src/freedreno/vulkan/tu_pass.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 #include "tu_private.h"
  28
  29 #include "vk_util.h"
  30 #include "vk_format.h"
  31
  32 static void update_samples(struct tu_subpass *subpass,
  33                            VkSampleCountFlagBits samples)
  34 {
  35    assert(subpass->samples == 0 || subpass->samples == samples);
  36    subpass->samples = samples;
  37 }
  38
  39 #define GMEM_ALIGN 0x4000
  40
  41 static void
  42 compute_gmem_offsets(struct tu_render_pass *pass, uint32_t gmem_size)
  43 {
  44    /* calculate total bytes per pixel */
  45    uint32_t cpp_total = 0;
  46    for (uint32_t i = 0; i < pass->attachment_count; i++) {
  47       struct tu_render_pass_attachment *att = &pass->attachments[i];
  48       if (att->gmem_offset >= 0)
  49          cpp_total += att->cpp;
  50    }
  51
  52    /* no gmem attachments */
  53    if (cpp_total == 0) {
  54       /* any value non-zero value so tiling config works with no attachments */
  55       pass->gmem_pixels = 1024*1024;
  56       return;
  57    }
  58
  59    /* TODO: this algorithm isn't optimal
  60     * for example, two attachments with cpp = {1, 4}
  61     * result:  nblocks = {12, 52}, pixels = 196608
  62     * optimal: nblocks = {13, 51}, pixels = 208896
  63     */
  64    uint32_t gmem_blocks = gmem_size / GMEM_ALIGN;
  65    uint32_t offset = 0, pixels = ~0u;
  66    for (uint32_t i = 0; i < pass->attachment_count; i++) {
  67       struct tu_render_pass_attachment *att = &pass->attachments[i];
  68       if (att->gmem_offset < 0)
  69          continue;
  70
  71       att->gmem_offset = offset;
  72
  73       /* Note: divide by 16 is for GMEM_ALIGN=16k, tile align w=64/h=16 */
  74       uint32_t align = MAX2(1, att->cpp / 16);
  75       uint32_t nblocks = MAX2((gmem_blocks * att->cpp / cpp_total) & ~(align - 1), align);
  76
  77       gmem_blocks -= nblocks;
  78       cpp_total -= att->cpp;
  79       offset += nblocks * GMEM_ALIGN;
  80       pixels = MIN2(pixels, nblocks * GMEM_ALIGN / att->cpp);
  81    }
  82
  83    pass->gmem_pixels = pixels;
  84 }
  85
  86 VkResult
  87 tu_CreateRenderPass(VkDevice _device,
  88                     const VkRenderPassCreateInfo *pCreateInfo,
  89                     const VkAllocationCallbacks *pAllocator,
  90                     VkRenderPass *pRenderPass)
  91 {
  92    TU_FROM_HANDLE(tu_device, device, _device);
  93    struct tu_render_pass *pass;
  94    size_t size;
  95    size_t attachments_offset;
  96
  97    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO);
  98
  99    size = sizeof(*pass);
 100    size += pCreateInfo->subpassCount * sizeof(pass->subpasses[0]);
 101    attachments_offset = size;
 102    size += pCreateInfo->attachmentCount * sizeof(pass->attachments[0]);
 103
 104    pass = vk_alloc2(&device->alloc, pAllocator, size, 8,
 105                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 106    if (pass == NULL)
 107       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 108
 109    memset(pass, 0, size);
 110    pass->attachment_count = pCreateInfo->attachmentCount;
 111    pass->subpass_count = pCreateInfo->subpassCount;
 112    pass->attachments = (void *) pass + attachments_offset;
 113
 114    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
 115       struct tu_render_pass_attachment *att = &pass->attachments[i];
 116
 117       att->format = pCreateInfo->pAttachments[i].format;
 118       att->cpp = vk_format_get_blocksize(att->format) *
 119                            pCreateInfo->pAttachments[i].samples;
 120       att->load_op = pCreateInfo->pAttachments[i].loadOp;
 121       att->stencil_load_op = pCreateInfo->pAttachments[i].stencilLoadOp;
 122       att->store_op = pCreateInfo->pAttachments[i].storeOp;
 123       if (pCreateInfo->pAttachments[i].stencilStoreOp == VK_ATTACHMENT_STORE_OP_STORE &&
 124           vk_format_has_stencil(att->format))
 125          att->store_op = VK_ATTACHMENT_STORE_OP_STORE;
 126       att->gmem_offset = -1;
 127    }
 128
 129    uint32_t subpass_attachment_count = 0;
 130    struct tu_subpass_attachment *p;
 131    for (uint32_t i = 0; i < pCreateInfo->subpassCount; i++) {
 132       const VkSubpassDescription *desc = &pCreateInfo->pSubpasses[i];
 133
 134       subpass_attachment_count +=
 135          desc->inputAttachmentCount + desc->colorAttachmentCount +
 136          (desc->pResolveAttachments ? desc->colorAttachmentCount : 0);
 137    }
 138
 139    if (subpass_attachment_count) {
 140       pass->subpass_attachments = vk_alloc2(
 141          &device->alloc, pAllocator,
 142          subpass_attachment_count * sizeof(struct tu_subpass_attachment), 8,
 143          VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 144       if (pass->subpass_attachments == NULL) {
 145          vk_free2(&device->alloc, pAllocator, pass);
 146          return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 147       }
 148    } else
 149       pass->subpass_attachments = NULL;
 150
 151    p = pass->subpass_attachments;
 152    for (uint32_t i = 0; i < pCreateInfo->subpassCount; i++) {
 153       const VkSubpassDescription *desc = &pCreateInfo->pSubpasses[i];
 154       struct tu_subpass *subpass = &pass->subpasses[i];
 155
 156       subpass->input_count = desc->inputAttachmentCount;
 157       subpass->color_count = desc->colorAttachmentCount;
 158       subpass->samples = 0;
 159
 160       if (desc->inputAttachmentCount > 0) {
 161          subpass->input_attachments = p;
 162          p += desc->inputAttachmentCount;
 163
 164          for (uint32_t j = 0; j < desc->inputAttachmentCount; j++) {
 165             uint32_t a = desc->pInputAttachments[j].attachment;
 166             subpass->input_attachments[j].attachment = a;
 167             if (a != VK_ATTACHMENT_UNUSED)
 168                pass->attachments[a].gmem_offset = 0;
 169          }
 170       }
 171
 172       if (desc->colorAttachmentCount > 0) {
 173          subpass->color_attachments = p;
 174          p += desc->colorAttachmentCount;
 175
 176          for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
 177             uint32_t a = desc->pColorAttachments[j].attachment;
 178             subpass->color_attachments[j].attachment = a;
 179
 180             if (a != VK_ATTACHMENT_UNUSED) {
 181                pass->attachments[a].gmem_offset = 0;
 182                update_samples(subpass, pCreateInfo->pAttachments[a].samples);
 183             }
 184          }
 185       }
 186
 187       subpass->resolve_attachments = desc->pResolveAttachments ? p : NULL;
 188       if (desc->pResolveAttachments) {
 189          p += desc->colorAttachmentCount;
 190          for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
 191             subpass->resolve_attachments[j].attachment =
 192                   desc->pResolveAttachments[j].attachment;
 193          }
 194       }
 195
 196       uint32_t a = desc->pDepthStencilAttachment ?
 197          desc->pDepthStencilAttachment->attachment : VK_ATTACHMENT_UNUSED;
 198       subpass->depth_stencil_attachment.attachment = a;
 199       if (a != VK_ATTACHMENT_UNUSED) {
 200             pass->attachments[a].gmem_offset = 0;
 201             update_samples(subpass, pCreateInfo->pAttachments[a].samples);
 202       }
 203
 204       subpass->samples = subpass->samples ?: 1;
 205    }
 206
 207    *pRenderPass = tu_render_pass_to_handle(pass);
 208
 209    compute_gmem_offsets(pass, device->physical_device->gmem_size);
 210
 211    return VK_SUCCESS;
 212 }
 213
 214 VkResult
 215 tu_CreateRenderPass2(VkDevice _device,
 216                      const VkRenderPassCreateInfo2KHR *pCreateInfo,
 217                      const VkAllocationCallbacks *pAllocator,
 218                      VkRenderPass *pRenderPass)
 219 {
 220    TU_FROM_HANDLE(tu_device, device, _device);
 221    struct tu_render_pass *pass;
 222    size_t size;
 223    size_t attachments_offset;
 224
 225    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO_2_KHR);
 226
 227    size = sizeof(*pass);
 228    size += pCreateInfo->subpassCount * sizeof(pass->subpasses[0]);
 229    attachments_offset = size;
 230    size += pCreateInfo->attachmentCount * sizeof(pass->attachments[0]);
 231
 232    pass = vk_alloc2(&device->alloc, pAllocator, size, 8,
 233                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 234    if (pass == NULL)
 235       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 236
 237    memset(pass, 0, size);
 238    pass->attachment_count = pCreateInfo->attachmentCount;
 239    pass->subpass_count = pCreateInfo->subpassCount;
 240    pass->attachments = (void *) pass + attachments_offset;
 241
 242    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
 243       struct tu_render_pass_attachment *att = &pass->attachments[i];
 244
 245       att->format = pCreateInfo->pAttachments[i].format;
 246       att->cpp = vk_format_get_blocksize(att->format) *
 247                            pCreateInfo->pAttachments[i].samples;
 248       att->load_op = pCreateInfo->pAttachments[i].loadOp;
 249       att->stencil_load_op = pCreateInfo->pAttachments[i].stencilLoadOp;
 250       att->store_op = pCreateInfo->pAttachments[i].storeOp;
 251       att->stencil_store_op = pCreateInfo->pAttachments[i].stencilStoreOp;
 252       if (pCreateInfo->pAttachments[i].stencilStoreOp == VK_ATTACHMENT_STORE_OP_STORE &&
 253           vk_format_has_stencil(att->format))
 254          att->store_op = VK_ATTACHMENT_STORE_OP_STORE;
 255       att->gmem_offset = -1;
 256    }
 257    uint32_t subpass_attachment_count = 0;
 258    struct tu_subpass_attachment *p;
 259    for (uint32_t i = 0; i < pCreateInfo->subpassCount; i++) {
 260       const VkSubpassDescription2KHR *desc = &pCreateInfo->pSubpasses[i];
 261
 262       subpass_attachment_count +=
 263          desc->inputAttachmentCount + desc->colorAttachmentCount +
 264          (desc->pResolveAttachments ? desc->colorAttachmentCount : 0);
 265    }
 266
 267    if (subpass_attachment_count) {
 268       pass->subpass_attachments = vk_alloc2(
 269          &device->alloc, pAllocator,
 270          subpass_attachment_count * sizeof(struct tu_subpass_attachment), 8,
 271          VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 272       if (pass->subpass_attachments == NULL) {
 273          vk_free2(&device->alloc, pAllocator, pass);
 274          return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 275       }
 276    } else
 277       pass->subpass_attachments = NULL;
 278
 279    p = pass->subpass_attachments;
 280    for (uint32_t i = 0; i < pCreateInfo->subpassCount; i++) {
 281       const VkSubpassDescription2KHR *desc = &pCreateInfo->pSubpasses[i];
 282       struct tu_subpass *subpass = &pass->subpasses[i];
 283
 284       subpass->input_count = desc->inputAttachmentCount;
 285       subpass->color_count = desc->colorAttachmentCount;
 286       subpass->samples = 0;
 287
 288       if (desc->inputAttachmentCount > 0) {
 289          subpass->input_attachments = p;
 290          p += desc->inputAttachmentCount;
 291
 292          for (uint32_t j = 0; j < desc->inputAttachmentCount; j++) {
 293             uint32_t a = desc->pInputAttachments[j].attachment;
 294             subpass->input_attachments[j].attachment = a;
 295             if (a != VK_ATTACHMENT_UNUSED)
 296                pass->attachments[a].gmem_offset = 0;
 297          }
 298       }
 299
 300       if (desc->colorAttachmentCount > 0) {
 301          subpass->color_attachments = p;
 302          p += desc->colorAttachmentCount;
 303
 304          for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
 305             uint32_t a = desc->pColorAttachments[j].attachment;
 306             subpass->color_attachments[j].attachment = a;
 307
 308             if (a != VK_ATTACHMENT_UNUSED) {
 309                pass->attachments[a].gmem_offset = 0;
 310                update_samples(subpass, pCreateInfo->pAttachments[a].samples);
 311             }
 312          }
 313       }
 314
 315       subpass->resolve_attachments = desc->pResolveAttachments ? p : NULL;
 316       if (desc->pResolveAttachments) {
 317          p += desc->colorAttachmentCount;
 318          for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
 319             subpass->resolve_attachments[j].attachment =
 320                   desc->pResolveAttachments[j].attachment;
 321          }
 322       }
 323
 324
 325       uint32_t a = desc->pDepthStencilAttachment ?
 326          desc->pDepthStencilAttachment->attachment : VK_ATTACHMENT_UNUSED;
 327       subpass->depth_stencil_attachment.attachment = a;
 328       if (a != VK_ATTACHMENT_UNUSED) {
 329             pass->attachments[a].gmem_offset = 0;
 330             update_samples(subpass, pCreateInfo->pAttachments[a].samples);
 331       }
 332
 333       subpass->samples = subpass->samples ?: 1;
 334    }
 335
 336    *pRenderPass = tu_render_pass_to_handle(pass);
 337
 338    compute_gmem_offsets(pass, device->physical_device->gmem_size);
 339
 340    return VK_SUCCESS;
 341 }
 342
 343 void
 344 tu_DestroyRenderPass(VkDevice _device,
 345                      VkRenderPass _pass,
 346                      const VkAllocationCallbacks *pAllocator)
 347 {
 348    TU_FROM_HANDLE(tu_device, device, _device);
 349    TU_FROM_HANDLE(tu_render_pass, pass, _pass);
 350
 351    if (!_pass)
 352       return;
 353
 354    vk_free2(&device->alloc, pAllocator, pass->subpass_attachments);
 355    vk_free2(&device->alloc, pAllocator, pass);
 356 }
 357
 358 void
 359 tu_GetRenderAreaGranularity(VkDevice _device,
 360                             VkRenderPass renderPass,
 361                             VkExtent2D *pGranularity)
 362 {
 363    TU_FROM_HANDLE(tu_device, device, _device);
 364
 365    pGranularity->width = device->physical_device->tile_align_w;
 366    pGranularity->height = device->physical_device->tile_align_h;
 367 }