Rebase to CTK 12.5

README.md

@@ -17,7 +17,7 @@ There're differences in each of these options that are described further in [Ins
 CUDA Python is supported on all platforms that CUDA is supported. Specific dependencies are as follows:
 
 * Driver: Linux (450.80.02 or later) Windows (456.38 or later)
-* CUDA Toolkit 12.0 to 12.4
+* CUDA Toolkit 12.0 to 12.5
 
 Only the NVRTC redistributable component is required from the CUDA Toolkit. [CUDA Toolkit Documentation](https://docs.nvidia.com/cuda/index.html) Installation Guides can be used for guidance. Note that the NVRTC component in the Toolkit can be obtained via PYPI, Conda or Local Installer.
 

cuda/_cuda/ccuda.pxd.in

@@ -147,6 +147,11 @@ cdef CUresult _cuCtxCreate_v2(CUcontext* pctx, unsigned int flags, CUdevice dev)
 cdef CUresult _cuCtxCreate_v3(CUcontext* pctx, CUexecAffinityParam* paramsArray, int numParams, unsigned int flags, CUdevice dev) except ?CUDA_ERROR_NOT_FOUND nogil
 {{endif}}
 
+{{if 'cuCtxCreate_v4' in found_functions}}
+
+cdef CUresult _cuCtxCreate_v4(CUcontext* pctx, CUctxCreateParams* ctxCreateParams, unsigned int flags, CUdevice dev) except ?CUDA_ERROR_NOT_FOUND nogil
+{{endif}}
+
 {{if 'cuCtxDestroy_v2' in found_functions}}
 
 cdef CUresult _cuCtxDestroy_v2(CUcontext ctx) except ?CUDA_ERROR_NOT_FOUND nogil
@@ -237,6 +242,16 @@ cdef CUresult _cuCtxResetPersistingL2Cache() except ?CUDA_ERROR_NOT_FOUND nogil
 cdef CUresult _cuCtxGetExecAffinity(CUexecAffinityParam* pExecAffinity, CUexecAffinityType typename) except ?CUDA_ERROR_NOT_FOUND nogil
 {{endif}}
 
+{{if 'cuCtxRecordEvent' in found_functions}}
+
+cdef CUresult _cuCtxRecordEvent(CUcontext hCtx, CUevent hEvent) except ?CUDA_ERROR_NOT_FOUND nogil
+{{endif}}
+
+{{if 'cuCtxWaitEvent' in found_functions}}
+
+cdef CUresult _cuCtxWaitEvent(CUcontext hCtx, CUevent hEvent) except ?CUDA_ERROR_NOT_FOUND nogil
+{{endif}}
+
 {{if 'cuCtxAttach' in found_functions}}
 
 cdef CUresult _cuCtxAttach(CUcontext* pctx, unsigned int flags) except ?CUDA_ERROR_NOT_FOUND nogil
@@ -382,6 +397,11 @@ cdef CUresult _cuLibraryGetModule(CUmodule* pMod, CUlibrary library) except ?CUD
 cdef CUresult _cuKernelGetFunction(CUfunction* pFunc, CUkernel kernel) except ?CUDA_ERROR_NOT_FOUND nogil
 {{endif}}
 
+{{if 'cuKernelGetLibrary' in found_functions}}
+
+cdef CUresult _cuKernelGetLibrary(CUlibrary* pLib, CUkernel kernel) except ?CUDA_ERROR_NOT_FOUND nogil
+{{endif}}
+
 {{if 'cuLibraryGetGlobal' in found_functions}}
 
 cdef CUresult _cuLibraryGetGlobal(CUdeviceptr* dptr, size_t* numbytes, CUlibrary library, const char* name) except ?CUDA_ERROR_NOT_FOUND nogil
@@ -1027,6 +1047,11 @@ cdef CUresult _cuStreamGetId(CUstream hStream, unsigned long long* streamId) exc
 cdef CUresult _cuStreamGetCtx(CUstream hStream, CUcontext* pctx) except ?CUDA_ERROR_NOT_FOUND nogil
 {{endif}}
 
+{{if 'cuStreamGetCtx_v2' in found_functions}}
+
+cdef CUresult _cuStreamGetCtx_v2(CUstream hStream, CUcontext* pCtx, CUgreenCtx* pGreenCtx) except ?CUDA_ERROR_NOT_FOUND nogil
+{{endif}}
+
 {{if 'cuStreamWaitEvent' in found_functions}}
 
 cdef CUresult _cuStreamWaitEvent(CUstream hStream, CUevent hEvent, unsigned int Flags) except ?CUDA_ERROR_NOT_FOUND nogil
@@ -2152,6 +2177,11 @@ cdef CUresult _cuGreenCtxWaitEvent(CUgreenCtx hCtx, CUevent hEvent) except ?CUDA
 cdef CUresult _cuStreamGetGreenCtx(CUstream hStream, CUgreenCtx* phCtx) except ?CUDA_ERROR_NOT_FOUND nogil
 {{endif}}
 
+{{if 'cuGreenCtxStreamCreate' in found_functions}}
+
+cdef CUresult _cuGreenCtxStreamCreate(CUstream* phStream, CUgreenCtx greenCtx, unsigned int flags, int priority) except ?CUDA_ERROR_NOT_FOUND nogil
+{{endif}}
+
 {{if 'cuProfilerStart' in found_functions}}
 
 cdef CUresult _cuProfilerStart() except ?CUDA_ERROR_NOT_FOUND nogil

cuda/_lib/ccudart/ccudart.pxd.in

@@ -308,3 +308,4 @@ from libcpp cimport bool
 {{if True}}cdef cudaError_t _getLocalRuntimeVersion(int* runtimeVersion) except ?cudaErrorCallRequiresNewerDriver nogil{{endif}}
 {{if 'cudaDeviceRegisterAsyncNotification' in found_functions}}cdef cudaError_t _cudaDeviceRegisterAsyncNotification(int device, cudaAsyncCallback callbackFunc, void* userData, cudaAsyncCallbackHandle_t* callback) except ?cudaErrorCallRequiresNewerDriver nogil{{endif}}
 {{if 'cudaDeviceUnregisterAsyncNotification' in found_functions}}cdef cudaError_t _cudaDeviceUnregisterAsyncNotification(int device, cudaAsyncCallbackHandle_t callback) except ?cudaErrorCallRequiresNewerDriver nogil{{endif}}
+{{if 'cudaGetDriverEntryPointByVersion' in found_functions}}cdef cudaError_t _cudaGetDriverEntryPointByVersion(const char* symbol, void** funcPtr, unsigned int cudaVersion, unsigned long long flags, cudaDriverEntryPointQueryResult* driverStatus) except ?cudaErrorCallRequiresNewerDriver nogil{{endif}}

cuda/_lib/ccudart/ccudart.pyx.in

@@ -4916,4 +4916,15 @@ cdef cudaError_t _cudaDeviceUnregisterAsyncNotification(int device, cudaAsyncCal
     if err != cudaSuccess:
         _setLastError(err)
     return err
+
+{{endif}}
+{{if 'cudaGetDriverEntryPointByVersion' in found_functions}}
+
+cdef cudaError_t _cudaGetDriverEntryPointByVersion(const char* symbol, void** funcPtr, unsigned int cudaVersion, unsigned long long flags, cudaDriverEntryPointQueryResult* driverStatus) except ?cudaErrorCallRequiresNewerDriver nogil:
+    cdef cudaError_t err = cudaSuccess
+    err = <cudaError_t>ccuda._cuGetProcAddress_v2(symbol, funcPtr, cudaVersion, flags, <ccuda.CUdriverProcAddressQueryResult*>driverStatus)
+    if err != cudaSuccess:
+        _setLastError(err)
+    return err
+
 {{endif}}

cuda/_lib/utils.pyx.in

@@ -142,6 +142,8 @@ cdef class HelperInputVoidPtr:
         elif isinstance(ptr, (int)):
             # Easy run, user gave us an already configured void** address
             self._cptr = <void*><void_ptr>ptr
+        elif isinstance(ptr, (cuda.CUdeviceptr)):
+            self._cptr = <void*><void_ptr>int(ptr)
         elif PyObject_CheckBuffer(ptr):
             # Easy run, get address from Python Buffer Protocol
             err_buffer = PyObject_GetBuffer(ptr, &self._pybuffer, PyBUF_SIMPLE | PyBUF_ANY_CONTIGUOUS)

cuda/ccuda.pxd.in

@@ -116,6 +116,10 @@ cdef extern from "cuda.h":
         pass
     ctypedef CUasyncCallbackEntry_st* CUasyncCallbackHandle
 
+    cdef struct CUgreenCtx_st:
+        pass
+    ctypedef CUgreenCtx_st* CUgreenCtx
+
     cdef struct CUuuid_st:
         char bytes[16]
 
@@ -359,7 +363,20 @@ cdef extern from "cuda.h":
         CU_AD_FORMAT_BC6H_SF16 = 156
         CU_AD_FORMAT_BC7_UNORM = 157
         CU_AD_FORMAT_BC7_UNORM_SRGB = 158
+        CU_AD_FORMAT_P010 = 159
+        CU_AD_FORMAT_P016 = 161
+        CU_AD_FORMAT_NV16 = 162
+        CU_AD_FORMAT_P210 = 163
+        CU_AD_FORMAT_P216 = 164
+        CU_AD_FORMAT_YUY2 = 165
+        CU_AD_FORMAT_Y210 = 166
+        CU_AD_FORMAT_Y216 = 167
+        CU_AD_FORMAT_AYUV = 168
+        CU_AD_FORMAT_Y410 = 169
         CU_AD_FORMAT_NV12 = 176
+        CU_AD_FORMAT_Y416 = 177
+        CU_AD_FORMAT_Y444_PLANAR8 = 178
+        CU_AD_FORMAT_Y444_PLANAR10 = 179
         CU_AD_FORMAT_UNORM_INT8X1 = 192
         CU_AD_FORMAT_UNORM_INT8X2 = 193
         CU_AD_FORMAT_UNORM_INT8X4 = 194
@@ -372,6 +389,7 @@ cdef extern from "cuda.h":
         CU_AD_FORMAT_SNORM_INT16X1 = 201
         CU_AD_FORMAT_SNORM_INT16X2 = 202
         CU_AD_FORMAT_SNORM_INT16X4 = 203
+        CU_AD_FORMAT_MAX = 2147483647
 
     ctypedef CUarray_format_enum CUarray_format
 
@@ -530,7 +548,8 @@ cdef extern from "cuda.h":
         CU_DEVICE_ATTRIBUTE_MULTICAST_SUPPORTED = 132
         CU_DEVICE_ATTRIBUTE_MPS_ENABLED = 133
         CU_DEVICE_ATTRIBUTE_HOST_NUMA_ID = 134
-        CU_DEVICE_ATTRIBUTE_MAX = 135
+        CU_DEVICE_ATTRIBUTE_D3D12_CIG_SUPPORTED = 135
+        CU_DEVICE_ATTRIBUTE_MAX = 136
 
     ctypedef CUdevice_attribute_enum CUdevice_attribute
 
@@ -778,7 +797,10 @@ cdef extern from "cuda.h":
         CU_LIMIT_DEV_RUNTIME_PENDING_LAUNCH_COUNT = 4
         CU_LIMIT_MAX_L2_FETCH_GRANULARITY = 5
         CU_LIMIT_PERSISTING_L2_CACHE_SIZE = 6
-        CU_LIMIT_MAX = 7
+        CU_LIMIT_SHMEM_SIZE = 7
+        CU_LIMIT_CIG_ENABLED = 8
+        CU_LIMIT_CIG_SHMEM_FALLBACK_ENABLED = 9
+        CU_LIMIT_MAX = 10
 
     ctypedef CUlimit_enum CUlimit
 
@@ -998,6 +1020,7 @@ cdef extern from "cuda.h":
         CU_LAUNCH_ATTRIBUTE_MEM_SYNC_DOMAIN = 10
         CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT = 12
         CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE = 13
+        CU_LAUNCH_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT = 14
 
     ctypedef CUlaunchAttributeID_enum CUlaunchAttributeID
 
@@ -1033,6 +1056,7 @@ cdef extern from "cuda.h":
         CUlaunchMemSyncDomainMap memSyncDomainMap
         CUlaunchMemSyncDomain memSyncDomain
         anon_struct4 deviceUpdatableKernelNode
+        unsigned int sharedMemCarveout
 
     ctypedef CUlaunchAttributeValue_union CUlaunchAttributeValue
 
@@ -1120,6 +1144,24 @@ cdef extern from "cuda.h":
 
     ctypedef CUexecAffinityParam_v1 CUexecAffinityParam
 
+    cdef enum CUcigDataType_enum:
+        CIG_DATA_TYPE_D3D12_COMMAND_QUEUE = 1
+
+    ctypedef CUcigDataType_enum CUcigDataType
+
+    cdef struct CUctxCigParam_st:
+        CUcigDataType sharedDataType
+        void* sharedData
+
+    ctypedef CUctxCigParam_st CUctxCigParam
+
+    cdef struct CUctxCreateParams_st:
+        CUexecAffinityParam* execAffinityParams
+        int numExecAffinityParams
+        CUctxCigParam* cigParams
+
+    ctypedef CUctxCreateParams_st CUctxCreateParams
+
     cdef enum CUlibraryOption_enum:
         CU_LIBRARY_HOST_UNIVERSAL_FUNCTION_AND_DATA_TABLE = 0
         CU_LIBRARY_BINARY_IS_PRESERVED = 1
@@ -2123,13 +2165,19 @@ cdef extern from "cuda.h":
         CU_COREDUMP_ENABLE_USER_TRIGGER = 4
         CU_COREDUMP_FILE = 5
         CU_COREDUMP_PIPE = 6
-        CU_COREDUMP_MAX = 7
+        CU_COREDUMP_GENERATION_FLAGS = 7
+        CU_COREDUMP_MAX = 8
 
     ctypedef CUcoredumpSettings_enum CUcoredumpSettings
 
-    cdef struct CUgreenCtx_st:
-        pass
-    ctypedef CUgreenCtx_st* CUgreenCtx
+    cdef enum CUCoredumpGenerationFlags:
+        CU_COREDUMP_DEFAULT_FLAGS = 0
+        CU_COREDUMP_SKIP_NONRELOCATED_ELF_IMAGES = 1
+        CU_COREDUMP_SKIP_GLOBAL_MEMORY = 2
+        CU_COREDUMP_SKIP_SHARED_MEMORY = 4
+        CU_COREDUMP_SKIP_LOCAL_MEMORY = 8
+        CU_COREDUMP_LIGHTWEIGHT_FLAGS = 15
+        CU_COREDUMP_SKIP_ABORT = 16
 
     cdef struct CUdevResourceDesc_st:
         pass
@@ -2138,6 +2186,10 @@ cdef extern from "cuda.h":
     ctypedef enum CUgreenCtxCreate_flags:
         CU_GREEN_CTX_DEFAULT_STREAM = 1
 
+    ctypedef enum CUdevSmResourceSplit_flags:
+        CU_DEV_SM_RESOURCE_SPLIT_IGNORE_SM_COSCHEDULING = 1
+        CU_DEV_SM_RESOURCE_SPLIT_MAX_POTENTIAL_CLUSTER_SIZE = 2
+
     ctypedef enum CUdevResourceType:
         CU_DEV_RESOURCE_TYPE_INVALID = 0
         CU_DEV_RESOURCE_TYPE_SM = 1
@@ -2504,6 +2556,11 @@ cdef CUresult cuCtxCreate(CUcontext* pctx, unsigned int flags, CUdevice dev) exc
 cdef CUresult cuCtxCreate_v3(CUcontext* pctx, CUexecAffinityParam* paramsArray, int numParams, unsigned int flags, CUdevice dev) except ?CUDA_ERROR_NOT_FOUND nogil
 {{endif}}
 
+{{if 'cuCtxCreate_v4' in found_functions}}
+
+cdef CUresult cuCtxCreate_v4(CUcontext* pctx, CUctxCreateParams* ctxCreateParams, unsigned int flags, CUdevice dev) except ?CUDA_ERROR_NOT_FOUND nogil
+{{endif}}
+
 {{if 'cuCtxDestroy_v2' in found_functions}}
 
 cdef CUresult cuCtxDestroy(CUcontext ctx) except ?CUDA_ERROR_NOT_FOUND nogil
@@ -2594,6 +2651,16 @@ cdef CUresult cuCtxResetPersistingL2Cache() except ?CUDA_ERROR_NOT_FOUND nogil
 cdef CUresult cuCtxGetExecAffinity(CUexecAffinityParam* pExecAffinity, CUexecAffinityType typename) except ?CUDA_ERROR_NOT_FOUND nogil
 {{endif}}
 
+{{if 'cuCtxRecordEvent' in found_functions}}
+
+cdef CUresult cuCtxRecordEvent(CUcontext hCtx, CUevent hEvent) except ?CUDA_ERROR_NOT_FOUND nogil
+{{endif}}
+
+{{if 'cuCtxWaitEvent' in found_functions}}
+
+cdef CUresult cuCtxWaitEvent(CUcontext hCtx, CUevent hEvent) except ?CUDA_ERROR_NOT_FOUND nogil
+{{endif}}
+
 {{if 'cuCtxAttach' in found_functions}}
 
 cdef CUresult cuCtxAttach(CUcontext* pctx, unsigned int flags) except ?CUDA_ERROR_NOT_FOUND nogil
@@ -2739,6 +2806,11 @@ cdef CUresult cuLibraryGetModule(CUmodule* pMod, CUlibrary library) except ?CUDA
 cdef CUresult cuKernelGetFunction(CUfunction* pFunc, CUkernel kernel) except ?CUDA_ERROR_NOT_FOUND nogil
 {{endif}}
 
+{{if 'cuKernelGetLibrary' in found_functions}}
+
+cdef CUresult cuKernelGetLibrary(CUlibrary* pLib, CUkernel kernel) except ?CUDA_ERROR_NOT_FOUND nogil
+{{endif}}
+
 {{if 'cuLibraryGetGlobal' in found_functions}}
 
 cdef CUresult cuLibraryGetGlobal(CUdeviceptr* dptr, size_t* numbytes, CUlibrary library, const char* name) except ?CUDA_ERROR_NOT_FOUND nogil
@@ -3384,6 +3456,11 @@ cdef CUresult cuStreamGetId(CUstream hStream, unsigned long long* streamId) exce
 cdef CUresult cuStreamGetCtx(CUstream hStream, CUcontext* pctx) except ?CUDA_ERROR_NOT_FOUND nogil
 {{endif}}
 
+{{if 'cuStreamGetCtx_v2' in found_functions}}
+
+cdef CUresult cuStreamGetCtx_v2(CUstream hStream, CUcontext* pCtx, CUgreenCtx* pGreenCtx) except ?CUDA_ERROR_NOT_FOUND nogil
+{{endif}}
+
 {{if 'cuStreamWaitEvent' in found_functions}}
 
 cdef CUresult cuStreamWaitEvent(CUstream hStream, CUevent hEvent, unsigned int Flags) except ?CUDA_ERROR_NOT_FOUND nogil
@@ -4509,6 +4586,11 @@ cdef CUresult cuGreenCtxWaitEvent(CUgreenCtx hCtx, CUevent hEvent) except ?CUDA_
 cdef CUresult cuStreamGetGreenCtx(CUstream hStream, CUgreenCtx* phCtx) except ?CUDA_ERROR_NOT_FOUND nogil
 {{endif}}
 
+{{if 'cuGreenCtxStreamCreate' in found_functions}}
+
+cdef CUresult cuGreenCtxStreamCreate(CUstream* phStream, CUgreenCtx greenCtx, unsigned int flags, int priority) except ?CUDA_ERROR_NOT_FOUND nogil
+{{endif}}
+
 {{if 'cuProfilerStart' in found_functions}}
 
 cdef CUresult cuProfilerStart() except ?CUDA_ERROR_NOT_FOUND nogil
@@ -4614,7 +4696,7 @@ cdef CUresult cuGraphicsVDPAURegisterVideoSurface(CUgraphicsResource* pCudaResou
 cdef CUresult cuGraphicsVDPAURegisterOutputSurface(CUgraphicsResource* pCudaResource, VdpOutputSurface vdpSurface, unsigned int flags) except ?CUDA_ERROR_NOT_FOUND nogil
 {{endif}}
 
-cdef enum: CUDA_VERSION = 12040
+cdef enum: CUDA_VERSION = 12050
 
 cdef enum: CU_IPC_HANDLE_SIZE = 64
 
@@ -4648,6 +4730,8 @@ cdef enum: CU_KERNEL_NODE_ATTRIBUTE_MEM_SYNC_DOMAIN = 10
 
 cdef enum: CU_KERNEL_NODE_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE = 13
 
+cdef enum: CU_KERNEL_NODE_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT = 14
+
 cdef enum: CU_STREAM_ATTRIBUTE_ACCESS_POLICY_WINDOW = 1
 
 cdef enum: CU_STREAM_ATTRIBUTE_SYNCHRONIZATION_POLICY = 3
@@ -4710,6 +4794,8 @@ cdef enum: CUDA_ARRAY3D_SPARSE = 64
 
 cdef enum: CUDA_ARRAY3D_DEFERRED_MAPPING = 128
 
+cdef enum: CUDA_ARRAY3D_VIDEO_ENCODE_DECODE = 256
+
 cdef enum: CU_TRSA_OVERRIDE_FORMAT = 1
 
 cdef enum: CU_TRSF_READ_AS_INTEGER = 1