external/vulkancts/modules/vulkan/sparse_resources/vktSparseResourcesBufferTests.cpp

   1 /*------------------------------------------------------------------------
   2  * Vulkan Conformance Tests
   3  * ------------------------
   4  *
   5  * Copyright (c) 2016 The Khronos Group Inc.
   6  *
   7  * Licensed under the Apache License, Version 2.0 (the "License");
   8  * you may not use this file except in compliance with the License.
   9  * You may obtain a copy of the License at
  10  *
  11  *      http://www.apache.org/licenses/LICENSE-2.0
  12  *
  13  * Unless required by applicable law or agreed to in writing, software
  14  * distributed under the License is distributed on an "AS IS" BASIS,
  15  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  16  * See the License for the specific language governing permissions and
  17  * limitations under the License.
  18  *
  19  *//*!
  20  * \file
  21  * \brief Sparse buffer tests
  22  *//*--------------------------------------------------------------------*/
  23
  24 #include "vktSparseResourcesBufferTests.hpp"
  25 #include "vktTestCaseUtil.hpp"
  26 #include "vktTestGroupUtil.hpp"
  27 #include "vktSparseResourcesTestsUtil.hpp"
  28 #include "vktSparseResourcesBase.hpp"
  29 #include "vktSparseResourcesBufferSparseBinding.hpp"
  30 #include "vktSparseResourcesBufferSparseResidency.hpp"
  31 #include "vktSparseResourcesBufferMemoryAliasing.hpp"
  32
  33 #include "vkRef.hpp"
  34 #include "vkRefUtil.hpp"
  35 #include "vkPlatform.hpp"
  36 #include "vkPrograms.hpp"
  37 #include "vkMemUtil.hpp"
  38 #include "vkBuilderUtil.hpp"
  39 #include "vkQueryUtil.hpp"
  40 #include "vkTypeUtil.hpp"
  41
  42 #include "tcuTestLog.hpp"
  43
  44 #include "deUniquePtr.hpp"
  45 #include "deSharedPtr.hpp"
  46 #include "deMath.h"
  47
  48 #include <string>
  49 #include <vector>
  50 #include <map>
  51
  52 using namespace vk;
  53 using de::MovePtr;
  54 using de::UniquePtr;
  55 using de::SharedPtr;
  56 using tcu::Vec4;
  57 using tcu::IVec2;
  58 using tcu::IVec4;
  59
  60 namespace vkt
  61 {
  62 namespace sparse
  63 {
  64 namespace
  65 {
  66
  67 typedef SharedPtr<UniquePtr<Allocation> > AllocationSp;
  68
  69 enum
  70 {
  71         RENDER_SIZE     = 128,                          //!< framebuffer size in pixels
  72         GRID_SIZE       = RENDER_SIZE / 8,      //!< number of grid tiles in a row
  73 };
  74
  75 enum TestFlagBits
  76 {
  77                                                                                                 //   sparseBinding is implied
  78         TEST_FLAG_ALIASED                               = 1u << 0,      //!< sparseResidencyAliased
  79         TEST_FLAG_RESIDENCY                             = 1u << 1,      //!< sparseResidencyBuffer
  80         TEST_FLAG_NON_RESIDENT_STRICT   = 1u << 2,      //!< residencyNonResidentStrict
  81 };
  82 typedef deUint32 TestFlags;
  83
  84 //! SparseAllocationBuilder output. Owns the allocated memory.
  85 struct SparseAllocation
  86 {
  87         deUint32                                                        numResourceChunks;
  88         VkDeviceSize                                            resourceSize;           //!< buffer size in bytes
  89         std::vector<AllocationSp>                       allocations;            //!< actual allocated memory
  90         std::vector<VkSparseMemoryBind>         memoryBinds;            //!< memory binds backing the resource
  91 };
  92
  93 //! Utility to lay out memory allocations for a sparse buffer, including holes and aliased regions.
  94 //! Will allocate memory upon building.
  95 class SparseAllocationBuilder
  96 {
  97 public:
  98                                                                 SparseAllocationBuilder (void);
  99
 100         // \note "chunk" is the smallest (due to alignment) bindable amount of memory
 101
 102         SparseAllocationBuilder&        addMemoryHole                   (const deUint32 numChunks = 1u);
 103         SparseAllocationBuilder&        addResourceHole                 (const deUint32 numChunks = 1u);
 104         SparseAllocationBuilder&        addMemoryBind                   (const deUint32 numChunks = 1u);
 105         SparseAllocationBuilder&        addAliasedMemoryBind    (const deUint32 allocationNdx, const deUint32 chunkOffset, const deUint32 numChunks = 1u);
 106         SparseAllocationBuilder&        addMemoryAllocation             (void);
 107
 108         MovePtr<SparseAllocation>       build                                   (const DeviceInterface&         vk,
 109                                                                                                                  const VkDevice                         device,
 110                                                                                                                  Allocator&                                     allocator,
 111                                                                                                                  VkBufferCreateInfo                     referenceCreateInfo,            //!< buffer size is ignored in this info
 112                                                                                                                  const VkDeviceSize                     minChunkSize = 0ull) const;     //!< make sure chunks are at least this big
 113
 114 private:
 115         struct MemoryBind
 116         {
 117                 deUint32        allocationNdx;
 118                 deUint32        resourceChunkNdx;
 119                 deUint32        memoryChunkNdx;
 120                 deUint32        numChunks;
 121         };
 122
 123         deUint32                                        m_allocationNdx;
 124         deUint32                                        m_resourceChunkNdx;
 125         deUint32                                        m_memoryChunkNdx;
 126         std::vector<MemoryBind>         m_memoryBinds;
 127         std::vector<deUint32>           m_chunksPerAllocation;
 128
 129 };
 130
 131 SparseAllocationBuilder::SparseAllocationBuilder (void)
 132         : m_allocationNdx               (0)
 133         , m_resourceChunkNdx    (0)
 134         , m_memoryChunkNdx              (0)
 135 {
 136         m_chunksPerAllocation.push_back(0);
 137 }
 138
 139 SparseAllocationBuilder& SparseAllocationBuilder::addMemoryHole (const deUint32 numChunks)
 140 {
 141         m_memoryChunkNdx                                                += numChunks;
 142         m_chunksPerAllocation[m_allocationNdx]  += numChunks;
 143
 144         return *this;
 145 }
 146
 147 SparseAllocationBuilder& SparseAllocationBuilder::addResourceHole (const deUint32 numChunks)
 148 {
 149         m_resourceChunkNdx += numChunks;
 150
 151         return *this;
 152 }
 153
 154 SparseAllocationBuilder& SparseAllocationBuilder::addMemoryAllocation (void)
 155 {
 156         DE_ASSERT(m_memoryChunkNdx != 0);       // doesn't make sense to have an empty allocation
 157
 158         m_allocationNdx  += 1;
 159         m_memoryChunkNdx  = 0;
 160         m_chunksPerAllocation.push_back(0);
 161
 162         return *this;
 163 }
 164
 165 SparseAllocationBuilder& SparseAllocationBuilder::addMemoryBind (const deUint32 numChunks)
 166 {
 167         const MemoryBind memoryBind =
 168         {
 169                 m_allocationNdx,
 170                 m_resourceChunkNdx,
 171                 m_memoryChunkNdx,
 172                 numChunks
 173         };
 174         m_memoryBinds.push_back(memoryBind);
 175
 176         m_resourceChunkNdx                                              += numChunks;
 177         m_memoryChunkNdx                                                += numChunks;
 178         m_chunksPerAllocation[m_allocationNdx]  += numChunks;
 179
 180         return *this;
 181 }
 182
 183 SparseAllocationBuilder& SparseAllocationBuilder::addAliasedMemoryBind  (const deUint32 allocationNdx, const deUint32 chunkOffset, const deUint32 numChunks)
 184 {
 185         DE_ASSERT(allocationNdx <= m_allocationNdx);
 186
 187         const MemoryBind memoryBind =
 188         {
 189                 allocationNdx,
 190                 m_resourceChunkNdx,
 191                 chunkOffset,
 192                 numChunks
 193         };
 194         m_memoryBinds.push_back(memoryBind);
 195
 196         m_resourceChunkNdx += numChunks;
 197
 198         return *this;
 199 }
 200
 201 inline VkMemoryRequirements requirementsWithSize (VkMemoryRequirements requirements, const VkDeviceSize size)
 202 {
 203         requirements.size = size;
 204         return requirements;
 205 }
 206
 207 MovePtr<SparseAllocation> SparseAllocationBuilder::build (const DeviceInterface&        vk,
 208                                                                                                                   const VkDevice                        device,
 209                                                                                                                   Allocator&                            allocator,
 210                                                                                                                   VkBufferCreateInfo            referenceCreateInfo,
 211                                                                                                                   const VkDeviceSize            minChunkSize) const
 212 {
 213
 214         MovePtr<SparseAllocation>       sparseAllocation                        (new SparseAllocation());
 215
 216                                                                 referenceCreateInfo.size        = sizeof(deUint32);
 217         const Unique<VkBuffer>          refBuffer                                       (createBuffer(vk, device, &referenceCreateInfo));
 218         const VkMemoryRequirements      memoryRequirements                      = getBufferMemoryRequirements(vk, device, *refBuffer);
 219         const VkDeviceSize                      chunkSize                                       = std::max(memoryRequirements.alignment, static_cast<VkDeviceSize>(deAlign64(minChunkSize, memoryRequirements.alignment)));
 220
 221         for (std::vector<deUint32>::const_iterator numChunksIter = m_chunksPerAllocation.begin(); numChunksIter != m_chunksPerAllocation.end(); ++numChunksIter)
 222         {
 223                 sparseAllocation->allocations.push_back(makeDeSharedPtr(
 224                         allocator.allocate(requirementsWithSize(memoryRequirements, *numChunksIter * chunkSize), MemoryRequirement::Any)));
 225         }
 226
 227         for (std::vector<MemoryBind>::const_iterator memBindIter = m_memoryBinds.begin(); memBindIter != m_memoryBinds.end(); ++memBindIter)
 228         {
 229                 const Allocation&                       alloc   = **sparseAllocation->allocations[memBindIter->allocationNdx];
 230                 const VkSparseMemoryBind        bind    =
 231                 {
 232                         memBindIter->resourceChunkNdx * chunkSize,                                                      // VkDeviceSize               resourceOffset;
 233                         memBindIter->numChunks * chunkSize,                                                                     // VkDeviceSize               size;
 234                         alloc.getMemory(),                                                                                                      // VkDeviceMemory             memory;
 235                         alloc.getOffset() + memBindIter->memoryChunkNdx * chunkSize,            // VkDeviceSize               memoryOffset;
 236                         (VkSparseMemoryBindFlags)0,                                                                                     // VkSparseMemoryBindFlags    flags;
 237                 };
 238                 sparseAllocation->memoryBinds.push_back(bind);
 239                 referenceCreateInfo.size = std::max(referenceCreateInfo.size, bind.resourceOffset + bind.size);
 240         }
 241
 242         sparseAllocation->resourceSize          = referenceCreateInfo.size;
 243         sparseAllocation->numResourceChunks = m_resourceChunkNdx;
 244
 245         return sparseAllocation;
 246 }
 247
 248 VkImageCreateInfo makeImageCreateInfo (const VkFormat format, const IVec2& size, const VkImageUsageFlags usage)
 249 {
 250         const VkImageCreateInfo imageParams =
 251         {
 252                 VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,                    // VkStructureType                      sType;
 253                 DE_NULL,                                                                                // const void*                          pNext;
 254                 (VkImageCreateFlags)0,                                                  // VkImageCreateFlags           flags;
 255                 VK_IMAGE_TYPE_2D,                                                               // VkImageType                          imageType;
 256                 format,                                                                                 // VkFormat                                     format;
 257                 makeExtent3D(size.x(), size.y(), 1),                    // VkExtent3D                           extent;
 258                 1u,                                                                                             // deUint32                                     mipLevels;
 259                 1u,                                                                                             // deUint32                                     arrayLayers;
 260                 VK_SAMPLE_COUNT_1_BIT,                                                  // VkSampleCountFlagBits        samples;
 261                 VK_IMAGE_TILING_OPTIMAL,                                                // VkImageTiling                        tiling;
 262                 usage,                                                                                  // VkImageUsageFlags            usage;
 263                 VK_SHARING_MODE_EXCLUSIVE,                                              // VkSharingMode                        sharingMode;
 264                 0u,                                                                                             // deUint32                                     queueFamilyIndexCount;
 265                 DE_NULL,                                                                                // const deUint32*                      pQueueFamilyIndices;
 266                 VK_IMAGE_LAYOUT_UNDEFINED,                                              // VkImageLayout                        initialLayout;
 267         };
 268         return imageParams;
 269 }
 270
 271 Move<VkRenderPass> makeRenderPass (const DeviceInterface&       vk,
 272                                                                    const VkDevice                       device,
 273                                                                    const VkFormat                       colorFormat)
 274 {
 275         const VkAttachmentDescription colorAttachmentDescription =
 276         {
 277                 (VkAttachmentDescriptionFlags)0,                                        // VkAttachmentDescriptionFlags         flags;
 278                 colorFormat,                                                                            // VkFormat                                                     format;
 279                 VK_SAMPLE_COUNT_1_BIT,                                                          // VkSampleCountFlagBits                        samples;
 280                 VK_ATTACHMENT_LOAD_OP_CLEAR,                                            // VkAttachmentLoadOp                           loadOp;
 281                 VK_ATTACHMENT_STORE_OP_STORE,                                           // VkAttachmentStoreOp                          storeOp;
 282                 VK_ATTACHMENT_LOAD_OP_DONT_CARE,                                        // VkAttachmentLoadOp                           stencilLoadOp;
 283                 VK_ATTACHMENT_STORE_OP_DONT_CARE,                                       // VkAttachmentStoreOp                          stencilStoreOp;
 284                 VK_IMAGE_LAYOUT_UNDEFINED,                                                      // VkImageLayout                                        initialLayout;
 285                 VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,                       // VkImageLayout                                        finalLayout;
 286         };
 287
 288         const VkAttachmentReference colorAttachmentRef =
 289         {
 290                 0u,                                                                                                     // deUint32                     attachment;
 291                 VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL                        // VkImageLayout        layout;
 292         };
 293
 294         const VkSubpassDescription subpassDescription =
 295         {
 296                 (VkSubpassDescriptionFlags)0,                                           // VkSubpassDescriptionFlags            flags;
 297                 VK_PIPELINE_BIND_POINT_GRAPHICS,                                        // VkPipelineBindPoint                          pipelineBindPoint;
 298                 0u,                                                                                                     // deUint32                                                     inputAttachmentCount;
 299                 DE_NULL,                                                                                        // const VkAttachmentReference*         pInputAttachments;
 300                 1u,                                                                                                     // deUint32                                                     colorAttachmentCount;
 301                 &colorAttachmentRef,                                                            // const VkAttachmentReference*         pColorAttachments;
 302                 DE_NULL,                                                                                        // const VkAttachmentReference*         pResolveAttachments;
 303                 DE_NULL,                                                                                        // const VkAttachmentReference*         pDepthStencilAttachment;
 304                 0u,                                                                                                     // deUint32                                                     preserveAttachmentCount;
 305                 DE_NULL                                                                                         // const deUint32*                                      pPreserveAttachments;
 306         };
 307
 308         const VkRenderPassCreateInfo renderPassInfo =
 309         {
 310                 VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,                      // VkStructureType                                      sType;
 311                 DE_NULL,                                                                                        // const void*                                          pNext;
 312                 (VkRenderPassCreateFlags)0,                                                     // VkRenderPassCreateFlags                      flags;
 313                 1u,                                                                                                     // deUint32                                                     attachmentCount;
 314                 &colorAttachmentDescription,                                            // const VkAttachmentDescription*       pAttachments;
 315                 1u,                                                                                                     // deUint32                                                     subpassCount;
 316                 &subpassDescription,                                                            // const VkSubpassDescription*          pSubpasses;
 317                 0u,                                                                                                     // deUint32                                                     dependencyCount;
 318                 DE_NULL                                                                                         // const VkSubpassDependency*           pDependencies;
 319         };
 320
 321         return createRenderPass(vk, device, &renderPassInfo);
 322 }
 323
 324 Move<VkPipeline> makeGraphicsPipeline (const DeviceInterface&                                   vk,
 325                                                                            const VkDevice                                                       device,
 326                                                                            const VkPipelineLayout                                       pipelineLayout,
 327                                                                            const VkRenderPass                                           renderPass,
 328                                                                            const IVec2                                                          renderSize,
 329                                                                            const VkPrimitiveTopology                            topology,
 330                                                                            const deUint32                                                       stageCount,
 331                                                                            const VkPipelineShaderStageCreateInfo*       pStages)
 332 {
 333         const VkVertexInputBindingDescription vertexInputBindingDescription =
 334         {
 335                 0u,                                                             // uint32_t                             binding;
 336                 sizeof(Vec4),                                   // uint32_t                             stride;
 337                 VK_VERTEX_INPUT_RATE_VERTEX,    // VkVertexInputRate    inputRate;
 338         };
 339
 340         const VkVertexInputAttributeDescription vertexInputAttributeDescription =
 341         {
 342                 0u,                                                                     // uint32_t                     location;
 343                 0u,                                                                     // uint32_t                     binding;
 344                 VK_FORMAT_R32G32B32A32_SFLOAT,          // VkFormat                     format;
 345                 0u,                                                                     // uint32_t                     offset;
 346         };
 347
 348         const VkPipelineVertexInputStateCreateInfo vertexInputStateInfo =
 349         {
 350                 VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,              // VkStructureType                             sType;
 351                 DE_NULL,                                                                                                                // const void*                                 pNext;
 352                 (VkPipelineVertexInputStateCreateFlags)0,                                               // VkPipelineVertexInputStateCreateFlags       flags;
 353                 1u,                                                                                                                             // uint32_t                                    vertexBindingDescriptionCount;
 354                 &vertexInputBindingDescription,                                                                 // const VkVertexInputBindingDescription*      pVertexBindingDescriptions;
 355                 1u,                                                                                                                             // uint32_t                                    vertexAttributeDescriptionCount;
 356                 &vertexInputAttributeDescription,                                                               // const VkVertexInputAttributeDescription*    pVertexAttributeDescriptions;
 357         };
 358
 359         const VkPipelineInputAssemblyStateCreateInfo pipelineInputAssemblyStateInfo =
 360         {
 361                 VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,    // VkStructureType                             sType;
 362                 DE_NULL,                                                                                                                // const void*                                 pNext;
 363                 (VkPipelineInputAssemblyStateCreateFlags)0,                                             // VkPipelineInputAssemblyStateCreateFlags     flags;
 364                 topology,                                                                                                               // VkPrimitiveTopology                         topology;
 365                 VK_FALSE,                                                                                                               // VkBool32                                    primitiveRestartEnable;
 366         };
 367
 368         const VkViewport viewport = makeViewport(
 369                 0.0f, 0.0f,
 370                 static_cast<float>(renderSize.x()), static_cast<float>(renderSize.y()),
 371                 0.0f, 1.0f);
 372
 373         const VkRect2D scissor = {
 374                 makeOffset2D(0, 0),
 375                 makeExtent2D(static_cast<deUint32>(renderSize.x()), static_cast<deUint32>(renderSize.y())),
 376         };
 377
 378         const VkPipelineViewportStateCreateInfo pipelineViewportStateInfo =
 379         {
 380                 VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,                  // VkStructureType                             sType;
 381                 DE_NULL,                                                                                                                // const void*                                 pNext;
 382                 (VkPipelineViewportStateCreateFlags)0,                                                  // VkPipelineViewportStateCreateFlags          flags;
 383                 1u,                                                                                                                             // uint32_t                                    viewportCount;
 384                 &viewport,                                                                                                              // const VkViewport*                           pViewports;
 385                 1u,                                                                                                                             // uint32_t                                    scissorCount;
 386                 &scissor,                                                                                                               // const VkRect2D*                             pScissors;
 387         };
 388
 389         const VkPipelineRasterizationStateCreateInfo pipelineRasterizationStateInfo =
 390         {
 391                 VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,             // VkStructureType                          sType;
 392                 DE_NULL,                                                                                                                // const void*                              pNext;
 393                 (VkPipelineRasterizationStateCreateFlags)0,                                             // VkPipelineRasterizationStateCreateFlags  flags;
 394                 VK_FALSE,                                                                                                               // VkBool32                                 depthClampEnable;
 395                 VK_FALSE,                                                                                                               // VkBool32                                 rasterizerDiscardEnable;
 396                 VK_POLYGON_MODE_FILL,                                                                                   // VkPolygonMode                                                        polygonMode;
 397                 VK_CULL_MODE_NONE,                                                                                              // VkCullModeFlags                                                      cullMode;
 398                 VK_FRONT_FACE_COUNTER_CLOCKWISE,                                                                // VkFrontFace                                                          frontFace;
 399                 VK_FALSE,                                                                                                               // VkBool32                                                                     depthBiasEnable;
 400                 0.0f,                                                                                                                   // float                                                                        depthBiasConstantFactor;
 401                 0.0f,                                                                                                                   // float                                                                        depthBiasClamp;
 402                 0.0f,                                                                                                                   // float                                                                        depthBiasSlopeFactor;
 403                 1.0f,                                                                                                                   // float                                                                        lineWidth;
 404         };
 405
 406         const VkPipelineMultisampleStateCreateInfo pipelineMultisampleStateInfo =
 407         {
 408                 VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,               // VkStructureType                                                      sType;
 409                 DE_NULL,                                                                                                                // const void*                                                          pNext;
 410                 (VkPipelineMultisampleStateCreateFlags)0,                                               // VkPipelineMultisampleStateCreateFlags        flags;
 411                 VK_SAMPLE_COUNT_1_BIT,                                                                                  // VkSampleCountFlagBits                                        rasterizationSamples;
 412                 VK_FALSE,                                                                                                               // VkBool32                                                                     sampleShadingEnable;
 413                 0.0f,                                                                                                                   // float                                                                        minSampleShading;
 414                 DE_NULL,                                                                                                                // const VkSampleMask*                                          pSampleMask;
 415                 VK_FALSE,                                                                                                               // VkBool32                                                                     alphaToCoverageEnable;
 416                 VK_FALSE                                                                                                                // VkBool32                                                                     alphaToOneEnable;
 417         };
 418
 419         const VkStencilOpState stencilOpState = makeStencilOpState(
 420                 VK_STENCIL_OP_KEEP,                             // stencil fail
 421                 VK_STENCIL_OP_KEEP,                             // depth & stencil pass
 422                 VK_STENCIL_OP_KEEP,                             // depth only fail
 423                 VK_COMPARE_OP_ALWAYS,                   // compare op
 424                 0u,                                                             // compare mask
 425                 0u,                                                             // write mask
 426                 0u);                                                    // reference
 427
 428         VkPipelineDepthStencilStateCreateInfo pipelineDepthStencilStateInfo =
 429         {
 430                 VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,             // VkStructureType                                                      sType;
 431                 DE_NULL,                                                                                                                // const void*                                                          pNext;
 432                 (VkPipelineDepthStencilStateCreateFlags)0,                                              // VkPipelineDepthStencilStateCreateFlags       flags;
 433                 VK_FALSE,                                                                                                               // VkBool32                                                                     depthTestEnable;
 434                 VK_FALSE,                                                                                                               // VkBool32                                                                     depthWriteEnable;
 435                 VK_COMPARE_OP_LESS,                                                                                             // VkCompareOp                                                          depthCompareOp;
 436                 VK_FALSE,                                                                                                               // VkBool32                                                                     depthBoundsTestEnable;
 437                 VK_FALSE,                                                                                                               // VkBool32                                                                     stencilTestEnable;
 438                 stencilOpState,                                                                                                 // VkStencilOpState                                                     front;
 439                 stencilOpState,                                                                                                 // VkStencilOpState                                                     back;
 440                 0.0f,                                                                                                                   // float                                                                        minDepthBounds;
 441                 1.0f,                                                                                                                   // float                                                                        maxDepthBounds;
 442         };
 443
 444         const VkColorComponentFlags                                     colorComponentsAll                                      = VK_COLOR_COMPONENT_R_BIT | VK_COLOR_COMPONENT_G_BIT | VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT;
 445         const VkPipelineColorBlendAttachmentState       pipelineColorBlendAttachmentState       =
 446         {
 447                 VK_FALSE,                                               // VkBool32                                     blendEnable;
 448                 VK_BLEND_FACTOR_ONE,                    // VkBlendFactor                        srcColorBlendFactor;
 449                 VK_BLEND_FACTOR_ZERO,                   // VkBlendFactor                        dstColorBlendFactor;
 450                 VK_BLEND_OP_ADD,                                // VkBlendOp                            colorBlendOp;
 451                 VK_BLEND_FACTOR_ONE,                    // VkBlendFactor                        srcAlphaBlendFactor;
 452                 VK_BLEND_FACTOR_ZERO,                   // VkBlendFactor                        dstAlphaBlendFactor;
 453                 VK_BLEND_OP_ADD,                                // VkBlendOp                            alphaBlendOp;
 454                 colorComponentsAll,                             // VkColorComponentFlags        colorWriteMask;
 455         };
 456
 457         const VkPipelineColorBlendStateCreateInfo pipelineColorBlendStateInfo =
 458         {
 459                 VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,               // VkStructureType                                                              sType;
 460                 DE_NULL,                                                                                                                // const void*                                                                  pNext;
 461                 (VkPipelineColorBlendStateCreateFlags)0,                                                // VkPipelineColorBlendStateCreateFlags                 flags;
 462                 VK_FALSE,                                                                                                               // VkBool32                                                                             logicOpEnable;
 463                 VK_LOGIC_OP_COPY,                                                                                               // VkLogicOp                                                                    logicOp;
 464                 1u,                                                                                                                             // deUint32                                                                             attachmentCount;
 465                 &pipelineColorBlendAttachmentState,                                                             // const VkPipelineColorBlendAttachmentState*   pAttachments;
 466                 { 0.0f, 0.0f, 0.0f, 0.0f },                                                                             // float                                                                                blendConstants[4];
 467         };
 468
 469         const VkGraphicsPipelineCreateInfo graphicsPipelineInfo =
 470         {
 471                 VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,        // VkStructureType                                                                      sType;
 472                 DE_NULL,                                                                                        // const void*                                                                          pNext;
 473                 (VkPipelineCreateFlags)0,                                                       // VkPipelineCreateFlags                                                        flags;
 474                 stageCount,                                                                                     // deUint32                                                                                     stageCount;
 475                 pStages,                                                                                        // const VkPipelineShaderStageCreateInfo*                       pStages;
 476                 &vertexInputStateInfo,                                                          // const VkPipelineVertexInputStateCreateInfo*          pVertexInputState;
 477                 &pipelineInputAssemblyStateInfo,                                        // const VkPipelineInputAssemblyStateCreateInfo*        pInputAssemblyState;
 478                 DE_NULL,                                                                                        // const VkPipelineTessellationStateCreateInfo*         pTessellationState;
 479                 &pipelineViewportStateInfo,                                                     // const VkPipelineViewportStateCreateInfo*                     pViewportState;
 480                 &pipelineRasterizationStateInfo,                                        // const VkPipelineRasterizationStateCreateInfo*        pRasterizationState;
 481                 &pipelineMultisampleStateInfo,                                          // const VkPipelineMultisampleStateCreateInfo*          pMultisampleState;
 482                 &pipelineDepthStencilStateInfo,                                         // const VkPipelineDepthStencilStateCreateInfo*         pDepthStencilState;
 483                 &pipelineColorBlendStateInfo,                                           // const VkPipelineColorBlendStateCreateInfo*           pColorBlendState;
 484                 DE_NULL,                                                                                        // const VkPipelineDynamicStateCreateInfo*                      pDynamicState;
 485                 pipelineLayout,                                                                         // VkPipelineLayout                                                                     layout;
 486                 renderPass,                                                                                     // VkRenderPass                                                                         renderPass;
 487                 0u,                                                                                                     // deUint32                                                                                     subpass;
 488                 DE_NULL,                                                                                        // VkPipeline                                                                           basePipelineHandle;
 489                 0,                                                                                                      // deInt32                                                                                      basePipelineIndex;
 490         };
 491
 492         return createGraphicsPipeline(vk, device, DE_NULL, &graphicsPipelineInfo);
 493 }
 494
 495 //! Return true if there are any red (or all zero) pixels in the image
 496 bool imageHasErrorPixels (const tcu::ConstPixelBufferAccess image)
 497 {
 498         const Vec4 errorColor   = Vec4(1.0f, 0.0f, 0.0f, 1.0f);
 499         const Vec4 blankColor   = Vec4();
 500
 501         for (int y = 0; y < image.getHeight(); ++y)
 502         for (int x = 0; x < image.getWidth(); ++x)
 503         {
 504                 const Vec4 color = image.getPixel(x, y);
 505                 if (color == errorColor || color == blankColor)
 506                         return true;
 507         }
 508
 509         return false;
 510 }
 511
 512 class Renderer
 513 {
 514 public:
 515         typedef std::map<VkShaderStageFlagBits, const VkSpecializationInfo*>    SpecializationMap;
 516
 517         //! Use the delegate to bind descriptor sets, vertex buffers, etc. and make a draw call
 518         struct Delegate
 519         {
 520                 virtual                 ~Delegate               (void) {}
 521                 virtual void    rendererDraw    (const VkPipelineLayout pipelineLayout, const VkCommandBuffer cmdBuffer) const = 0;
 522         };
 523
 524         Renderer (const DeviceInterface&                                        vk,
 525                           const VkDevice                                                        device,
 526                           Allocator&                                                            allocator,
 527                           const deUint32                                                        queueFamilyIndex,
 528                           const VkDescriptorSetLayout                           descriptorSetLayout,    //!< may be NULL, if no descriptors are used
 529                           ProgramCollection<vk::ProgramBinary>&         binaryCollection,
 530                           const std::string&                                            vertexName,
 531                           const std::string&                                            fragmentName,
 532                           const VkBuffer                                                        colorBuffer,
 533                           const IVec2&                                                          renderSize,
 534                           const VkFormat                                                        colorFormat,
 535                           const Vec4&                                                           clearColor,
 536                           const VkPrimitiveTopology                                     topology,
 537                           SpecializationMap                                                     specMap = SpecializationMap())
 538                 : m_colorBuffer                         (colorBuffer)
 539                 , m_renderSize                          (renderSize)
 540                 , m_colorFormat                         (colorFormat)
 541                 , m_colorSubresourceRange       (makeImageSubresourceRange(VK_IMAGE_ASPECT_COLOR_BIT, 0u, 1u, 0u, 1u))
 542                 , m_clearColor                          (clearColor)
 543                 , m_topology                            (topology)
 544                 , m_descriptorSetLayout         (descriptorSetLayout)
 545         {
 546                 m_colorImage            = makeImage             (vk, device, makeImageCreateInfo(m_colorFormat, m_renderSize, VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT));
 547                 m_colorImageAlloc       = bindImage             (vk, device, allocator, *m_colorImage, MemoryRequirement::Any);
 548                 m_colorAttachment       = makeImageView (vk, device, *m_colorImage, VK_IMAGE_VIEW_TYPE_2D, m_colorFormat, m_colorSubresourceRange);
 549
 550                 m_vertexModule          = createShaderModule    (vk, device, binaryCollection.get(vertexName), 0u);
 551                 m_fragmentModule        = createShaderModule    (vk, device, binaryCollection.get(fragmentName), 0u);
 552
 553                 const VkPipelineShaderStageCreateInfo pShaderStages[] =
 554                 {
 555                         {
 556                                 VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,            // VkStructureType                                              sType;
 557                                 DE_NULL,                                                                                                        // const void*                                                  pNext;
 558                                 (VkPipelineShaderStageCreateFlags)0,                                            // VkPipelineShaderStageCreateFlags             flags;
 559                                 VK_SHADER_STAGE_VERTEX_BIT,                                                                     // VkShaderStageFlagBits                                stage;
 560                                 *m_vertexModule,                                                                                        // VkShaderModule                                               module;
 561                                 "main",                                                                                                         // const char*                                                  pName;
 562                                 specMap[VK_SHADER_STAGE_VERTEX_BIT],                                            // const VkSpecializationInfo*                  pSpecializationInfo;
 563                         },
 564                         {
 565                                 VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,            // VkStructureType                                              sType;
 566                                 DE_NULL,                                                                                                        // const void*                                                  pNext;
 567                                 (VkPipelineShaderStageCreateFlags)0,                                            // VkPipelineShaderStageCreateFlags             flags;
 568                                 VK_SHADER_STAGE_FRAGMENT_BIT,                                                           // VkShaderStageFlagBits                                stage;
 569                                 *m_fragmentModule,                                                                                      // VkShaderModule                                               module;
 570                                 "main",                                                                                                         // const char*                                                  pName;
 571                                 specMap[VK_SHADER_STAGE_FRAGMENT_BIT],                                          // const VkSpecializationInfo*                  pSpecializationInfo;
 572                         }
 573                 };
 574
 575                 m_renderPass            = makeRenderPass                (vk, device, m_colorFormat);
 576                 m_framebuffer           = makeFramebuffer               (vk, device, *m_renderPass, 1u, &m_colorAttachment.get(),
 577                                                                                                          static_cast<deUint32>(m_renderSize.x()), static_cast<deUint32>(m_renderSize.y()));
 578                 m_pipelineLayout        = makePipelineLayout    (vk, device, m_descriptorSetLayout);
 579                 m_pipeline                      = makeGraphicsPipeline  (vk, device, *m_pipelineLayout, *m_renderPass, m_renderSize, m_topology, DE_LENGTH_OF_ARRAY(pShaderStages), pShaderStages);
 580                 m_cmdPool                       = makeCommandPool               (vk, device, queueFamilyIndex);
 581                 m_cmdBuffer                     = allocateCommandBuffer (vk, device, *m_cmdPool, VK_COMMAND_BUFFER_LEVEL_PRIMARY);
 582         }
 583
 584         void draw (const DeviceInterface&       vk,
 585                            const VkDevice                       device,
 586                            const VkQueue                        queue,
 587                            const Delegate&                      drawDelegate) const
 588         {
 589                 beginCommandBuffer(vk, *m_cmdBuffer);
 590
 591                 const VkClearValue                      clearValue      = makeClearValueColor(m_clearColor);
 592                 const VkRect2D                          renderArea      =
 593                 {
 594                         makeOffset2D(0, 0),
 595                         makeExtent2D(m_renderSize.x(), m_renderSize.y()),
 596                 };
 597                 const VkRenderPassBeginInfo renderPassBeginInfo =
 598                 {
 599                         VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,               // VkStructureType         sType;
 600                         DE_NULL,                                                                                // const void*             pNext;
 601                         *m_renderPass,                                                                  // VkRenderPass            renderPass;
 602                         *m_framebuffer,                                                                 // VkFramebuffer           framebuffer;
 603                         renderArea,                                                                             // VkRect2D                renderArea;
 604                         1u,                                                                                             // uint32_t                clearValueCount;
 605                         &clearValue,                                                                    // const VkClearValue*     pClearValues;
 606                 };
 607                 vk.cmdBeginRenderPass(*m_cmdBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);
 608
 609                 vk.cmdBindPipeline(*m_cmdBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, *m_pipeline);
 610                 drawDelegate.rendererDraw(*m_pipelineLayout, *m_cmdBuffer);
 611
 612                 vk.cmdEndRenderPass(*m_cmdBuffer);
 613
 614                 // Prepare color image for copy
 615                 {
 616                         const VkImageMemoryBarrier barriers[] =
 617                         {
 618                                 {
 619                                         VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,                                         // VkStructureType                      sType;
 620                                         DE_NULL,                                                                                                        // const void*                          pNext;
 621                                         VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,                                           // VkAccessFlags                        outputMask;
 622                                         VK_ACCESS_TRANSFER_READ_BIT,                                                            // VkAccessFlags                        inputMask;
 623                                         VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,                                       // VkImageLayout                        oldLayout;
 624                                         VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,                                           // VkImageLayout                        newLayout;
 625                                         VK_QUEUE_FAMILY_IGNORED,                                                                        // deUint32                                     srcQueueFamilyIndex;
 626                                         VK_QUEUE_FAMILY_IGNORED,                                                                        // deUint32                                     destQueueFamilyIndex;
 627                                         *m_colorImage,                                                                                          // VkImage                                      image;
 628                                         m_colorSubresourceRange,                                                                        // VkImageSubresourceRange      subresourceRange;
 629                                 },
 630                         };
 631
 632                         vk.cmdPipelineBarrier(*m_cmdBuffer, VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0u,
 633                                 0u, DE_NULL, 0u, DE_NULL, DE_LENGTH_OF_ARRAY(barriers), barriers);
 634                 }
 635                 // Color image -> host buffer
 636                 {
 637                         const VkBufferImageCopy region =
 638                         {
 639                                 0ull,                                                                                                                                           // VkDeviceSize                bufferOffset;
 640                                 0u,                                                                                                                                                     // uint32_t                    bufferRowLength;
 641                                 0u,                                                                                                                                                     // uint32_t                    bufferImageHeight;
 642                                 makeImageSubresourceLayers(VK_IMAGE_ASPECT_COLOR_BIT, 0u, 0u, 1u),                      // VkImageSubresourceLayers    imageSubresource;
 643                                 makeOffset3D(0, 0, 0),                                                                                                          // VkOffset3D                  imageOffset;
 644                                 makeExtent3D(m_renderSize.x(), m_renderSize.y(), 1u),                                           // VkExtent3D                  imageExtent;
 645                         };
 646
 647                         vk.cmdCopyImageToBuffer(*m_cmdBuffer, *m_colorImage, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, m_colorBuffer, 1u, &region);
 648                 }
 649                 // Buffer write barrier
 650                 {
 651                         const VkBufferMemoryBarrier barriers[] =
 652                         {
 653                                 {
 654                                         VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,                // VkStructureType    sType;
 655                                         DE_NULL,                                                                                // const void*        pNext;
 656                                         VK_ACCESS_TRANSFER_WRITE_BIT,                                   // VkAccessFlags      srcAccessMask;
 657                                         VK_ACCESS_HOST_READ_BIT,                                                // VkAccessFlags      dstAccessMask;
 658                                         VK_QUEUE_FAMILY_IGNORED,                                                // uint32_t           srcQueueFamilyIndex;
 659                                         VK_QUEUE_FAMILY_IGNORED,                                                // uint32_t           dstQueueFamilyIndex;
 660                                         m_colorBuffer,                                                                  // VkBuffer           buffer;
 661                                         0ull,                                                                                   // VkDeviceSize       offset;
 662                                         VK_WHOLE_SIZE,                                                                  // VkDeviceSize       size;
 663                                 },
 664                         };
 665
 666                         vk.cmdPipelineBarrier(*m_cmdBuffer, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_HOST_BIT, 0u,
 667                                 0u, DE_NULL, DE_LENGTH_OF_ARRAY(barriers), barriers, DE_NULL, 0u);
 668                 }
 669
 670                 VK_CHECK(vk.endCommandBuffer(*m_cmdBuffer));
 671                 submitCommandsAndWait(vk, device, queue, *m_cmdBuffer);
 672         }
 673
 674 private:
 675         const VkBuffer                                  m_colorBuffer;
 676         const IVec2                                             m_renderSize;
 677         const VkFormat                                  m_colorFormat;
 678         const VkImageSubresourceRange   m_colorSubresourceRange;
 679         const Vec4                                              m_clearColor;
 680         const VkPrimitiveTopology               m_topology;
 681         const VkDescriptorSetLayout             m_descriptorSetLayout;
 682
 683         Move<VkImage>                                   m_colorImage;
 684         MovePtr<Allocation>                             m_colorImageAlloc;
 685         Move<VkImageView>                               m_colorAttachment;
 686         Move<VkShaderModule>                    m_vertexModule;
 687         Move<VkShaderModule>                    m_fragmentModule;
 688         Move<VkRenderPass>                              m_renderPass;
 689         Move<VkFramebuffer>                             m_framebuffer;
 690         Move<VkPipelineLayout>                  m_pipelineLayout;
 691         Move<VkPipeline>                                m_pipeline;
 692         Move<VkCommandPool>                             m_cmdPool;
 693         Move<VkCommandBuffer>                   m_cmdBuffer;
 694
 695         // "deleted"
 696                                 Renderer        (const Renderer&);
 697         Renderer&       operator=       (const Renderer&);
 698 };
 699
 700 void bindSparseBuffer (const DeviceInterface& vk, const VkDevice device, const VkQueue sparseQueue, const VkBuffer buffer, const SparseAllocation& sparseAllocation)
 701 {
 702         const VkSparseBufferMemoryBindInfo sparseBufferMemoryBindInfo =
 703         {
 704                 buffer,                                                                                                         // VkBuffer                     buffer;
 705                 static_cast<deUint32>(sparseAllocation.memoryBinds.size()),     // uint32_t                     bindCount;
 706                 &sparseAllocation.memoryBinds[0],                                                       // const VkSparseMemoryBind*    pBinds;
 707         };
 708
 709         const VkBindSparseInfo bindInfo =
 710         {
 711                 VK_STRUCTURE_TYPE_BIND_SPARSE_INFO,                                     // VkStructureType                             sType;
 712                 DE_NULL,                                                                                        // const void*                                 pNext;
 713                 0u,                                                                                                     // uint32_t                                    waitSemaphoreCount;
 714                 DE_NULL,                                                                                        // const VkSemaphore*                          pWaitSemaphores;
 715                 1u,                                                                                                     // uint32_t                                    bufferBindCount;
 716                 &sparseBufferMemoryBindInfo,                                            // const VkSparseBufferMemoryBindInfo*         pBufferBinds;
 717                 0u,                                                                                                     // uint32_t                                    imageOpaqueBindCount;
 718                 DE_NULL,                                                                                        // const VkSparseImageOpaqueMemoryBindInfo*    pImageOpaqueBinds;
 719                 0u,                                                                                                     // uint32_t                                    imageBindCount;
 720                 DE_NULL,                                                                                        // const VkSparseImageMemoryBindInfo*          pImageBinds;
 721                 0u,                                                                                                     // uint32_t                                    signalSemaphoreCount;
 722                 DE_NULL,                                                                                        // const VkSemaphore*                          pSignalSemaphores;
 723         };
 724
 725         const Unique<VkFence> fence(createFence(vk, device));
 726
 727         VK_CHECK(vk.queueBindSparse(sparseQueue, 1u, &bindInfo, *fence));
 728         VK_CHECK(vk.waitForFences(device, 1u, &fence.get(), VK_TRUE, ~0ull));
 729 }
 730
 731 class SparseBufferTestInstance : public SparseResourcesBaseInstance, Renderer::Delegate
 732 {
 733 public:
 734         SparseBufferTestInstance (Context& context, const TestFlags flags)
 735                 : SparseResourcesBaseInstance   (context)
 736                 , m_aliased                                             ((flags & TEST_FLAG_ALIASED)   != 0)
 737                 , m_residency                                   ((flags & TEST_FLAG_RESIDENCY) != 0)
 738                 , m_nonResidentStrict                   ((flags & TEST_FLAG_NON_RESIDENT_STRICT) != 0)
 739                 , m_renderSize                                  (RENDER_SIZE, RENDER_SIZE)
 740                 , m_colorFormat                                 (VK_FORMAT_R8G8B8A8_UNORM)
 741                 , m_colorBufferSize                             (m_renderSize.x() * m_renderSize.y() * tcu::getPixelSize(mapVkFormat(m_colorFormat)))
 742         {
 743                 const VkPhysicalDeviceFeatures  features        = getPhysicalDeviceFeatures(m_context.getInstanceInterface(), m_context.getPhysicalDevice());
 744
 745                 if (!features.sparseBinding)
 746                         TCU_THROW(NotSupportedError, "Missing feature: sparseBinding");
 747
 748                 if (m_residency && !features.sparseResidencyBuffer)
 749                         TCU_THROW(NotSupportedError, "Missing feature: sparseResidencyBuffer");
 750
 751                 if (m_aliased && !features.sparseResidencyAliased)
 752                         TCU_THROW(NotSupportedError, "Missing feature: sparseResidencyAliased");
 753
 754                 if (m_nonResidentStrict && !m_context.getDeviceProperties().sparseProperties.residencyNonResidentStrict)
 755                         TCU_THROW(NotSupportedError, "Missing sparse property: residencyNonResidentStrict");
 756
 757                 {
 758                         QueueRequirementsVec requirements;
 759                         requirements.push_back(QueueRequirements(VK_QUEUE_SPARSE_BINDING_BIT, 1u));
 760                         requirements.push_back(QueueRequirements(VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT, 1u));
 761
 762                         createDeviceSupportingQueues(requirements);
 763                 }
 764
 765                 const DeviceInterface& vk               = getDeviceInterface();
 766                 m_sparseQueue                                   = getQueue(VK_QUEUE_SPARSE_BINDING_BIT, 0u);
 767                 m_universalQueue                                = getQueue(VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT, 0u);
 768
 769                 m_sharedQueueFamilyIndices[0]   = m_sparseQueue.queueFamilyIndex;
 770                 m_sharedQueueFamilyIndices[1]   = m_universalQueue.queueFamilyIndex;
 771
 772                 m_colorBuffer                                   = makeBuffer(vk, getDevice(), makeBufferCreateInfo(m_colorBufferSize, VK_BUFFER_USAGE_TRANSFER_DST_BIT));
 773                 m_colorBufferAlloc                              = bindBuffer(vk, getDevice(), getAllocator(), *m_colorBuffer, MemoryRequirement::HostVisible);
 774
 775                 deMemset(m_colorBufferAlloc->getHostPtr(), 0, static_cast<std::size_t>(m_colorBufferSize));
 776                 flushMappedMemoryRange(vk, getDevice(), m_colorBufferAlloc->getMemory(), m_colorBufferAlloc->getOffset(), m_colorBufferSize);
 777         }
 778
 779 protected:
 780         VkBufferCreateInfo getSparseBufferCreateInfo (const VkBufferUsageFlags usage) const
 781         {
 782                 VkBufferCreateFlags     flags = VK_BUFFER_CREATE_SPARSE_BINDING_BIT;
 783                 if (m_residency)
 784                         flags |= VK_BUFFER_CREATE_SPARSE_RESIDENCY_BIT;
 785                 if (m_aliased)
 786                         flags |= VK_BUFFER_CREATE_SPARSE_ALIASED_BIT;
 787
 788                 VkBufferCreateInfo referenceBufferCreateInfo =
 789                 {
 790                         VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,                           // VkStructureType        sType;
 791                         DE_NULL,                                                                                        // const void*            pNext;
 792                         flags,                                                                                          // VkBufferCreateFlags    flags;
 793                         0u,     // override later                                                               // VkDeviceSize           size;
 794                         VK_BUFFER_USAGE_TRANSFER_DST_BIT | usage,                       // VkBufferUsageFlags     usage;
 795                         VK_SHARING_MODE_EXCLUSIVE,                                                      // VkSharingMode          sharingMode;
 796                         0u,                                                                                                     // uint32_t               queueFamilyIndexCount;
 797                         DE_NULL,                                                                                        // const uint32_t*        pQueueFamilyIndices;
 798                 };
 799
 800                 if (m_sparseQueue.queueFamilyIndex != m_universalQueue.queueFamilyIndex)
 801                 {
 802                         referenceBufferCreateInfo.sharingMode                   = VK_SHARING_MODE_CONCURRENT;
 803                         referenceBufferCreateInfo.queueFamilyIndexCount = DE_LENGTH_OF_ARRAY(m_sharedQueueFamilyIndices);
 804                         referenceBufferCreateInfo.pQueueFamilyIndices   = m_sharedQueueFamilyIndices;
 805                 }
 806
 807                 return referenceBufferCreateInfo;
 808         }
 809
 810         void draw (const VkPrimitiveTopology    topology,
 811                            const VkDescriptorSetLayout  descriptorSetLayout     = DE_NULL,
 812                            Renderer::SpecializationMap  specMap                         = Renderer::SpecializationMap())
 813         {
 814                 const UniquePtr<Renderer> renderer(new Renderer(
 815                         getDeviceInterface(), getDevice(), getAllocator(), m_universalQueue.queueFamilyIndex, descriptorSetLayout,
 816                         m_context.getBinaryCollection(), "vert", "frag", *m_colorBuffer, m_renderSize, m_colorFormat, Vec4(1.0f, 0.0f, 0.0f, 1.0f), topology, specMap));
 817
 818                 renderer->draw(getDeviceInterface(), getDevice(), m_universalQueue.queueHandle, *this);
 819         }
 820
 821         tcu::TestStatus verifyDrawResult (void) const
 822         {
 823                 invalidateMappedMemoryRange(getDeviceInterface(), getDevice(), m_colorBufferAlloc->getMemory(), 0ull, m_colorBufferSize);
 824
 825                 const tcu::ConstPixelBufferAccess resultImage (mapVkFormat(m_colorFormat), m_renderSize.x(), m_renderSize.y(), 1u, m_colorBufferAlloc->getHostPtr());
 826
 827                 m_context.getTestContext().getLog()
 828                         << tcu::LogImageSet("Result", "Result") << tcu::LogImage("color0", "", resultImage) << tcu::TestLog::EndImageSet;
 829
 830                 if (imageHasErrorPixels(resultImage))
 831                         return tcu::TestStatus::fail("Some buffer values were incorrect");
 832                 else
 833                         return tcu::TestStatus::pass("Pass");
 834         }
 835
 836         const bool                                                      m_aliased;
 837         const bool                                                      m_residency;
 838         const bool                                                      m_nonResidentStrict;
 839
 840         Queue                                                           m_sparseQueue;
 841         Queue                                                           m_universalQueue;
 842
 843 private:
 844         const IVec2                                                     m_renderSize;
 845         const VkFormat                                          m_colorFormat;
 846         const VkDeviceSize                                      m_colorBufferSize;
 847
 848         Move<VkBuffer>                                          m_colorBuffer;
 849         MovePtr<Allocation>                                     m_colorBufferAlloc;
 850
 851         deUint32                                                        m_sharedQueueFamilyIndices[2];
 852 };
 853
 854 void initProgramsDrawWithUBO (vk::SourceCollections& programCollection, const TestFlags flags)
 855 {
 856         // Vertex shader
 857         {
 858                 std::ostringstream src;
 859                 src << glu::getGLSLVersionDeclaration(glu::GLSL_VERSION_450) << "\n"
 860                         << "\n"
 861                         << "layout(location = 0) in vec4 in_position;\n"
 862                         << "\n"
 863                         << "out gl_PerVertex {\n"
 864                         << "    vec4 gl_Position;\n"
 865                         << "};\n"
 866                         << "\n"
 867                         << "void main(void)\n"
 868                         << "{\n"
 869                         << "    gl_Position = in_position;\n"
 870                         << "}\n";
 871
 872                 programCollection.glslSources.add("vert") << glu::VertexSource(src.str());
 873         }
 874
 875         // Fragment shader
 876         {
 877                 const bool                      aliased                         = (flags & TEST_FLAG_ALIASED) != 0;
 878                 const bool                      residency                       = (flags & TEST_FLAG_RESIDENCY) != 0;
 879                 const bool                      nonResidentStrict       = (flags & TEST_FLAG_NON_RESIDENT_STRICT) != 0;
 880                 const std::string       valueExpr                       = (aliased ? "ivec4(3*(ndx % nonAliasedSize) ^ 127, 0, 0, 0)" : "ivec4(3*ndx ^ 127, 0, 0, 0)");
 881
 882                 std::ostringstream src;
 883                 src << glu::getGLSLVersionDeclaration(glu::GLSL_VERSION_450) << "\n"
 884                         << "\n"
 885                         << "layout(location = 0) out vec4 o_color;\n"
 886                         << "\n"
 887                         << "layout(constant_id = 1) const int dataSize  = 1;\n"
 888                         << "layout(constant_id = 2) const int chunkSize = 1;\n"
 889                         << "\n"
 890                         << "layout(set = 0, binding = 0, std140) uniform SparseBuffer {\n"
 891                         << "    ivec4 data[dataSize];\n"
 892                         << "} ubo;\n"
 893                         << "\n"
 894                         << "void main(void)\n"
 895                         << "{\n"
 896                         << "    const int fragNdx        = int(gl_FragCoord.x) + " << RENDER_SIZE << " * int(gl_FragCoord.y);\n"
 897                         << "    const int pageSize       = " << RENDER_SIZE << " * " << RENDER_SIZE << ";\n"
 898                         << "    const int numChunks      = dataSize / chunkSize;\n";
 899
 900                 if (aliased)
 901                         src << "    const int nonAliasedSize = (numChunks > 1 ? dataSize - chunkSize : dataSize);\n";
 902
 903                 src << "    bool      ok             = true;\n"
 904                         << "\n"
 905                         << "    for (int ndx = fragNdx; ndx < dataSize; ndx += pageSize)\n"
 906                         << "    {\n";
 907
 908                 if (residency && nonResidentStrict)
 909                 {
 910                         src << "        if (ndx >= chunkSize && ndx < 2*chunkSize)\n"
 911                                 << "            ok = ok && (ubo.data[ndx] == ivec4(0));\n"
 912                                 << "        else\n"
 913                                 << "            ok = ok && (ubo.data[ndx] == " + valueExpr + ");\n";
 914                 }
 915                 else if (residency)
 916                 {
 917                         src << "        if (ndx >= chunkSize && ndx < 2*chunkSize)\n"
 918                                 << "            continue;\n"
 919                                 << "        ok = ok && (ubo.data[ndx] == " << valueExpr << ");\n";
 920                 }
 921                 else
 922                         src << "        ok = ok && (ubo.data[ndx] == " << valueExpr << ");\n";
 923
 924                 src << "    }\n"
 925                         << "\n"
 926                         << "    if (ok)\n"
 927                         << "        o_color = vec4(0.0, 1.0, 0.0, 1.0);\n"
 928                         << "    else\n"
 929                         << "        o_color = vec4(1.0, 0.0, 0.0, 1.0);\n"
 930                         << "}\n";
 931
 932                 programCollection.glslSources.add("frag") << glu::FragmentSource(src.str());
 933         }
 934 }
 935
 936 //! Sparse buffer backing a UBO
 937 class UBOTestInstance : public SparseBufferTestInstance
 938 {
 939 public:
 940         UBOTestInstance (Context& context, const TestFlags flags)
 941                 : SparseBufferTestInstance      (context, flags)
 942         {
 943         }
 944
 945         void rendererDraw (const VkPipelineLayout pipelineLayout, const VkCommandBuffer cmdBuffer) const
 946         {
 947                 const DeviceInterface&  vk                              = getDeviceInterface();
 948                 const VkDeviceSize              vertexOffset    = 0ull;
 949
 950                 vk.cmdBindVertexBuffers (cmdBuffer, 0u, 1u, &m_vertexBuffer.get(), &vertexOffset);
 951                 vk.cmdBindDescriptorSets(cmdBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0u, 1u, &m_descriptorSet.get(), 0u, DE_NULL);
 952                 vk.cmdDraw                              (cmdBuffer, 4u, 1u, 0u, 0u);
 953         }
 954
 955         tcu::TestStatus iterate (void)
 956         {
 957                 const DeviceInterface&          vk                                      = getDeviceInterface();
 958                 MovePtr<SparseAllocation>       sparseAllocation;
 959                 Move<VkBuffer>                          sparseBuffer;
 960                 Move<VkBuffer>                          sparseBufferAliased;
 961
 962                 // Set up the sparse buffer
 963                 {
 964                         VkBufferCreateInfo      referenceBufferCreateInfo       = getSparseBufferCreateInfo(VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT);
 965                         const VkDeviceSize      minChunkSize                            = 512u; // make sure the smallest allocation is at least this big
 966                         deUint32                        numMaxChunks                            = 0u;
 967
 968                         // Check how many chunks we can allocate given the alignment and size requirements of UBOs
 969                         {
 970                                 const UniquePtr<SparseAllocation> minAllocation(SparseAllocationBuilder()
 971                                         .addMemoryBind()
 972                                         .build(vk, getDevice(), getAllocator(), referenceBufferCreateInfo, minChunkSize));
 973
 974                                 numMaxChunks = deMaxu32(static_cast<deUint32>(m_context.getDeviceProperties().limits.maxUniformBufferRange / minAllocation->resourceSize), 1u);
 975                         }
 976
 977                         if (numMaxChunks < 4)
 978                         {
 979                                 sparseAllocation = SparseAllocationBuilder()
 980                                         .addMemoryBind()
 981                                         .build(vk, getDevice(), getAllocator(), referenceBufferCreateInfo, minChunkSize);
 982                         }
 983                         else
 984                         {
 985                                 // Try to use a non-trivial memory allocation scheme to make it different from a non-sparse binding
 986                                 SparseAllocationBuilder builder;
 987                                 builder.addMemoryBind();
 988
 989                                 if (m_residency)
 990                                         builder.addResourceHole();
 991
 992                                 builder
 993                                         .addMemoryAllocation()
 994                                         .addMemoryHole()
 995                                         .addMemoryBind();
 996
 997                                 if (m_aliased)
 998                                         builder.addAliasedMemoryBind(0u, 0u);
 999
1000                                 sparseAllocation = builder.build(vk, getDevice(), getAllocator(), referenceBufferCreateInfo, minChunkSize);
1001                                 DE_ASSERT(sparseAllocation->resourceSize <= m_context.getDeviceProperties().limits.maxUniformBufferRange);
1002                         }
1003
1004                         // Create the buffer
1005                         referenceBufferCreateInfo.size  = sparseAllocation->resourceSize;
1006                         sparseBuffer                                    = makeBuffer(vk, getDevice(), referenceBufferCreateInfo);
1007                         bindSparseBuffer(vk, getDevice(), m_sparseQueue.queueHandle, *sparseBuffer, *sparseAllocation);
1008
1009                         if (m_aliased)
1010                         {
1011                                 sparseBufferAliased = makeBuffer(vk, getDevice(), referenceBufferCreateInfo);
1012                                 bindSparseBuffer(vk, getDevice(), m_sparseQueue.queueHandle, *sparseBufferAliased, *sparseAllocation);
1013                         }
1014                 }
1015
1016                 // Set uniform data
1017                 {
1018                         const bool                                      hasAliasedChunk         = (m_aliased && sparseAllocation->memoryBinds.size() > 1u);
1019                         const VkDeviceSize                      chunkSize                       = sparseAllocation->resourceSize / sparseAllocation->numResourceChunks;
1020                         const VkDeviceSize                      stagingBufferSize       = sparseAllocation->resourceSize - (hasAliasedChunk ? chunkSize : 0);
1021                         const deUint32                          numBufferEntries        = static_cast<deUint32>(stagingBufferSize / sizeof(IVec4));
1022
1023                         const Unique<VkBuffer>          stagingBuffer           (makeBuffer(vk, getDevice(), makeBufferCreateInfo(stagingBufferSize, VK_BUFFER_USAGE_TRANSFER_SRC_BIT)));
1024                         const UniquePtr<Allocation>     stagingBufferAlloc      (bindBuffer(vk, getDevice(), getAllocator(), *stagingBuffer, MemoryRequirement::HostVisible));
1025
1026                         {
1027                                 // If aliased chunk is used, the staging buffer is smaller than the sparse buffer and we don't overwrite the last chunk
1028                                 IVec4* const pData = static_cast<IVec4*>(stagingBufferAlloc->getHostPtr());
1029                                 for (deUint32 i = 0; i < numBufferEntries; ++i)
1030                                         pData[i] = IVec4(3*i ^ 127, 0, 0, 0);
1031
1032                                 flushMappedMemoryRange(vk, getDevice(), stagingBufferAlloc->getMemory(), stagingBufferAlloc->getOffset(), stagingBufferSize);
1033
1034                                 const VkBufferCopy copyRegion =
1035                                 {
1036                                         0ull,                                           // VkDeviceSize    srcOffset;
1037                                         0ull,                                           // VkDeviceSize    dstOffset;
1038                                         stagingBufferSize,                      // VkDeviceSize    size;
1039                                 };
1040
1041                                 const Unique<VkCommandPool>             cmdPool         (makeCommandPool(vk, getDevice(), m_universalQueue.queueFamilyIndex));
1042                                 const Unique<VkCommandBuffer>   cmdBuffer       (allocateCommandBuffer(vk, getDevice(), *cmdPool, VK_COMMAND_BUFFER_LEVEL_PRIMARY));
1043
1044                                 beginCommandBuffer      (vk, *cmdBuffer);
1045                                 vk.cmdCopyBuffer        (*cmdBuffer, *stagingBuffer, *sparseBuffer, 1u, &copyRegion);
1046                                 endCommandBuffer        (vk, *cmdBuffer);
1047
1048                                 submitCommandsAndWait(vk, getDevice(), m_universalQueue.queueHandle, *cmdBuffer);
1049                                 // Once the fence is signaled, the write is also available to the aliasing buffer.
1050                         }
1051                 }
1052
1053                 // Make sure that we don't try to access a larger range than is allowed. This only applies to a single chunk case.
1054                 const deUint32 maxBufferRange = deMinu32(static_cast<deUint32>(sparseAllocation->resourceSize), m_context.getDeviceProperties().limits.maxUniformBufferRange);
1055
1056                 // Descriptor sets
1057                 {
1058                         m_descriptorSetLayout = DescriptorSetLayoutBuilder()
1059                                 .addSingleBinding(VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER, VK_SHADER_STAGE_FRAGMENT_BIT)
1060                                 .build(vk, getDevice());
1061
1062                         m_descriptorPool = DescriptorPoolBuilder()
1063                                 .addType(VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER)
1064                                 .build(vk, getDevice(), VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT, 1u);
1065
1066                         m_descriptorSet = makeDescriptorSet(vk, getDevice(), *m_descriptorPool, *m_descriptorSetLayout);
1067
1068                         const VkBuffer                                  buffer                          = (m_aliased ? *sparseBufferAliased : *sparseBuffer);
1069                         const VkDescriptorBufferInfo    sparseBufferInfo        = makeDescriptorBufferInfo(buffer, 0ull, maxBufferRange);
1070
1071                         DescriptorSetUpdateBuilder()
1072                                 .writeSingle(*m_descriptorSet, DescriptorSetUpdateBuilder::Location::binding(0u), VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER, &sparseBufferInfo)
1073                                 .update(vk, getDevice());
1074                 }
1075
1076                 // Vertex data
1077                 {
1078                         const Vec4 vertexData[] =
1079                         {
1080                                 Vec4(-1.0f, -1.0f, 0.0f, 1.0f),
1081                                 Vec4(-1.0f,  1.0f, 0.0f, 1.0f),
1082                                 Vec4( 1.0f, -1.0f, 0.0f, 1.0f),
1083                                 Vec4( 1.0f,  1.0f, 0.0f, 1.0f),
1084                         };
1085
1086                         const VkDeviceSize      vertexBufferSize        = sizeof(vertexData);
1087
1088                         m_vertexBuffer          = makeBuffer(vk, getDevice(), makeBufferCreateInfo(vertexBufferSize, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT));
1089                         m_vertexBufferAlloc     = bindBuffer(vk, getDevice(), getAllocator(), *m_vertexBuffer, MemoryRequirement::HostVisible);
1090
1091                         deMemcpy(m_vertexBufferAlloc->getHostPtr(), &vertexData[0], vertexBufferSize);
1092                         flushMappedMemoryRange(vk, getDevice(), m_vertexBufferAlloc->getMemory(), m_vertexBufferAlloc->getOffset(), vertexBufferSize);
1093                 }
1094
1095                 // Draw
1096                 {
1097                         std::vector<deInt32> specializationData;
1098                         {
1099                                 const deUint32  numBufferEntries        = maxBufferRange / static_cast<deUint32>(sizeof(IVec4));
1100                                 const deUint32  numEntriesPerChunk      = numBufferEntries / sparseAllocation->numResourceChunks;
1101
1102                                 specializationData.push_back(numBufferEntries);
1103                                 specializationData.push_back(numEntriesPerChunk);
1104                         }
1105
1106                         const VkSpecializationMapEntry  specMapEntries[] =
1107                         {
1108                                 {
1109                                         1u,                                     // uint32_t    constantID;
1110                                         0u,                                     // uint32_t    offset;
1111                                         sizeof(deInt32),        // size_t      size;
1112                                 },
1113                                 {
1114                                         2u,                                     // uint32_t    constantID;
1115                                         sizeof(deInt32),        // uint32_t    offset;
1116                                         sizeof(deInt32),        // size_t      size;
1117                                 },
1118                         };
1119
1120                         const VkSpecializationInfo specInfo =
1121                         {
1122                                 DE_LENGTH_OF_ARRAY(specMapEntries),             // uint32_t                           mapEntryCount;
1123                                 specMapEntries,                                                 // const VkSpecializationMapEntry*    pMapEntries;
1124                                 sizeInBytes(specializationData),                // size_t                             dataSize;
1125                                 getDataOrNullptr(specializationData),   // const void*                        pData;
1126                         };
1127
1128                         Renderer::SpecializationMap     specMap;
1129                         specMap[VK_SHADER_STAGE_FRAGMENT_BIT] = &specInfo;
1130
1131                         draw(VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP, *m_descriptorSetLayout, specMap);
1132                 }
1133
1134                 return verifyDrawResult();
1135         }
1136
1137 private:
1138         Move<VkBuffer>                                  m_vertexBuffer;
1139         MovePtr<Allocation>                             m_vertexBufferAlloc;
1140
1141         Move<VkDescriptorSetLayout>             m_descriptorSetLayout;
1142         Move<VkDescriptorPool>                  m_descriptorPool;
1143         Move<VkDescriptorSet>                   m_descriptorSet;
1144 };
1145
1146 void initProgramsDrawGrid (vk::SourceCollections& programCollection, const TestFlags flags)
1147 {
1148         DE_UNREF(flags);
1149
1150         // Vertex shader
1151         {
1152                 std::ostringstream src;
1153                 src << glu::getGLSLVersionDeclaration(glu::GLSL_VERSION_450) << "\n"
1154                         << "\n"
1155                         << "layout(location = 0) in  vec4 in_position;\n"
1156                         << "layout(location = 0) out int  out_ndx;\n"
1157                         << "\n"
1158                         << "out gl_PerVertex {\n"
1159                         << "    vec4 gl_Position;\n"
1160                         << "};\n"
1161                         << "\n"
1162                         << "void main(void)\n"
1163                         << "{\n"
1164                         << "    gl_Position = in_position;\n"
1165                         << "    out_ndx     = gl_VertexIndex;\n"
1166                         << "}\n";
1167
1168                 programCollection.glslSources.add("vert") << glu::VertexSource(src.str());
1169         }
1170
1171         // Fragment shader
1172         {
1173                 std::ostringstream src;
1174                 src << glu::getGLSLVersionDeclaration(glu::GLSL_VERSION_450) << "\n"
1175                         << "\n"
1176                         << "layout(location = 0) flat in  int  in_ndx;\n"
1177                         << "layout(location = 0)      out vec4 o_color;\n"
1178                         << "\n"
1179                         << "void main(void)\n"
1180                         << "{\n"
1181                         << "    if (in_ndx % 2 == 0)\n"
1182                         << "        o_color = vec4(vec3(1.0), 1.0);\n"
1183                         << "    else\n"
1184                         << "        o_color = vec4(vec3(0.75), 1.0);\n"
1185                         << "}\n";
1186
1187                 programCollection.glslSources.add("frag") << glu::FragmentSource(src.str());
1188         }
1189 }
1190
1191 //! Generate vertex positions for a grid of tiles composed of two triangles each (6 vertices)
1192 void generateGrid (void* pRawData, const float step, const float ox, const float oy, const deUint32 numX, const deUint32 numY, const float z = 0.0f)
1193 {
1194         typedef Vec4 (*TilePtr)[6];
1195
1196         TilePtr const pData = static_cast<TilePtr>(pRawData);
1197         {
1198                 for (deUint32 iy = 0; iy < numY; ++iy)
1199                 for (deUint32 ix = 0; ix < numX; ++ix)
1200                 {
1201                         const deUint32  ndx     = ix + numX * iy;
1202                         const float             x       = ox + step * static_cast<float>(ix);
1203                         const float             y       = oy + step * static_cast<float>(iy);
1204
1205                         pData[ndx][0] = Vec4(x + step,  y,                      z, 1.0f);
1206                         pData[ndx][1] = Vec4(x,                 y,                      z, 1.0f);
1207                         pData[ndx][2] = Vec4(x,                 y + step,       z, 1.0f);
1208
1209                         pData[ndx][3] = Vec4(x,                 y + step,       z, 1.0f);
1210                         pData[ndx][4] = Vec4(x + step,  y + step,       z, 1.0f);
1211                         pData[ndx][5] = Vec4(x + step,  y,                      z, 1.0f);
1212                 }
1213         }
1214 }
1215
1216 //! Base test for a sparse buffer backing a vertex/index buffer
1217 class DrawGridTestInstance : public SparseBufferTestInstance
1218 {
1219 public:
1220         DrawGridTestInstance (Context& context, const TestFlags flags, const VkBufferUsageFlags usage, const VkDeviceSize minChunkSize)
1221                 : SparseBufferTestInstance      (context, flags)
1222         {
1223                 const DeviceInterface&  vk                                                      = getDeviceInterface();
1224                 VkBufferCreateInfo              referenceBufferCreateInfo       = getSparseBufferCreateInfo(usage);
1225
1226                 {
1227                         // Allocate two chunks, each covering half of the viewport
1228                         SparseAllocationBuilder builder;
1229                         builder.addMemoryBind();
1230
1231                         if (m_residency)
1232                                 builder.addResourceHole();
1233
1234                         builder
1235                                 .addMemoryAllocation()
1236                                 .addMemoryHole()
1237                                 .addMemoryBind();
1238
1239                         if (m_aliased)
1240                                 builder.addAliasedMemoryBind(0u, 0u);
1241
1242                         m_sparseAllocation      = builder.build(vk, getDevice(), getAllocator(), referenceBufferCreateInfo, minChunkSize);
1243                 }
1244
1245                 // Create the buffer
1246                 referenceBufferCreateInfo.size  = m_sparseAllocation->resourceSize;
1247                 m_sparseBuffer                                  = makeBuffer(vk, getDevice(), referenceBufferCreateInfo);
1248
1249                 // Bind the memory
1250                 bindSparseBuffer(vk, getDevice(), m_sparseQueue.queueHandle, *m_sparseBuffer, *m_sparseAllocation);
1251
1252                 m_perDrawBufferOffset   = m_sparseAllocation->resourceSize / m_sparseAllocation->numResourceChunks;
1253                 m_stagingBufferSize             = 2 * m_perDrawBufferOffset;
1254                 m_stagingBuffer                 = makeBuffer(vk, getDevice(), makeBufferCreateInfo(m_stagingBufferSize, VK_BUFFER_USAGE_TRANSFER_SRC_BIT));
1255                 m_stagingBufferAlloc    = bindBuffer(vk, getDevice(), getAllocator(), *m_stagingBuffer, MemoryRequirement::HostVisible);
1256         }
1257
1258         tcu::TestStatus iterate (void)
1259         {
1260                 initializeBuffers();
1261
1262                 const DeviceInterface&  vk      = getDeviceInterface();
1263
1264                 // Upload to the sparse buffer
1265                 {
1266                         flushMappedMemoryRange(vk, getDevice(), m_stagingBufferAlloc->getMemory(), m_stagingBufferAlloc->getOffset(), m_stagingBufferSize);
1267
1268                         VkDeviceSize    firstChunkOffset        = 0ull;
1269                         VkDeviceSize    secondChunkOffset       = m_perDrawBufferOffset;
1270
1271                         if (m_residency)
1272                                 secondChunkOffset += m_perDrawBufferOffset;
1273
1274                         if (m_aliased)
1275                                 firstChunkOffset = secondChunkOffset + m_perDrawBufferOffset;
1276
1277                         const VkBufferCopy copyRegions[] =
1278                         {
1279                                 {
1280                                         0ull,                                           // VkDeviceSize    srcOffset;
1281                                         firstChunkOffset,                       // VkDeviceSize    dstOffset;
1282                                         m_perDrawBufferOffset,          // VkDeviceSize    size;
1283                                 },
1284                                 {
1285                                         m_perDrawBufferOffset,          // VkDeviceSize    srcOffset;
1286                                         secondChunkOffset,                      // VkDeviceSize    dstOffset;
1287                                         m_perDrawBufferOffset,          // VkDeviceSize    size;
1288                                 },
1289                         };
1290
1291                         const Unique<VkCommandPool>             cmdPool         (makeCommandPool(vk, getDevice(), m_universalQueue.queueFamilyIndex));
1292                         const Unique<VkCommandBuffer>   cmdBuffer       (allocateCommandBuffer(vk, getDevice(), *cmdPool, VK_COMMAND_BUFFER_LEVEL_PRIMARY));
1293
1294                         beginCommandBuffer      (vk, *cmdBuffer);
1295                         vk.cmdCopyBuffer        (*cmdBuffer, *m_stagingBuffer, *m_sparseBuffer, DE_LENGTH_OF_ARRAY(copyRegions), copyRegions);
1296                         endCommandBuffer        (vk, *cmdBuffer);
1297
1298                         submitCommandsAndWait(vk, getDevice(), m_universalQueue.queueHandle, *cmdBuffer);
1299                 }
1300
1301                 draw(VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST);
1302
1303                 return verifyDrawResult();
1304         }
1305
1306 protected:
1307         virtual void                            initializeBuffers               (void) = 0;
1308
1309         VkDeviceSize                            m_perDrawBufferOffset;
1310
1311         VkDeviceSize                            m_stagingBufferSize;
1312         Move<VkBuffer>                          m_stagingBuffer;
1313         MovePtr<Allocation>                     m_stagingBufferAlloc;
1314
1315         MovePtr<SparseAllocation>       m_sparseAllocation;
1316         Move<VkBuffer>                          m_sparseBuffer;
1317 };
1318
1319 //! Sparse buffer backing a vertex input buffer
1320 class VertexBufferTestInstance : public DrawGridTestInstance
1321 {
1322 public:
1323         VertexBufferTestInstance (Context& context, const TestFlags flags)
1324                 : DrawGridTestInstance  (context,
1325                                                                  flags,
1326                                                                  VK_BUFFER_USAGE_VERTEX_BUFFER_BIT,
1327                                                                  GRID_SIZE * GRID_SIZE * 6 * sizeof(Vec4))
1328         {
1329         }
1330
1331         void rendererDraw (const VkPipelineLayout pipelineLayout, const VkCommandBuffer cmdBuffer) const
1332         {
1333                 DE_UNREF(pipelineLayout);
1334
1335                 m_context.getTestContext().getLog()
1336                         << tcu::TestLog::Message << "Drawing a grid of triangles backed by a sparse vertex buffer. There should be no red pixels visible." << tcu::TestLog::EndMessage;
1337
1338                 const DeviceInterface&  vk                              = getDeviceInterface();
1339                 const deUint32                  vertexCount             = 6 * (GRID_SIZE * GRID_SIZE) / 2;
1340                 VkDeviceSize                    vertexOffset    = 0ull;
1341
1342                 vk.cmdBindVertexBuffers (cmdBuffer, 0u, 1u, &m_sparseBuffer.get(), &vertexOffset);
1343                 vk.cmdDraw                              (cmdBuffer, vertexCount, 1u, 0u, 0u);
1344
1345                 vertexOffset += m_perDrawBufferOffset * (m_residency ? 2 : 1);
1346
1347                 vk.cmdBindVertexBuffers (cmdBuffer, 0u, 1u, &m_sparseBuffer.get(), &vertexOffset);
1348                 vk.cmdDraw                              (cmdBuffer, vertexCount, 1u, 0u, 0u);
1349         }
1350
1351         void initializeBuffers (void)
1352         {
1353                 deUint8*        pData   = static_cast<deUint8*>(m_stagingBufferAlloc->getHostPtr());
1354                 const float     step    = 2.0f / static_cast<float>(GRID_SIZE);
1355
1356                 // Prepare data for two draw calls
1357                 generateGrid(pData,                                                     step, -1.0f, -1.0f, GRID_SIZE, GRID_SIZE/2);
1358                 generateGrid(pData + m_perDrawBufferOffset,     step, -1.0f,  0.0f, GRID_SIZE, GRID_SIZE/2);
1359         }
1360 };
1361
1362 //! Sparse buffer backing an index buffer
1363 class IndexBufferTestInstance : public DrawGridTestInstance
1364 {
1365 public:
1366         IndexBufferTestInstance (Context& context, const TestFlags flags)
1367                 : DrawGridTestInstance  (context,
1368                                                                  flags,
1369                                                                  VK_BUFFER_USAGE_INDEX_BUFFER_BIT,
1370                                                                  GRID_SIZE * GRID_SIZE * 6 * sizeof(deUint32))
1371                 , m_halfVertexCount             (6 * (GRID_SIZE * GRID_SIZE) / 2)
1372         {
1373         }
1374
1375         void rendererDraw (const VkPipelineLayout pipelineLayout, const VkCommandBuffer cmdBuffer) const
1376         {
1377                 DE_UNREF(pipelineLayout);
1378
1379                 m_context.getTestContext().getLog()
1380                         << tcu::TestLog::Message << "Drawing a grid of triangles from a sparse index buffer. There should be no red pixels visible." << tcu::TestLog::EndMessage;
1381
1382                 const DeviceInterface&  vk                              = getDeviceInterface();
1383                 const VkDeviceSize              vertexOffset    = 0ull;
1384                 VkDeviceSize                    indexOffset             = 0ull;
1385
1386                 vk.cmdBindVertexBuffers (cmdBuffer, 0u, 1u, &m_vertexBuffer.get(), &vertexOffset);
1387
1388                 vk.cmdBindIndexBuffer   (cmdBuffer, *m_sparseBuffer, indexOffset, VK_INDEX_TYPE_UINT32);
1389                 vk.cmdDrawIndexed               (cmdBuffer, m_halfVertexCount, 1u, 0u, 0, 0u);
1390
1391                 indexOffset += m_perDrawBufferOffset * (m_residency ? 2 : 1);
1392
1393                 vk.cmdBindIndexBuffer   (cmdBuffer, *m_sparseBuffer, indexOffset, VK_INDEX_TYPE_UINT32);
1394                 vk.cmdDrawIndexed               (cmdBuffer, m_halfVertexCount, 1u, 0u, 0, 0u);
1395         }
1396
1397         void initializeBuffers (void)
1398         {
1399                 // Vertex buffer
1400                 const DeviceInterface&  vk                                      = getDeviceInterface();
1401                 const VkDeviceSize              vertexBufferSize        = 2 * m_halfVertexCount * sizeof(Vec4);
1402                                                                 m_vertexBuffer          = makeBuffer(vk, getDevice(), makeBufferCreateInfo(vertexBufferSize, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT));
1403                                                                 m_vertexBufferAlloc     = bindBuffer(vk, getDevice(), getAllocator(), *m_vertexBuffer, MemoryRequirement::HostVisible);
1404
1405                 {
1406                         const float     step = 2.0f / static_cast<float>(GRID_SIZE);
1407
1408                         generateGrid(m_vertexBufferAlloc->getHostPtr(), step, -1.0f, -1.0f, GRID_SIZE, GRID_SIZE);
1409
1410                         flushMappedMemoryRange(vk, getDevice(), m_vertexBufferAlloc->getMemory(), m_vertexBufferAlloc->getOffset(), vertexBufferSize);
1411                 }
1412
1413                 // Sparse index buffer
1414                 for (deUint32 chunkNdx = 0u; chunkNdx < 2; ++chunkNdx)
1415                 {
1416                         deUint8* const  pData           = static_cast<deUint8*>(m_stagingBufferAlloc->getHostPtr()) + chunkNdx * m_perDrawBufferOffset;
1417                         deUint32* const pIndexData      = reinterpret_cast<deUint32*>(pData);
1418                         const deUint32  ndxBase         = chunkNdx * m_halfVertexCount;
1419
1420                         for (deUint32 i = 0u; i < m_halfVertexCount; ++i)
1421                                 pIndexData[i] = ndxBase + i;
1422                 }
1423         }
1424
1425 private:
1426         const deUint32                  m_halfVertexCount;
1427         Move<VkBuffer>                  m_vertexBuffer;
1428         MovePtr<Allocation>             m_vertexBufferAlloc;
1429 };
1430
1431 //! Draw from a sparse indirect buffer
1432 class IndirectBufferTestInstance : public DrawGridTestInstance
1433 {
1434 public:
1435         IndirectBufferTestInstance (Context& context, const TestFlags flags)
1436                 : DrawGridTestInstance  (context,
1437                                                                  flags,
1438                                                                  VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT,
1439                                                                  sizeof(VkDrawIndirectCommand))
1440         {
1441         }
1442
1443         void rendererDraw (const VkPipelineLayout pipelineLayout, const VkCommandBuffer cmdBuffer) const
1444         {
1445                 DE_UNREF(pipelineLayout);
1446
1447                 m_context.getTestContext().getLog()
1448                         << tcu::TestLog::Message << "Drawing two triangles covering the whole viewport. There should be no red pixels visible." << tcu::TestLog::EndMessage;
1449
1450                 const DeviceInterface&  vk                              = getDeviceInterface();
1451                 const VkDeviceSize              vertexOffset    = 0ull;
1452                 VkDeviceSize                    indirectOffset  = 0ull;
1453
1454                 vk.cmdBindVertexBuffers (cmdBuffer, 0u, 1u, &m_vertexBuffer.get(), &vertexOffset);
1455                 vk.cmdDrawIndirect              (cmdBuffer, *m_sparseBuffer, indirectOffset, 1u, 0u);
1456
1457                 indirectOffset += m_perDrawBufferOffset * (m_residency ? 2 : 1);
1458
1459                 vk.cmdDrawIndirect              (cmdBuffer, *m_sparseBuffer, indirectOffset, 1u, 0u);
1460         }
1461
1462         void initializeBuffers (void)
1463         {
1464                 // Vertex buffer
1465                 const DeviceInterface&  vk                                      = getDeviceInterface();
1466                 const VkDeviceSize              vertexBufferSize        = 2 * 3 * sizeof(Vec4);
1467                                                                 m_vertexBuffer          = makeBuffer(vk, getDevice(), makeBufferCreateInfo(vertexBufferSize, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT));
1468                                                                 m_vertexBufferAlloc     = bindBuffer(vk, getDevice(), getAllocator(), *m_vertexBuffer, MemoryRequirement::HostVisible);
1469
1470                 {
1471                         generateGrid(m_vertexBufferAlloc->getHostPtr(), 2.0f, -1.0f, -1.0f, 1, 1);
1472                         flushMappedMemoryRange(vk, getDevice(), m_vertexBufferAlloc->getMemory(), m_vertexBufferAlloc->getOffset(), vertexBufferSize);
1473                 }
1474
1475                 // Indirect buffer
1476                 for (deUint32 chunkNdx = 0u; chunkNdx < 2; ++chunkNdx)
1477                 {
1478                         deUint8* const                                  pData           = static_cast<deUint8*>(m_stagingBufferAlloc->getHostPtr()) + chunkNdx * m_perDrawBufferOffset;
1479                         VkDrawIndirectCommand* const    pCmdData        = reinterpret_cast<VkDrawIndirectCommand*>(pData);
1480
1481                         pCmdData->firstVertex   = 3u * chunkNdx;
1482                         pCmdData->firstInstance = 0u;
1483                         pCmdData->vertexCount   = 3u;
1484                         pCmdData->instanceCount = 1u;
1485                 }
1486         }
1487
1488 private:
1489         Move<VkBuffer>                  m_vertexBuffer;
1490         MovePtr<Allocation>             m_vertexBufferAlloc;
1491 };
1492
1493 //! Similar to the class in vktTestCaseUtil.hpp, but uses Arg0 directly rather than through a InstanceFunction1
1494 template<typename Arg0>
1495 class FunctionProgramsSimple1
1496 {
1497 public:
1498         typedef void    (*Function)                             (vk::SourceCollections& dst, Arg0 arg0);
1499                                         FunctionProgramsSimple1 (Function func) : m_func(func)                                                  {}
1500         void                    init                                    (vk::SourceCollections& dst, const Arg0& arg0) const    { m_func(dst, arg0); }
1501
1502 private:
1503         const Function  m_func;
1504 };
1505
1506 //! Convenience function to create a TestCase based on a freestanding initPrograms and a TestInstance implementation
1507 template<typename TestInstanceT, typename Arg0>
1508 TestCase* createTestInstanceWithPrograms (tcu::TestContext&                                                                     testCtx,
1509                                                                                   const std::string&                                                            name,
1510                                                                                   const std::string&                                                            desc,
1511                                                                                   typename FunctionProgramsSimple1<Arg0>::Function      initPrograms,
1512                                                                                   Arg0                                                                                          arg0)
1513 {
1514         return new InstanceFactory1<TestInstanceT, Arg0, FunctionProgramsSimple1<Arg0> >(
1515                 testCtx, tcu::NODETYPE_SELF_VALIDATE, name, desc, FunctionProgramsSimple1<Arg0>(initPrograms), arg0);
1516 }
1517
1518 void populateTestGroup (tcu::TestCaseGroup* parentGroup)
1519 {
1520         const struct
1521         {
1522                 std::string             name;
1523                 TestFlags               flags;
1524         } groups[] =
1525         {
1526                 { "sparse_binding",                                                     0u                                                                                                              },
1527                 { "sparse_binding_aliased",                                     TEST_FLAG_ALIASED,                                                                              },
1528                 { "sparse_residency",                                           TEST_FLAG_RESIDENCY,                                                                    },
1529                 { "sparse_residency_aliased",                           TEST_FLAG_RESIDENCY | TEST_FLAG_ALIASED,                                },
1530                 { "sparse_residency_non_resident_strict",       TEST_FLAG_RESIDENCY | TEST_FLAG_NON_RESIDENT_STRICT,    },
1531         };
1532
1533         const int numGroupsIncludingNonResidentStrict   = DE_LENGTH_OF_ARRAY(groups);
1534         const int numGroupsDefaultList                                  = numGroupsIncludingNonResidentStrict - 1;
1535
1536         // Transfer
1537         {
1538                 MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(parentGroup->getTestContext(), "transfer", ""));
1539                 {
1540                         MovePtr<tcu::TestCaseGroup> subGroup(new tcu::TestCaseGroup(parentGroup->getTestContext(), "sparse_binding", ""));
1541                         addBufferSparseBindingTests(subGroup.get());
1542                         group->addChild(subGroup.release());
1543                 }
1544                 parentGroup->addChild(group.release());
1545         }
1546
1547         // SSBO
1548         {
1549                 MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(parentGroup->getTestContext(), "ssbo", ""));
1550                 {
1551                         MovePtr<tcu::TestCaseGroup> subGroup(new tcu::TestCaseGroup(parentGroup->getTestContext(), "sparse_binding_aliased", ""));
1552                         addBufferSparseMemoryAliasingTests(subGroup.get());
1553                         group->addChild(subGroup.release());
1554                 }
1555                 {
1556                         MovePtr<tcu::TestCaseGroup> subGroup(new tcu::TestCaseGroup(parentGroup->getTestContext(), "sparse_residency", ""));
1557                         addBufferSparseResidencyTests(subGroup.get());
1558                         group->addChild(subGroup.release());
1559                 }
1560                 parentGroup->addChild(group.release());
1561         }
1562
1563         // UBO
1564         {
1565                 MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(parentGroup->getTestContext(), "ubo", ""));
1566
1567                 for (int groupNdx = 0u; groupNdx < numGroupsIncludingNonResidentStrict; ++groupNdx)
1568                         group->addChild(createTestInstanceWithPrograms<UBOTestInstance>(group->getTestContext(), groups[groupNdx].name.c_str(), "", initProgramsDrawWithUBO, groups[groupNdx].flags));
1569
1570                 parentGroup->addChild(group.release());
1571         }
1572
1573         // Vertex buffer
1574         {
1575                 MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(parentGroup->getTestContext(), "vertex_buffer", ""));
1576
1577                 for (int groupNdx = 0u; groupNdx < numGroupsDefaultList; ++groupNdx)
1578                         group->addChild(createTestInstanceWithPrograms<VertexBufferTestInstance>(group->getTestContext(), groups[groupNdx].name.c_str(), "", initProgramsDrawGrid, groups[groupNdx].flags));
1579
1580                 parentGroup->addChild(group.release());
1581         }
1582
1583         // Index buffer
1584         {
1585                 MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(parentGroup->getTestContext(), "index_buffer", ""));
1586
1587                 for (int groupNdx = 0u; groupNdx < numGroupsDefaultList; ++groupNdx)
1588                         group->addChild(createTestInstanceWithPrograms<IndexBufferTestInstance>(group->getTestContext(), groups[groupNdx].name.c_str(), "", initProgramsDrawGrid, groups[groupNdx].flags));
1589
1590                 parentGroup->addChild(group.release());
1591         }
1592
1593         // Indirect buffer
1594         {
1595                 MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(parentGroup->getTestContext(), "indirect_buffer", ""));
1596
1597                 for (int groupNdx = 0u; groupNdx < numGroupsDefaultList; ++groupNdx)
1598                         group->addChild(createTestInstanceWithPrograms<IndirectBufferTestInstance>(group->getTestContext(), groups[groupNdx].name.c_str(), "", initProgramsDrawGrid, groups[groupNdx].flags));
1599
1600                 parentGroup->addChild(group.release());
1601         }
1602 }
1603
1604 } // anonymous ns
1605
1606 tcu::TestCaseGroup* createSparseBufferTests (tcu::TestContext& testCtx)
1607 {
1608         return createTestGroup(testCtx, "buffer", "Sparse buffer usage tests", populateTestGroup);
1609 }
1610
1611 } // sparse
1612 } // vkt