As someone who writes CUDA code professionally, these are my two cents on the matter: The reported speed enhancements, particularly the claimed 117.83x speedup, might be somewhat misleading. Consider, for example, the comparison of CUDA speedups. The authors contrast their CUDA Fast Feed Forward (CUDA fff) implementation with their own highly unoptimized version of a CUDA Fast Forward (CUDA ff).
In an effort to ensure a fair comparison, they maintained the same code structure for both CUDA fff and CUDA ff. However, this approach resulted in the CUDA ff not utilizing any shared memory and caused significant memory divergence due to the use of threadIdx.x for indexing the outer dimensions of matrices.
As someone who writes CUDA code professionally, these are my two cents on the matter: The reported speed enhancements, particularly the claimed 117.83x speedup, might be somewhat misleading. Consider, for example, the comparison of CUDA speedups. The authors contrast their CUDA Fast Feed Forward (CUDA fff) implementation with their own highly unoptimized version of a CUDA Fast Forward (CUDA ff).
In an effort to ensure a fair comparison, they maintained the same code structure for both CUDA fff and CUDA ff. However, this approach resulted in the CUDA ff not utilizing any shared memory and caused significant memory divergence due to the use of threadIdx.x for indexing the outer dimensions of matrices.