An adaptive performance modeling tool for GPU architectures

High Performance Datacenter Networks: Architectures, Algorithms, and Opportunities

“…The value, ITILP, models the possibility of inter-thread instruction-level parallelism in GPGPUs. The concept of ITILP was introduced in Baghsorkhi et al [11]. In particular, instructions may issue from multiple warps on a GPGPU; thus, we consider global ILP (i.e., ILP among warps) rather than warp-local ILP (i.e., the ILP of one warp).…”

“…To provide a formal framework to study this problem, Baghsorkhi et al introduced the concept of balanced GPGPU computation [11]. This model represents a GPGPU computation using the computation carried by an average warp.…”

“…To identify bottlenecks that result in an imbalanced computation, Baghsorkhi et al combined the amount of available concurrency in a kernel (TLP, DLP, and ILP)-considering the effect of coexisting threads-and the latencies of the SIMD pipeline and the memory system such that the interactive effects of different performance factors are preserved [11]. For this task, they introduced the work flow graph (WFG).…”

Performance models for asynchronous data transfers on consumer Graphics Processing Units

“…GPU performance modeling has been tackled in some valuable research works [1,5,20], but none of them deals with data transfers between CPU and GPU and the use of streams. To the best of our knowledge, there is only one research work focused on CUDA streams performance [8].…”

Fast filter bank convolution for three-dimensional wavelet transform by shared memory on mobile GPU computing

“…Acceleration of applications by GPU shared memory is developed on single GPU [6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21][22], multi-GPU [23] and GPU cluster [24][25][26]. Since the capacity of GPU shared memory is restricted, accessing them by multiple threads often leads to the problem of bank conflict, which is one of key factors to deteriorate the performance of CUDA kernels.…”