Skip to content

Commit b9f40f6

Browse files
vzhurba01vzhurba01
committed
Rebase to CTK 12.6
1 parent 6923211 commit b9f40f6

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

63 files changed

+1852
-542
lines changed

‎README.md‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -17,7 +17,7 @@ There're differences in each of these options that are described further in [Ins
1717
CUDA Python is supported on all platforms that CUDA is supported. Specific dependencies are as follows:
1818

1919
* Driver: Linux (450.80.02 or later) Windows (456.38 or later)
20-
* CUDA Toolkit 12.0 to 12.5
20+
* CUDA Toolkit 12.0 to 12.6
2121

2222
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.
2323

‎cuda/_lib/ccudart/utils.pyx.in‎

Lines changed: 45 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -555,6 +555,51 @@ cdef cudaError_t initDevice(cudaPythonDevice *device, int deviceOrdinal) except
555555
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
556556
return cudaErrorInitializationError
557557

558+
err = ccuda._cuDeviceGetAttribute(&(device[0].deviceProperties.hostRegisterSupported), ccuda.CU_DEVICE_ATTRIBUTE_HOST_REGISTER_SUPPORTED, <ccuda.CUdevice>(deviceOrdinal))
559+
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
560+
return cudaErrorInitializationError
561+
err = ccuda._cuDeviceGetAttribute(&(device[0].deviceProperties.sparseCudaArraySupported), ccuda.CU_DEVICE_ATTRIBUTE_SPARSE_CUDA_ARRAY_SUPPORTED, <ccuda.CUdevice>(deviceOrdinal))
562+
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
563+
return cudaErrorInitializationError
564+
err = ccuda._cuDeviceGetAttribute(&(device[0].deviceProperties.hostRegisterReadOnlySupported), ccuda.CU_DEVICE_ATTRIBUTE_READ_ONLY_HOST_REGISTER_SUPPORTED, <ccuda.CUdevice>(deviceOrdinal))
565+
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
566+
return cudaErrorInitializationError
567+
err = ccuda._cuDeviceGetAttribute(&(device[0].deviceProperties.timelineSemaphoreInteropSupported), ccuda.CU_DEVICE_ATTRIBUTE_TIMELINE_SEMAPHORE_INTEROP_SUPPORTED, <ccuda.CUdevice>(deviceOrdinal))
568+
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
569+
return cudaErrorInitializationError
570+
err = ccuda._cuDeviceGetAttribute(&(device[0].deviceProperties.memoryPoolsSupported), ccuda.CU_DEVICE_ATTRIBUTE_MEMORY_POOLS_SUPPORTED, <ccuda.CUdevice>(deviceOrdinal))
571+
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
572+
return cudaErrorInitializationError
573+
err = ccuda._cuDeviceGetAttribute(&(device[0].deviceProperties.gpuDirectRDMASupported), ccuda.CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_SUPPORTED, <ccuda.CUdevice>(deviceOrdinal))
574+
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
575+
return cudaErrorInitializationError
576+
577+
cdef int gpuDirectRDMAFlushWritesOptions
578+
err = ccuda._cuDeviceGetAttribute(&gpuDirectRDMAFlushWritesOptions, ccuda.CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_FLUSH_WRITES_OPTIONS, <ccuda.CUdevice>(deviceOrdinal))
579+
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
580+
return cudaErrorInitializationError
581+
device[0].deviceProperties.gpuDirectRDMAFlushWritesOptions = gpuDirectRDMAFlushWritesOptions
582+
583+
err = ccuda._cuDeviceGetAttribute(&(device[0].deviceProperties.gpuDirectRDMAWritesOrdering), ccuda.CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WRITES_ORDERING, <ccuda.CUdevice>(deviceOrdinal))
584+
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
585+
return cudaErrorInitializationError
586+
587+
cdef int memoryPoolSupportedHandleTypes
588+
err = ccuda._cuDeviceGetAttribute(&memoryPoolSupportedHandleTypes, ccuda.CU_DEVICE_ATTRIBUTE_MEMPOOL_SUPPORTED_HANDLE_TYPES, <ccuda.CUdevice>(deviceOrdinal))
589+
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
590+
return cudaErrorInitializationError
591+
device[0].deviceProperties.memoryPoolSupportedHandleTypes = memoryPoolSupportedHandleTypes;
592+
593+
err = ccuda._cuDeviceGetAttribute(&(device[0].deviceProperties.deferredMappingCudaArraySupported), ccuda.CU_DEVICE_ATTRIBUTE_DEFERRED_MAPPING_CUDA_ARRAY_SUPPORTED, <ccuda.CUdevice>(deviceOrdinal))
594+
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
595+
return cudaErrorInitializationError
596+
err = ccuda._cuDeviceGetAttribute(&(device[0].deviceProperties.ipcEventSupported), ccuda.CU_DEVICE_ATTRIBUTE_IPC_EVENT_SUPPORTED, <ccuda.CUdevice>(deviceOrdinal))
597+
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
598+
return cudaErrorInitializationError
599+
err = ccuda._cuDeviceGetAttribute(&(device[0].deviceProperties.clusterLaunch), ccuda.CU_DEVICE_ATTRIBUTE_CLUSTER_LAUNCH, <ccuda.CUdevice>(deviceOrdinal))
600+
if err != ccuda.cudaError_enum.CUDA_SUCCESS:
601+
return cudaErrorInitializationError
602+
558603
cdef int reservedSharedMemPerBlock
559604
err = ccuda._cuDeviceGetAttribute(&reservedSharedMemPerBlock, ccuda.CU_DEVICE_ATTRIBUTE_RESERVED_SHARED_MEMORY_PER_BLOCK, <ccuda.CUdevice>(deviceOrdinal))
560605
if err != ccuda.cudaError_enum.CUDA_SUCCESS:

‎cuda/ccuda.pxd.in‎

Lines changed: 5 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -1995,7 +1995,8 @@ cdef extern from "cuda.h":
19951995
CUmemLocation location
19961996
void* win32SecurityAttributes
19971997
size_t maxSize
1998-
unsigned char reserved[56]
1998+
unsigned short usage
1999+
unsigned char reserved[54]
19992000

20002001
ctypedef CUmemPoolProps_st CUmemPoolProps_v1
20012002

@@ -2176,8 +2177,9 @@ cdef extern from "cuda.h":
21762177
CU_COREDUMP_SKIP_GLOBAL_MEMORY = 2
21772178
CU_COREDUMP_SKIP_SHARED_MEMORY = 4
21782179
CU_COREDUMP_SKIP_LOCAL_MEMORY = 8
2179-
CU_COREDUMP_LIGHTWEIGHT_FLAGS = 15
21802180
CU_COREDUMP_SKIP_ABORT = 16
2181+
CU_COREDUMP_SKIP_CONSTBANK_MEMORY = 32
2182+
CU_COREDUMP_LIGHTWEIGHT_FLAGS = 47
21812183

21822184
cdef struct CUdevResourceDesc_st:
21832185
pass
@@ -4696,7 +4698,7 @@ cdef CUresult cuGraphicsVDPAURegisterVideoSurface(CUgraphicsResource* pCudaResou
46964698
cdef CUresult cuGraphicsVDPAURegisterOutputSurface(CUgraphicsResource* pCudaResource, VdpOutputSurface vdpSurface, unsigned int flags) except ?CUDA_ERROR_NOT_FOUND nogil
46974699
{{endif}}
46984700

4699-
cdef enum: CUDA_VERSION = 12050
4701+
cdef enum: CUDA_VERSION = 12060
47004702

47014703
cdef enum: CU_IPC_HANDLE_SIZE = 64
47024704

‎cuda/ccudart.pxd.in‎

Lines changed: 7 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -145,6 +145,9 @@ cdef extern from "driver_types.h":
145145
cudaErrorGraphExecUpdateFailure = 910
146146
cudaErrorExternalDevice = 911
147147
cudaErrorInvalidClusterSize = 912
148+
cudaErrorFunctionNotLoaded = 913
149+
cudaErrorInvalidResourceType = 914
150+
cudaErrorInvalidResourceConfiguration = 915
148151
cudaErrorUnknown = 999
149152
cudaErrorApiFailureBase = 10000
150153

@@ -342,7 +345,8 @@ cdef extern from "driver_types.h":
342345
cudaMemLocation location
343346
void* win32SecurityAttributes
344347
size_t maxSize
345-
unsigned char reserved[56]
348+
unsigned short usage
349+
unsigned char reserved[54]
346350

347351
cdef struct cudaMemPoolPtrExportData:
348352
unsigned char reserved[64]
@@ -3301,8 +3305,8 @@ cdef enum: cudaTextureType2DLayered = 242
33013305

33023306
cdef enum: cudaTextureTypeCubemapLayered = 252
33033307

3304-
cdef enum: CUDART_VERSION = 12050
3308+
cdef enum: CUDART_VERSION = 12060
33053309

3306-
cdef enum: __CUDART_API_VERSION = 12050
3310+
cdef enum: __CUDART_API_VERSION = 12060
33073311

33083312
cdef enum: CUDA_EGL_MAX_PLANES = 3

‎cuda/ccudart.pyx.in‎

Lines changed: 9 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -556,6 +556,15 @@ cdef const char* cudaGetErrorName(cudaError_t error) except ?NULL nogil:
556556
{{if 'cudaErrorInvalidClusterSize' in found_values}}
557557
if error == cudaErrorInvalidClusterSize:
558558
return "cudaErrorInvalidClusterSize"{{endif}}
559+
{{if 'cudaErrorFunctionNotLoaded' in found_values}}
560+
if error == cudaErrorFunctionNotLoaded:
561+
return "cudaErrorFunctionNotLoaded"{{endif}}
562+
{{if 'cudaErrorInvalidResourceType' in found_values}}
563+
if error == cudaErrorInvalidResourceType:
564+
return "cudaErrorInvalidResourceType"{{endif}}
565+
{{if 'cudaErrorInvalidResourceConfiguration' in found_values}}
566+
if error == cudaErrorInvalidResourceConfiguration:
567+
return "cudaErrorInvalidResourceConfiguration"{{endif}}
559568
{{if 'cudaErrorUnknown' in found_values}}
560569
if error == cudaErrorUnknown:
561570
return "cudaErrorUnknown"{{endif}}

‎cuda/cuda.pxd.in‎

Lines changed: 96 additions & 18 deletions
Original file line numberDiff line numberDiff line change
@@ -1457,10 +1457,18 @@ cdef class CUlaunchAttributeValue_union:
14571457
Value of launch attribute
14581458
CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION.
14591459
programmaticEvent : anon_struct2
1460-
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_EVENT.
1460+
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_EVENT
1461+
with the following fields: - `CUevent` event - Event to fire when
1462+
all blocks trigger it. - `Event` record flags, see
1463+
cuEventRecordWithFlags. Does not accept :CU_EVENT_RECORD_EXTERNAL.
1464+
- `triggerAtBlockStart` - If this is set to non-0, each block
1465+
launch will automatically trigger the event.
14611466
launchCompletionEvent : anon_struct3
14621467
Value of launch attribute
1463-
CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT.
1468+
CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT with the following
1469+
fields: - `CUevent` event - Event to fire when the last block
1470+
launches - `int` flags; - Event record flags, see
1471+
cuEventRecordWithFlags. Does not accept CU_EVENT_RECORD_EXTERNAL.
14641472
priority : int
14651473
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PRIORITY. Execution
14661474
priority of the kernel.
@@ -1472,7 +1480,11 @@ cdef class CUlaunchAttributeValue_union:
14721480
See::CUlaunchMemSyncDomain
14731481
deviceUpdatableKernelNode : anon_struct4
14741482
Value of launch attribute
1475-
CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE.
1483+
CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE. with the
1484+
following fields: - `int` deviceUpdatable - Whether or not the
1485+
resulting kernel node should be device-updatable. -
1486+
`CUgraphDeviceNode` devNode - Returns a handle to pass to the
1487+
various device-side update functions.
14761488
sharedMemCarveout : unsigned int
14771489
Value of launch attribute
14781490
CU_LAUNCH_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT.
@@ -3255,6 +3267,8 @@ cdef class CUmemPoolProps_st:
32553267
maxSize : size_t
32563268
Maximum pool size. When set to 0, defaults to a system dependent
32573269
value.
3270+
usage : unsigned short
3271+
Bitmask indicating intended usage for the pool.
32583272
reserved : bytes
32593273
reserved for future use, must be 0
32603274

@@ -4582,10 +4596,18 @@ cdef class CUlaunchAttributeValue(CUlaunchAttributeValue_union):
45824596
Value of launch attribute
45834597
CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION.
45844598
programmaticEvent : anon_struct2
4585-
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_EVENT.
4599+
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_EVENT
4600+
with the following fields: - `CUevent` event - Event to fire when
4601+
all blocks trigger it. - `Event` record flags, see
4602+
cuEventRecordWithFlags. Does not accept :CU_EVENT_RECORD_EXTERNAL.
4603+
- `triggerAtBlockStart` - If this is set to non-0, each block
4604+
launch will automatically trigger the event.
45864605
launchCompletionEvent : anon_struct3
45874606
Value of launch attribute
4588-
CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT.
4607+
CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT with the following
4608+
fields: - `CUevent` event - Event to fire when the last block
4609+
launches - `int` flags; - Event record flags, see
4610+
cuEventRecordWithFlags. Does not accept CU_EVENT_RECORD_EXTERNAL.
45894611
priority : int
45904612
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PRIORITY. Execution
45914613
priority of the kernel.
@@ -4597,7 +4619,11 @@ cdef class CUlaunchAttributeValue(CUlaunchAttributeValue_union):
45974619
See::CUlaunchMemSyncDomain
45984620
deviceUpdatableKernelNode : anon_struct4
45994621
Value of launch attribute
4600-
CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE.
4622+
CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE. with the
4623+
following fields: - `int` deviceUpdatable - Whether or not the
4624+
resulting kernel node should be device-updatable. -
4625+
`CUgraphDeviceNode` devNode - Returns a handle to pass to the
4626+
various device-side update functions.
46014627
sharedMemCarveout : unsigned int
46024628
Value of launch attribute
46034629
CU_LAUNCH_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT.
@@ -4703,10 +4729,18 @@ cdef class CUkernelNodeAttrValue_v1(CUlaunchAttributeValue):
47034729
Value of launch attribute
47044730
CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION.
47054731
programmaticEvent : anon_struct2
4706-
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_EVENT.
4732+
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_EVENT
4733+
with the following fields: - `CUevent` event - Event to fire when
4734+
all blocks trigger it. - `Event` record flags, see
4735+
cuEventRecordWithFlags. Does not accept :CU_EVENT_RECORD_EXTERNAL.
4736+
- `triggerAtBlockStart` - If this is set to non-0, each block
4737+
launch will automatically trigger the event.
47074738
launchCompletionEvent : anon_struct3
47084739
Value of launch attribute
4709-
CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT.
4740+
CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT with the following
4741+
fields: - `CUevent` event - Event to fire when the last block
4742+
launches - `int` flags; - Event record flags, see
4743+
cuEventRecordWithFlags. Does not accept CU_EVENT_RECORD_EXTERNAL.
47104744
priority : int
47114745
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PRIORITY. Execution
47124746
priority of the kernel.
@@ -4718,7 +4752,11 @@ cdef class CUkernelNodeAttrValue_v1(CUlaunchAttributeValue):
47184752
See::CUlaunchMemSyncDomain
47194753
deviceUpdatableKernelNode : anon_struct4
47204754
Value of launch attribute
4721-
CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE.
4755+
CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE. with the
4756+
following fields: - `int` deviceUpdatable - Whether or not the
4757+
resulting kernel node should be device-updatable. -
4758+
`CUgraphDeviceNode` devNode - Returns a handle to pass to the
4759+
various device-side update functions.
47224760
sharedMemCarveout : unsigned int
47234761
Value of launch attribute
47244762
CU_LAUNCH_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT.
@@ -4766,10 +4804,18 @@ cdef class CUkernelNodeAttrValue(CUkernelNodeAttrValue_v1):
47664804
Value of launch attribute
47674805
CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION.
47684806
programmaticEvent : anon_struct2
4769-
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_EVENT.
4807+
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_EVENT
4808+
with the following fields: - `CUevent` event - Event to fire when
4809+
all blocks trigger it. - `Event` record flags, see
4810+
cuEventRecordWithFlags. Does not accept :CU_EVENT_RECORD_EXTERNAL.
4811+
- `triggerAtBlockStart` - If this is set to non-0, each block
4812+
launch will automatically trigger the event.
47704813
launchCompletionEvent : anon_struct3
47714814
Value of launch attribute
4772-
CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT.
4815+
CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT with the following
4816+
fields: - `CUevent` event - Event to fire when the last block
4817+
launches - `int` flags; - Event record flags, see
4818+
cuEventRecordWithFlags. Does not accept CU_EVENT_RECORD_EXTERNAL.
47734819
priority : int
47744820
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PRIORITY. Execution
47754821
priority of the kernel.
@@ -4781,7 +4827,11 @@ cdef class CUkernelNodeAttrValue(CUkernelNodeAttrValue_v1):
47814827
See::CUlaunchMemSyncDomain
47824828
deviceUpdatableKernelNode : anon_struct4
47834829
Value of launch attribute
4784-
CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE.
4830+
CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE. with the
4831+
following fields: - `int` deviceUpdatable - Whether or not the
4832+
resulting kernel node should be device-updatable. -
4833+
`CUgraphDeviceNode` devNode - Returns a handle to pass to the
4834+
various device-side update functions.
47854835
sharedMemCarveout : unsigned int
47864836
Value of launch attribute
47874837
CU_LAUNCH_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT.
@@ -4829,10 +4879,18 @@ cdef class CUstreamAttrValue_v1(CUlaunchAttributeValue):
48294879
Value of launch attribute
48304880
CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION.
48314881
programmaticEvent : anon_struct2
4832-
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_EVENT.
4882+
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_EVENT
4883+
with the following fields: - `CUevent` event - Event to fire when
4884+
all blocks trigger it. - `Event` record flags, see
4885+
cuEventRecordWithFlags. Does not accept :CU_EVENT_RECORD_EXTERNAL.
4886+
- `triggerAtBlockStart` - If this is set to non-0, each block
4887+
launch will automatically trigger the event.
48334888
launchCompletionEvent : anon_struct3
48344889
Value of launch attribute
4835-
CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT.
4890+
CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT with the following
4891+
fields: - `CUevent` event - Event to fire when the last block
4892+
launches - `int` flags; - Event record flags, see
4893+
cuEventRecordWithFlags. Does not accept CU_EVENT_RECORD_EXTERNAL.
48364894
priority : int
48374895
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PRIORITY. Execution
48384896
priority of the kernel.
@@ -4844,7 +4902,11 @@ cdef class CUstreamAttrValue_v1(CUlaunchAttributeValue):
48444902
See::CUlaunchMemSyncDomain
48454903
deviceUpdatableKernelNode : anon_struct4
48464904
Value of launch attribute
4847-
CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE.
4905+
CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE. with the
4906+
following fields: - `int` deviceUpdatable - Whether or not the
4907+
resulting kernel node should be device-updatable. -
4908+
`CUgraphDeviceNode` devNode - Returns a handle to pass to the
4909+
various device-side update functions.
48484910
sharedMemCarveout : unsigned int
48494911
Value of launch attribute
48504912
CU_LAUNCH_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT.
@@ -4892,10 +4954,18 @@ cdef class CUstreamAttrValue(CUstreamAttrValue_v1):
48924954
Value of launch attribute
48934955
CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION.
48944956
programmaticEvent : anon_struct2
4895-
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_EVENT.
4957+
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_EVENT
4958+
with the following fields: - `CUevent` event - Event to fire when
4959+
all blocks trigger it. - `Event` record flags, see
4960+
cuEventRecordWithFlags. Does not accept :CU_EVENT_RECORD_EXTERNAL.
4961+
- `triggerAtBlockStart` - If this is set to non-0, each block
4962+
launch will automatically trigger the event.
48964963
launchCompletionEvent : anon_struct3
48974964
Value of launch attribute
4898-
CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT.
4965+
CU_LAUNCH_ATTRIBUTE_LAUNCH_COMPLETION_EVENT with the following
4966+
fields: - `CUevent` event - Event to fire when the last block
4967+
launches - `int` flags; - Event record flags, see
4968+
cuEventRecordWithFlags. Does not accept CU_EVENT_RECORD_EXTERNAL.
48994969
priority : int
49004970
Value of launch attribute CU_LAUNCH_ATTRIBUTE_PRIORITY. Execution
49014971
priority of the kernel.
@@ -4907,7 +4977,11 @@ cdef class CUstreamAttrValue(CUstreamAttrValue_v1):
49074977
See::CUlaunchMemSyncDomain
49084978
deviceUpdatableKernelNode : anon_struct4
49094979
Value of launch attribute
4910-
CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE.
4980+
CU_LAUNCH_ATTRIBUTE_DEVICE_UPDATABLE_KERNEL_NODE. with the
4981+
following fields: - `int` deviceUpdatable - Whether or not the
4982+
resulting kernel node should be device-updatable. -
4983+
`CUgraphDeviceNode` devNode - Returns a handle to pass to the
4984+
various device-side update functions.
49114985
sharedMemCarveout : unsigned int
49124986
Value of launch attribute
49134987
CU_LAUNCH_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT.
@@ -6843,6 +6917,8 @@ cdef class CUmemPoolProps_v1(CUmemPoolProps_st):
68436917
maxSize : size_t
68446918
Maximum pool size. When set to 0, defaults to a system dependent
68456919
value.
6920+
usage : unsigned short
6921+
Bitmask indicating intended usage for the pool.
68466922
reserved : bytes
68476923
reserved for future use, must be 0
68486924

@@ -6878,6 +6954,8 @@ cdef class CUmemPoolProps(CUmemPoolProps_v1):
68786954
maxSize : size_t
68796955
Maximum pool size. When set to 0, defaults to a system dependent
68806956
value.
6957+
usage : unsigned short
6958+
Bitmask indicating intended usage for the pool.
68816959
reserved : bytes
68826960
reserved for future use, must be 0
68836961

0 commit comments

Comments
 (0)