Fixes in defragmentation tests for integrated GPUs Fixes #339
diff --git a/src/Tests.cpp b/src/Tests.cpp index 71ed3af..b568009 100644 --- a/src/Tests.cpp +++ b/src/Tests.cpp
@@ -760,6 +760,7 @@ VkBufferCreateInfo m_BufferInfo; VkImageCreateInfo m_ImageInfo; }; + bool m_DefragmentationMovable = true; // After defragmentation. VkBuffer m_NewBuffer = VK_NULL_HANDLE; @@ -1447,13 +1448,14 @@ for (uint32_t i = 0; i < stepInfo.moveCount; ++i) { - if (stepInfo.pMoves[i].operation == VMA_DEFRAGMENTATION_MOVE_OPERATION_COPY) + VmaAllocationInfo info; + vmaGetAllocationInfo(g_hAllocator, stepInfo.pMoves[i].srcAllocation, &info); + + AllocInfo* allocInfo = (AllocInfo*)info.pUserData; + // Allocation comes from this test and it is movable. + if(stepInfo.pMoves[i].operation == VMA_DEFRAGMENTATION_MOVE_OPERATION_COPY && + allocInfo != nullptr && allocInfo->m_DefragmentationMovable) { - VmaAllocationInfo info; - vmaGetAllocationInfo(g_hAllocator, stepInfo.pMoves[i].srcAllocation, &info); - - AllocInfo* allocInfo = (AllocInfo*)info.pUserData; - if (allocInfo->m_Image) { VkImage newImage; @@ -1531,6 +1533,11 @@ wantsMemoryBarrier = true; } } + else + { + // Some unrelated allocation not from this test or non-movable. + stepInfo.pMoves[i].operation = VMA_DEFRAGMENTATION_MOVE_OPERATION_IGNORE; + } } if (!beginImageBarriers.empty() || wantsMemoryBarrier) @@ -1626,6 +1633,8 @@ VmaDefragmentationPassMoveInfo pass = {}; while ((res = vmaBeginDefragmentationPass(g_hAllocator, defragCtx, &pass)) == VK_INCOMPLETE) { + wprintf(L" Pass: moveCount=%u\n", pass.moveCount); + BeginSingleTimeCommands(); ProcessDefragmentationPass(pass); EndSingleTimeCommands(); @@ -1637,23 +1646,26 @@ VmaAllocationInfo vmaAllocInfo; vmaGetAllocationInfo(g_hAllocator, alloc, &vmaAllocInfo); AllocInfo* allocInfo = (AllocInfo*)vmaAllocInfo.pUserData; - - if(allocInfo->m_Buffer) + if(pass.pMoves[i].operation != VMA_DEFRAGMENTATION_MOVE_OPERATION_IGNORE) { - assert(allocInfo->m_NewBuffer && !allocInfo->m_Image && !allocInfo->m_NewImage); - vkDestroyBuffer(g_hDevice, allocInfo->m_Buffer, g_Allocs); - allocInfo->m_Buffer = allocInfo->m_NewBuffer; - allocInfo->m_NewBuffer = VK_NULL_HANDLE; + TEST(allocInfo != nullptr && allocInfo->m_DefragmentationMovable); + if (allocInfo->m_Buffer) + { + TEST(allocInfo->m_NewBuffer && !allocInfo->m_Image && !allocInfo->m_NewImage); + vkDestroyBuffer(g_hDevice, allocInfo->m_Buffer, g_Allocs); + allocInfo->m_Buffer = allocInfo->m_NewBuffer; + allocInfo->m_NewBuffer = VK_NULL_HANDLE; + } + else if (allocInfo->m_Image) + { + TEST(allocInfo->m_NewImage && !allocInfo->m_Buffer && !allocInfo->m_NewBuffer); + vkDestroyImage(g_hDevice, allocInfo->m_Image, g_Allocs); + allocInfo->m_Image = allocInfo->m_NewImage; + allocInfo->m_NewImage = VK_NULL_HANDLE; + } + else + assert(0); } - else if(allocInfo->m_Image) - { - assert(allocInfo->m_NewImage && !allocInfo->m_Buffer && !allocInfo->m_NewBuffer); - vkDestroyImage(g_hDevice, allocInfo->m_Image, g_Allocs); - allocInfo->m_Image = allocInfo->m_NewImage; - allocInfo->m_NewImage = VK_NULL_HANDLE; - } - else - assert(0); } if ((res = vmaEndDefragmentationPass(g_hAllocator, defragCtx, &pass)) == VK_SUCCESS) break; @@ -1897,7 +1909,6 @@ // persistentlyMappedOption = 1 - persistently mapped. for (uint32_t persistentlyMappedOption = 0; persistentlyMappedOption < 2; ++persistentlyMappedOption) { - wprintf(L" Persistently mapped option = %u\n", persistentlyMappedOption); const bool persistentlyMapped = persistentlyMappedOption != 0; // # Test 1 @@ -1905,6 +1916,8 @@ // Fill 2 blocks. Remove odd buffers. Defragment everything. // Expected result: at least 1 block freed. { + wprintf(L" Persistently mapped option = %u test 1\n", persistentlyMappedOption); + for (size_t i = 0; i < BLOCK_SIZE / BUF_SIZE * 2; ++i) { AllocInfo allocInfo; @@ -1935,6 +1948,8 @@ // Fill 2 blocks. Remove odd buffers. Defragment one buffer at time. // Expected result: Each of 4 interations makes some progress. { + wprintf(L" Persistently mapped option = %u test 2\n", persistentlyMappedOption); + for (size_t i = 0; i < BLOCK_SIZE / BUF_SIZE * 2; ++i) { AllocInfo allocInfo; @@ -1964,6 +1979,7 @@ VmaDefragmentationPassMoveInfo pass = {}; res = vmaBeginDefragmentationPass(g_hAllocator, defragCtx, &pass); TEST(res == VK_INCOMPLETE); + wprintf(L" Pass: moveCount=%u\n", pass.moveCount); BeginSingleTimeCommands(); ProcessDefragmentationPass(pass); @@ -1977,6 +1993,7 @@ vmaGetAllocationInfo(g_hAllocator, alloc, &vmaAllocInfo); AllocInfo* allocInfo = (AllocInfo*)vmaAllocInfo.pUserData; + if(allocInfo != nullptr && allocInfo->m_DefragmentationMovable) if (allocInfo->m_Buffer) { assert(allocInfo->m_NewBuffer && !allocInfo->m_Image && !allocInfo->m_NewImage); @@ -2013,6 +2030,8 @@ // Defragment while having some percent of them unmovable. // Expected result: Just simple validation. { + wprintf(L" Persistently mapped option = %u test 3\n", persistentlyMappedOption); + for (size_t i = 0; i < 100; ++i) { VkBufferCreateInfo localBufCreateInfo = bufCreateInfo; @@ -2039,7 +2058,10 @@ { size_t indexNonMovable = i + rand.Generate() % (uint32_t)(allocations.size() - i); if (indexNonMovable != i) + { std::swap(allocations[i], allocations[indexNonMovable]); + allocations[i].m_DefragmentationMovable = false; + } } // Set data for defragmentation retrieval @@ -2056,13 +2078,7 @@ VmaDefragmentationPassMoveInfo pass = {}; while ((res = vmaBeginDefragmentationPass(g_hAllocator, defragCtx, &pass)) == VK_INCOMPLETE) { - VmaDefragmentationMove* end = pass.pMoves + pass.moveCount; - for (uint32_t i = 0; i < numberNonMovable; ++i) - { - VmaDefragmentationMove* move = std::find_if(pass.pMoves, end, [&](VmaDefragmentationMove& move) { return move.srcAllocation == allocations[i].m_Allocation; }); - if (move != end) - move->operation = VMA_DEFRAGMENTATION_MOVE_OPERATION_IGNORE; - } + wprintf(L" Pass: moveCount=%u\n", pass.moveCount); BeginSingleTimeCommands(); ProcessDefragmentationPass(pass); @@ -2191,8 +2207,7 @@ if(res == VK_SUCCESS) break; TEST(res == VK_INCOMPLETE); - - wprintf(L" Pass %u moving %u allocations\n", passIndex, passInfo.moveCount); + wprintf(L" Pass: moveCount=%u\n", passInfo.moveCount); for(uint32_t moveIndex = 0; moveIndex < passInfo.moveCount; ++moveIndex) { @@ -2382,6 +2397,8 @@ VmaDefragmentationPassMoveInfo pass = {}; while ((res = vmaBeginDefragmentationPass(g_hAllocator, defragCtx, &pass)) == VK_INCOMPLETE) { + wprintf(L" Pass: moveCount=%u\n", pass.moveCount); + VmaDefragmentationMove* end = pass.pMoves + pass.moveCount; for (uint32_t i = 0; i < numberNonMovable; ++i) { @@ -2402,13 +2419,13 @@ // Destroy old buffers/images and replace them with new handles. for (size_t i = 0; i < pass.moveCount; ++i) { - if (pass.pMoves[i].operation == VMA_DEFRAGMENTATION_MOVE_OPERATION_COPY) + if (pass.pMoves[i].operation != VMA_DEFRAGMENTATION_MOVE_OPERATION_IGNORE) { VmaAllocation const alloc = pass.pMoves[i].srcAllocation; VmaAllocationInfo vmaAllocInfo; vmaGetAllocationInfo(g_hAllocator, alloc, &vmaAllocInfo); AllocInfo* allocInfo = (AllocInfo*)vmaAllocInfo.pUserData; - + TEST(allocInfo != nullptr); if (allocInfo->m_Buffer) { assert(allocInfo->m_NewBuffer && !allocInfo->m_Image && !allocInfo->m_NewImage); @@ -2433,7 +2450,7 @@ TEST(res == VK_INCOMPLETE); } TEST(res == VK_SUCCESS); - + VmaDefragmentationStats defragStats; vmaEndDefragmentation(g_hAllocator, defragCtx, &defragStats); @@ -2448,6 +2465,8 @@ void TestDefragmentationFull() { + wprintf(L"Test defragmentation full\n"); + std::vector<AllocInfo> allocations; // Create initial allocations. @@ -2487,39 +2506,23 @@ for (auto& alloc : allocations) vmaSetAllocationUserData(g_hAllocator, alloc.m_Allocation, &alloc); - const uint32_t defragCount = 1; - for(uint32_t defragIndex = 0; defragIndex < defragCount; ++defragIndex) - { - std::vector<VkBool32> allocationsChanged(vmaAllocations.size()); + VmaDefragmentationInfo defragmentationInfo = {}; + defragmentationInfo.flags = VMA_DEFRAGMENTATION_FLAG_ALGORITHM_FULL_BIT; - VmaDefragmentationInfo defragmentationInfo = {}; - defragmentationInfo.flags = VMA_DEFRAGMENTATION_FLAG_ALGORITHM_FULL_BIT; + time_point begTime = std::chrono::high_resolution_clock::now(); - wprintf(L"Defragmentation #%u\n", defragIndex); + VmaDefragmentationStats stats; + Defragment(defragmentationInfo, &stats); - time_point begTime = std::chrono::high_resolution_clock::now(); + float defragmentDuration = ToFloatSeconds(std::chrono::high_resolution_clock::now() - begTime); - VmaDefragmentationStats stats; - Defragment(defragmentationInfo, &stats); + wprintf(L" Moved allocations %u, bytes %llu\n", stats.allocationsMoved, stats.bytesMoved); + wprintf(L" Freed blocks %u, bytes %llu\n", stats.deviceMemoryBlocksFreed, stats.bytesFreed); + wprintf(L" Time: %.2f s\n", defragmentDuration); - float defragmentDuration = ToFloatSeconds(std::chrono::high_resolution_clock::now() - begTime); - - wprintf(L"Moved allocations %u, bytes %llu\n", stats.allocationsMoved, stats.bytesMoved); - wprintf(L"Freed blocks %u, bytes %llu\n", stats.deviceMemoryBlocksFreed, stats.bytesFreed); - wprintf(L"Time: %.2f s\n", defragmentDuration); - - for(size_t i = 0; i < vmaAllocations.size(); ++i) - { - if(allocationsChanged[i]) - { - RecreateAllocationResource(allocations[i]); - } - } - - //wchar_t fileName[MAX_PATH]; - //swprintf(fileName, MAX_PATH, L"After_%02u.csv", defragIndex); - //SaveAllocatorStatsToFile(fileName); - } + //wchar_t fileName[MAX_PATH]; + //swprintf(fileName, MAX_PATH, L"After_%02u.csv", defragIndex); + //SaveAllocatorStatsToFile(fileName); } // Destroy all remaining allocations. @@ -2527,12 +2530,16 @@ DestroyAllAllocations(allocations); } +static void PrintDefragmentationStats(const VmaDefragmentationStats& stats) +{ + wprintf(L" Stats: bytesMoved=%llu, bytesFreed=%llu, allocationsMoved=%u, deviceMemoryBlocksFreed=%u\n", + stats.bytesMoved, stats.bytesFreed, stats.allocationsMoved, stats.deviceMemoryBlocksFreed); +} + static void TestDefragmentationGpu() { wprintf(L"Test defragmentation GPU\n"); - std::vector<AllocInfo> allocations; - // Create that many allocations to surely fill 3 new blocks of 256 MB. const VkDeviceSize bufSizeMin = 5ull * 1024 * 1024; const VkDeviceSize bufSizeMax = 10ull * 1024 * 1024; @@ -2542,6 +2549,9 @@ const size_t percentNonMovable = 3; RandomNumberGenerator rand = { 234522 }; + std::vector<AllocInfo> allocations; + allocations.reserve(bufCount); + VkBufferCreateInfo bufCreateInfo = { VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO }; VmaAllocationCreateInfo allocCreateInfo = {}; @@ -2552,12 +2562,14 @@ { bufCreateInfo.size = align_up(rand.Generate() % (bufSizeMax - bufSizeMin) + bufSizeMin, 32ull); + AllocInfo alloc; + if(rand.Generate() % 100 < percentNonMovable) { bufCreateInfo.usage = VK_BUFFER_USAGE_VERTEX_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_TRANSFER_SRC_BIT; - allocCreateInfo.pUserData = (void*)(uintptr_t)2; + alloc.m_DefragmentationMovable = false; } else { @@ -2566,10 +2578,8 @@ VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_TRANSFER_SRC_BIT; // And in JSON dump. - allocCreateInfo.pUserData = (void*)(uintptr_t)1; } - AllocInfo alloc; alloc.CreateBuffer(bufCreateInfo, allocCreateInfo); alloc.m_StartValue = rand.Generate(); allocations.push_back(alloc); @@ -2599,38 +2609,10 @@ // Defragment using GPU only. { - const size_t allocCount = allocations.size(); - - std::vector<VmaAllocation> allocationPtrs; - std::vector<VkBool32> allocationChanged; - std::vector<size_t> allocationOriginalIndex; - - for(size_t i = 0; i < allocCount; ++i) - { - VmaAllocationInfo allocInfo = {}; - vmaGetAllocationInfo(g_hAllocator, allocations[i].m_Allocation, &allocInfo); - if((uintptr_t)allocInfo.pUserData == 1) // Movable - { - allocationPtrs.push_back(allocations[i].m_Allocation); - allocationChanged.push_back(VK_FALSE); - allocationOriginalIndex.push_back(i); - } - } - - const size_t movableAllocCount = allocationPtrs.size(); - VmaDefragmentationInfo defragInfo = {}; VmaDefragmentationStats stats; Defragment(defragInfo, &stats); - - for(size_t i = 0; i < movableAllocCount; ++i) - { - if(allocationChanged[i]) - { - const size_t origAllocIndex = allocationOriginalIndex[i]; - RecreateAllocationResource(allocations[origAllocIndex]); - } - } + PrintDefragmentationStats(stats); // If corruption detection is enabled, GPU defragmentation may not work on // memory types that have this detection active, e.g. on Intel. @@ -2749,6 +2731,8 @@ VmaDefragmentationPassMoveInfo pass = {}; while ((res = vmaBeginDefragmentationPass(g_hAllocator, ctx, &pass)) == VK_INCOMPLETE) { + wprintf(L" Pass: moveCount=%u\n", pass.moveCount); + // Ignore data outside of test for (uint32_t i = 0; i < pass.moveCount; ++i) { @@ -2890,11 +2874,9 @@ } } - { - // Set our user data pointers. A real application should probably be more clever here - for (auto& alloc : allocations) - vmaSetAllocationUserData(g_hAllocator, alloc.m_Allocation, &alloc); - } + // Set our user data pointers. A real application should probably be more clever here + for (auto& alloc : allocations) + vmaSetAllocationUserData(g_hAllocator, alloc.m_Allocation, &alloc); // Fill them with meaningful data. UploadGpuData(allocations.data(), allocations.size()); @@ -2936,19 +2918,9 @@ VmaDefragmentationPassMoveInfo pass = {}; while((res = vmaBeginDefragmentationPass(g_hAllocator, ctx, &pass)) == VK_INCOMPLETE) { - makeAdditionalAllocation(); + wprintf(L" Pass: moveCount=%u\n", pass.moveCount); - // Ignore data outside of test - for (uint32_t i = 0; i < pass.moveCount; ++i) - { - auto it = std::find_if(allocations.begin(), allocations.end(), [&](const AllocInfo& info) { return pass.pMoves[i].srcAllocation == info.m_Allocation; }); - if (it == allocations.end()) - { - auto it = std::find_if(additionalAllocations.begin(), additionalAllocations.end(), [&](const AllocInfo& info) { return pass.pMoves[i].srcAllocation == info.m_Allocation; }); - if (it == additionalAllocations.end()) - pass.pMoves[i].operation = VMA_DEFRAGMENTATION_MOVE_OPERATION_IGNORE; - } - } + makeAdditionalAllocation(); BeginSingleTimeCommands(); ProcessDefragmentationPass(pass); @@ -5373,6 +5345,8 @@ VmaDefragmentationPassMoveInfo pass = {}; while ((res = vmaBeginDefragmentationPass(g_hAllocator, defragCtx, &pass)) == VK_INCOMPLETE) { + wprintf(L" Pass: moveCount=%u\n", pass.moveCount); + if ((res = vmaEndDefragmentationPass(g_hAllocator, defragCtx, &pass)) == VK_SUCCESS) break; TEST(res == VK_INCOMPLETE);