curand host APIs hurt the performance of the following cublas APIs

kangshiyin · December 20, 2012, 6:17am

I found cublasDgemm() slowing down 2 times after invoking curandGenerateUniformDouble().

It seems curand host APIs change the Shared Mem from 48KB to 16KB and never change it back,
which hurt the performance of all callings following the curand APIs.

Any idea to solve these issues ?

njuffa · December 20, 2012, 7:37pm

I would suggest filing a bug via the registered developer website, attaching a self-contained repro case that demonstrates the issue, so the library team can have a look. Thanks!

Topic		Replies	Views
The cublasGemmGroupedBatchedEx API results in an additional cudaMemcpyAsync H2D GPU-Accelerated Libraries cuda	0	20	February 27, 2025
About Host API of curand curandGenerateUniform() function CUDA Programming and Performance	3	1645	June 15, 2016
Slow execution of simpleCublas CUDA Programming and Performance	0	4879	March 25, 2010
CUBLAS slow down after calling CUFFT real to complex transform CUDA Programming and Performance	2	915	February 18, 2013
cuBLAS GEMM 2.5 times slower on 4090 than on 3090? GPU-Accelerated Libraries cublas , curand	0	414	December 25, 2023
Why performance is worse with CUBLAS- than with kernel-function GPU-Accelerated Libraries	3	928	September 5, 2019
cuBLAS sgemm is slow CUDA Programming and Performance	4	2436	June 26, 2017
Poor accuracy of curand Sobol at high dimensions GPU-Accelerated Libraries curand	2	664	May 28, 2021
Timeline mis-aligned between MemCpy and Compute Kernels in Visual profile CUDA Programming and Performance	0	793	December 24, 2012
Is it correct that my Pascal card is calling Maxwell_gemm kernels through cublas? And if so, why is cublas unusably slow for me? CUDA Programming and Performance	6	933	August 23, 2018

curand host APIs hurt the performance of the following cublas APIs

Related topics