cuBLASConfiguration of atomic mode in

2.4.20. cublasSetAtomicsMode()

cublasStatus_t cublasSetAtomicsMode(cublasHandlet handle, cublasAtomicsMode_t mode)

一些方法,如 cublas<t>symv 和 cublas<t>hemv Has an alternative implementation that uses atoms to accumulate results. This implementation is usually much faster,But the results that can be generated from one run to another are not exactly the same. 从数学上讲,These different results were not significant,But biases can arise when debugging these differences.

This function allows or disallows cuBLAS The library uses atomics for all routines with alternate implementations. 如果没有在任何 cuBLAS The routine's documentation is explicitly specified,It means that the routine does not use an alternative implementation of atoms. When atomic mode is disabled,When called with the same parameters on the same hardware,每个 cuBLAS The routine should produce the same results from one run to another.

默认初始化的 cublasHandle_t The default atomic mode of an object is CUBLAS_ATOMICS_NOT_ALLOWED. 有关详细信息,See the section on types.

2.4.21. cublasGetAtomicsMode()

cublasStatus_t cublasGetAtomicsMode(cublasHandle_t handle, cublasAtomicsMode_t *mode)

This function is query specific cuBLAS Atomic mode for context.

默认初始化的 cublasHandle_t The default atomic mode of an object is CUBLAS_ATOMICS_NOT_ALLOWED. 有关详细信息,See the section on types.

2.4.22. cublasSetMathMode()

cublasStatus_t cublasSetMathMode(cublasHandle_t handle, cublasMath_t mode)

cublasSetMathMode The function enables you to choose from cublasMath_t The defined computational precision mode（请参阅 cublasMath_t）. Allows the user to set computational precision modes to their logical combination（不推荐使用的 CUBLAS_TENSOR_OP_MATH 除外）. 例如,cublasSetMathMode(handle, CUBLAS_DEFAULT_MATH | CUBLAS_MATH_DISALLOW_REDUCED_PRECISION_REDUCTION). 请注意,The default math mode is CUBLAS_DEFAULT_MATH.

有关 cublasGemmEx() 和 cublasLtMatmul() API and the matrix and computational precision allowed by its stride variant,请参阅：cublasGemmEx()、cublasGemmBatchedEx()、cublasGemmStridedBatchedEx() 和 cublasLtMatmul().

2.4.23. cublasGetMathMode()

cublasStatus_t cublasGetMathMode(cublasHandle_t handle, cublasMath_t *mode)

This function returns the math mode used by the library routines.

2.4.24. cublasSetSmCountTarget()

cublasStatus_t cublasSetSmCountTarget(cublasHandle_t handle, int smCountTarget)

cublasSetSmCountTarget The function allows to override the number of multiprocessors available to the library during kernel execution.

当已知 cuBLAS routine and different CUDA While other work on the stream is running concurrently,This option can be used to improve library performance. 例如. NVIDIA A100 GPU 有 108 个 SM,And there is a concurrency kenrel 运行,网格大小为 8,Values ​​can be used 100 的 cublasSetSmCountTarget to override the library heuristics,以优化在 100 run on multiple processors.

当设置为 0 时,The library will return to its default behavior. The input value should not exceed the device's multiprocessor count,可使用 cudaDeviceGetAttribute 获取. 不接受负值.

Users must ensure thread safety when using this routine to modify library handles,类似于使用 cublasSetStream 时.

2.4.25. cublasGetSmCountTarget()

cublasStatus_t cublasGetSmCountTarget(cublasHandle_t handle, int *smCountTarget)

This function gets the value previously programmed into the library handle.