Closed willow-ahrens closed 3 months ago
uses a more unified style for SparseDict, in keeping with the rest of Finch.
also, fixes #579 and fixes #540. We also remove SparseHashLevel, because it's too much work to keep that from bitrotting.
Benchmarks (note permutedims(Dense(Sparse)))
| | dirty | main | |:-----------------------------------------------------|:------------------:|:------------------:| | compile/compile_SpGeMM | 0.0369 ± 0.00067 s | 0.037 ± 0.0011 s | | compile/compile_pretty_triangle | 2.28 s | 2.3 s | | compile/time_to_first_SpGeMM | 4.96 s | 4.95 s | | graphs/bellmanford/Newman/netscience | 0.0811 ± 0.0057 ms | 0.0792 ± 0.0048 ms | | graphs/bellmanford/SNAP/roadNet-CA | 0.137 s | 0.145 s | | graphs/bfs/SNAP/soc-Epinions1 | 3.11 ± 0.12 ms | 3.09 ± 0.11 ms | | graphs/bfs/SNAP/soc-LiveJournal1 | 0.4 ± 0.007 s | 0.399 ± 0.018 s | | graphs/pagerank/SNAP/soc-Epinions1 | 17.7 ± 0.37 ms | 17.7 ± 0.43 ms | | graphs/pagerank/SNAP/soc-LiveJournal1 | 3.4 s | 3.5 s | | high-level/einsum_spmv_baremetal | 4.46 ± 0.083 ns | 4.42 ± 0.041 ns | | high-level/einsum_spmv_call_overhead | 25.2 ± 2 μs | 23.8 ± 1.5 μs | | high-level/einsum_spmv_compile_overhead | 0.256 ± 0.035 s | 0.272 ± 0.045 s | | high-level/permutedims(Dense(Dense())) | 0.276 ± 0.01 s | 0.286 ± 0.015 s | | high-level/permutedims(Dense(Sparse())) | 0.0846 ± 0.064 s | 0.0937 ± 0.0033 s | | indices/SpMV_32/SNAP/soc-Epinions1 | 1.6 ± 0.044 ms | 1.56 ± 0.093 ms | | indices/SpMV_64/SNAP/soc-Epinions1 | 0.797 ± 0.016 ms | 0.797 ± 0.017 ms | | indices/SpMV_p1/SNAP/soc-Epinions1 | 0.831 ± 0.014 ms | 0.833 ± 0.02 ms | | matrices/ATA_spgemm_gustavson/SNAP/soc-Epinions1 | 0.819 ± 0.17 s | 0.802 ± 0.2 s | | matrices/ATA_spgemm_outer/SNAP/soc-Epinions1 | 3.97 s | 4.06 s | | parallel/SpMV_serial/SNAP/soc-Epinions1 | 0.798 ± 0.02 ms | 0.804 ± 0.025 ms | | parallel/SpMV_serial/fsprand(10_000, 10_000, 0.01) | 0.724 ± 0.022 ms | 0.715 ± 0.018 ms | | parallel/SpMV_threaded/SNAP/soc-Epinions1 | 0.844 ± 0.036 ms | 0.837 ± 0.027 ms | | parallel/SpMV_threaded/fsprand(10_000, 10_000, 0.01) | 0.744 ± 0.026 ms | 0.737 ± 0.02 ms | | time_to_load | 0.605 ± 0.014 s | 0.609 ± 0.024 s |
uses a more unified style for SparseDict, in keeping with the rest of Finch.
also, fixes #579 and fixes #540. We also remove SparseHashLevel, because it's too much work to keep that from bitrotting.
Benchmarks (note permutedims(Dense(Sparse)))