Index Expression Experiment

[rachitnigam]

Experiment

Figure out if index expression analysis can catastrophically hurt DSE.

• Two instances of the gemm kernel.
  • 1D Gemm kernel that takes two extra parameters x and out[1]. It performs the computation out[0] += x ^ addr where addr is the matrix address in the innermost loop and also performs the normal computation for GeMM.
  • 1D Gemm kernel that takes two parameters x and out[1] and indexes into the arrays by doing M[x ^ addr] where addr is the normal indexing expression while also doing out[0] += x ^ addr.
  • Both programs are run with parameters: partitioning = 16, 1 ported memories and unrolling factors [2, 4, 8] for both the two innermost loops 18 configurations.
  • See if the second program uses more resources and is slower than the first one.

[rachitnigam]

RFC from @sampsyo and @tissue3.

[tissue3]

Sorry I still don't know what DSE is.

[rachitnigam]

Design space exploration

[sampsyo]

Again, sounds just about perfect!

[rachitnigam]

@sampsyo comments on the current heatmaps (permalink):

Wow; pretty weird outlier in the execution time results, huh? But it’s again odd that the execution time is so stable among the other points, even as the unrolling and partitioning changes…

The resource usage indeed goes up as expected but the runtime does not go down. One possible hypothesis is that the benchmark is memory bound -- the data transfer cost outweigh the total runtime of the gemm kernel. Figure out a way to validate this.

[rachitnigam]

Also, note that unlike the misaligned-partition-and-unroll experiment where the unrolling and partitioning factors increase together and the runtime changes more predictably, this experiment uses single ported memories.