From 57cba3f3dac9c1a12dd15d11c71abc36539b0d40 Mon Sep 17 00:00:00 2001 From: RXGottlieb <36906916+RXGottlieb@users.noreply.github.com> Date: Fri, 8 Nov 2024 13:17:51 -0500 Subject: [PATCH] Bug fixes (#10) Shifted incrementation of _iteration_count to match current version of EAGO, added Reexport to deps, re-ran README examples. --- Project.toml | 4 +++- README.md | 37 +++++++++++++++++------------------ examples/ParBB/subroutines.jl | 7 ++++--- src/SourceCodeMcCormick.jl | 7 +++---- 4 files changed, 28 insertions(+), 27 deletions(-) diff --git a/Project.toml b/Project.toml index 3f91bbe..1a59902 100644 --- a/Project.toml +++ b/Project.toml @@ -1,7 +1,7 @@ name = "SourceCodeMcCormick" uuid = "a7283dc5-4ecf-47fb-a95b-1412723fc960" authors = ["Robert Gottlieb "] -version = "0.4.0" +version = "0.4.1" [deps] CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba" @@ -9,6 +9,7 @@ Dates = "ade2ca70-3891-5945-98fb-dc099432e06a" DocStringExtensions = "ffbed154-4ef7-542d-bbb7-c09d3a79fcae" Graphs = "86223c79-3864-5bf0-83f7-82e725a168b6" IfElse = "615f187c-cbe4-4ef1-ba3b-2fcf58d6d173" +Reexport = "189a3867-3050-52da-a836-e630ba90ab69" SymbolicUtils = "d1185830-fcd6-423d-90d6-eec64667417b" Symbolics = "0c5d862f-8b57-4792-8d23-62f2024744c7" @@ -17,6 +18,7 @@ CUDA = "5" DocStringExtensions = "0.8 - 0.9" Graphs = "1" IfElse = "0.1.0 - 0.1.1" +Reexport = "~1" SymbolicUtils = "3" Symbolics = "6" julia = "1.10" diff --git a/README.md b/README.md index 1c373c8..f232e06 100644 --- a/README.md +++ b/README.md @@ -524,15 +524,15 @@ Running this code (after an initial compilation run) generates the following out --------------------------------------------------------------------------------------------------------------------------------- | Iteration # | Nodes | Lower Bound | Upper Bound | Gap | Ratio | Timer | Time Left | --------------------------------------------------------------------------------------------------------------------------------- -| 10 | 47818 | -8.337E-01 | -7.048E-01 | 1.289E-01 | 1.546E-01 | 0.74 | 3599.26 | -| 20 | 117316 | -7.703E-01 | -7.048E-01 | 6.549E-02 | 8.503E-02 | 1.47 | 3598.53 | -| 30 | 162052 | -7.608E-01 | -7.048E-01 | 5.604E-02 | 7.365E-02 | 1.90 | 3598.10 | -| 40 | 177018 | -7.473E-01 | -7.048E-01 | 4.254E-02 | 5.693E-02 | 2.38 | 3597.62 | -| 50 | 161712 | -7.411E-01 | -7.048E-01 | 3.632E-02 | 4.901E-02 | 3.24 | 3596.76 | -| 60 | 128062 | -7.293E-01 | -7.048E-01 | 2.453E-02 | 3.363E-02 | 4.10 | 3595.90 | -| 70 | 73936 | -7.211E-01 | -7.048E-01 | 1.631E-02 | 2.262E-02 | 5.03 | 3594.97 | -| 80 | 11916 | -7.180E-01 | -7.048E-01 | 1.321E-02 | 1.841E-02 | 5.59 | 3594.41 | -| 83 | 4560 | -7.051E-01 | -7.048E-01 | 3.648E-04 | 5.173E-04 | 5.64 | 3594.36 | +| 10 | 47818 | -8.337E-01 | -7.048E-01 | 1.289E-01 | 1.546E-01 | 0.47 | 3599.53 | +| 20 | 117316 | -7.703E-01 | -7.048E-01 | 6.549E-02 | 8.503E-02 | 0.83 | 3599.17 | +| 30 | 162052 | -7.608E-01 | -7.048E-01 | 5.604E-02 | 7.365E-02 | 1.26 | 3598.74 | +| 40 | 177018 | -7.473E-01 | -7.048E-01 | 4.254E-02 | 5.693E-02 | 1.74 | 3598.26 | +| 50 | 161712 | -7.411E-01 | -7.048E-01 | 3.632E-02 | 4.901E-02 | 2.62 | 3597.38 | +| 60 | 128062 | -7.293E-01 | -7.048E-01 | 2.453E-02 | 3.363E-02 | 3.17 | 3596.83 | +| 70 | 73936 | -7.211E-01 | -7.048E-01 | 1.631E-02 | 2.262E-02 | 3.76 | 3596.24 | +| 80 | 11916 | -7.180E-01 | -7.048E-01 | 1.321E-02 | 1.841E-02 | 4.36 | 3595.64 | +| 82 | 4560 | -7.051E-01 | -7.048E-01 | 3.648E-04 | 5.173E-04 | 4.42 | 3595.58 | --------------------------------------------------------------------------------------------------------------------------------- Relative Tolerance Achieved @@ -550,13 +550,13 @@ Solution: p[8] = 0.9999999998842802 ``` -Using the subgradient-free method, and processing 8192 nodes per iteration, this problem converged in 83 -iterations (roughly 679,936 nodes explored), with a total time of 5.64 seconds. These times were obtained +Using the subgradient-free method, and processing 8192 nodes per iteration, this problem converged in 82 +iterations (roughly 671,744 nodes explored), with a total time of 4.42 seconds. These times were obtained on a workstation with an Intel i7-9850H processor and an NVIDIA Quadro T2000 GPU. Using a GPU with a greater capacity for double-precision floating point calculations will, of course, improve the overall performance of the algorithm, but even with a fairly "typical" GPU such as this, relatively competitive speed can be obtained. As compared to the base version of EAGO (which makes use of subgradients), this -example ran roughly 9x faster. +example ran roughly 11x faster. It is also important to note that, because a subgradient-free method was used, the lower bounds for any individual branch-and-bound node are not as tight as those that can be obtained using subgradient-based @@ -662,16 +662,15 @@ Running this code (after an initial compilation run) generates the following out --------------------------------------------------------------------------------------------------------------------------------- | Iteration # | Nodes | Lower Bound | Upper Bound | Gap | Ratio | Timer | Time Left | --------------------------------------------------------------------------------------------------------------------------------- -| 10 | 16280 | -7.754E-01 | -7.048E-01 | 7.067E-02 | 9.113E-02 | 0.59 | 3599.41 | -| 20 | 59342 | -7.353E-01 | -7.048E-01 | 3.055E-02 | 4.154E-02 | 1.16 | 3598.84 | -| 30 | 46048 | -7.108E-01 | -7.048E-01 | 5.983E-03 | 8.418E-03 | 1.82 | 3598.18 | -| 40 | 2240 | -7.058E-01 | -7.048E-01 | 9.852E-04 | 1.396E-03 | 2.30 | 3597.70 | -| 41 | 2240 | -7.058E-01 | -7.048E-01 | 9.852E-04 | 1.396E-03 | 2.30 | 3597.70 | +| 10 | 16280 | -7.754E-01 | -7.048E-01 | 7.067E-02 | 9.113E-02 | 0.45 | 3599.55 | +| 20 | 59342 | -7.353E-01 | -7.048E-01 | 3.055E-02 | 4.154E-02 | 1.09 | 3598.91 | +| 30 | 46048 | -7.108E-01 | -7.048E-01 | 5.983E-03 | 8.418E-03 | 1.85 | 3598.15 | +| 40 | 2240 | -7.058E-01 | -7.048E-01 | 9.852E-04 | 1.396E-03 | 2.33 | 3597.67 | --------------------------------------------------------------------------------------------------------------------------------- Absolute Tolerance Achieved Optimal Solution Found at Node 1 -Lower Bound: -0.7057619894307734 +Lower Bound: -0.7057619894307733 Upper Bound: -0.7047768067379733 Solution: p[1] = -0.9999999998903418 @@ -688,7 +687,7 @@ As in the previous examples, these results were generated using an Intel i7-9850 Quadro T2000 GPU. Effectively, ParBB is running precisely the same lower-bounding routine as the base version of the global solver EAGO, except that the routine is being performed in parallel on a GPU rather than serially on the CPU. This example converges in 40 iterations of at most 8192 nodes per iteration, -for a total of roughly 327,680 nodes explored, and reaches its solution in 2.30 seconds. As compared to the +for a total of roughly 327,680 nodes explored, and reaches its solution in 2.33 seconds. As compared to the base version of EAGO, this implementation of ParBB solves the problem roughly 22x faster. As in the previous subsection, it should be noted that a GPU with better double-precision floating-point calculation throughput would yield even faster results. diff --git a/examples/ParBB/subroutines.jl b/examples/ParBB/subroutines.jl index 83569c5..c5dd4cb 100644 --- a/examples/ParBB/subroutines.jl +++ b/examples/ParBB/subroutines.jl @@ -31,8 +31,7 @@ function solve_gpu!(m::EAGO.GlobalOptimizer) # Identify the extension ext = EAGO._ext(m) - # Set counts to 1 - m._iteration_count = 1 + # Set node count to 1 m._node_count = 1 # Prepare to run branch-and-bound @@ -54,6 +53,9 @@ function solve_gpu!(m::EAGO.GlobalOptimizer) # Run branch and bound; terminate when the stack is empty or when some # tolerance or limit is hit while !EAGO.termination_check(m) + + # Update iteration counter + m._iteration_count += 1 # Garbage collect every gc_freq iterations if mod(m._iteration_count, EAGO._ext(m).gc_freq)==0 @@ -105,7 +107,6 @@ function solve_gpu!(m::EAGO.GlobalOptimizer) m._time_left = m._parameters.time_limit - m._run_time EAGO.log_iteration!(m) EAGO.print_iteration!(m, false) - m._iteration_count += 1 end EAGO.print_iteration!(m, true) diff --git a/src/SourceCodeMcCormick.jl b/src/SourceCodeMcCormick.jl index 0f68330..46edd44 100644 --- a/src/SourceCodeMcCormick.jl +++ b/src/SourceCodeMcCormick.jl @@ -1,7 +1,8 @@ module SourceCodeMcCormick -using Symbolics +using Reexport +@reexport using Symbolics using SymbolicUtils.Code using IfElse using DocStringExtensions @@ -51,8 +52,6 @@ export McCormickIntervalTransform, IntervalTransform export apply_transform, all_evaluators, convex_evaluator, extract_terms, genvar, genparam, get_name, factor, binarize!, pull_vars, shrink_eqs, grad, shrink_grad!, convex_subgradient, all_subgradients, grad_transform!, - levels, eqn_edges, eval_generator, grad_eval_generator, fgen_cv, fgen_cvgrad, - fgen, fgen2 -export @variables, Num + levels, eqn_edges, eval_generator, grad_eval_generator, fgen end \ No newline at end of file