Types¶
The following types are used by the NCCL library.
ncclResult_t¶
-
ncclResult_t¶ Return values for all NCCL functions. Possible values are :
-
ncclSuccess¶ (
0) Function succeeded.
-
ncclUnhandledCudaError¶ (
1) A call to a CUDA function failed.
-
ncclSystemError¶ (
2) A call to the system failed.
-
ncclInternalError¶ (
3) An internal check failed. This is either a bug in NCCL or due to memory corruption.
-
ncclInvalidArgument¶ (
4) One argument has an invalid value.
-
ncclInvalidUsage¶ (
5) The call to NCCL is incorrect. This is usually reflecting a programming error.
Whenever a function returns an error (not ncclSuccess), NCCL should print a more detailed message when the environment variable NCCL_DEBUG is set to “WARN”.
-
ncclDataType_t¶
-
ncclDataType_t¶ NCCL defines the following integral and floating data-types.
-
ncclInt8¶ Signed 8-bits integer
-
ncclChar¶ Signed 8-bits integer
-
ncclUint8¶ Unsigned 8-bits integer
-
ncclInt32¶ Signed 32-bits integer
-
ncclInt¶ Signed 32-bits integer
-
ncclUint32¶ Unsigned 32-bits integer
-
ncclInt64¶ Signed 64-bits integer
-
ncclUint64¶ Unsigned 64-bits integer
-
ncclFloat16¶ 16-bits floating point number (half precision)
-
ncclHalf¶ 16-bits floating point number (half precision)
-
ncclFloat32¶ 32-bits floating point number (single precision)
-
ncclFloat¶ 32-bits floating point number (single precision)
-
ncclFloat64¶ 64-bits floating point number (double precision)
-
ncclDouble¶ 64-bits floating point number (double precision)
-
ncclBfloat16¶ 16-bits floating point number (truncated precision in bfloat16 format, CUDA 11 or later)
-
ncclRedOp_t¶
ncclScalarResidence_t¶
-
ncclScalarResidence_t¶ Indicates where (memory space) scalar arguments reside and when they can be dereferenced.
-
ncclScalarHostImmediate¶ The scalar resides in host memory and should be derefenced in the most immediate way.
-
ncclScalarDevice¶ The scalar resides on device visible memory and should be dereferenced once needed.
-