CUDAで画像処理ができない

実現したいこと

デスクトップをキャプチャした画像を共有ハンドルを用いてCUDAに登録したソースSurfaceとNV１２フォーマットから作ったY面とUV面のsurfaceにカーネルを用いてRGBAからNV12への変換とリサイズを同時に行いたい

発生している問題・分からないこと

CUDAカーネルでY面とUV面にデータがかきこまれない

エラーメッセージ

error
1エラーメッセージは特になし
2
3エラハンドリングで特にエラーをはいていない

該当のソースコード

C++
1            D3D12_RESOURCE_DESC desc12NV12 = {};
2            desc12NV12.Dimension = D3D12_RESOURCE_DIMENSION_TEXTURE2D;
3            desc12NV12.Alignment = 0;
4            desc12NV12.Width = targetWidth;
5            desc12NV12.Height = targetHeight;
6            desc12NV12.DepthOrArraySize = 1;
7            desc12NV12.MipLevels = 1;
8            desc12NV12.Format = DXGI_FORMAT_NV12;
9            desc12NV12.SampleDesc.Count = 1;
10            desc12NV12.Layout = D3D12_TEXTURE_LAYOUT_UNKNOWN;
11            desc12NV12.Flags = D3D12_RESOURCE_FLAG_ALLOW_SIMULTANEOUS_ACCESS;
12
13            // ヒープ設定は同一 GPU 内共有のみ
14            D3D12_HEAP_PROPERTIES heapNV12 = {};
15            heapNV12.Type = D3D12_HEAP_TYPE_DEFAULT;
16            heapNV12.CPUPageProperty = D3D12_CPU_PAGE_PROPERTY_UNKNOWN;
17            heapNV12.MemoryPoolPreference = D3D12_MEMORY_POOL_UNKNOWN;
18
19            D3D11_RESOURCE_FLAGS d3d11FlagsNV12 = {};
20            d3d11FlagsNV12.MiscFlags = D3D11_RESOURCE_MISC_SHARED | D3D11_RESOURCE_MISC_SHARED_NTHANDLE;
21
22            hr = compatDev->CreateSharedResource(
23                &heapNV12,
24                D3D12_HEAP_FLAG_SHARED,
25                &desc12NV12,
26                D3D12_RESOURCE_STATE_COMMON,
27                nullptr,
28                &d3d11FlagsNV12,
29                D3D12_COMPATIBILITY_SHARED_FLAG_NONE,
30                nullptr, nullptr,
31                __uuidof(ID3D12Resource),
32                reinterpret_cast<void**>(context.d3d12NV12Res.GetAddressOf())
33            );
34            if (FAILED(hr)) {
35                DebugLog(L"Failed to Create SharedResource(NV12).");
36                return;
37            }
38
39            HANDLE sharedHandleNV12 = nullptr;
40            SECURITY_ATTRIBUTES saNV12{};
41            saNV12.nLength = sizeof(saNV12);
42            saNV12.bInheritHandle = FALSE;
43            saNV12.lpSecurityDescriptor = nullptr;
44
45            hr = d3d12Device->CreateSharedHandle(
46                context.d3d12NV12Res.Get(),
47                &saNV12,
48                GENERIC_ALL,
49                nullptr,
50                &sharedHandleNV12
51            );
52            if (FAILED(hr)) {
53                DebugLog(L"Failed to Create sharedHandle(NV12).");
54                return;
55            }
56
57            D3D12_RESOURCE_DESC desc = context.d3d12NV12Res->GetDesc();
58
59            // 2) 実際のアロケーションサイズを取得
60            D3D12_RESOURCE_ALLOCATION_INFO allocInfo =
61                d3d12Device->GetResourceAllocationInfo(
62                    0,              // visibleMask（通常 0）
63                    1,              // リソース個数
64                    &desc           // リソース記述へのポインタ
65                );
66            UINT64 resourceSize = allocInfo.SizeInBytes;
67            
68            DebugLog(L"external buffer size : " + std::to_wstring(resourceSize));
69
70            CUDA_EXTERNAL_MEMORY_HANDLE_DESC extDescNV12 = {};
71            extDescNV12.type = CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE;
72            extDescNV12.handle.win32.handle = sharedHandleNV12;
73            extDescNV12.size = resourceSize;
74            extDescNV12.flags = CUDA_EXTERNAL_MEMORY_DEDICATED;
75
76            err = cuImportExternalMemory(&context.extMemNV12, &extDescNV12);
77            if (err != CUDA_SUCCESS) {
78                DebugLog(L"cuImportExternalMemory(NV12) failed");
79                return;
80            }
81
82            CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC descY = {};
83            descY.offset    = 0;                          // Y は先頭
84            descY.numLevels = 1;
85            descY.arrayDesc.Width       = targetWidth;
86            descY.arrayDesc.Height      = targetHeight;
87            descY.arrayDesc.Depth       = 0;              
88            descY.arrayDesc.Format      = CU_AD_FORMAT_UNSIGNED_INT8;
89            descY.arrayDesc.NumChannels = 1;              // Y のみ
90            descY.arrayDesc.Flags       = CUDA_ARRAY3D_SURFACE_LDST;
91
92            CUmipmappedArray mipY;
93            err = cuExternalMemoryGetMappedMipmappedArray(&mipY, context.extMemNV12, &descY);
94            if (err != CUDA_SUCCESS) {
95                DebugLog(L"cuExternalMemoryGetMappedMipmappedArray(Y) failed");
96                cuDestroyExternalMemory(context.extMem);
97                CloseHandle(sharedHandle);
98                context.sharedD3D12Tex.Reset();
99                cuDestroyExternalMemory(context.extMemNV12);
100                CloseHandle(sharedHandleNV12);
101                context.d3d12NV12Res.Reset();
102                return;
103            }
104
105            CUarray arrY;
106            err = cuMipmappedArrayGetLevel(&arrY, mipY, 0);
107            if (err != CUDA_SUCCESS) {
108                DebugLog(L"cuMipmappedArrayGetLevel(Y) failed");
109                cuDestroyExternalMemory(context.extMem);
110                CloseHandle(sharedHandle);
111                context.sharedD3D12Tex.Reset();
112                cuDestroyExternalMemory(context.extMemNV12);
113                CloseHandle(sharedHandleNV12);
114                context.d3d12NV12Res.Reset();
115                return;
116            }
117
118            CUDA_RESOURCE_DESC resDescY = {};
119            resDescY.resType               = CU_RESOURCE_TYPE_ARRAY;
120            resDescY.res.array.hArray      = arrY;
121            resDescY.flags                 = 0;
122
123            err = cuSurfObjectCreate(&context.surfObjY, &resDescY);
124            if (err != CUDA_SUCCESS) {
125                DebugLog(L"cuSurfObjectCreate(Y) failed");
126                cuDestroyExternalMemory(context.extMem);
127                CloseHandle(sharedHandle);
128                context.sharedD3D12Tex.Reset();
129                cuDestroyExternalMemory(context.extMemNV12);
130                CloseHandle(sharedHandleNV12);
131                context.d3d12NV12Res.Reset();
132                return;
133            }
134
135            // 2) UV プレーン用サーフェス
136
137            CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC descUV = {};
138            descUV.offset = 0;
139            descUV.numLevels = 1;
140            descUV.arrayDesc.Width = targetWidth / 2;
141            descUV.arrayDesc.Height = targetHeight / 2;
142            descUV.arrayDesc.Depth = 0;
143            descUV.arrayDesc.Format = CU_AD_FORMAT_UNSIGNED_INT8;
144            descUV.arrayDesc.NumChannels = 2;
145            descUV.arrayDesc.Flags = CUDA_ARRAY3D_SURFACE_LDST;             // U,V インターリーブ
146
147            CUmipmappedArray mipUV;
148            err = cuExternalMemoryGetMappedMipmappedArray(&mipUV, context.extMemNV12, &descUV);
149            if (err != CUDA_SUCCESS) {
150                const char* name = nullptr;
151                cuGetErrorName(err, &name);
152                DebugLog(L"cuExternalMemoryGetMappedMipmappedArray(UV) failed: " + std::wstring(name, name + strlen(name)));
153                cuDestroyExternalMemory(context.extMem);
154                CloseHandle(sharedHandle);
155                context.sharedD3D12Tex.Reset();
156                cuDestroyExternalMemory(context.extMemNV12);
157                CloseHandle(sharedHandleNV12);
158                context.d3d12NV12Res.Reset();
159                return;
160            }
161
162            CUarray arrUV;
163            err = cuMipmappedArrayGetLevel(&arrUV, mipUV, 0);
164            if (err != CUDA_SUCCESS) {
165                DebugLog(L"cuMipmappedArrayGetLevel(UV) failed");
166                cuDestroyExternalMemory(context.extMem);
167                CloseHandle(sharedHandle);
168                context.sharedD3D12Tex.Reset();
169                cuDestroyExternalMemory(context.extMemNV12);
170                CloseHandle(sharedHandleNV12);
171                context.d3d12NV12Res.Reset();
172                return;
173            }
174
175            CUDA_RESOURCE_DESC resDescUV = resDescY;
176            resDescUV.res.array.hArray    = arrUV;
177;
178            err = cuSurfObjectCreate(&context.surfObjUV, &resDescUV);
179            if (err != CUDA_SUCCESS) {
180                DebugLog(L"cuSurfObjectCreate(UV) failed");
181                cuDestroyExternalMemory(context.extMem);
182                CloseHandle(sharedHandle);
183                context.sharedD3D12Tex.Reset();
184                cuDestroyExternalMemory(context.extMemNV12);
185                CloseHandle(sharedHandleNV12);
186                context.d3d12NV12Res.Reset();
187                return;
188            }