This covers:
- What
acpphintis a variation on . . . - Where are the various README.md files?
- Licensing notes . . .
- Changelog . . .
acpphint is a variation on the original HINT serial and pthread benchmarks,
but written in C++ and with various choices specific to my use. The benchmark
program still generates files appropriate for gnuplot use, but not fully
matching the original HINT benchmark's output files.
I am not trying to collect/archive reports of results.
An example gnuplot from OverDrive 1000 example data was:
QUality ImProvement/sec (QUIPS) is based on Quality in a hierarchical
integration problem, Quality being 1/(high_bound-low_bound). So higher is
better. As the total size of the kernel vectors grows noticeably beyond a cache
size, the cache tends to become fairly ineffective. So the right hand side tends
to be more memory subsystem bandwidth limited for the stopping constraints that
I've used. (I used stopping-constraints that avoided using kernel vector sizes
that would cause paging, for example.)
_ull is for unsigned long long, _ul is for unsigned long. DSIZE is for
the Data Type. ISIZE is for the Indexing Type (not double or other floating
point types).
For more examples of this type of plot with some notes about the plots, see:
Some acpphint curves with notes, QUIPS vs. RAM
Another plot from the same data files was:
This tends to show the latency involved for small sized kernel vectors, where single threaded does not measure any thread creation time or the like but the multi-threaded cases do.
The various README.md files are:
README.md (this file)
acpphint_src/README.md (the main README.md)
acpphint_example_data/README.md (.gp gnulot usage too)
other_src_used/README.md (other source code used)
(I normally have a directory called acpphint as the top of the tree of the
acpphint source distribution.) It is the acpphint_src/README.md that covers
how to build and how to run for the way the examples are set up. Material about
using the example data files is in: acpphint_example_data/README.md .
The various LICENSE.md files are:
LICENSE.md (the overall license: GPLv2)
acpphint_src/LICENSE.md (GPLv2 for HINT code variations)
acpphint_example_data/LICENSE.md (None for the example data)
other_src_used/LICENSE.md (OpenBSD-vintage-ISC source)
Overall, each acpphint*-NAME program is licensed as GPLv2 (not earlier, not
later). But building the acpphint related code does build some
OpenBSD-vintage-ISC licensed source files, those under other_src_used/ instead
of under acpphint_src/. Some resulting object files are linked into the
acpphint*-NAME related programs (GPLv2 programs).
See the acpphint_src/LICENSE.md for the GPLv2 license text that applies to the
files in acpphint_src/ and to the binary files containing code from them. As I
have the things structured, programs that do not have names that start with
acpphint do not contain code from acpphint_src/ ( just from other_src_used/ ).
See the other_src_used/LICENSE.md for the OpenBSD-vintage-ISC license text
that applies to just the source files in other_src_used/ and to any binaries
not containing code from acpphint_src/ source files.
Note: The original HINT's copyright holder, Iowa State University Research
Foundation, Inc., reported that, despite how HINT was mistakenly distributed
back then, the intended license was GPLv2 without any "or later" status: exactly
GPLv2. This is what sets the license for the source code in acpphint_src/. (I
did not bother to make distinct licensing for the acpphint_src/makefile.*
examples.) That code's license in turn sets the overall license to also be GLPv2
(no earlier, no later), despite OpenBSD-vintage-ISC licensed source in
other_src_used/ also being built and linked in.
Changes
- Visual Studio experiments inspired C++ source code adjustments, not that Visual Studio is supported for how I want things to work as far as builds go.
Changes
- clang-tidy inspired adjustments.
Changes
-
Add based on timing of parallel threads only once they complete a barrier and only until there is a result (not wating for join and such). This allows getting an idea of the thread creation and destruction overhead consequnces. Adds *_thread_creation_untimed*.txt data files to what is generated, an addition vs. the original *_thread*.txt named files.
-
2024-Jan-26: adjust a *.gp file's details.
Additions (2024-Jan-26)
- Add 2 RPi5 FreeBSD makefile.* 's. Add the RPi5-FBSD/ example data directory and the contained prelimiary cpulockdown *.txt files. Add 2 *.gp files that involve use of the RPi5 *.txt files.
Additions (2024-Jan-28)
- Add 1 RPi5 RasPiOS64 for 2712 makefile.* . Add the RPi5-RasPiOS64_2712/ example data directory and the contained prelimiary cpulockdown *.txt files. Add 1 *.gp file for comparison to FreeBSD, based on not timing the thread creation/join times in order to avoid the extra source of varaibility.
Additions
- Add a *.pg file for comparing/contrasting using 8 threads (avoiding SMT pair use) for the 7950X3D vs. the ThreadRipper 1950X. The 1950X data are from an older context.
Later Changes
-
Adjust that *.pg to show log-QUIPS y-axis scaling.
-
Modernize varous *.[hc]* file copyright notice date-range ends. Also fix a couple of lines that had leading tabs instead of leading spaces.
More Additions
- Add some other such *.gp files for comparison/contrast use.
Changes
- Update handling of includes, generally following include-what-you-use . No functional change intended.
Additions
- Add 7950X3D makefile.*'s for libstdc++ use (via g++13 for now).
Changes
-
Use long double with std::to_chars when: 201611L <= __cpp_lib_to_chars Otherwise use just double with std::to_chars. (llvm's libc++ supports double but treats long double like double so far, even for amd64 and aarch64 were the long double's are distinct from double.)
This change is outside the code that is benchmarked.
Added later:
- Add *.gp files for the 192GiByte RAM tests.
Changes
- Rename the *.gp files for the 96GiByte RAM tests to be explicit about the 96GiByte RAM context.
Additions
-
Add makefile.7950X3D-* FreeBSD files for Ryzen 9 7950X3D with 192 GiBytes of lightly slower RAM instead of 96 GiBytes of slightly faster RAM. Also: FreeBSD main is now llvm17 based, including the system libc++.
-
Add example *.txt benchmark run files.
Additions
- Add makefile.7950X3D-* FreeBSD files for Ryzen 9 7950X3D benchmarking.
Changes to those additions
- Various typo fixes to the makefile.7950X3D-* files. Various additions of more benchmarking cases.
Additions of material for results
-
Add the *.txt benchmark run files to a new directory for the FreeBSD 7950X3D data.
-
Add RAM and time *.gp files for plotting some of the benchmarking cases.
Added later:
-
Add the *.gp and *.png for a basic 7950X3D plot for 8 locked down threads (and single-threaded for scale).
-
Add material in Some_accphint_curves_with_notes.md material that uses the added plot (.png file).
Additions
-
Add makefile.HoneyComb*-IPL32-FreeBSD_main_64bit-*clang++* that will produce acpphint programs that use the new lib32 for armv7 code. (2023-Jul-26)
-
Add *.txt file output for modern LP64 and ILP32 for aarch64(/armv7) FreeBSD main on the HoneyComb system, with ILP32 *.txt files going into the new:
acpphint_example_data/HoneyComb-FBSD-ILP32/
-
Add *.txt file output for modern LP64 and ILP32 for aarch64(/armv7) FreeBSD main on the WDK23 system, with ILP32 *.txt files going into the new:
acpphint_example_data/WDK23_CA78C-FBSD-ILP32/
(Added: 2023-Jul-27)
-
Add *.gp files for the new HoneyComb and WDK23 data. (Added: 2023-Jul-29)
Changes/removals
-
Update other_src_used/cpp_clockinfo.cpp to no longer use async, avoiding its potential policy consequences. This is associated with progressing to clang 16's libc++ that finally allows the alternative code in the form used.
Somewhat later: fix typos in instances of 2 variable names in some comments in some source files.
-
Update acpphint_src/acpphint_kernelrunners.cpp to no longer use async, avoiding its potential policy consequences. This is associated with progressing to clang 16's libc++ that finally allows the alternative code in the form used.
-
Where there has been FreeBSD clang++ 16 or g++ 13 coverage already, remove makefile.* that were for earlier versions. But for other platforms, keep the most recent around for reference.
-
Update the FreeBSD clang++ 16 makefile.* to reference a FreeBSD main-vintage that is LLVM 16 based, which allows avoiding async use for clang++ because of the system libc++ update involved. (async could impose threading managment policy that is not desired.) On FreeBSD, devel/llvm* ports use the system libc++ instead of having their own. So, prior clang++16 results were based on llvm 15's libc++ being in use.
Additions
-
Add makefile.*'s and such for looking at time measurement for full-feature, nolse, nolse and norcpc, norcpc, and CA72 style code (kernel/world/any-ports-used intended to be matching). Add an acpphint_example_data/WDK23_cpp_clockinfo_examples/ with the example cpp_clockinfo outputs.
-
Note: I settled on keeping most data just for: CA78Cnolse (for normal WDK23 use in my context), CA72 (use of media also used for cortex-a72 based systems), and CA78C (+lse and implicitly +rcpc that I started with but did not keep media around after gathering the data). In each case there was a kernel, world, ports (such as llvm16), and benchmark build involved: a full context matched for lse and (and rcpc) status. I've not done such for involving +norcpc in builds.
-
Add modern HoneyComb example data and supporting material for: clang++ 16 & libc++, g++ 13 with that libc++, g++ 13 with its libstdc++ . Both with CPU lock down and without it.
-
Add/update modern HoneyComb related *.gp files, including for comparison to WDK23_CA72 and WDK23_CA78Cnolse .
-
Update various *.gp files for HoneyComb vs. WDK23 CA78C* to also include 8 thread examples. This involved renaming the files to indicate the addition. complete the coverage of g++ 13 libc++ and libstdc++ combinations.
Changes/removals
-
Update CA78Cnolse examples based on the UFS context that now normally is based on +nolse for the context (kernel, world) and the benchmark.
-
Update the cpp_clockinfo code to have better output for the comparison's I was doing.
-
Adjust acpphint_src/makefile.common_part for doing some just cpp_clockinfo activity rather than the full benchmarking.
-
Remove materials that mixed kernels vs. world that had distinct compiling configurations: avoid dealing with the volume of stuff tied to the combinatorics.
-
Remove existing g++ examples as I discovered that the kernel+world configuration measured was far from optimal for the FreeBSD activity (buildworld, buildkernel, building ports). Have the makefile.* 's for g++ use be just for CA72 and CA78Cnolse (+nolse use). My normal environments are based on +nolse now.
-
Build and run the clang 16 WDK23_CA72 tests in a UFS cortex-a72 targeted kernel+world context and update to the results. The prior results were odd but I've not managed to ever reproduce the oddity. Some form of operator oddity/error for which I do not know the details.
-
Build and run the g++ 13 WDK23_CA72 tests in a UFS cortex-a72 targeted kernel+world context and update to the results, for both libc++ and libstdc++ .
-
Build and run the g++ 13 WDK23_CA78Cnolse tests in a UFS cortex-x1c/a78c+nolse targeted kernel+world context and update to the results, for both libc++ and libstdc++ . This involved adding a missing character to a makefile.* filename as well. (I also added missing "-cpulockdown" text in various makefile.* files.)
-
Have WDK23 *.gp file names be explicit about when libc++ and libstd++ are both involved. Add 2 g++ related WDK23 RAM vs. time *.gp files (for -cpulockdown contexts).
-
Update the HoneyComb *-g++_13_O3lto-libc++* materials after fixing the makefile.* files involved to have the intended lists of tests.
-
Fix the g++ CA78Cnolse makefile.*'s to actually specify the +nolse . The g++ tests will all need to be rerun and the data updated. This is likely what explains some of the odd example data curve structure.
-
Remove *.gp files that mixed both HoneyComb vs. WDK23_CA78C*'s and the compilers/libraries used.
-
Update to have g++ 13 WSDK23_CA78Cnolse example data that is actually based on using +nolse.
Additions
-
Start to add Windows Dev Kit 2023 based examples (WDK23 naming here), with code generation tailored to the cortex-a78c/cortex-x1c involved. The kernel and world for the context were also tailored so tailored is the overall type context. The prepliminary data was from a ZFS context (unusual procedure) instead of UFS (normal). For just clang++16 based, I'm updating to a UFS base test context for now. (g++13 likely will be updated later.) No .gp files yet.
-
Set up to later get some modern HoneyComb based data for comparison and contrast use. (Not built or run yet: build in/for a cortex-a72 context.) Similarly for testing cortex-a72 code on the WDK23.
-
Added later: UFS WDK23 CA72 (kernel/world/acpphint). No replacement of ZFS context example data with UFS context data yet for CA78C(/X1C).
-
Added later: Start to explore the kernel being targetted one way and the rest (world/acpphint) being the other. First example: UFS WDK23 with CA78C(/X1C) kernel and CA72 world. This adds makefile.* files with appropriate naming. Then add UFS WDK23 with CA72 kernel and CA78C(/X1C) world.
-
Added later: UFS cpu clockdown CA78C+nolse example data.
-
Added later: UFS cpu clockdown world CA78C+nolse_norcpc example data for CA72 kernel.
Changes
-
Update to using lang/gcc13 and devel/clang16 based builds for the ones I intend to test. (libc++ is still from llvm15 in the system).
-
Later: fix naming error in 2 makefile.* files. Even later: fix more typos in makefile.* files.
-
Later: I've rerun the 4 CA78C cpu cpuset-restricted examples that were for lack of cpu lockdowns --because the original data looked to possibly have had competing non-benchmark activity that made for problematical data. The result was somewhat less messy for the plot curves --but still messy. No messes observed for any cpu clock down testing so far. Looks like the wide variability is tied to thread migration activity among the cpuset restricted set of cores.
-
Later: Various WDK23 related removals of likely unused material. Looks like I'll just be looking at time measuring for a while and will likely need to redo various things afterwards and possibly delete even more as not really appropraite long term.
Notes:
-
The LLVM15 libc++ include __type_traits/is_convertible.h needed handling of __is_convertible being a builtin, which g++13 has. g++13 -stdlib=libc++ does not compile successfully for libc++ use otherwise. FreeBSD main [so: 14] is still using LLVM 15.0.7 .
-
clang++15, clang++16, and g++13 all have issues with mistaken or incomplete/inaccurate default features for cortex-x1c and/or cortex-a78c, not even treating the 2 the same in places where ARM documents them as matching --with FreeBSD's boot reporting agreeing. (Mostly [completely?] they match.)
Changes
- Have quips_peak and quips_to_peak_ratio updated in more cases.
Additions
- Have the std::cout quips_to_peak_ratio line also report the numerator and denominator for the ratio so the quips and peak quips are explicitly visible as well.
Changes
-
Later: Adjust/rename most g++_11 *.gp's to be g++_12 ones instead. Also, mostly avoid the runs involving a command line cpuset or schedctl style of run.
-
Later: Adjust HoneyComb FreeBSD clang++ 15 make.*'s to identify modern FBSD vintage. Add files from runs to acpphint_example_data/HoneyComb-FBSD/ . Update *.gp's too. (One unrelated rename as well.)
Removals
- Later: Remove most older makefile.*'s that do not match one of the now 12 *.gp files at all. That includes clang++_12 files.
Additions
-
Add Fedora makefile.*'s for Running on a HoneyComb. Add files from a run in a new acpphint_example_data/HoneyComb-Fedora/ directory for this combination.
-
Later: add 2 such *.pg files. Adjust/rename a couple of others, adding time variants to the existing RAM ones as well.
-
Later: Add a missing time *.gp file to go with a RAM one.
-
Later: Add FreeBSD libstdc++ data, no *.gp files yet.
Removals
- Delete various unused HoneyComb FreeBSD g++ 11 based data files, now that a more modern FreeBSD/g++ combination has data files.
Changes
-
Adjust a FreeBSD libstdc++ makefile.*'s for Running on a HoneyComb. Add files from a run into acpphint_example_data/HoneyComb-FBSD/ . Also add *.gp files for this.
-
Adjust FreeBSD libc++ makefile.*'s for Running on a HoneyComb. (The variant for still using g++12.) No data files or *.gp files yet.
-
Later: Add data files for libc++ .
-
Later: Adjust libc++ related cpulockdown make.* file. Add files from a run into acpphint_example_data/HoneyComb-FBSD/ . No *.gp files for it yet.
-
Later: Only have 3 NetBSD related *.gp files: Just for showing the NetBSD data oddities. Also use 10 for the time upper bounds in general for *.gp files, not 1.6.
-
Later: Adjust/rename 3 g++_11 *.gp files to instead be g++_12 ones.
-
Later: Modernize the FreeBSD libstdc++ make.* files to match the current context.
Additions
-
Add NetBSD makefile.*'s for Running on a HoneyComb. Add files from a run in a new acpphint_example_data/HoneyComb-NBSD/ directory for this combination.
-
Later: add such *.pg files. Also rescale in various *.gp files.
Changes
-
Update other_src_used/sys_cpubinding.cpp to have support for cpu lock down for NetBSD.
-
Update a couple of the HoneyComb FreeBSD make.*'s to add 2 additional test runs, matching what NetBSD has.
Additions
-
Add OpenBSD makefile.* for Running on a HoneyComb with a kernel without DIAGNOSTIC variant of GENERIC. Add files from a run for this type of context.
-
Add vectors_total_bytes/median_mean_sec_per_lap std::cout line.
Additions
-
Add OpenBSD makefile.*'s for Running on a HoneyComb. Add files from a run in a new acpphint_example_data/HoneyComb-OBSD/ directory for this combination (but there are not any *.gp files using them yet).
-
Later: add a *.pg file that uses the 1, 2, and 4 thread OpenBSD data compared/contrasted with cpuset based FreeBSD data (but not with cpu lockdowns).
Changes
- Correct acpphint_src/Usage_differences_with_original_HINTs.md notes about how the command line arguments to acpphint are handled.
Additions
-
Add new Fedora 37 Server makefile.*'s for Running on a 2018 macMini under what macOS provides for a VM. Add files from runs in a new acpphint_example_data/2018macMiniVM-Fedora/ directory for this combination (but there are not any *.gp files using them yet).
-
Later: Add the 2018macMniVM *.gp files. Also add the missing HoneyComb RAM vs_cpulockdown *.gp file.
-
Later: Add makefiles, data files, and *.gp files for using g++/libc++ for the 2018macMiniVM . (g++ 12 as it turns out.)
Changes
-
Make some *.gp content more uniform.
-
Update the .gitignore files, in part for future C++ module experimentation.
-
Later: removal of deleted data files from being gnerated by the makefile files for the 2018macMiniVM.
-
Later: fix the libstdc++ spelling in the makefile.2018macMiniVM* names.
-
Remove 2 misnamed 2018macMIniVM data files that are not to be regenerated.
Additions
- Add some new HoneyComb makefile.*'s, matching the more recent context (preliminary).
Changes
- Update the FreeBSD code in other_src_used/sys_cpubinding.cpp to use the newer FreeBSD CPU_AND macro (now 3 arguments, not 2 arguments).
Removals
-
Delete most of the *.gp files and update to have just 12 of the historical g++ 11 and clang++ 12 ones and 2 of the older OverDrive ones. MacchDblShot and RPi4B are only involved in Cortex-A72 comparisons, not by themselves. The ThreadRipper1950X data, HoneyComb data, and old OverDrive data are primary now.
-
Delete lots of old *.txt data files --and even some whole directories of such. Keep at least what is referenced by the 12 *.gp files.
Additions
-
Add HoneyComb tests and related files.
-
Add some code based on C++20 --and some C++17 that is now available: The output of the scales vs. the known approximation to the result now uses std::to_chars for the floating point output when available.
Changes
-
A round of updating to more modern FreeBSD and, for the HoneyComb and the MACCHIATObin Double Shot, Fedora 34. Only FreeBSD is tested in other contexts this time.
-
Add/replace gnuplot .gp files (for exploring just the new data and the older OverDrive 1000 data that I've kept).
-
Update to g++ 11 use.
-
For just HoneyComb, also provide tests and supporting files for the main FreeBSD system clang++ (12).
-
Update minor things in the sources to silence new warnings and the like.
-
Update the FreeBSD "cpu lock down" code in other_src_used to work within just whatever cpuset is active for the process.
-
Update various *.md files to track what has been going on.
Removals
-
The OverDrive 1000 and the other PowerMac11,2 "bit the dust". I've kept some plots around that involve the OverDrive 1000, with just the matching data files, but not the makefile.OverDrive* files. The PowerMac files have all been removed.
-
Other than the OverDrive 1000 files that I've kept, older example data files, makefile.* files, *.gp files, and the like have been removed.
Additions
-
Actually get the FreeBSD and Linux cpu lock down code to be part of such builds and make the code actually not be rejected. Also have it work. (I do not distribute binaries at all --nor that linux specific source, avoiding spending time on the additional licensing criteria that may be involved. But I've checked that such linux code is possible.)
-
Add some FreeBSD head -r363590 data, g++10 -O3 LTO based examples. (Exploring -O2, -O3, and -O3 -flto did not find great differences, but I stuck with the last that I tried.) Matching makefile.* files and the like are involved. I still founds that g++ produced more performant benchmark code compared to clang++ for what I looked at.
-
Add some Fedora 32 ThreadRipper 1950X data, g++10 -O3 LTO based. Add some ubuntu 20.04.1 RPi4B data, g++9 -O3 LTO based.
-
Add some gnuplot .gp files for exploring the new data.
Changes
- Update minor things in the sources to silence warnings after having added -Wpedantic -Wall -Wextra . Echo more information in the cout material, mostly explicit values in involved in the reported relationships.
Removals
-
Older example data files, makefile.* files, *.gp files, and the like have mostly been removed. Some of the old makefile.* files are around as examples.
-
The PowerMac11,2 that allowed all memory slots to be used "bit the dust", as did the Pine64+ 2GB. As such I'm reverting to no coverage of old PowerMacs or the Pine64+ 2GB. (FreeBSD has 32-bit power pc messed up currently.)
Additions
- Add FreeBSD head -r356426 data for Rock64, MacchDblShot, and OverDrive 1000. Add some gnuplot .gp files for comparing older examples with -r356426 based examples.
Changes
- Update various gnuplot .gp files to use FreeBSD head -r356426 based data.
Removals
- None.
Additions
- Add FreeBSD head -r356426 powerpc64 (PowerMac11,2) results. This is an ELFv2 ABI context, instead of the ELFv1 context that the prior results were for. (No g++9 vs. system-libc++ ABI mismatches showed up, unlike for the earlier 32-bit powerpc test of -r356246 . 32-bit was not trying to update its ABI but did switch to the system being clang/clang++ based --a new mismatch resulted.)
Changes
- None.
Removals
- None.
Additions
- Add RPi4B Ubuntu 19.10.1 server results. Added some ThreadRipper .gp files that had not not been checked in.
Changes
- Also updated to have the removal of a comment that for which the update had not been checked in.
Removals
- None.
Additions
- Add dual socket PowerMac3,6 G4 data, but based on -maix-struct-return to match the odd(?) clang++ code generation for 32-bit powerpc. (May well have to replace later if clang++ is "corrected" (but that would be another ABI change).
Changes
- Fix various typos and out of date comments.
Removals
- None.
Additions
- Add FreeBSD pg_ps enabled example of stopping once having reached 6sec+ for the median trial time on the ThreadRipper 1950X. The previous examples were for pg_ps disabled (so large page mappings disabled). The 6+ seconds gets to a little under 95.77 GiBytes.
Changes
- None.
Removals
- None.
Additions
- Copy prior ThreadRipper 1950X data to a directory for it being a re-run and put in place new data from a just-after-boot run for comparison.
Changes
- Where the original data is found.
Removals
- None.
Additions
- Add STOPTM_ALT macro for using -DSTOPTM_ALT=... to compile-time control the median trial time stopping criterion. Also add NSAMP_ALT for using -DNSAMP_ALT=... in case it is needed to reach the STOPTM_ALT value. Add more ThreadRipper1950X example material based on using these. Still no .gp or .png for the ThreadRipper additions.
Changes
- Of course, the STOPTM_ALT and NSAMP_ALT additions are changes to acpphint_kernelsamplers.cpp .
Removals
- None.
Additions
- Add examples from a FreeBSD ThreadRipper 1950X with 128 GiByte of RAM. BIOS (well, Ryzen Master) NUMA vs. nonNUMA for the same head -r355777 based build, also pg_ps enabled (large page mappings allowed) vs. disabled. No exploration of round robin or other such for NUMA: just defaults.
Changes
- None.
Removals
- None.
Additions
- Add FreeBSD MACCHIATObin Double Shot vs. OverDrive 1000 and PowerMac11,2 .png files.
Changes
- None.
Removals
- None.
Additions
- Add FreeBSD MACCHIATObin Double Shot vs. PowerMac11,2 .gp file.
Changes
- Correct 0.1.12 date.
Removals
- None.
Additions
- Add FreeBSD MACCHIATObin Double Shot files, based on running in the 0.1.11 context.
Changes
- Adjust makefile.common_part to indicate v0.1.11 .
Removals
- None.
Additions
- None.
Changes
- Have the FreeBSD vs. ArchLinuxARM comparison .gp file list FreeBSD first, ArchLinuxARM second in the plots. With sysctl dev.cpu.0.freq=1296 FreeBSD is faster in some contexts. Also rename the file to list FreeeBSD apects before the matching ArchLinuxARM aspects.
- Adjust makefile.common_part to indicate v0.1.11 .
Removals
- None.
Additions
- None.
Changes
- Update Rock64 benchmark data to be based on manually having set: sysctl dev.cpu.0.freq=1296
- Adjust makefile.common_part to indicate v0.1.10 .
Removals
- None.
Additions
- Add Rock64 .gp files. Note: The benchmark data indicates that FreeBSD is not changing from 600MHz for the CPU to faster, such as around 1200MHz. This looks to be manually adjustable so I'll likely get new benchmark data later.
Changes
- Adjust makefile.common_part to indicate v0.1.9 .
Removals
- None.
Additions
- Add Rock64 FreeBSD 13 files, but not yet .gp files. These were counted as running under v0.1.7 .
Changes
- Adjust makefile.common_part to indicate v0.1.8 .
Removals
- None.
Additions
- Add Rock64 .gp files.
Changes
- Adjust makefile.common_part to indicate v0.1.7 .
Removals
- None.
Additions
- Add Rock64 example's material, no .gp files yet.
Changes
- Update future compiler options for MACCHIATObin and RPi4B to be explicit about the cpu types to target. Other minor text edits in the 2 makefiles. Materials were not regenerated.
- Adjust makefile.common_part to indicate v0.1.6 .
Removals
- None.
Additions
- Add a .gp file (for gnuplot use) that compares a MACCHIATObin Double Shot to a Raspberry Pi 4B: Both have an A72, but with very different memory subsystems, as well as different cpu clock rates.
Changes
- Adjust makefile.common_part to indicate v0.1.5 .
Removals
- None.
Additions
- Add iMac18,3 Corei5 example's material. (The .gp file, for gnuplot use, is untested.)
Changes
- Adjust makefile.common_part to indicate v0.1.4 .
Removals
- None.
Additions
- Add MACCHIATObin double shot example's material. (The .gp file, for gnuplot use, is untested.)
Changes
- Adjust makefile.common_part to indicate v0.1.3 .
Removals
- None.
Additions
- None beyond initial addition of files.
Changes
- Fix a typo in README.md to explore making a change.
- Adjust makefile.common_part to indicate v0.1.2 .
Removals
- None.