WIP: Let's discuss single-precision tolerances

This MR is a preliminary output of my study on the single-precision portability of ACTS. It should probably not be merged as is, but is only meant as a strawman to get a discussion started.

It builds on top of !490 (merged), so you may want to explore it commit by commit and/or diff it locally against that branch in order to get a clean diff.

In this branch, I've been running the ACTS test suite under Verrou's single-precision emulation mode, and tuning up test tolerances until the test suite mostly passes. At this point, the test suite does not fully pass because it hits the same off-diagonal covariance term imprecision issues as !490 (merged), but it gets close enough to get the discussion started. => That problem is solved (also, found some room for doing with narrower tolerances, under investigation)

Here is my question before I go further in my single-precision studies: which of these tolerance changes would be acceptable (in the sense that we would still produce precise enough physics), and which wouldn't?

Depending on the answer to this question, I will next investigate specific areas of unacceptable imprecision, in order to figure out whether the relevant parts of ACTS should stay in double precision, or have a "precision bottleneck" somewhere that can easily be resolved to make single precision viable.

changed the description

added 43 commits

• a865f33f...4cf178f5 - 18 commits from branch acts:master
• 36d1b4c7 - Reform checkCloseXyz into unified CHECK_XYZ macros
• 814763f5 - Replace BOOST_CHECK_CLOSE[_FRACTION] with CHECK_CLOSE
• 48449cf1 - Replace Eigen's isApprox with CHECK_CLOSE
• 90c96163 - Switch from CHECK_CLOSE to more appropriate comparisons
• 05b3d96f - Instrument remaining propagation test failures
• 07fe683b - Replace BOOST_CHECK_SMALL with CHECK_SMALL or CHECK_CLOSE_ABS
• a12c4ea0 - Try to eradicate BOOST_TEST from ACTS tests
• 7bc0c2b2 - Apply clang-format patch
• ded9d62a - Rework comparison error reporting
• 89e720b1 - Prefer CHECK_CLOSE_ABS(a, b, ...) over CHECK_SMALL(a-b, ...)
• dd76d2af - Reduce precision of DiscTrapezoidalBounds output
• 417eab3c - Tune up the on-surface tolerance a bit
• d5d215c8 - Tune up SolenoidBField tolerance + express it more logically
• 4f8f5fc8 - Tune up Jacobian tolerance a bit
• 8a3d3ac5 - Tune up Propagator tolerance a bit
• 8bceb28f - Adapt and homogeneize unit conversion tolerances
• 29138bd6 - Tune up tolerances and homogeneize comparisons in BFieldMapUtilsTest
• 180e489f - Tune up tolerances of CylinderVolumeBounds
• e00a9f5c - 0.1µm is impossible in f32
• 5d7df7c7 - 1e-6 relative precision can be too optimistic in f32
• c72aa241 - 1 keV is unrealistic in f32
• 3a225bde - Tune up ATLAS/Eigen comparison tolerance
• 250f9fd5 - Let's get real
• 2140bebd - 1 µm seems out of reach
• 1e641d90 - Reduce covariance ambitions

Compare with previous version

added 23 commits

• 46dcd0c3 - Replace Eigen's isApprox with CHECK_CLOSE
• 9d502f41 - Switch from CHECK_CLOSE to more appropriate comparisons
• 4934aa12 - Instrument remaining propagation test failures
• 5ca4e127 - Replace BOOST_CHECK_SMALL with CHECK_SMALL or CHECK_CLOSE_ABS
• ad381f92 - Try to eradicate BOOST_TEST from ACTS tests
• bebf92bb - Apply clang-format patch
• cb4b45a5 - Rework comparison error reporting
• dd41bee1 - Prefer CHECK_CLOSE_ABS(a, b, ...) over CHECK_SMALL(a-b, ...)
• af934bed - Reduce precision of DiscTrapezoidalBounds output
• 5bad5ac3 - Tune up the on-surface tolerance a bit
• e13449a5 - Tune up SolenoidBField tolerance + express it more logically
• ea8cfb7f - Tune up Jacobian tolerance a bit
• ef1bd1c6 - Tune up Propagator tolerance a bit
• 4fa8be06 - Adapt and homogeneize unit conversion tolerances
• e0c677d1 - Tune up tolerances and homogeneize comparisons in BFieldMapUtilsTest
• dbddf084 - Tune up tolerances of CylinderVolumeBounds
• b1c123a0 - 0.1µm is impossible in f32
• 81dff9a7 - 1e-6 relative precision can be too optimistic in f32
• 60cf3d17 - 1 keV is unrealistic in f32
• 1f6f6746 - Tune up ATLAS/Eigen comparison tolerance
• 20a49f08 - Let's get real
• 5eb02ff3 - 1 µm seems out of reach
• b94a3abe - Reduce covariance ambitions

Compare with previous version

added 18 commits

• 66cdcfa4 - Apply clang-format patch
• 605c025c - Rework comparison error reporting
• 1133c437 - Prefer CHECK_CLOSE_ABS(a, b, ...) over CHECK_SMALL(a-b, ...)
• 1dee3b19 - Reduce precision of DiscTrapezoidalBounds output
• 85657eda - Tune up the on-surface tolerance a bit
• 3c60ba7a - Tune up SolenoidBField tolerance + express it more logically
• aee7079e - Tune up Jacobian tolerance a bit
• cdfc617b - Tune up Propagator tolerance a bit
• b29014f2 - Adapt and homogeneize unit conversion tolerances
• ab04f565 - Tune up tolerances and homogeneize comparisons in BFieldMapUtilsTest
• 9f67dac2 - Tune up tolerances of CylinderVolumeBounds
• 6022edec - 0.1µm is impossible in f32
• 72d79853 - 1e-6 relative precision can be too optimistic in f32
• 39263719 - 1 keV is unrealistic in f32
• 3cd03c9b - Tune up ATLAS/Eigen comparison tolerance
• 81b7be11 - Let's get real
• 81572df9 - 1 µm seems out of reach
• 95717cfe - Reduce covariance ambitions

Compare with previous version

added 122 commits

• 95717cfe...98f2bdc5 - 101 commits from branch acts:master
• f333e4e2 - Reform checkCloseXyz into unified CHECK_XYZ macros
• c9007f6f - Replace BOOST_CHECK_CLOSE[_FRACTION] with CHECK_XYZ
• 1b3adef9 - Replace Eigen's isApprox with CHECK_XYZ
• 5a0aa684 - Replace BOOST_CHECK_SMALL with CHECK_SMALL or CHECK_CLOSE_ABS
• c93ff56b - Reduce usage of binary BOOST_CHECK
• 112849e1 - Try to eradicate BOOST_TEST from ACTS tests
• ea781052 - Reduce precision of DiscTrapezoidalBounds output
• 0b20983e - Tune up the on-surface tolerance a bit
• 5bfb819e - Tune up SolenoidBField tolerance + express it more logically
• a02c0312 - Tune up Jacobian tolerance a bit
• 38e2fbe7 - Tune up Propagator tolerance a bit
• bb249bca - Adapt and homogeneize unit conversion tolerances
• f17223ec - Tune up tolerances and homogeneize comparisons in BFieldMapUtilsTest
• 33299efd - Tune up tolerances of CylinderVolumeBounds
• 5befd60a - 0.1µm is impossible in f32
• 372de550 - 1e-6 relative precision can be too optimistic in f32
• b8770297 - 1 keV is unrealistic in f32
• 4c3e0e7a - Tune up ATLAS/Eigen comparison tolerance
• 9af60559 - Let's get real
• 7f50b225 - 1 µm seems out of reach
• 7e477910 - Reduce covariance ambitions

Compare with previous version

added 252 commits

• 7e477910...b8e95b22 - 232 commits from branch acts:master
• 684c1ec4 - Reform checkCloseXyz into unified CHECK_XYZ macros
• 6fed8f07 - Replace BOOST_CHECK_CLOSE[_FRACTION] with CHECK_XYZ
• 39f563c5 - Replace Eigen's isApprox with CHECK_XYZ
• 1f88b139 - Replace BOOST_CHECK_SMALL with CHECK_SMALL or CHECK_CLOSE_ABS
• a15cc8a2 - Reduce usage of binary BOOST_CHECK
• e6a32573 - Try to eradicate BOOST_TEST from ACTS tests
• fedd357d - Reduce precision of DiscTrapezoidalBounds output
• e9fa6f28 - Tune up the on-surface tolerance a bit
• 6ab4b424 - Tune up SolenoidBField tolerance + express it more logically
• 4886317b - Tune up Jacobian tolerance a bit
• 7b4a6529 - Tune up Propagator tolerance a bit
• 8a4a4ede - Adapt and homogeneize unit conversion tolerances
• 7648adb8 - Tune up tolerances and homogeneize comparisons in BFieldMapUtilsTest
• 20947f43 - Tune up tolerances of CylinderVolumeBounds
• 4f915032 - 0.1µm is impossible in f32
• d37e5385 - 1e-6 relative precision can be too optimistic in f32
• 28ab8854 - 1 keV is unrealistic in f32
• 91c8c6eb - Tune up ATLAS/Eigen comparison tolerance
• aaccfa8a - Let's get real
• e770cfe7 - Reduce covariance ambitions

Compare with previous version

added 13 commits

• fafa74d1 - Add a dedicated comparison function for covariance matrices
• 23d9531b - Reduce precision of DiscTrapezoidalBounds output
• 1914864e - Tune up the on-surface tolerance a bit
• 8feae16f - Tune up SolenoidBField tolerance + express it more logically
• 963d04ff - Tune up Jacobian tolerance a bit
• 5c70d3c2 - Adapt and homogeneize unit conversion tolerances
• 3900da14 - Tune up tolerances and homogeneize comparisons in BFieldMapUtilsTest
• b6dbb53c - Tune up tolerances of CylinderVolumeBounds
• f70d18bd - 0.1µm is impossible in f32
• b9225550 - 1e-6 relative precision can be too optimistic in f32
• 214ff41a - 1 keV is unrealistic in f32
• 3d71528e - Tune up ATLAS/Eigen comparison tolerance
• 1fb4058f - Let's get real

Compare with previous version

added 2 commits

• 25d71ce3 - Single-eV tolerance over GeV quantities is a bit too optimistic
• 3671fcb5 - Tune up propagation tolerances a little bit

Compare with previous version

added 2 commits

• a3e41c6c - Tune up covariance tolerance to match on-surface tolerance increase
• 814efae7 - Tune up spatial tolerances a bit more too

Compare with previous version

changed the description

For job clang_tidy on commit: 814efae7

Static analysis results: Accepted

ok	pattern	count	limit
✅	readability-inconsistent-declaration-parameter-name	0	0
✅	readability-named-parameter	0	0
✅	readability-container-size-empty	0	0
✅	modernize-use-using	0	0
✅	readability-braces-around-statements	0	0
✅	modernize-use-override	0	0
✅	readability-implicit-bool-cast	0	0
✅	modernize-use-default-member-init	0	0
✅	performance-unnecessary-value-param	0	0
✅	modernize-use-equals-default	0	0
✅	modernize-use-nullptr	0	0

Analysis results at: https://gitlab.cern.ch/api/v4/projects/14559/jobs/2986897/artifacts

added 30 commits

• 814efae7...95174080 - 5 commits from branch acts:master
• 0a8d7611 - Reform checkCloseXyz into unified CHECK_XYZ macros
• aa77d20b - Replace BOOST_CHECK_CLOSE[_FRACTION] with CHECK_XYZ
• 87642fe9 - Replace Eigen's isApprox with CHECK_XYZ
• ab58d5a3 - Replace BOOST_CHECK_SMALL with CHECK_SMALL or CHECK_CLOSE_ABS
• 3a58b25c - Reduce usage of binary BOOST_CHECK
• 9686ff75 - Try to eradicate BOOST_TEST from ACTS tests
• 7dc434fb - Add a dedicated comparison function for covariance matrices
• 15cd8f9a - Tune down overly optimistic Kalman fitter tolerances
• 1afc96dd - Add indices in test error output
• 3eb2c500 - Remove useless dTheta << 0 assignments
• dd14512e - Reduce precision of DiscTrapezoidalBounds output
• f8834de1 - Tune up the on-surface tolerance a bit
• f16dd703 - Tune up SolenoidBField tolerance + express it more logically
• 41784e65 - Tune up Jacobian tolerance a bit
• 89c7b40f - Adapt and homogeneize unit conversion tolerances
• 51c6aacf - Tune up tolerances and homogeneize comparisons in BFieldMapUtilsTest
• 9a8b3495 - Tune up tolerances of CylinderVolumeBounds
• 4111f1b7 - 0.1µm is impossible in f32
• f3a7daf2 - 1e-6 relative precision can be too optimistic in f32
• ec9b2a66 - 1 keV is unrealistic in f32
• c8f890b4 - Tune up ATLAS/Eigen comparison tolerance
• 33d08258 - Single-eV tolerance over GeV quantities is a bit too optimistic
• dfa3bc24 - Tune up propagation tolerances a little bit
• f6628f5b - Tune up covariance tolerance to match on-surface tolerance increase
• 73d8e138 - Tune up spatial tolerances a bit more too

Compare with previous version

added 17 commits

• 9128e69f - Add indices in test error output
• c7240832 - Remove useless dTheta << 0 assignments
• d63301fc - Reduce precision of DiscTrapezoidalBounds output
• 36e0ed42 - Tune up the on-surface tolerance a bit
• 3dcabc55 - Tune up SolenoidBField tolerance + express it more logically
• efee84bc - Tune up Jacobian tolerance a bit
• 91b041d5 - Adapt and homogeneize unit conversion tolerances
• 80e040d7 - Tune up tolerances and homogeneize comparisons in BFieldMapUtilsTest
• 5102775f - Tune up tolerances of CylinderVolumeBounds
• a88ad280 - 0.1µm is impossible in f32
• 773c2e3d - 1e-6 relative precision can be too optimistic in f32
• 591cfa95 - 1 keV is unrealistic in f32
• efbc7753 - Tune up ATLAS/Eigen comparison tolerance
• c13e616e - Single-eV tolerance over GeV quantities is a bit too optimistic
• dcbae578 - Tune up propagation tolerances a little bit
• 7c700db6 - Tune up covariance tolerance to match on-surface tolerance increase
• d34761dd - Tune up spatial tolerances a bit more too

Compare with previous version

For job clang_tidy on commit: d34761dd

Static analysis results: Accepted

ok	pattern	count	limit
✅	readability-inconsistent-declaration-parameter-name	0	0
✅	readability-named-parameter	0	0
✅	readability-container-size-empty	0	0
✅	modernize-use-using	0	0
✅	readability-braces-around-statements	0	0
✅	modernize-use-override	0	0
✅	readability-implicit-bool-cast	0	0
✅	modernize-use-default-member-init	0	0
✅	performance-unnecessary-value-param	0	0
✅	modernize-use-equals-default	0	0
✅	modernize-use-nullptr	0	0

Analysis results at: https://gitlab.cern.ch/api/v4/projects/14559/jobs/3082047/artifacts

added 39 commits

• d34761dd...1c53a20f - 14 commits from branch acts:master
• 65e429f1 - Reform checkCloseXyz into unified CHECK_XYZ macros
• 5fb59398 - Replace BOOST_CHECK_CLOSE[_FRACTION] with CHECK_XYZ
• 80d27691 - Replace Eigen's isApprox with CHECK_XYZ
• d4d67280 - Replace BOOST_CHECK_SMALL with CHECK_SMALL or CHECK_CLOSE_ABS
• f5f11f00 - Reduce usage of binary BOOST_CHECK
• 5c5c9d44 - Try to eradicate BOOST_TEST from ACTS tests
• 52e589db - Add a dedicated comparison function for covariance matrices
• 68a0e437 - Tune down overly optimistic Kalman fitter tolerances
• 94c4bd76 - Add indices in test error output
• 0c66667e - Remove useless dTheta << 0 assignments
• e0905e04 - Reduce precision of DiscTrapezoidalBounds output
• e68f4297 - Tune up the on-surface tolerance a bit
• 8033e1e0 - Tune up SolenoidBField tolerance + express it more logically
• c0e11be2 - Tune up Jacobian tolerance a bit
• 7786619b - Adapt and homogeneize unit conversion tolerances
• dce7acbf - Tune up tolerances and homogeneize comparisons in BFieldMapUtilsTest
• 337ee4b9 - Tune up tolerances of CylinderVolumeBounds
• 675d8920 - 0.1µm is impossible in f32
• 6f657ffe - 1e-6 relative precision can be too optimistic in f32
• a31039db - 1 keV is unrealistic in f32
• 8f40683f - Tune up ATLAS/Eigen comparison tolerance
• 4571529c - Single-eV tolerance over GeV quantities is a bit too optimistic
• 3d7b0178 - Tune up propagation tolerances a little bit
• 10001b42 - Tune up covariance tolerance to match on-surface tolerance increase
• f49f0901 - Tune up spatial tolerances a bit more too

Compare with previous version

For job clang_tidy on commit: f49f0901

Static analysis results: Accepted

ok	pattern	count	limit
✅	readability-inconsistent-declaration-parameter-name	0	0
✅	readability-named-parameter	0	0
✅	readability-container-size-empty	0	0
✅	modernize-use-using	0	0
✅	readability-braces-around-statements	0	0
✅	modernize-use-override	0	0
✅	readability-implicit-bool-cast	0	0
✅	modernize-use-default-member-init	0	0
✅	performance-unnecessary-value-param	0	0
✅	modernize-use-equals-default	0	0
✅	modernize-use-nullptr	0	0

Analysis results at: https://gitlab.cern.ch/api/v4/projects/14559/jobs/3090679/artifacts

added 20 commits

• 5b87858f - Replace exact float comparison with approximate one
• 88336499 - Tune down max step count
• 9e3b0362 - Use an exact computation
• 48614087 - Add some tolerance for end-of-world detection
• 833a360d - Clarify which volumes are involved
• fc3f8780 - Reduce precision of DiscTrapezoidalBounds output
• a8f71933 - Tune up the on-surface tolerance a bit
• 0008fce1 - Tune up SolenoidBField tolerance + express it more logically
• dff12721 - Tune up Jacobian tolerance a bit
• 0f231ffa - Adapt and homogeneize unit conversion tolerances
• 31999821 - Tune up tolerances and homogeneize comparisons in BFieldMapUtilsTest
• 67ae3a8f - Tune up tolerances of CylinderVolumeBounds
• e1e169f9 - 0.1µm is impossible in f32
• fe8795b4 - 1e-6 relative precision can be too optimistic in f32
• adb30bda - 1 keV is unrealistic in f32
• 3a81ba2e - Tune up ATLAS/Eigen comparison tolerance
• bf73f886 - Single-eV tolerance over GeV quantities is a bit too optimistic
• d5205425 - Tune up propagation tolerances a little bit
• 74b732b5 - Tune up covariance tolerance to match on-surface tolerance increase
• 48eeead9 - Tune up spatial tolerances a bit more too

Compare with previous version

added 36 commits

• 48eeead9...90aa1ce1 - 3 commits from branch acts:master
• 12216f79 - Reform checkCloseXyz into unified CHECK_XYZ macros
• 477e1a0c - Replace BOOST_CHECK_CLOSE[_FRACTION] with CHECK_XYZ
• 9b8f18bb - Replace Eigen's isApprox with CHECK_XYZ
• 176b2721 - Replace BOOST_CHECK_SMALL with CHECK_SMALL or CHECK_CLOSE_ABS
• f5b83623 - Reduce usage of binary BOOST_CHECK
• 55c34364 - Try to eradicate BOOST_TEST from ACTS tests
• 8e4343ec - Add a dedicated comparison function for covariance matrices
• 6418874c - Tune down overly optimistic Kalman fitter tolerances
• 80642d8f - Add indices in test error output
• 51e05711 - Remove useless dTheta << 0 assignments
• ccc87c3f - Replace exact float comparison with approximate one
• d196d080 - Tune up covariance tolerance
• ab1900e6 - Tune down max step count
• 2fedcac1 - Use an exact computation
• c631d2bb - Add some tolerance for end-of-world detection
• 08682b73 - Clarify which volumes are involved
• 2e7329b1 - Clean up stepper a bit
• db43b14d - Avoid endless oscillation
• 396d8cf0 - Reduce precision of DiscTrapezoidalBounds output
• c9c9b625 - Tune up the on-surface tolerance a bit
• 9723461a - Tune up SolenoidBField tolerance + express it more logically
• 47240cc8 - Tune up Jacobian tolerance a bit
• 634b3dc0 - Adapt and homogeneize unit conversion tolerances
• 71740fd0 - Tune up tolerances and homogeneize comparisons in BFieldMapUtilsTest
• 3036185a - Tune up tolerances of CylinderVolumeBounds
• 6fd912a7 - 0.1µm is impossible in f32
• 8028dade - 1e-6 relative precision can be too optimistic in f32
• 07d96b11 - 1 keV is unrealistic in f32
• 0f19f045 - Tune up ATLAS/Eigen comparison tolerance
• 2235cfe9 - Single-eV tolerance over GeV quantities is a bit too optimistic
• 38e1185a - Tune up propagation tolerances a little bit
• 10d9f9d5 - Tune up covariance tolerance to match on-surface tolerance increase
• 072160b2 - Tune up spatial tolerances a bit more too

Compare with previous version

added 18 commits

• e3c7c5d3 - Clarify which volumes are involved
• 512e2d65 - Clean up stepper a bit
• b014eeb7 - Avoid endless oscillation
• 75671498 - Reduce precision of DiscTrapezoidalBounds output
• e8d95b32 - Tune up the on-surface tolerance a bit
• 0a478fe7 - Tune up SolenoidBField tolerance + express it more logically
• e6652226 - Tune up Jacobian tolerance a bit
• d3a90ed9 - Adapt and homogeneize unit conversion tolerances
• bd5b23bc - Tune up tolerances and homogeneize comparisons in BFieldMapUtilsTest
• 768eadcb - Tune up tolerances of CylinderVolumeBounds
• 9c10fadb - 0.1µm is impossible in f32
• f21ba5b6 - 1e-6 relative precision can be too optimistic in f32
• 8a3b1e1e - 1 keV is unrealistic in f32
• b5964afe - Tune up ATLAS/Eigen comparison tolerance
• 7baf4541 - Single-eV tolerance over GeV quantities is a bit too optimistic
• 39760de1 - Tune up propagation tolerances a little bit
• c5e209d8 - Tune up covariance tolerance to match on-surface tolerance increase
• 46425e74 - Tune up spatial tolerances a bit more too

Compare with previous version