198
4199    if (use_sysmem_rendering(cmd_buffer))
4200       tu_cmd_render_sysmem(cmd_buffer);
4201    else
4202       tu_cmd_render_tiles(cmd_buffer);
4203
4204    /* discard draw_cs and draw_epilogue_cs entries now that the tiles are
4205       rendered */
4206    tu_cs_discard_entries(&cmd_buffer->draw_cs);
4207    tu_cs_begin(&cmd_buffer->draw_cs);
4208    tu_cs_discard_entries(&cmd_buffer->draw_epilogue_cs);
4209    tu_cs_begin(&cmd_buffer->draw_epilogue_cs);
4210
4211    cmd_buffer->state.pass = NULL;
4212    cmd_buffer->state.subpass = NULL;
4213    cmd_buffer->state.framebuffer = NULL;
4214 }
4215
4216 void
4217 tu_CmdEndRenderPass2(VkCommandBuffer commandBuffer,
4218                      const VkSubpassEndInfoKHR *pSubpassEndInfo)
4219 {
4220    tu_CmdEndRenderPass(commandBuffer);
4221 }
4222
4223 struct tu_barrier_info
4224 {
4225    uint32_t eventCount;
4226    const VkEvent *pEvents;
4227    VkPipelineStageFlags srcStageMask;
4228 };
4229
4230 static void
4231 tu_barrier(struct tu_cmd_buffer *cmd_buffer,
4232            uint32_t memoryBarrierCount,
4233            const VkMemoryBarrier *pMemoryBarriers,
4234            uint32_t bufferMemoryBarrierCount,
4235            const VkBufferMemoryBarrier *pBufferMemoryBarriers,
4236            uint32_t imageMemoryBarrierCount,
4237            const VkImageMemoryBarrier *pImageMemoryBarriers,
4238            const struct tu_barrier_info *info)
4239 {
4240 }
4241
4242 void
4243 tu_CmdPipelineBarrier(VkCommandBuffer commandBuffer,
4244                       VkPipelineStageFlags srcStageMask,
4245                       VkPipelineStageFlags destStageMask,
4246                       VkBool32 byRegion,
4247                       uint32_t memoryBarrierCount,
4248                       const VkMemoryBarrier *pMemoryBarriers,
4249                       uint32_t bufferMemoryBarrierCount,
4250                       const VkBufferMemoryBarrier *pBufferMemoryBarriers,
4251                       uint32_t imageMemoryBarrierCount,
4252                       const VkImageMemoryBarrier *pImageMemoryBarriers)
4253 {
4254    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
4255    struct tu_barrier_info info;
4256
4257    info.eventCount = 0;
4258    info.pEvents = NULL;
4259    info.srcStageMask = srcStageMask;
4260
4261    tu_barrier(cmd_buffer, memoryBarrierCount, pMemoryBarriers,
4262               bufferMemoryBarrierCount, pBufferMemoryBarriers,
4263               imageMemoryBarrierCount, pImageMemoryBarriers, &info);
4264 }
4265
4266 static void
4267 write_event(struct tu_cmd_buffer *cmd, struct tu_event *event, unsigned value)
4268 {
4269    struct tu_cs *cs = &cmd->cs;
4270
4271    tu_bo_list_add(&cmd->bo_list, &event->bo, MSM_SUBMIT_BO_WRITE);
4272
4273    /* TODO: any flush required before/after ? */
4274
4275    tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 3);
4276    tu_cs_emit_qw(cs, event->bo.iova); /* ADDR_LO/HI */
4277    tu_cs_emit(cs, value);
4278 }
4279
4280 void
4281 tu_CmdSetEvent(VkCommandBuffer commandBuffer,
4282                VkEvent _event,
4283                VkPipelineStageFlags stageMask)
4284 {
4285    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
4286    TU_FROM_HANDLE(tu_event, event, _event);
4287
4288    write_event(cmd, event, 1);
4289 }
4290
4291 void
4292 tu_CmdResetEvent(VkCommandBuffer commandBuffer,
4293                  VkEvent _event,
4294                  VkPipelineStageFlags stageMask)
4295 {
4296    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
4297    TU_FROM_HANDLE(tu_event, event, _event);
4298
4299    write_event(cmd, event, 0);
4300 }
4301
4302 void
4303 tu_CmdWaitEvents(VkCommandBuffer commandBuffer,
4304                  uint32_t eventCount,
4305                  const VkEvent *pEvents,
4306                  VkPipelineStageFlags srcStageMask,
4307                  VkPipelineStageFlags dstStageMask,
4308                  uint32_t memoryBarrierCount,
4309                  const VkMemoryBarrier *pMemoryBarriers,
4310                  uint32_t bufferMemoryBarrierCount,
4311                  const VkBufferMemoryBarrier *pBufferMemoryBarriers,
4312                  uint32_t imageMemoryBarrierCount,
4313                  const VkImageMemoryBarrier *pImageMemoryBarriers)
4314 {
4315    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
4316    struct tu_cs *cs = &cmd->cs;
4317
4318    /* TODO: any flush required before/after? (CP_WAIT_FOR_ME?) */
4319
4320    for (uint32_t i = 0; i < eventCount; i++) {
4321       TU_FROM_HANDLE(tu_event, event, pEvents[i]);
4322
4323       tu_bo_list_add(&cmd->bo_list, &event->bo, MSM_SUBMIT_BO_READ);
4324
4325       tu_cs_emit_pkt7(cs, CP_WAIT_REG_MEM, 6);
4326       tu_cs_emit(cs, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ) |
4327                      CP_WAIT_REG_MEM_0_POLL_MEMORY);
4328       tu_cs_emit_qw(cs, event->bo.iova); /* POLL_ADDR_LO/HI */
4329       tu_cs_emit(cs, CP_WAIT_REG_MEM_3_REF(1));
4330       tu_cs_emit(cs, CP_WAIT_REG_MEM_4_MASK(~0u));
4331       tu_cs_emit(cs, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(20));
4332    }
4333 }
4334
4335 void
4336 tu_CmdSetDeviceMask(VkCommandBuffer commandBuffer, uint32_t deviceMask)
4337 {
4338    /* No-op */
4339 }