Move CUDA interop behind native opt-in by AnastaZIuk · Pull Request #1067 · Devsh-Graphics-Programming/Nabla

AnastaZIuk · 2026-05-07T15:58:44Z

Moves CUDA interop behind SDK-free Nabla headers with explicit Nabla::ext::CUDAInterop native opt-in. Keeps raw CUDA/NVRTC access available for consumers that ask for native opt-in while avoiding default public SDK requirements.

devshgraphicsprogramming · 2026-05-07T16:24:26Z

+// Opt-in native CUDA API. The declarations below are implemented by the Nabla library.
+// This header is intentionally the only public path that includes CUDA SDK types.
+class NBL_API2 CCUDAHandlerAccessor
+{
+	public:
+		static const CUDA& getCUDAFunctionTable(const CCUDAHandler& handler);
+		static const NVRTC& getNVRTCFunctionTable(const CCUDAHandler& handler);
+		static bool defaultHandleResult(CUresult result, const system::logger_opt_ptr& logger);
+		static bool defaultHandleResult(const CCUDAHandler& handler, CUresult result);
+		static bool defaultHandleResult(const CCUDAHandler& handler, nvrtcResult result);
+		static const core::vector<SCUDADeviceInfo>& getAvailableDevices(const CCUDAHandler& handler);
+		static nvrtcResult createProgram(CCUDAHandler& handler, nvrtcProgram* prog, std::string&& source, const char* name, const int headerCount=0, const char* const* headerContents=nullptr, const char* const* includeNames=nullptr);
+		static nvrtcResult compileProgram(const CCUDAHandler& handler, nvrtcProgram prog, core::SRange<const char* const> options);
+		static nvrtcResult getProgramLog(const CCUDAHandler& handler, nvrtcProgram prog, std::string& log);
+		static SPTXResult getPTX(const CCUDAHandler& handler, nvrtcProgram prog);
+		static SPTXResult compileDirectlyToPTX(
+			CCUDAHandler& handler, std::string&& source, const char* filename, core::SRange<const char* const> nvrtcOptions,
+			std::string& log, const int headerCount=0, const char* const* headerContents=nullptr, const char* const* includeNames=nullptr
+		);
+};
+
+class NBL_API2 CCUDADeviceAccessor
+{
+	public:
+		static CUdevice getInternalObject(const CCUDADevice& device);
+		static CUcontext getContext(const CCUDADevice& device);
+		static size_t roundToGranularity(const CCUDADevice& device, CUmemLocationType location, size_t size);
+		static core::smart_refctd_ptr<CCUDAExportableMemory> createExportableMemory(CCUDADevice& device, SExportableMemoryCreationParams&& params);
+};
+
+class NBL_API2 CCUDAExportableMemoryAccessor
+{
+	public:
+		static CUdeviceptr getDeviceptr(const CCUDAExportableMemory& memory);
+};
+
+class NBL_API2 CCUDAImportedMemoryAccessor
+{
+	public:
+		static CUexternalMemory getInternalObject(const CCUDAImportedMemory& memory);
+		static CUresult getMappedBuffer(const CCUDAImportedMemory& memory, CUdeviceptr* mappedBuffer);
+};
+
+class NBL_API2 CCUDAImportedSemaphoreAccessor
+{
+	public:
+		static CUexternalSemaphore getInternalObject(const CCUDAImportedSemaphore& semaphore);
+};


accessors make no sense just move all the nbl/video/CCUDA*.h to the extension

devshgraphicsprogramming · 2026-05-07T16:25:27Z

-#define ASSERT_CUDA_SUCCESS(expr, handler) \
-    do { \
-        const auto cudaResult = (expr); \
-        if (!((handler)->defaultHandleResult(cudaResult))) { \
-            assert(false); \
-        } \
-    } while(0)
-


that macro was useful just needs a rename

devshgraphicsprogramming · 2026-05-07T16:27:50Z

-inline bool CloseExternalHandle(external_handle_t handle)
-{
-#ifdef _WIN32
-    return CloseHandle(handle);
-#else
-    return (close(handle) == 0);
-#endif
-}
-
-inline external_handle_t DuplicateExternalHandle(external_handle_t handle)
-{
-#ifdef _WIN32
-    HANDLE re = ExternalHandleNull;
-
-    const HANDLE cur = GetCurrentProcess();
-    if (!DuplicateHandle(cur, handle, cur, &re, GENERIC_ALL, 0, DUPLICATE_SAME_ACCESS))
-        return ExternalHandleNull;
-
-    return re;
-#else
-    return dup(handle);
-#endif


you may want to keep that inline, these are OS calls, and when they're inline they'll work BEFORE Nabla.dll is delay loaded, which is useful

devshgraphicsprogramming · 2026-05-07T16:31:07Z

+#include "nbl/video/CUDAInterop.h"
+#include "nbl/system/IApplicationFramework.h"
+
+#include <type_traits>
+
+#ifdef _NBL_COMPILE_WITH_CUDA_
+#error "Nabla::Nabla must not propagate the CUDA build define."
+#endif
+
+#ifdef CUDA_VERSION
+#error "Nabla::Nabla must not require CUDA SDK headers."
+#endif
+
+namespace
+{
+
+class CUDAInteropCleanOptInSmoke final : public nbl::system::IApplicationFramework
+{
+	using base_t = nbl::system::IApplicationFramework;
+
+public:
+	using base_t::base_t;
+
+	bool onAppInitialized(nbl::core::smart_refctd_ptr<nbl::system::ISystem>&&) override
+	{
+		static_assert(std::is_class_v<nbl::video::CCUDADevice>);
+		static_assert(std::is_class_v<nbl::video::CCUDAExportableMemory>);
+		static_assert(std::is_class_v<nbl::video::CCUDAImportedMemory>);
+		static_assert(std::is_class_v<nbl::video::CCUDAImportedSemaphore>);


it would make more sense to not have anything CUDA related in Nabla itself

devshgraphicsprogramming · 2026-05-07T16:34:42Z

+	const auto& granularity = SAccess::native(device).allocationGranularity[location];
+	return ((size - 1) / granularity + 1) * granularity;
+}
+


I mentioned in the original PR, this should be inline

devshgraphicsprogramming · 2026-05-07T16:35:43Z

+		if (!cuda_native::CCUDAHandlerAccessor::defaultHandleResult(*handler, cu.pcuMemAddressFree(ptr, size)))
+			assert(false);
 		return err;
 	}

 	CUmemAccessDesc accessDesc = {
-		.location = { .type = location, .id = m_handle },
+		.location = { .type = location, .id = native.handle },
 		.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE,
 	};

 	if (auto err = cu.pcuMemSetAccess(ptr, size, &accessDesc, 1); CUDA_SUCCESS != err)
 	{
-		ASSERT_CUDA_SUCCESS(cu.pcuMemUnmap(ptr, size), m_handler);
-		ASSERT_CUDA_SUCCESS(cu.pcuMemAddressFree(ptr, size), m_handler);
+		if (!cuda_native::CCUDAHandlerAccessor::defaultHandleResult(*handler, cu.pcuMemUnmap(ptr, size)))
+			assert(false);
+		if (!cuda_native::CCUDAHandlerAccessor::defaultHandleResult(*handler, cu.pcuMemAddressFree(ptr, size)))
+			assert(false);


@kevyuu thinking of it we shouldn't crash an entire program vecause of failure here :s

AnastaZIuk added 27 commits May 6, 2026 09:28

Move CUDA interop behind extension target

f4ce3dc

Address CUDA interop review cleanup

78845ae

Simplify CUDA interop smoke CMake

ab9a7e5

Clean CUDA interop smoke usage requirements

bf8eeb3

Export CUDA interop package target

f701ac6

Use CUDAToolkit package targets

a520d57

Require CUDA version via CMake

4bddc57

Split CUDA interop native surface

6f68e66

Add native CUDA accessor overloads

49bcb2c

Document CUDA interop target split

d85657e

Trim CUDA interop README wording

6e8c4f9

Move CUDA interop into Nabla

881e9b8

Document CUDA interop accessor model

5dd1134

Inline CUDA interop stubs

e514df7

Refine CUDA interop boundary

e53c838

Add CUDA interop runtime header discovery

1417905

Tighten CUDA interop native helpers

045432e

Hide CUDA interop native state construction

8a119dd

Clean up CUDA runtime header discovery

e018545

Move CUDA interop API back into video

c6ef6ee

Move smart pointer helpers into core

d559a2c

Use CUDA interop accessors

38705b9

Use explicit CUDA compile log

23e6ef5

Trim CUDA interop API surface

a640183

Keep CUDA SDK layouts private

5bf0e2d

Simplify CUDA interop helper

d745421

Update CUDA interop examples pointer

ffba3d4

devshgraphicsprogramming reviewed May 7, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Move CUDA interop behind native opt-in#1067

Move CUDA interop behind native opt-in#1067
AnastaZIuk wants to merge 27 commits intovk_cuda_interopfrom
cuInteropBS

AnastaZIuk commented May 7, 2026

Uh oh!

devshgraphicsprogramming May 7, 2026

Uh oh!

devshgraphicsprogramming May 7, 2026

Uh oh!

devshgraphicsprogramming May 7, 2026

Uh oh!

devshgraphicsprogramming May 7, 2026

Uh oh!

devshgraphicsprogramming May 7, 2026

Uh oh!

devshgraphicsprogramming May 7, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

AnastaZIuk commented May 7, 2026

Uh oh!

devshgraphicsprogramming May 7, 2026

Choose a reason for hiding this comment

Uh oh!

devshgraphicsprogramming May 7, 2026

Choose a reason for hiding this comment

Uh oh!

devshgraphicsprogramming May 7, 2026

Choose a reason for hiding this comment

Uh oh!

devshgraphicsprogramming May 7, 2026

Choose a reason for hiding this comment

Uh oh!

devshgraphicsprogramming May 7, 2026

Choose a reason for hiding this comment

Uh oh!

devshgraphicsprogramming May 7, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants