1D vs 2D array access experiments

[rachitnigam]

Experiment

To analyze the performance and resource usage characteristics of multidimensional array accesses and partitioning strategies.

Questions:

• For 2D arrays w/ various partitioning configurations without any unrolling, how does the resource count and runtime change?

  • Run Gemm ncubed kernel with 2D memories partitioned [1,2,4] * [1,2,4] 9 configurations.
  • Run Gemm ncubed kernel with 1D memories partitioned [1, 4, 8, 16]. Only need to run the 16 case and re-use numbers from previous experiment 1 configuration.
  • See if there is difference between the (1, 2) and (2, 1) cases for the 2D array.
  • See how (1, 2), (2, 1) cases compare with 2 case for 1D memory.

• For 2D arrays, does unrolling the accessor for the dimension which is partitioned help more?

  • Run 2D Gemm with memories partitioned [1,2,4] * [1,2,4] and loop unrolling factors [1, 2, 4] * [1, 2, 4]. Ignore configurations where both the memories are partitioned the same or both the unrolling factors are the same and where part1 * part2 != unroll1 * unroll2. 12 configurations.
  • Compare performance to 1D array partition [1,2,4,16] and unrolling factor [1,2,4,16] with the same restriction (unroll factor = partition factor). Use numbers from previous experiment. 0 configurations.
  • See if partitioning the right dimension (the one being unrolled) makes performance better.
  • Compare the configurations (1, 2, 1, 2), (1, 2, 2, 1) and (2, 2) and see if the former are better/worse.

[rachitnigam]

RFC from @sampsyo and @tissue3.

[sampsyo]

Sounds about right to me!