Ramping up on parallel programming basics and gpu architecture experiments.
Approach: Execute a synthetic vector operation that performs a fixed number of floating-point operations (FLOPs) per memory element accessed. For a given number of FLOPs per element (FLOP_COUNT), each thread executes:
We read from A[i], B[i], D[i], E[i] and write to C[i]:
The
vec_opNkernel is used to sweep arithmetic intensity by increasingFLOP_COUNTand analyzing how performance shifts from memory-bound to compute-bound behavior.
We investigated standard dense matrix multiplication:
We read matrices A, B and write C, each of size
The
matmulkernel evaluates a more realistic compute-bound workload where intensity increases with matrix size.
These two benchmarks allow us to sweep across a range of arithmetic intensities, visualizing memory- and compute-bound regions on the roofline plot of the RTX 3090. Red dots represent vec_opN, and black/gray triangles represent matmul.