src/intel/vulkan/anv_pass.c

   1 /*
   2  * Copyright © 2015 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include "anv_private.h"
  25
  26 #include "vk_util.h"
  27
  28 static void
  29 anv_render_pass_add_subpass_dep(struct anv_render_pass *pass,
  30                                 const VkSubpassDependency2KHR *dep)
  31 {
  32    if (dep->dstSubpass == VK_SUBPASS_EXTERNAL) {
  33       pass->subpass_flushes[pass->subpass_count] |=
  34          anv_pipe_invalidate_bits_for_access_flags(dep->dstAccessMask);
  35    } else {
  36       assert(dep->dstSubpass < pass->subpass_count);
  37       pass->subpass_flushes[dep->dstSubpass] |=
  38          anv_pipe_invalidate_bits_for_access_flags(dep->dstAccessMask);
  39    }
  40
  41    if (dep->srcSubpass == VK_SUBPASS_EXTERNAL) {
  42       pass->subpass_flushes[0] |=
  43          anv_pipe_flush_bits_for_access_flags(dep->srcAccessMask);
  44    } else {
  45       assert(dep->srcSubpass < pass->subpass_count);
  46       pass->subpass_flushes[dep->srcSubpass + 1] |=
  47          anv_pipe_flush_bits_for_access_flags(dep->srcAccessMask);
  48    }
  49 }
  50
  51 /* Do a second "compile" step on a render pass */
  52 static void
  53 anv_render_pass_compile(struct anv_render_pass *pass)
  54 {
  55    /* The CreateRenderPass code zeros the entire render pass and also uses a
  56     * designated initializer for filling these out.  There's no need for us to
  57     * do it again.
  58     *
  59     * for (uint32_t i = 0; i < pass->attachment_count; i++) {
  60     *    pass->attachments[i].usage = 0;
  61     *    pass->attachments[i].first_subpass_layout = VK_IMAGE_LAYOUT_UNDEFINED;
  62     * }
  63     */
  64
  65    VkImageUsageFlags all_usage = 0;
  66    for (uint32_t i = 0; i < pass->subpass_count; i++) {
  67       struct anv_subpass *subpass = &pass->subpasses[i];
  68
  69       /* We don't allow depth_stencil_attachment to be non-NULL and be
  70        * VK_ATTACHMENT_UNUSED.  This way something can just check for NULL
  71        * and be guaranteed that they have a valid attachment.
  72        */
  73       if (subpass->depth_stencil_attachment &&
  74           subpass->depth_stencil_attachment->attachment == VK_ATTACHMENT_UNUSED)
  75          subpass->depth_stencil_attachment = NULL;
  76
  77       for (uint32_t j = 0; j < subpass->attachment_count; j++) {
  78          struct anv_subpass_attachment *subpass_att = &subpass->attachments[j];
  79          if (subpass_att->attachment == VK_ATTACHMENT_UNUSED)
  80             continue;
  81
  82          struct anv_render_pass_attachment *pass_att =
  83             &pass->attachments[subpass_att->attachment];
  84
  85          assert(__builtin_popcount(subpass_att->usage) == 1);
  86          pass_att->usage |= subpass_att->usage;
  87          pass_att->last_subpass_idx = i;
  88
  89          all_usage |= subpass_att->usage;
  90
  91          if (pass_att->first_subpass_layout == VK_IMAGE_LAYOUT_UNDEFINED) {
  92             pass_att->first_subpass_layout = subpass_att->layout;
  93             assert(pass_att->first_subpass_layout != VK_IMAGE_LAYOUT_UNDEFINED);
  94          }
  95
  96          if (subpass_att->usage == VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT &&
  97              subpass->depth_stencil_attachment &&
  98              subpass_att->attachment == subpass->depth_stencil_attachment->attachment)
  99             subpass->has_ds_self_dep = true;
 100       }
 101
 102       /* We have to handle resolve attachments specially */
 103       subpass->has_resolve = false;
 104       if (subpass->resolve_attachments) {
 105          for (uint32_t j = 0; j < subpass->color_count; j++) {
 106             struct anv_subpass_attachment *color_att =
 107                &subpass->color_attachments[j];
 108             struct anv_subpass_attachment *resolve_att =
 109                &subpass->resolve_attachments[j];
 110             if (resolve_att->attachment == VK_ATTACHMENT_UNUSED)
 111                continue;
 112
 113             subpass->has_resolve = true;
 114
 115             assert(resolve_att->usage == VK_IMAGE_USAGE_TRANSFER_DST_BIT);
 116             color_att->usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
 117          }
 118       }
 119    }
 120
 121    /* From the Vulkan 1.0.39 spec:
 122     *
 123     *    If there is no subpass dependency from VK_SUBPASS_EXTERNAL to the
 124     *    first subpass that uses an attachment, then an implicit subpass
 125     *    dependency exists from VK_SUBPASS_EXTERNAL to the first subpass it is
 126     *    used in. The subpass dependency operates as if defined with the
 127     *    following parameters:
 128     *
 129     *    VkSubpassDependency implicitDependency = {
 130     *        .srcSubpass = VK_SUBPASS_EXTERNAL;
 131     *        .dstSubpass = firstSubpass; // First subpass attachment is used in
 132     *        .srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
 133     *        .dstStageMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
 134     *        .srcAccessMask = 0;
 135     *        .dstAccessMask = VK_ACCESS_INPUT_ATTACHMENT_READ_BIT |
 136     *                         VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
 137     *                         VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
 138     *                         VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
 139     *                         VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
 140     *        .dependencyFlags = 0;
 141     *    };
 142     *
 143     *    Similarly, if there is no subpass dependency from the last subpass
 144     *    that uses an attachment to VK_SUBPASS_EXTERNAL, then an implicit
 145     *    subpass dependency exists from the last subpass it is used in to
 146     *    VK_SUBPASS_EXTERNAL. The subpass dependency operates as if defined
 147     *    with the following parameters:
 148     *
 149     *    VkSubpassDependency implicitDependency = {
 150     *        .srcSubpass = lastSubpass; // Last subpass attachment is used in
 151     *        .dstSubpass = VK_SUBPASS_EXTERNAL;
 152     *        .srcStageMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
 153     *        .dstStageMask = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT;
 154     *        .srcAccessMask = VK_ACCESS_INPUT_ATTACHMENT_READ_BIT |
 155     *                         VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
 156     *                         VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
 157     *                         VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
 158     *                         VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
 159     *        .dstAccessMask = 0;
 160     *        .dependencyFlags = 0;
 161     *    };
 162     *
 163     * We could implement this by walking over all of the attachments and
 164     * subpasses and checking to see if any of them don't have an external
 165     * dependency.  Or, we could just be lazy and add a couple extra flushes.
 166     * We choose to be lazy.
 167     */
 168    if (all_usage & VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT) {
 169       pass->subpass_flushes[0] |=
 170          ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT;
 171    }
 172    if (all_usage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
 173       pass->subpass_flushes[pass->subpass_count] |=
 174          ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT;
 175    }
 176    if (all_usage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
 177       pass->subpass_flushes[pass->subpass_count] |=
 178          ANV_PIPE_DEPTH_CACHE_FLUSH_BIT;
 179    }
 180 }
 181
 182 static unsigned
 183 num_subpass_attachments(const VkSubpassDescription *desc)
 184 {
 185    return desc->inputAttachmentCount +
 186           desc->colorAttachmentCount +
 187           (desc->pResolveAttachments ? desc->colorAttachmentCount : 0) +
 188           (desc->pDepthStencilAttachment != NULL);
 189 }
 190
 191 VkResult anv_CreateRenderPass(
 192     VkDevice                                    _device,
 193     const VkRenderPassCreateInfo*               pCreateInfo,
 194     const VkAllocationCallbacks*                pAllocator,
 195     VkRenderPass*                               pRenderPass)
 196 {
 197    ANV_FROM_HANDLE(anv_device, device, _device);
 198
 199    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO);
 200
 201    struct anv_render_pass *pass;
 202    struct anv_subpass *subpasses;
 203    struct anv_render_pass_attachment *attachments;
 204    enum anv_pipe_bits *subpass_flushes;
 205
 206    ANV_MULTIALLOC(ma);
 207    anv_multialloc_add(&ma, &pass, 1);
 208    anv_multialloc_add(&ma, &subpasses, pCreateInfo->subpassCount);
 209    anv_multialloc_add(&ma, &attachments, pCreateInfo->attachmentCount);
 210    anv_multialloc_add(&ma, &subpass_flushes, pCreateInfo->subpassCount + 1);
 211
 212    struct anv_subpass_attachment *subpass_attachments;
 213    uint32_t subpass_attachment_count = 0;
 214    for (uint32_t i = 0; i < pCreateInfo->subpassCount; i++) {
 215       subpass_attachment_count +=
 216          num_subpass_attachments(&pCreateInfo->pSubpasses[i]);
 217    }
 218    anv_multialloc_add(&ma, &subpass_attachments, subpass_attachment_count);
 219
 220    if (!anv_multialloc_alloc2(&ma, &device->alloc, pAllocator,
 221                               VK_SYSTEM_ALLOCATION_SCOPE_OBJECT))
 222       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 223
 224    /* Clear the subpasses along with the parent pass. This required because
 225     * each array member of anv_subpass must be a valid pointer if not NULL.
 226     */
 227    memset(pass, 0, ma.size);
 228    pass->attachment_count = pCreateInfo->attachmentCount;
 229    pass->subpass_count = pCreateInfo->subpassCount;
 230    pass->attachments = attachments;
 231    pass->subpass_flushes = subpass_flushes;
 232
 233    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
 234       pass->attachments[i] = (struct anv_render_pass_attachment) {
 235          .format                 = pCreateInfo->pAttachments[i].format,
 236          .samples                = pCreateInfo->pAttachments[i].samples,
 237          .load_op                = pCreateInfo->pAttachments[i].loadOp,
 238          .store_op               = pCreateInfo->pAttachments[i].storeOp,
 239          .stencil_load_op        = pCreateInfo->pAttachments[i].stencilLoadOp,
 240          .initial_layout         = pCreateInfo->pAttachments[i].initialLayout,
 241          .final_layout           = pCreateInfo->pAttachments[i].finalLayout,
 242       };
 243    }
 244
 245    for (uint32_t i = 0; i < pCreateInfo->subpassCount; i++) {
 246       const VkSubpassDescription *desc = &pCreateInfo->pSubpasses[i];
 247       struct anv_subpass *subpass = &pass->subpasses[i];
 248
 249       subpass->input_count = desc->inputAttachmentCount;
 250       subpass->color_count = desc->colorAttachmentCount;
 251       subpass->attachment_count = num_subpass_attachments(desc);
 252       subpass->attachments = subpass_attachments;
 253       subpass->view_mask = 0;
 254
 255       if (desc->inputAttachmentCount > 0) {
 256          subpass->input_attachments = subpass_attachments;
 257          subpass_attachments += desc->inputAttachmentCount;
 258
 259          for (uint32_t j = 0; j < desc->inputAttachmentCount; j++) {
 260             subpass->input_attachments[j] = (struct anv_subpass_attachment) {
 261                .usage =       VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT,
 262                .attachment =  desc->pInputAttachments[j].attachment,
 263                .layout =      desc->pInputAttachments[j].layout,
 264             };
 265          }
 266       }
 267
 268       if (desc->colorAttachmentCount > 0) {
 269          subpass->color_attachments = subpass_attachments;
 270          subpass_attachments += desc->colorAttachmentCount;
 271
 272          for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
 273             subpass->color_attachments[j] = (struct anv_subpass_attachment) {
 274                .usage =       VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT,
 275                .attachment =  desc->pColorAttachments[j].attachment,
 276                .layout =      desc->pColorAttachments[j].layout,
 277             };
 278          }
 279       }
 280
 281       if (desc->pResolveAttachments) {
 282          subpass->resolve_attachments = subpass_attachments;
 283          subpass_attachments += desc->colorAttachmentCount;
 284
 285          for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
 286             subpass->resolve_attachments[j] = (struct anv_subpass_attachment) {
 287                .usage =       VK_IMAGE_USAGE_TRANSFER_DST_BIT,
 288                .attachment =  desc->pResolveAttachments[j].attachment,
 289                .layout =      desc->pResolveAttachments[j].layout,
 290             };
 291          }
 292       }
 293
 294       if (desc->pDepthStencilAttachment) {
 295          subpass->depth_stencil_attachment = subpass_attachments++;
 296
 297          *subpass->depth_stencil_attachment = (struct anv_subpass_attachment) {
 298             .usage =       VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT,
 299             .attachment =  desc->pDepthStencilAttachment->attachment,
 300             .layout =      desc->pDepthStencilAttachment->layout,
 301          };
 302       }
 303    }
 304
 305    for (uint32_t i = 0; i < pCreateInfo->dependencyCount; i++) {
 306       /* Convert to a Dependency2KHR */
 307       struct VkSubpassDependency2KHR dep2 = {
 308          .srcSubpass       = pCreateInfo->pDependencies[i].srcSubpass,
 309          .dstSubpass       = pCreateInfo->pDependencies[i].dstSubpass,
 310          .srcStageMask     = pCreateInfo->pDependencies[i].srcStageMask,
 311          .dstStageMask     = pCreateInfo->pDependencies[i].dstStageMask,
 312          .srcAccessMask    = pCreateInfo->pDependencies[i].srcAccessMask,
 313          .dstAccessMask    = pCreateInfo->pDependencies[i].dstAccessMask,
 314          .dependencyFlags  = pCreateInfo->pDependencies[i].dependencyFlags,
 315       };
 316       anv_render_pass_add_subpass_dep(pass, &dep2);
 317    }
 318
 319    vk_foreach_struct(ext, pCreateInfo->pNext) {
 320       switch (ext->sType) {
 321       case VK_STRUCTURE_TYPE_RENDER_PASS_MULTIVIEW_CREATE_INFO_KHR: {
 322          VkRenderPassMultiviewCreateInfoKHR *mv = (void *)ext;
 323
 324          for (uint32_t i = 0; i < mv->subpassCount; i++) {
 325             pass->subpasses[i].view_mask = mv->pViewMasks[i];
 326          }
 327          break;
 328       }
 329
 330       default:
 331          anv_debug_ignored_stype(ext->sType);
 332       }
 333    }
 334
 335    anv_render_pass_compile(pass);
 336
 337    *pRenderPass = anv_render_pass_to_handle(pass);
 338
 339    return VK_SUCCESS;
 340 }
 341
 342 void anv_DestroyRenderPass(
 343     VkDevice                                    _device,
 344     VkRenderPass                                _pass,
 345     const VkAllocationCallbacks*                pAllocator)
 346 {
 347    ANV_FROM_HANDLE(anv_device, device, _device);
 348    ANV_FROM_HANDLE(anv_render_pass, pass, _pass);
 349
 350    vk_free2(&device->alloc, pAllocator, pass);
 351 }
 352
 353 void anv_GetRenderAreaGranularity(
 354     VkDevice                                    device,
 355     VkRenderPass                                renderPass,
 356     VkExtent2D*                                 pGranularity)
 357 {
 358    ANV_FROM_HANDLE(anv_render_pass, pass, renderPass);
 359
 360    /* This granularity satisfies HiZ fast clear alignment requirements
 361     * for all sample counts.
 362     */
 363    for (unsigned i = 0; i < pass->subpass_count; ++i) {
 364       if (pass->subpasses[i].depth_stencil_attachment) {
 365          *pGranularity = (VkExtent2D) { .width = 8, .height = 4 };
 366          return;
 367       }
 368    }
 369
 370    *pGranularity = (VkExtent2D) { 1, 1 };
 371 }