math icon indicating copy to clipboard operation
math copied to clipboard
verified publisher icon stan-dev

Add simplex row and column matrix constraints

Open SteveBronder opened this issue 1 year ago • 11 comments

Summary

This adds the functions simplex_row_constrain and simplex_column_constrain to be used by the stan language for matrices with simplexes along the rows or columns of the matrix. This is just the constrain version which is already a lot of code. I'm going to add the free functions in a separate PR

Tests

Tests were added in mix and prim and can be run with

python ./runTests.py -j3 ./test/unit/math/mix/fun/simplex_column_constrain_test.cpp \
 ./test/unit/math/mix/fun/simplex_row_constrain_test.cpp \
./test/unit/math/prim/fun/simplex_row_constrain_test.cpp  \
./test/unit/math/prim/fun/simplex_column_constrain_test.cpp

A prim test for simplex_column_constrain was not written because under the hood simplex_column_constrain for doubles just calls simplex_constrain

Side Effects

There is one interesting thing to point out in the reverse mode versio nof simplex_column_constrain. I think this is one of the only places in the Stan math library we use Eigen::RowMajor to have row major underlying matrices.

For each of these algorithms we want to access memory in a contiguous manner. For rows of simplex matrices we start with a (N, M) and get back a (N, M + 1) where we operate on each column of the matrix from left to right. For columns of simplex matrices we start with an (N, M) and get back a (N + 1, M) where we need to operate on the row of the matrix from top to bottom. So for the row simplex everything is nice with column major order because that's how our access happens. But for columns of simplex vectors for reverse mode, we make a hard copy of the input changing it from column major to row major so that the rest of our memory access is contiguous. That means that for the rest of the reverse mode code we access things nicely, but then pay the cost of different storage orders when copy the input to the arena matrix in forward mode and then when we copy the adjoint from the reverse mode into the input's adjoint. But reading from different storage orders twice I think is better than reading from the noncontiguous memory for the entire algorithm.

Release notes

Adds constrain function for creating matrices holding row or column simplexes

Checklist

  • [x] Copyright holder: Simons foundation

    The copyright holder is typically you or your assignee, such as a university or company. By submitting this pull request, the copyright holder is agreeing to the license the submitted work under the following licenses: - Code: BSD 3-clause (https://opensource.org/licenses/BSD-3-Clause) - Documentation: CC-BY 4.0 (https://creativecommons.org/licenses/by/4.0/)

  • [x] the basic tests are passing

    • unit tests pass (to run, use: ./runTests.py test/unit)
    • header checks pass, (make test-headers)
    • dependencies checks pass, (make test-math-dependencies)
    • docs build, (make doxygen)
    • code passes the built in C++ standards checks (make cpplint)

  • [x] the code is written in idiomatic C++ and changes are documented in the doxygen

  • [x] the new changes are tested

SteveBronder avatar Dec 22 '23 18:12 SteveBronder

@andrjohns could you take a look at this?

SteveBronder avatar Jan 02 '24 17:01 SteveBronder

Name Old Result New Result Ratio Performance change( 1 - new / old )
arma/arma.stan 0.19 0.18 1.06 6.0% faster
low_dim_corr_gauss/low_dim_corr_gauss.stan 0.01 0.01 1.04 3.92% faster
gp_regr/gen_gp_data.stan 0.02 0.02 1.06 5.22% faster
gp_regr/gp_regr.stan 0.11 0.1 1.05 4.77% faster
sir/sir.stan 76.94 75.78 1.02 1.5% faster
irt_2pl/irt_2pl.stan 3.81 3.82 1.0 -0.44% slower
eight_schools/eight_schools.stan 0.05 0.05 1.02 1.78% faster
pkpd/sim_one_comp_mm_elim_abs.stan 0.24 0.24 0.98 -2.1% slower
pkpd/one_comp_mm_elim_abs.stan 17.54 17.68 0.99 -0.83% slower
garch/garch.stan 0.44 0.46 0.96 -4.21% slower
low_dim_gauss_mix/low_dim_gauss_mix.stan 2.75 2.78 0.99 -1.1% slower
arK/arK.stan 1.6 1.63 0.98 -2.01% slower
gp_pois_regr/gp_pois_regr.stan 2.49 2.49 1.0 0.14% faster
low_dim_gauss_mix_collapse/low_dim_gauss_mix_collapse.stan 9.02 9.06 1.0 -0.48% slower
performance.compilation 172.25 175.88 0.98 -2.11% slower
Mean result: 1.0076686160263006

Jenkins Console Log Blue Ocean Commit hash: e0626a595a3bddd83ac6d0ec401fba4e39c8c5ea

Machine information No LSB modules are available. Distributor ID: Ubuntu Description: Ubuntu 20.04.3 LTS Release: 20.04 Codename: focal

CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian Address sizes: 46 bits physical, 48 bits virtual CPU(s): 80 On-line CPU(s) list: 0-79 Thread(s) per core: 2 Core(s) per socket: 20 Socket(s): 2 NUMA node(s): 2 Vendor ID: GenuineIntel CPU family: 6 Model: 85 Model name: Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz Stepping: 4 CPU MHz: 2400.000 CPU max MHz: 3700.0000 CPU min MHz: 1000.0000 BogoMIPS: 4800.00 Virtualization: VT-x L1d cache: 1.3 MiB L1i cache: 1.3 MiB L2 cache: 40 MiB L3 cache: 55 MiB NUMA node0 CPU(s): 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78 NUMA node1 CPU(s): 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63,65,67,69,71,73,75,77,79 Vulnerability Gather data sampling: Mitigation; Microcode Vulnerability Itlb multihit: KVM: Mitigation: VMX disabled Vulnerability L1tf: Mitigation; PTE Inversion; VMX conditional cache flushes, SMT vulnerable Vulnerability Mds: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Meltdown: Mitigation; PTI Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Retbleed: Mitigation; IBRS Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; IBRS, IBPB conditional, STIBP conditional, RSB filling, PBRSB-eIBRS Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Mitigation; Clear CPU buffers; SMT vulnerable Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single pti intel_ppin ssbd mba ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req pku ospke md_clear flush_l1d arch_capabilities

G++: g++ (Ubuntu 9.4.0-1ubuntu1~20.04) 9.4.0 Copyright (C) 2019 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Clang: clang version 10.0.0-4ubuntu1 Target: x86_64-pc-linux-gnu Thread model: posix InstalledDir: /usr/bin

stan-buildbot avatar Feb 16 '24 01:02 stan-buildbot

@syclik could you give this anotherl look?

SteveBronder avatar Feb 20 '24 20:02 SteveBronder

@SteveBronder, sorry, I just saw a comment that I missed. Thought I'd put it here so you can track it. (It's more of a question about the PR than a single line of code.)

@syclik Question: would it now be worth writing a function that accepts a pointer and length? And then using that function instead of reimplementing the same logic.

@SteveBronder Sorry I'm not following, where would you want to use this?

The question I was asking was whether both simplex_row_constrain and simplex_column_constrain could call a function that consolidates the implementation. The chunk of code that looked like duplication was:

stan/math/prim/fun/simplex_row_constrain.hpp

  eigen_arr stick_len = eigen_arr::Constant(N, 1.0);
  for (Eigen::Index k = 0; k < Km1; ++k) {
    auto z_k = inv_logit(y_ref.array().col(k) - log(Km1 - k));
    x.array().col(k) = stick_len * z_k;
    stick_len -= x.array().col(k);
  }
  x.array().col(Km1) = stick_len;

stan/math/rev/fun/simplex_column_constrain.hpp

  arr_vec stick_len = arr_vec::Constant(M, 1.0);
  for (Eigen::Index k = 0; k < N; ++k) {
    const double log_N_minus_k = std::log(N - k);
    arena_z.row(k)
        = inv_logit(arena_y.array().row(k).val_op() - log_N_minus_k).matrix();
    x_val.row(k) = stick_len.array() * arena_z.array().row(k);
    stick_len -= x_val.array().row(k);
  }
  x_val.row(N) = stick_len;

The version with the jacobian also repeats. I guess I was wondering if it's better to have a C-like function that takes the pointer to data and length and consolidates the computation. We'd still have the same functions from the outside to handle types and everything.

It's a question, not even a suggestion. I was just reading the code and found the logic to be complicated enough and duplicated where it made me wonder. (Repeating code twice isn't that big of a deal in the long run.)

syclik avatar Feb 24 '24 03:02 syclik

@SteveBronder, minor changes requested. Once addressed, we can merge!

syclik avatar Feb 24 '24 04:02 syclik

Thanks @syclik! Just updated everything

SteveBronder avatar Feb 27 '24 16:02 SteveBronder

Actually @syclik do you mind if I just add the free functions to this PR? I think I'm just going to have them call the current simplex_free over the rows / columns so it shouldn't be that much extra code

SteveBronder avatar Feb 27 '24 16:02 SteveBronder

@SteveBronder, go for it! I think it makes sense.

syclik avatar Feb 27 '24 16:02 syclik

Name Old Result New Result Ratio Performance change( 1 - new / old )
arma/arma.stan 0.21 0.18 1.18 15.31% faster
low_dim_corr_gauss/low_dim_corr_gauss.stan 0.01 0.01 1.06 5.96% faster
gp_regr/gen_gp_data.stan 0.02 0.02 1.07 6.63% faster
gp_regr/gp_regr.stan 0.11 0.1 1.03 3.24% faster
sir/sir.stan 77.62 74.35 1.04 4.21% faster
irt_2pl/irt_2pl.stan 3.89 3.81 1.02 2.13% faster
eight_schools/eight_schools.stan 0.06 0.05 1.06 5.27% faster
pkpd/sim_one_comp_mm_elim_abs.stan 0.26 0.24 1.06 5.25% faster
pkpd/one_comp_mm_elim_abs.stan 18.28 19.53 0.94 -6.82% slower
garch/garch.stan 0.45 0.44 1.03 2.66% faster
low_dim_gauss_mix/low_dim_gauss_mix.stan 2.77 2.82 0.98 -1.84% slower
arK/arK.stan 1.64 1.59 1.04 3.5% faster
gp_pois_regr/gp_pois_regr.stan 2.51 2.48 1.01 1.33% faster
low_dim_gauss_mix_collapse/low_dim_gauss_mix_collapse.stan 9.15 8.92 1.03 2.45% faster
performance.compilation 168.25 173.63 0.97 -3.2% slower
Mean result: 1.0343112712202833

Jenkins Console Log Blue Ocean Commit hash: f70fb84db501edcbf822691c02664c43b76287b2

Machine information No LSB modules are available. Distributor ID: Ubuntu Description: Ubuntu 20.04.3 LTS Release: 20.04 Codename: focal

CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian Address sizes: 46 bits physical, 48 bits virtual CPU(s): 80 On-line CPU(s) list: 0-79 Thread(s) per core: 2 Core(s) per socket: 20 Socket(s): 2 NUMA node(s): 2 Vendor ID: GenuineIntel CPU family: 6 Model: 85 Model name: Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz Stepping: 4 CPU MHz: 2400.000 CPU max MHz: 3700.0000 CPU min MHz: 1000.0000 BogoMIPS: 4800.00 Virtualization: VT-x L1d cache: 1.3 MiB L1i cache: 1.3 MiB L2 cache: 40 MiB L3 cache: 55 MiB NUMA node0 CPU(s): 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78 NUMA node1 CPU(s): 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63,65,67,69,71,73,75,77,79 Vulnerability Gather data sampling: Mitigation; Microcode Vulnerability Itlb multihit: KVM: Mitigation: VMX disabled Vulnerability L1tf: Mitigation; PTE Inversion; VMX conditional cache flushes, SMT vulnerable Vulnerability Mds: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Meltdown: Mitigation; PTI Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Retbleed: Mitigation; IBRS Vulnerability Spec rstack overflow: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; IBRS, IBPB conditional, STIBP conditional, RSB filling, PBRSB-eIBRS Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Mitigation; Clear CPU buffers; SMT vulnerable Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single pti intel_ppin ssbd mba ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req pku ospke md_clear flush_l1d arch_capabilities

G++: g++ (Ubuntu 9.4.0-1ubuntu1~20.04) 9.4.0 Copyright (C) 2019 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Clang: clang version 10.0.0-4ubuntu1 Target: x86_64-pc-linux-gnu Thread model: posix InstalledDir: /usr/bin

stan-buildbot avatar Feb 28 '24 04:02 stan-buildbot

Name Old Result New Result Ratio Performance change( 1 - new / old )
arma/arma.stan 0.2 0.19 1.03 3.08% faster
low_dim_corr_gauss/low_dim_corr_gauss.stan 0.01 0.01 0.95 -5.29% slower
gp_regr/gen_gp_data.stan 0.02 0.02 0.97 -2.88% slower
gp_regr/gp_regr.stan 0.13 0.12 1.12 10.89% faster
sir/sir.stan 80.29 78.65 1.02 2.04% faster
irt_2pl/irt_2pl.stan 4.56 3.62 1.26 20.63% faster
eight_schools/eight_schools.stan 0.06 0.05 1.11 10.07% faster
pkpd/sim_one_comp_mm_elim_abs.stan 0.25 0.24 1.07 6.74% faster
pkpd/one_comp_mm_elim_abs.stan 18.41 17.37 1.06 5.64% faster
garch/garch.stan 0.44 0.42 1.04 4.18% faster
low_dim_gauss_mix/low_dim_gauss_mix.stan 2.94 2.66 1.11 9.65% faster
arK/arK.stan 1.63 1.55 1.05 4.86% faster
gp_pois_regr/gp_pois_regr.stan 2.54 2.37 1.07 6.76% faster
low_dim_gauss_mix_collapse/low_dim_gauss_mix_collapse.stan 9.23 8.73 1.06 5.39% faster
performance.compilation 175.54 169.48 1.04 3.46% faster
Mean result: 1.0644859127821111

Jenkins Console Log Blue Ocean Commit hash: f70fb84db501edcbf822691c02664c43b76287b2

Machine information No LSB modules are available. Distributor ID: Ubuntu Description: Ubuntu 20.04.3 LTS Release: 20.04 Codename: focal

CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian Address sizes: 46 bits physical, 48 bits virtual CPU(s): 80 On-line CPU(s) list: 0-79 Thread(s) per core: 2 Core(s) per socket: 20 Socket(s): 2 NUMA node(s): 2 Vendor ID: GenuineIntel CPU family: 6 Model: 85 Model name: Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz Stepping: 4 CPU MHz: 3399.999 CPU max MHz: 3700.0000 CPU min MHz: 1000.0000 BogoMIPS: 4800.00 Virtualization: VT-x L1d cache: 1.3 MiB L1i cache: 1.3 MiB L2 cache: 40 MiB L3 cache: 55 MiB NUMA node0 CPU(s): 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78 NUMA node1 CPU(s): 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63,65,67,69,71,73,75,77,79 Vulnerability Gather data sampling: Mitigation; Microcode Vulnerability Itlb multihit: KVM: Mitigation: VMX disabled Vulnerability L1tf: Mitigation; PTE Inversion; VMX conditional cache flushes, SMT vulnerable Vulnerability Mds: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Meltdown: Mitigation; PTI Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Retbleed: Mitigation; IBRS Vulnerability Spec rstack overflow: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; IBRS, IBPB conditional, STIBP conditional, RSB filling, PBRSB-eIBRS Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Mitigation; Clear CPU buffers; SMT vulnerable Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single pti intel_ppin ssbd mba ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req pku ospke md_clear flush_l1d arch_capabilities

G++: g++ (Ubuntu 9.4.0-1ubuntu1~20.04) 9.4.0 Copyright (C) 2019 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Clang: clang version 10.0.0-4ubuntu1 Target: x86_64-pc-linux-gnu Thread model: posix InstalledDir: /usr/bin

stan-buildbot avatar Mar 01 '24 03:03 stan-buildbot

@syclik I think this is ready to merge!

SteveBronder avatar Mar 01 '24 15:03 SteveBronder