Environment Variables
The following environment variables are used by KML_BLAS.
Environment Variable |
Description |
Valid Value |
Default Value |
|---|---|---|---|
BLAS_VERBOSE |
Alarm severity. If the value is 1, info messages (such as component version and interface name) are output. If the value is greater than 1, debugging messages (such as component version, interface name, interface parameters, and execution duration) are output. |
Integer |
0 |
BLAS_THREAD_TIMEOUT |
Maximum loop wait time for thread synchronization in PTHREAD multi-thread backend mode. |
4 to 30 |
0 |
BLAS_NUM_THREADS |
Number of internal BLAS threads used for parallel execution in BLAS. This setting takes precedence over OMP_NUM_THREADS. |
Integer |
0 |
OMP_NUM_THREADS |
Number of internal OMP threads used for parallel execution in BLAS. |
Integer |
0 |
BLAS_JC_NT |
Block tiling strategy in the n direction during matrix operations in multi-thread scenarios. The product of BLAS_JC_NT and BLAS_IC_NT must equal the number of threads specified. Otherwise, the default block tiling strategy is used. |
Integer |
0 |
BLAS_IC_NT |
Block tiling strategy in the m direction during matrix operations in multi-thread scenarios. The product of BLAS_JC_NT and BLAS_IC_NT must equal the number of threads specified. Otherwise, the default block tiling strategy is used. |
Integer |
0 |
BLAS_NUMA_THREADS |
Number of cores per NUMA node, which controls thread NUMA affinity in PTHREAD multi-thread backend mode. |
Positive integer |
0 |
KML_BLAS_THREAD_TYPE |
Backend type for thread management. |
PTHREAD\OMP\KUPL |
OMP |
SGEMM_P |
Size of tiles in the m direction for single-precision General Matrix Multiply (GEMM). This parameter is valid only on Kunpeng 920 7270, and 7280 processors. |
A positive integer. For Kunpeng 920 7270 and 7280 processors, the value must be a multiple of 48. Ensure that SGEMM_P x SGEMM_Q x 4 x 2 ≤ 37748736. |
128 |
SGEMM_Q |
Size of tiles in the k direction for single-precision GEMM. This parameter is valid only on Kunpeng 920 7270, and 7280 processors. |
A positive integer. For Kunpeng 920 7270 and 7280 processors, the value must be a multiple of 48. Ensure that SGEMM_P x SGEMM_Q x 8 x 2 ≤ 37748736. |
512 |
SGEMM_R |
Size of tiles in the n direction for single-precision GEMM. This parameter is valid only on Kunpeng 920 7270, and 7280 processors. |
A positive integer. For Kunpeng 920 7270 and 7280 processors, the value must be a multiple of 48. Ensure that SGEMM_Q x SGEMM_R x 4 x 2 ≤ 37748736. |
4096 |
DGEMM_P |
Size of tiles in the m direction for double-precision GEMM. This parameter is valid only on Kunpeng 920 7270, and 7280 processors. |
A positive integer. For Kunpeng 920 7270 and 7280 processors, the value must be a multiple of 24. Ensure that DGEMM_P x DGEMM_Q x 8 x 2 ≤ 37748736. |
128 |
DGEMM_Q |
Size of tiles in the k direction for double-precision GEMM. This parameter is valid only on Kunpeng 920 7270, and 7280 processors. |
A positive integer. For Kunpeng 920 7270 and 7280 processors, the value must be a multiple of 24. Ensure that DGEMM_Q x DGEMM_R x 8 x 2 ≤ 37748736. |
512 |
DGEMM_R |
Size of tiles in the n direction for double-precision GEMM. This parameter is valid only on Kunpeng 920 7270, and 7280 processors. |
A positive integer. For Kunpeng 920 7270 and 7280 processors, the value must be a multiple of 24. Ensure that DGEMM_Q x DGEMM_R x 8 x 2 ≤ 37748736. |
4096 |
CGEMM_P |
Size of tiles in the m direction for complex single-precision GEMM. This parameter is valid only on Kunpeng 920 7270, and 7280 processors. |
A positive integer. For Kunpeng 920 7270 and 7280 processors, the value must be a multiple of 8. Ensure that CGEMM_P x CGEMM_Q x 8 x 2 ≤ 37748736. |
64 |
CGEMM_Q |
Size of tiles in the k direction for complex single-precision GEMM. This parameter is valid only on Kunpeng 920 7270, and 7280 processors. |
A positive integer. For Kunpeng 920 7270 and 7280 processors, the value must be a multiple of 8. Ensure that CGEMM_Q x CGEMM_R x 8 x 2 ≤ 37748736. |
512 |
CGEMM_R |
Size of tiles in the n direction for complex single-precision GEMM. This parameter is valid only on Kunpeng 920 7270, and 7280 processors. |
A positive integer. For Kunpeng 920 7270 and 7280 processors, the value must be a multiple of 8. Ensure that CGEMM_Q x CGEMM_R x 8 x 2 ≤ 37748736. |
4096 |
ZGEMM_P |
Size of tiles in the m direction for complex double-precision GEMM. This parameter is valid only on Kunpeng 920 7270, and 7280 processors. |
A positive integer. For Kunpeng 920 7270 and 7280 processors, the value must be a multiple of 8. Ensure that ZGEMM_P x ZGEMM_Q x 16 x 2 ≤ 37748736. |
64 |
ZGEMM_Q |
Size of tiles in the k direction for complex double-precision GEMM. This parameter is valid only on Kunpeng 920 7270, and 7280 processors. |
A positive integer. For Kunpeng 920 7270 and 7280 processors, the value must be a multiple of 8. Ensure that ZGEMM_Q x ZGEMM_R x 16 x 2 ≤ 37748736. |
320 |
ZGEMM_R |
Size of tiles in the n direction for complex double-precision GEMM. This parameter is valid only on Kunpeng 920 7270, and 7280 processors. |
A positive integer. For Kunpeng 920 7270 and 7280 processors, the value must be a multiple of 8. Ensure that ZGEMM_Q x ZGEMM_R x 16 x 2 ≤ 37748736. |
2048 |