Turbo findmax

[chriselrod]

After:

julia> @benchmark RecursiveFactorization.lu!(copyto!($B,$A), $ipiv,Val(true),Val(false), blocksize=8)
BenchmarkTools.Trial: 2562 samples with 1 evaluation.
 Range (min … max):  384.506 μs … 399.233 μs  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     386.841 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   387.488 μs ±   1.524 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

                ▃▂▄▄▅█▅▂▁▂                                       
  ▂▁▁▁▂▂▂▃▃▄▆▅▇████████████▇▅▄▄▃▃▃▃▃▃▃▄▄▅▄▆▆▇▇▇▇▇▇█▆▇▆▇▅▅▄▃▃▃▃▃ ▄
  385 μs           Histogram: frequency by time          391 μs <

 Memory estimate: 0 bytes, allocs estimate: 0.

Before:

julia> @benchmark RecursiveFactorization.lu!(copyto!($B,$A), $ipiv,Val(true),Val(false), blocksize=8)
BenchmarkTools.Trial: 2318 samples with 1 evaluation.
 Range (min … max):  426.292 μs … 445.566 μs  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     428.158 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   428.777 μs ±   1.491 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

           ▁▂▄▆▆█▅▆▆▆▁▃▁                    ▁▁▁▁▃▁ ▂▁            
  ▂▂▂▂▃▅▄▅██████████████▇▇▄▃▃▂▁▂▂▂▂▂▃▅▄▄▄▅▇▆██████▆██▇▆▆▄▄▃▃▃▃▃ ▄
  426 μs           Histogram: frequency by time          432 μs <

 Memory estimate: 0 bytes, allocs estimate: 0.

Setup:

@time using RecursiveFactorization
A = rand(300,300);
B = similar(A);
ipiv = Vector{Int}(undef,300);

Benchmarks across all sizes 4:500, RF 0.2.19:


versus PR on bottom.

Summary stats of relative differences:

julia> StatsBase.summarystats(respr ./ res19)
Summary Stats:
Length:         497
Missing Count:  0
Mean:           1.086197
Minimum:        0.889275
1st Quartile:   1.059688
Median:         1.092518
3rd Quartile:   1.120472
Maximum:        1.254754

So, 8.6% increase in GFLOPS on average on

julia> versioninfo()
Julia Version 1.9.3-DEV.0
Commit 6fc1be04ee (2023-07-06 14:55 UTC)
Platform Info:
  OS: Linux (x86_64-generic-linux)
  CPU: 36 × Intel(R) Core(TM) i9-10980XE CPU @ 3.00GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, cascadelake)
  Threads: 1 on 36 virtual cores

[YingboMa]

This looks pretty good, but the test failure looks real.

[chriselrod]

This looks pretty good, but the test failure looks real.

It is. It's specific to x86_64 CPUs without AVX2.

[chriselrod]

For a long time, I tried to get good performance on them when using integers by using Int32 instead of Int64, but that introduced bugs. I think I should just not care about performance and use Int64.

[chriselrod]

@YingboMa fixed upstream; all green