Add external resource interoperability

warp/__init__.py

@@ -82,6 +82,7 @@
 from warp.context import Kernel, Function, Launch
 from warp.context import Stream, get_stream, set_stream, wait_stream, synchronize_stream
 from warp.context import Event, record_event, wait_event, synchronize_event, get_event_elapsed_time
+from warp.context import ExternalMemoryBuffer, ExternalSemaphore, signal_external_semaphore, wait_external_semaphore
 from warp.context import RegisteredGLBuffer
 from warp.context import is_mempool_supported, is_mempool_enabled, set_mempool_enabled
 from warp.context import (

warp/context.py

@@ -2669,6 +2669,20 @@ def wait_event(self, event: Event):
         """
         runtime.core.cuda_stream_wait_event(self.cuda_stream, event.cuda_event)
 
+    def signal_external_semaphore(self, semaphore: ExternalSemaphore, value: int = 0):
+        """TODO: docs"""
+
+        runtime.core.cuda_signal_external_semaphore_async(
+            semaphore.device.context, semaphore.external_semaphore, value, self.cuda_stream
+        )
+
+    def wait_external_semaphore(self, semaphore: ExternalSemaphore, value: int = 0):
+        """TODO: docs"""
+
+        runtime.core.cuda_wait_external_semaphore_async(
+            semaphore.device.context, semaphore.external_semaphore, value, self.cuda_stream
+        )
+
     def wait_stream(self, other_stream: "Stream", event: Optional[Event] = None):
         """Records an event on `other_stream` and makes this stream wait on it.
 
@@ -3809,6 +3823,50 @@ def __init__(self):
             self.core.cuda_graphics_unregister_resource.argtypes = [ctypes.c_void_p, ctypes.c_void_p]
             self.core.cuda_graphics_unregister_resource.restype = None
 
+            self.core.cuda_import_external_memory.argtypes = [
+                ctypes.c_void_p,
+                ctypes.c_uint,
+                ctypes.c_void_p,
+                ctypes.c_uint64,
+                ctypes.c_uint,
+            ]
+            self.core.cuda_import_external_memory.restype = ctypes.c_void_p
+            self.core.cuda_external_memory_get_mapped_buffer.argtypes = [
+                ctypes.c_void_p,
+                ctypes.c_void_p,
+                ctypes.c_uint64,
+                ctypes.c_uint64,
+                ctypes.c_uint,
+                ctypes.POINTER(ctypes.c_uint64),
+            ]
+            self.core.cuda_external_memory_get_mapped_buffer.restype = None
+            self.core.cuda_destroy_external_memory.argtypes = [ctypes.c_void_p, ctypes.c_void_p]
+            self.core.cuda_destroy_external_memory.restype = None
+
+            self.core.cuda_import_external_semaphore.argtypes = [
+                ctypes.c_void_p,
+                ctypes.c_uint,
+                ctypes.c_void_p,
+                ctypes.c_uint,
+            ]
+            self.core.cuda_import_external_semaphore.restype = ctypes.c_void_p
+            self.core.cuda_destroy_external_semaphore.argtypes = [ctypes.c_void_p, ctypes.c_void_p]
+            self.core.cuda_destroy_external_semaphore.restype = None
+            self.core.cuda_signal_external_semaphore_async.argtypes = [
+                ctypes.c_void_p,
+                ctypes.c_void_p,
+                ctypes.c_uint64,
+                ctypes.c_void_p,
+            ]
+            self.core.cuda_signal_external_semaphore_async.restype = None
+            self.core.cuda_wait_external_semaphore_async.argtypes = [
+                ctypes.c_void_p,
+                ctypes.c_void_p,
+                ctypes.c_uint64,
+                ctypes.c_void_p,
+            ]
+            self.core.cuda_wait_external_semaphore_async.restype = None
+
             self.core.cuda_timing_begin.argtypes = [ctypes.c_int]
             self.core.cuda_timing_begin.restype = None
             self.core.cuda_timing_get_result_count.argtypes = []
@@ -4732,6 +4790,52 @@ def wait_event(event: Event):
     get_stream().wait_event(event)
 
 
+class ExternalSemaphore:
+    """TODO: docs"""
+
+    HANDLE_TYPE_OPAQUEFD = 1
+    HANDLE_TYPE_OPAQUEWIN32 = 2
+    HANDLE_TYPE_OPAQUEWIN32KMT = 3
+    HANDLE_TYPE_D3D12HEAP = 4
+    HANDLE_TYPE_D3D12RESOURCE = 5
+    HANDLE_TYPE_D3D11RESOURCE = 6
+    HANDLE_TYPE_D3D11RESOURCEKMT = 7
+    HANDLE_TYPE_NVSCIBUF = 8
+
+    def __init__(
+        self, handle: Union[ctypes.c_void_p, int], handle_type: int, flags: int = 0, device: Devicelike = None
+    ):
+        """TODO: docs"""
+
+        self.device = get_device(device)
+        self.context = self.device.context
+        self.external_semaphore = runtime.core.cuda_import_external_semaphore(self.context, handle_type, handle, flags)
+        if self.external_semaphore is None:
+            raise RuntimeError(f"Failed to import external semaphore {handle} with CUDA")
+
+    def __del__(self):
+        """TODO: docs"""
+
+        if not self.external_semaphore:
+            return
+
+        # use CUDA context guard to avoid side effects during garbage collection
+        with self.device.context_guard:
+            runtime.core.cuda_destroy_external_semaphore(self.context, self.external_semaphore)
+
+
+def signal_external_semaphore(semaphore: ExternalSemaphore, value: int = 0):
+    """TODO: docs"""
+
+    return get_stream().signal_external_semaphore(semaphore, value)
+
+
+def wait_external_semaphore(semaphore: ExternalSemaphore, value: int = 0):
+    """TODO: docs"""
+
+    get_stream().wait_external_semaphore(semaphore, value)
+
+
 def get_event_elapsed_time(start_event: Event, end_event: Event, synchronize: bool = True):
     """Get the elapsed time between two recorded events.
 
@@ -4772,6 +4876,57 @@ def wait_stream(other_stream: Stream, event: Optional[Event] = None):
     get_stream().wait_stream(other_stream, event=event)
 
 
+class ExternalMemoryBuffer:
+    """TODO: docs"""
+
+    HANDLE_TYPE_OPAQUEFD = 1
+    HANDLE_TYPE_OPAQUEWIN32 = 2
+    HANDLE_TYPE_OPAQUEWIN32KMT = 3
+    HANDLE_TYPE_D3D12HEAP = 4
+    HANDLE_TYPE_D3D12RESOURCE = 5
+    HANDLE_TYPE_D3D11RESOURCE = 6
+    HANDLE_TYPE_D3D11RESOURCEKMT = 7
+    HANDLE_TYPE_NVSCIBUF = 8
+
+    FLAG_DEDICATED = 1
+
+    def __init__(
+        self,
+        handle: Union[ctypes.c_void_p, int],
+        handle_type: int,
+        size: int,
+        flags: int = 0,
+        device: Devicelike = None,
+    ):
+        """TODO: docs"""
+
+        self.device = get_device(device)
+        self.context = self.device.context
+        self.external_memory = runtime.core.cuda_import_external_memory(self.context, handle_type, handle, size, flags)
+        self.size = size
+        if self.external_memory is None:
+            raise RuntimeError(f"Failed to import external memory {handle} with CUDA")
+
+    def map(self, dtype: type, shape: Sequence[int]) -> warp.array:
+        """TODO: docs"""
+
+        ptr = ctypes.c_uint64(0)
+        runtime.core.cuda_external_memory_get_mapped_buffer(
+            self.context, self.external_memory, 0, self.size, 0, ctypes.byref(ptr)
+        )
+        return warp.array(ptr=ptr.value, dtype=dtype, shape=shape, device=self.device, capacity=self.size)
+
+    def __del__(self):
+        """TODO: docs"""
+
+        if not self.external_memory:
+            return
+
+        # use CUDA context guard to avoid side effects during garbage collection
+        with self.device.context_guard:
+            runtime.core.cuda_destroy_external_memory(self.context, self.external_memory)
+
+
 class RegisteredGLBuffer:
     """
     Helper class to register a GL buffer with CUDA so that it can be mapped to a Warp array.

warp/native/warp.cpp

@@ -1076,6 +1076,15 @@ WP_API void cuda_graphics_device_ptr_and_size(void* context, void* resource, uin
 WP_API void* cuda_graphics_register_gl_buffer(void* context, uint32_t gl_buffer, unsigned int flags) { return NULL; }
 WP_API void cuda_graphics_unregister_resource(void* context, void* resource) {}
 
+WP_API void* cuda_import_external_memory(void* context, unsigned int type, void* handle, uint64_t size, unsigned int flags) { return NULL}
+WP_API void cuda_destroy_external_memory(void* context, void* external_memory) {}
+WP_API void cuda_external_memory_get_mapped_buffer(void* context, void* external_memory, uint64_t offset, uint64_t size, unsigned int flags, uint64_t* ptr) {}
+
+WP_API void* cuda_import_external_semaphore(void* context, unsigned int type, void* handle, unsigned int flags) { return NULL }
+WP_API void cuda_destroy_external_semaphore(void* context, void* external_semaphore) {}
+WP_API void cuda_signal_external_semaphore_async(void* context, void* external_semaphore, uint64_t value, void* stream) {}
+WP_API void cuda_wait_external_semaphore_async(void* context, void* external_semaphore, uint64_t value, void* stream) {}
+
 WP_API void cuda_timing_begin(int flags) {}
 WP_API int cuda_timing_get_result_count() { return 0; }
 WP_API void cuda_timing_end(timing_result_t* results, int size) {}

warp/native/warp.cu

@@ -3584,6 +3584,100 @@ void cuda_graphics_unregister_resource(void* context, void* resource)
     delete res;
 }
 
+void* cuda_import_external_memory(void* context, unsigned int type, void* handle, uint64_t size, unsigned int flags)
+{
+    ContextGuard guard(context);
+
+    cudaExternalMemory_t *external_memory = new cudaExternalMemory_t;
+    cudaExternalMemoryHandleDesc desc = {};
+    desc.type = (cudaExternalMemoryHandleType)type;
+    desc.handle.win32.handle = handle;
+    desc.size = size;
+    desc.flags = flags;
+
+    bool success = check_cuda(cudaImportExternalMemory(external_memory, &desc));
+    if (!success)
+    {
+        delete external_memory;
+        return NULL;
+    }
+
+    return external_memory;
+}
+
+void cuda_external_memory_get_mapped_buffer(void* context, void* external_memory, uint64_t offset, uint64_t size, unsigned int flags, uint64_t* ptr)
+{
+    ContextGuard guard(context);
+
+    cudaExternalMemory_t *memory = (cudaExternalMemory_t*)external_memory;
+    cudaExternalMemoryBufferDesc desc = {};
+    desc.offset = offset;
+    desc.size = size;
+    desc.flags = flags;
+
+    void* device_ptr;
+    check_cuda(cudaExternalMemoryGetMappedBuffer(&device_ptr, *memory, &desc));
+    *ptr = (uint64_t)device_ptr;
+}
+
+void cuda_destroy_external_memory(void* context, void* external_memory)
+{
+    ContextGuard guard(context);
+
+    cudaExternalMemory_t *memory = (cudaExternalMemory_t*)external_memory;
+    check_cuda(cudaDestroyExternalMemory(*memory));
+    delete memory;
+}
+
+void* cuda_import_external_semaphore(void* context, unsigned int type, void* handle, unsigned int flags)
+{
+    ContextGuard guard(context);
+
+    cudaExternalSemaphore_t *external_semaphore = new cudaExternalSemaphore_t;
+    cudaExternalSemaphoreHandleDesc desc = {};
+    desc.type = (cudaExternalSemaphoreHandleType)type;
+    desc.handle.win32.handle = handle;
+    desc.flags = flags;
+
+    bool success = check_cuda(cudaImportExternalSemaphore(external_semaphore, &desc));
+    if (!success)
+    {
+        delete external_semaphore;
+        return NULL;
+    }
+
+    return external_semaphore;
+}
+
+void cuda_destroy_external_semaphore(void* context, void* external_semaphore)
+{
+    ContextGuard guard(context);
+
+    cudaExternalSemaphore_t *semaphore = (cudaExternalSemaphore_t*)external_semaphore;
+    check_cuda(cudaDestroyExternalSemaphore(*semaphore));
+    delete semaphore;
+}
+
+void cuda_signal_external_semaphore_async(void* context, void* external_semaphore, uint64_t value, void* stream)
+{
+    ContextGuard guard(context);
+
+    cudaExternalSemaphore_t *semaphore = (cudaExternalSemaphore_t*)external_semaphore;
+    cudaExternalSemaphoreSignalParams params = {};
+    params.params.fence.value = value;
+    check_cuda(cudaSignalExternalSemaphoresAsync(semaphore, &params, 1, static_cast<CUstream>(stream)));
+}
+
+void cuda_wait_external_semaphore_async(void* context, void* external_semaphore, uint64_t value, void* stream)
+{
+    ContextGuard guard(context);
+
+    cudaExternalSemaphore_t *semaphore = (cudaExternalSemaphore_t*)external_semaphore;
+    cudaExternalSemaphoreWaitParams params = {};
+    params.params.fence.value = value;
+    check_cuda(cudaWaitExternalSemaphoresAsync(semaphore, &params, 1, static_cast<CUstream>(stream)));
+}
+
 void cuda_timing_begin(int flags)
 {
     g_cuda_timing_state = new CudaTimingState(flags, g_cuda_timing_state);

warp/native/warp.h

@@ -370,6 +370,16 @@ extern "C"
     WP_API void* cuda_graphics_register_gl_buffer(void* context, uint32_t gl_buffer, unsigned int flags);
     WP_API void cuda_graphics_unregister_resource(void* context, void* resource);
 
+    // external resource interoperability
+    WP_API void* cuda_import_external_memory(void* context, unsigned int type, void* handle, uint64_t size, unsigned int flags);
+    WP_API void cuda_destroy_external_memory(void* context, void* external_memory);
+    WP_API void cuda_external_memory_get_mapped_buffer(void* context, void* external_memory, uint64_t offset, uint64_t size, unsigned int flags, uint64_t* ptr);
+
+    WP_API void* cuda_import_external_semaphore(void* context, unsigned int type, void* handle, unsigned int flags);
+    WP_API void cuda_destroy_external_semaphore(void* context, void* external_semaphore);
+    WP_API void cuda_signal_external_semaphore_async(void* context, void* external_semaphore, uint64_t value, void* stream);
+    WP_API void cuda_wait_external_semaphore_async(void* context, void* external_semaphore, uint64_t value, void* stream);
+
     // CUDA timing
     WP_API void cuda_timing_begin(int flags);
     WP_API int cuda_timing_get_result_count();