Developing adaptive multi-device applications with the Heterogeneous Programming Library

Abstract

The usage of heterogeneous devices presents two main problems. One is their complex programming, a problem that grows when multiple devices are used. The second issue is that even if the codes for these devices can be portable on top of OpenCL, they lack performance portability, effectively requiring specialized implementations for each device to get good performance. In this paper we extend the Heterogeneous Programming Library (HPL), which improves the usability of heterogeneous systems on top of OpenCL, to better handle both issues. First, we provide HPL with mechanisms to support the implementation of any multi-device application that requires arbitrary patterns of communication between several devices and a host memory. In a second stage HPL is improved with an adaptive scheme to optimize communications between devices depending on the execution environment. An evaluation using benchmarks with very different nature shows that HPL reduces the SLOCs and programming effort of OpenCL applications by 27 and 43 %, respectively, while improving the performance of applications that exchange data between devices by 28 % on average.

Facilitating the development of stencil applications using the Heterogeneous Programming Library

“…HPL has proven to be a good alternative to low level approaches like OpenCL or CUDA for programmability and performance as we can see in Viñas et al 8 and Viñas et al 12 Also, its evaluation in Viñas et al 9 showed that the usage of subarrays largely improved the programmability while having a negligible impact on performance when compared to native OpenCL. HPL has proven to be a good alternative to low level approaches like OpenCL or CUDA for programmability and performance as we can see in Viñas et al 8 and Viñas et al 12 Also, its evaluation in Viñas et al 9 showed that the usage of subarrays largely improved the programmability while having a negligible impact on performance when compared to native OpenCL.…”

“…We will first focus on the programmability advantages of the mechanism to automatically synchronize ghost regions. HPL has proven to be a good alternative to low level approaches like OpenCL or CUDA for programmability and performance as we can see in Viñas et al 8 and Viñas et al 12 Also, its evaluation in Viñas et al 9 showed that the usage of subarrays largely improved the programmability while having a negligible impact on performance when compared to native OpenCL.…”

“…12 Coherency is maintained, as if during its execution the host program, or the execution of a new kernel in a device, requires the contents of an Array, HPL checks whether the associated memory has an updated copy according to sequential consistency. This allows HPL to launch multiple kernels that can execute in parallel in different devices.…”

“…This allows HPL to launch multiple kernels that can execute in parallel in different devices. 12 Coherency is maintained, as if during its execution the host program, or the execution of a new kernel in a device, requires the contents of an Array, HPL checks whether the associated memory has an updated copy according to sequential consistency. If this is not the case, the host program creates or updates that copy with the most recent version, which could imply a wait period if the Array is written by a kernel in execution or whose execution is pending.…”

“…In our example, this involves manually updating the ghost regions of input in each iteration, as lines 36 to 40 in Figure 4 show. Another issue is the correct computation of the limits of each subarray needed in any multi-device problem, which is performed in function obtain_subarrays_size() (lines [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19][20]).…”