|
| 1 | +// D3D12HelloTightAlignment.cpp : This file contains the 'main' function. Program execution begins and ends there. |
| 2 | +// |
| 3 | + |
| 4 | +#include <windows.h> |
| 5 | +#include <iostream> |
| 6 | +#include <atlbase.h> |
| 7 | +#include <vector> |
| 8 | +#include <initguid.h> |
| 9 | +#include "dxcapi.h" |
| 10 | +#include "d3d12.h" |
| 11 | +#include "d3dx12.h" |
| 12 | +#include <dxgi1_6.h> |
| 13 | + |
| 14 | +extern "C" { __declspec(dllexport) extern const UINT D3D12SDKVersion = 716; } |
| 15 | +extern "C" { __declspec(dllexport) extern const char* D3D12SDKPath = u8".\\D3D12\\"; } |
| 16 | +extern "C" { __declspec(dllexport) extern const char* WarpPath = u8".\\WARP\\"; } |
| 17 | + |
| 18 | +// use a warp device instead of a hardware device |
| 19 | +bool g_useWarpDevice = false; |
| 20 | + |
| 21 | +#pragma region helper_and_setup |
| 22 | +//================================================================================================================================= |
| 23 | +// Helper / setup code, not specific to work graphs |
| 24 | +// Look for "Start of interesting code" further below or just collapse this region. |
| 25 | +//================================================================================================================================= |
| 26 | +#define PRINT(text) std::cout << (char*)text << "\n" << std::flush; |
| 27 | +#define VERIFY_SUCCEEDED(hr) {HRESULT hrLocal = hr; if(FAILED(hrLocal)) {PRINT("Error at: " << __FILE__ << ", line: " << __LINE__ << ", HRESULT: 0x" << std::hex << hrLocal); throw E_FAIL;} } |
| 28 | + |
| 29 | +struct D3DContext |
| 30 | +{ |
| 31 | + CComPtr<ID3D12Device14> spDevice; |
| 32 | + CComPtr<IDXGIAdapter3> spAdapter; |
| 33 | + bool bIsHeapTier1; |
| 34 | +}; |
| 35 | + |
| 36 | +//================================================================================================================================= |
| 37 | +void InitDeviceAndContext(D3DContext& D3D, bool useWarp) |
| 38 | +{ |
| 39 | + D3D.spDevice.Release(); |
| 40 | + D3D.spAdapter.Release(); |
| 41 | + |
| 42 | + CComPtr<ID3D12Debug1> pDebug; |
| 43 | + VERIFY_SUCCEEDED(D3D12GetDebugInterface(IID_PPV_ARGS(&pDebug))); |
| 44 | + pDebug->EnableDebugLayer(); |
| 45 | + |
| 46 | + D3D_FEATURE_LEVEL FL = D3D_FEATURE_LEVEL_11_0; |
| 47 | + CComPtr<ID3D12Device> spDevice; |
| 48 | + CComPtr<IDXGIFactory4> factory; |
| 49 | + VERIFY_SUCCEEDED(CreateDXGIFactory2(0, IID_PPV_ARGS(&factory))); |
| 50 | + |
| 51 | + if (useWarp) |
| 52 | + { |
| 53 | + VERIFY_SUCCEEDED(factory->EnumWarpAdapter(IID_PPV_ARGS(&D3D.spAdapter))); |
| 54 | + VERIFY_SUCCEEDED(D3D12CreateDevice(D3D.spAdapter, FL, IID_PPV_ARGS(&spDevice))); |
| 55 | + } |
| 56 | + else |
| 57 | + { |
| 58 | + VERIFY_SUCCEEDED(D3D12CreateDevice(NULL, FL, IID_PPV_ARGS(&spDevice))); |
| 59 | + LUID luid = spDevice->GetAdapterLuid(); |
| 60 | + factory->EnumAdapterByLuid(luid, IID_PPV_ARGS(&D3D.spAdapter)); |
| 61 | + } |
| 62 | + D3D.spDevice = spDevice; |
| 63 | + D3D.bIsHeapTier1 = false; |
| 64 | +} |
| 65 | + |
| 66 | +template <UINT _Size> |
| 67 | +LPSTR FormatMemoryUsage(UINT64 usage, CHAR(&result)[_Size]) |
| 68 | +{ |
| 69 | + const UINT64 mb = 1 << 20; |
| 70 | + const UINT64 kb = 1 << 10; |
| 71 | + if (usage > mb) |
| 72 | + { |
| 73 | + sprintf_s(result, "%.2f MB", static_cast<float>(usage) / mb); |
| 74 | + } |
| 75 | + else if (usage > kb) |
| 76 | + { |
| 77 | + sprintf_s(result, "%.2f KB", static_cast<float>(usage) / kb); |
| 78 | + } |
| 79 | + else |
| 80 | + { |
| 81 | + sprintf_s(result, "%I64d B", usage); |
| 82 | + } |
| 83 | + return result; |
| 84 | +} |
| 85 | + |
| 86 | +void PrintGPUMemoryUsage(D3DContext& D3D, UINT64 baselineUsage = 0) |
| 87 | +{ |
| 88 | + DXGI_QUERY_VIDEO_MEMORY_INFO memoryInfo; |
| 89 | + D3D.spAdapter->QueryVideoMemoryInfo(0, DXGI_MEMORY_SEGMENT_GROUP_LOCAL, &memoryInfo); |
| 90 | + |
| 91 | + char text[100]; |
| 92 | + char usageString[20]; |
| 93 | + UINT64 usage = memoryInfo.CurrentUsage - baselineUsage; |
| 94 | + if (baselineUsage > memoryInfo.CurrentUsage) |
| 95 | + { |
| 96 | + PRINT(" ## Error - baseline usage was greater than current usage"); |
| 97 | + return; |
| 98 | + } |
| 99 | + sprintf_s(text, " Memory Used: %s", FormatMemoryUsage(usage, usageString)); |
| 100 | + PRINT(text); |
| 101 | +} |
| 102 | + |
| 103 | +#pragma endregion |
| 104 | + |
| 105 | +template<UINT32 numBuffers> |
| 106 | +void CreateManyPlacedBuffers(D3DContext& D3D, const CD3DX12_RESOURCE_DESC1& rDesc) |
| 107 | +{ |
| 108 | + DXGI_QUERY_VIDEO_MEMORY_INFO baslineMemoryInfo; |
| 109 | + D3D.spAdapter->QueryVideoMemoryInfo(0, DXGI_MEMORY_SEGMENT_GROUP_LOCAL, &baslineMemoryInfo); |
| 110 | + |
| 111 | + D3D12_RESOURCE_ALLOCATION_INFO1 infos[numBuffers]; |
| 112 | + D3D12_RESOURCE_DESC1 pResourceDescs[numBuffers]; |
| 113 | + std::fill_n(pResourceDescs, numBuffers, rDesc); |
| 114 | + D3D12_RESOURCE_ALLOCATION_INFO aggregateInfo = D3D.spDevice->GetResourceAllocationInfo2(0, numBuffers, pResourceDescs, infos); |
| 115 | + |
| 116 | + CComPtr<ID3D12Heap> pHeap; |
| 117 | + CComPtr<ID3D12Resource> buffers[numBuffers]; |
| 118 | + CD3DX12_HEAP_DESC heapDesc = CD3DX12_HEAP_DESC(aggregateInfo.SizeInBytes, D3D12_HEAP_TYPE_DEFAULT); |
| 119 | + heapDesc.Flags = D3D.bIsHeapTier1 ? D3D12_HEAP_FLAG_ALLOW_ONLY_BUFFERS : D3D12_HEAP_FLAG_NONE; |
| 120 | + VERIFY_SUCCEEDED(D3D.spDevice->CreateHeap(&heapDesc, IID_PPV_ARGS(&pHeap))); |
| 121 | + for (UINT32 i = 0; i < numBuffers; i++) |
| 122 | + { |
| 123 | + VERIFY_SUCCEEDED(D3D.spDevice->CreatePlacedResource1(pHeap, infos[i].Offset, &rDesc, D3D12_RESOURCE_STATE_COMMON, nullptr, IID_PPV_ARGS(&buffers[i]))); |
| 124 | + } |
| 125 | + PrintGPUMemoryUsage(D3D, baslineMemoryInfo.CurrentUsage); |
| 126 | +} |
| 127 | + |
| 128 | +template<UINT32 numBuffers> |
| 129 | +void ComparePlacedBufferMemoryFootprint(D3DContext& D3D) |
| 130 | +{ |
| 131 | + PRINT("Comparing memory footprint for " << numBuffers << " placed buffers...\n"); |
| 132 | + CD3DX12_RESOURCE_DESC1 bufferDesc; |
| 133 | + //Beyond the obvious benefits for small resources, you can also see savings for large buffers that go beyond the 64KiB mark |
| 134 | + UINT bufferSizes[] = { 1, 16, 256, 4096, 32000, 65540, 80000 }; |
| 135 | + for (UINT bufferSize : bufferSizes) |
| 136 | + { |
| 137 | + PRINT("Size: " << bufferSize << " B"); |
| 138 | + PRINT(" Classic buffer alignment:"); |
| 139 | + bufferDesc = CD3DX12_RESOURCE_DESC1::Buffer(bufferSize); |
| 140 | + CreateManyPlacedBuffers<numBuffers>(D3D, bufferDesc); |
| 141 | + |
| 142 | + PRINT(" Tight buffer alignment:"); |
| 143 | + bufferDesc = CD3DX12_RESOURCE_DESC1::Buffer(bufferSize, D3D12_RESOURCE_FLAG_USE_TIGHT_ALIGNMENT); |
| 144 | + CreateManyPlacedBuffers<numBuffers>(D3D, bufferDesc); |
| 145 | + |
| 146 | + PRINT("----------"); |
| 147 | + } |
| 148 | + PRINT("===================="); |
| 149 | +} |
| 150 | + |
| 151 | +template<UINT32 numBuffers> |
| 152 | +void CreateManyCommittedBuffers(D3DContext& D3D, const CD3DX12_RESOURCE_DESC1& rDesc) |
| 153 | +{ |
| 154 | + DXGI_QUERY_VIDEO_MEMORY_INFO baslineMemoryInfo; |
| 155 | + D3D.spAdapter->QueryVideoMemoryInfo(0, DXGI_MEMORY_SEGMENT_GROUP_LOCAL, &baslineMemoryInfo); |
| 156 | + |
| 157 | + CD3DX12_HEAP_PROPERTIES heapProperties = CD3DX12_HEAP_PROPERTIES(D3D12_HEAP_TYPE_DEFAULT); |
| 158 | + CComPtr<ID3D12Resource> buffers[numBuffers]; |
| 159 | + for (UINT32 i = 0; i < numBuffers; i++) |
| 160 | + { |
| 161 | + VERIFY_SUCCEEDED(D3D.spDevice->CreateCommittedResource2(&heapProperties, D3D12_HEAP_FLAG_NONE, &rDesc, D3D12_RESOURCE_STATE_COMMON, nullptr, nullptr, IID_PPV_ARGS(&buffers[i]))); |
| 162 | + } |
| 163 | + |
| 164 | + PrintGPUMemoryUsage(D3D, baslineMemoryInfo.CurrentUsage); |
| 165 | +} |
| 166 | + |
| 167 | +template<UINT32 numBuffers> |
| 168 | +void CompareCommittedBufferMemoryFootprint(D3DContext& D3D) |
| 169 | +{ |
| 170 | + PRINT("Comparing memory footprint for " << numBuffers << " committed buffers...\n"); |
| 171 | + CD3DX12_RESOURCE_DESC1 bufferDesc; |
| 172 | + |
| 173 | + UINT bufferSizes[] = { 1, 16, 256, 4096, 65536, 65540, 80000 }; |
| 174 | + for (UINT bufferSize : bufferSizes) |
| 175 | + { |
| 176 | + PRINT("Size: " << bufferSize << " B"); |
| 177 | + PRINT(" Classic buffer alignment:"); |
| 178 | + bufferDesc = CD3DX12_RESOURCE_DESC1::Buffer(bufferSize); |
| 179 | + CreateManyCommittedBuffers<numBuffers>(D3D, bufferDesc); |
| 180 | + |
| 181 | + PRINT(" Tight buffer alignment:"); |
| 182 | + bufferDesc = CD3DX12_RESOURCE_DESC1::Buffer(bufferSize, D3D12_RESOURCE_FLAG_USE_TIGHT_ALIGNMENT); |
| 183 | + CreateManyCommittedBuffers<numBuffers>(D3D, bufferDesc); |
| 184 | + |
| 185 | + PRINT("----------"); |
| 186 | + } |
| 187 | + PRINT("===================="); |
| 188 | +} |
| 189 | + |
| 190 | +int main() |
| 191 | +{ |
| 192 | + try |
| 193 | + { |
| 194 | + PRINT("\n" << |
| 195 | + "==================================================================================\n" << |
| 196 | + " Hello Tight Alignment\n" << |
| 197 | + "=================================================================================="); |
| 198 | + D3DContext D3D; |
| 199 | + InitDeviceAndContext(D3D, g_useWarpDevice); |
| 200 | + bool warpFallback = false; |
| 201 | + |
| 202 | + CD3DX12FeatureSupport featureSupport; |
| 203 | + featureSupport.Init(D3D.spDevice); |
| 204 | + if (featureSupport.TightAlignmentSupportTier() == D3D12_TIGHT_ALIGNMENT_TIER_NOT_SUPPORTED && !g_useWarpDevice) |
| 205 | + { |
| 206 | + PRINT("Tight alignment is not supported on this driver. Retrying with Warp..."); |
| 207 | + InitDeviceAndContext(D3D, true); |
| 208 | + featureSupport.Init(D3D.spDevice); |
| 209 | + warpFallback = true; |
| 210 | + } |
| 211 | + if (featureSupport.TightAlignmentSupportTier() == D3D12_TIGHT_ALIGNMENT_TIER_NOT_SUPPORTED) |
| 212 | + { |
| 213 | + PRINT("Tight alignment is not supported on this driver or the currently installed version of Warp. Aborting."); |
| 214 | + return 0; |
| 215 | + } |
| 216 | + DXGI_ADAPTER_DESC desc; |
| 217 | + VERIFY_SUCCEEDED(D3D.spAdapter->GetDesc(&desc)); |
| 218 | + std::wcout << L"Running sample on " << desc.Description << L"\n\n" << std::flush; |
| 219 | + |
| 220 | + D3D12_FEATURE_DATA_D3D12_OPTIONS options{}; |
| 221 | + VERIFY_SUCCEEDED(D3D.spDevice->CheckFeatureSupport(D3D12_FEATURE_D3D12_OPTIONS, &options, sizeof(options))); |
| 222 | + D3D.bIsHeapTier1 = options.ResourceHeapTier == D3D12_RESOURCE_HEAP_TIER_1; |
| 223 | + |
| 224 | + ComparePlacedBufferMemoryFootprint<1>(D3D); // No benefit due to the heap allocations being made in 64KiB chuncks |
| 225 | + ComparePlacedBufferMemoryFootprint<8>(D3D); // Can already see some benefit being realized |
| 226 | + ComparePlacedBufferMemoryFootprint<4096>(D3D); // drastic savings in most cases |
| 227 | + |
| 228 | + PRINT("\n========================================\n========================================\n"); |
| 229 | + |
| 230 | + CompareCommittedBufferMemoryFootprint<1>(D3D); |
| 231 | + CompareCommittedBufferMemoryFootprint<8>(D3D); |
| 232 | + CompareCommittedBufferMemoryFootprint<4096>(D3D); |
| 233 | + } |
| 234 | + catch (HRESULT) |
| 235 | + { |
| 236 | + PRINT("Aborting."); |
| 237 | + return -1; |
| 238 | + } |
| 239 | + return 0; |
| 240 | + |
| 241 | +} |
0 commit comments