feat: SoA layout for vector #95

niermann999 · 2023-03-13T01:59:08Z

This adds a new vector wrapper class that allows for an SoA layout, in particular, to enable us to use the Vc::Vector type efficiently. The new storage/vector keeps an array-like data structure, that holds the vector elements (say, x, y, z). Both the array type, as well as the value type of the elements are templated (storage/vector< algebraic vector dim, aos/soa value type (scalar vs e.g SIMD vector), array-like storage for the algebraic vector elements >), so that combinations like these are possible:

3-dim AoS std::array based vector (not necessary, since this would essentially duplicate the array plugin, while begin more complicated): storage::vector<3, scalar_t, std::array>
3-dim AoS vertical vectorized (Not there, yet, but would re-implement the current vc_vc plugin, while making the extra vc_array4 wrapper superfluous): storage::vector<3, scalar_t, Vc::SimdArray>
3-dim SoA of size N, std::array based: storage::vector<3, std::array<scalar_t, N>, std::array>
3-dim vetorized SoA (in this PR): storage::vector<3, Vc::Vector<scalar_t>, std::array>

Also adds benchmarks to compare to the std::array and Eigen AoS plugins. The benchmarks contain a base class that holds the basic configuration (e.g. number of warmup samples) and a derived vector benchmark class that holds the samples of random inititialized vectors and is templated on the kind of vector operation to be benchmarked (unary and binary are possible). The benchmarks for the getter namespace also make use of the vector benchmark class. The benchmarks for single and double precision are then instantiated per plugin, including a prescription on how to produce a random vector.

niermann999 · 2023-03-13T02:01:59Z

~~This is still VERY WIP~~

niermann999 · 2023-03-16T13:38:11Z

Since this PR is already huge, the SoA layout for std::array will be done in a separate PR

krasznaa

There are tons that I don't understand in here... 😦 Could you possibly post a more detailed explanation of the goals into the PR? SoA is of course good. I just don't see yet how this code here is really meant to work... 😕

.gitignore

benchmarks/common/include/benchmark/common/benchmark_getter.hpp

math/vc_soa/include/algebra/math/impl/vc_soa_getter.hpp

niermann999 · 2023-03-17T12:38:04Z

OK, I added a bit more of an explanation :)

krasznaa

Are the issues on macOS really insurmountable? I really don't like how that code is disabled right now. 😦

At the same time, let's discuss this face to face at some point. I'd love to understand better where this thing is meant to be heading.

math/cmath/CMakeLists.txt

math/vc_soa/include/algebra/math/vc_soa.hpp

storage/common/include/algebra/storage/array_operators.hpp

storage/common/include/algebra/storage/vector.hpp

niermann999 · 2023-03-31T12:01:11Z

Are the issues on macOS really insurmountable? I really don't like how that code is disabled right now.

Me neither, but I am not sure how to debug it. Basically, as soon as I call e.g. Vc::atan2 the release build on xcode throws Illegal while running the tests. The debug test is fine though

benchmarks/common/include/benchmark/common/benchmark_getter.hpp

storage/common/include/algebra/storage/vector.hpp

krasznaa · 2023-10-07T13:59:41Z

P.S. I saw some "new" Eigen warnings from the CUDA build in my tests. Something to follow up on separately at one point. (Those may be only showing up on Windows, didn't test it rigorously.)

krasznaa · 2023-10-09T09:27:58Z

Note that unrelated to the templating issues I also had to make these changes:

diff --git a/benchmarks/common/include/benchmark/common/benchmark_vector.hpp b/benchmarks/common/include/benchmark/common/benchmark_vector.hpp
index 4000dda..2ddd055 100644
--- a/benchmarks/common/include/benchmark/common/benchmark_vector.hpp
+++ b/benchmarks/common/include/benchmark/common/benchmark_vector.hpp
@@ -71,7 +71,7 @@ struct vector_unaryOP_bm : public vector_bm<vector_t<scalar_t>> {
     const std::size_t n_warmup{this->m_cfg.n_warmup()};

     // Spin down before benchmark (Thread zero is counting the clock)
-    if (state.thread_index() == 0 and this->m_cfg.do_sleep()) {
+    if (state.thread_index() == 0 && this->m_cfg.do_sleep()) {
       std::this_thread::sleep_for(std::chrono::seconds(this->m_cfg.n_sleep()));
     }

@@ -113,7 +113,7 @@ struct vector_binaryOP_bm : public vector_bm<vector_t<scalar_t>> {
     const std::size_t n_warmup{this->m_cfg.n_warmup()};

     // Spin down before benchmark (Thread zero is counting the clock)
-    if ((state.thread_index() == 0) and this->m_cfg.do_sleep()) {
+    if ((state.thread_index() == 0) && this->m_cfg.do_sleep()) {
       std::this_thread::sleep_for(std::chrono::seconds(this->m_cfg.n_sleep()));
     }

MSVC just does not like the and keyword. 😦

niermann999 · 2023-10-09T09:46:19Z

Note that unrelated to the templating issues I also had to make these changes:

diff --git a/benchmarks/common/include/benchmark/common/benchmark_vector.hpp b/benchmarks/common/include/benchmark/common/benchmark_vector.hpp
index 4000dda..2ddd055 100644
--- a/benchmarks/common/include/benchmark/common/benchmark_vector.hpp
+++ b/benchmarks/common/include/benchmark/common/benchmark_vector.hpp
@@ -71,7 +71,7 @@ struct vector_unaryOP_bm : public vector_bm<vector_t<scalar_t>> {
     const std::size_t n_warmup{this->m_cfg.n_warmup()};

     // Spin down before benchmark (Thread zero is counting the clock)
-    if (state.thread_index() == 0 and this->m_cfg.do_sleep()) {
+    if (state.thread_index() == 0 && this->m_cfg.do_sleep()) {
       std::this_thread::sleep_for(std::chrono::seconds(this->m_cfg.n_sleep()));
     }

@@ -113,7 +113,7 @@ struct vector_binaryOP_bm : public vector_bm<vector_t<scalar_t>> {
     const std::size_t n_warmup{this->m_cfg.n_warmup()};

     // Spin down before benchmark (Thread zero is counting the clock)
-    if ((state.thread_index() == 0) and this->m_cfg.do_sleep()) {
+    if ((state.thread_index() == 0) && this->m_cfg.do_sleep()) {
       std::this_thread::sleep_for(std::chrono::seconds(this->m_cfg.n_sleep()));
     }

MSVC just does not like the and keyword. 😦

Yes, I realized :) It is already gone from the PR #97 , but I have to put it here as well

niermann999 · 2023-10-09T10:23:30Z

P.S. I saw some "new" Eigen warnings from the CUDA build in my tests. Something to follow up on separately at one point. (Those may be only showing up on Windows, didn't test it rigorously.)

Btw, what were those warnings about? I started seeing some in detray, too, although it had previously stopped after we rolled back to version 34.0 (and now started again)

krasznaa

With the Windows/MSVC issues gone, I'll try to find some time in the evening to have a look at the Mac issue. 🤔

In the future at one point the benchmarking code could also be simplified/unified a bit, but for this PR this setup should be fine as is...

storage/common/include/algebra/storage/vector.hpp

benchmarks/CMakeLists.txt

tests/vc_soa/vc_soa.cpp

krasznaa

Code-wise I don't have objections. Just some smaller cmake updates, and I'm happy for this to go in. 😄

benchmarks/CMakeLists.txt

storage/vc/CMakeLists.txt

tests/CMakeLists.txt

…n particular, to enable us to use the Vc::Vector type efficiently. The new storage/vector keeps an array-like data structure, that holds the vector elements (say, x, y, z). Both the array type, as well as the value type of the elements are templated (storage/vector< algebraic vector dim, aos/soa value type (scalar vs e.g SIMD vector), array-like storage for the algebraic vector elements >), so that combinations like these are possible: 3-dim AoS std::array based vector (not necessary, since this would essentially duplicate the array plugin, while begin more complicated): storage::vector<3, scalar_t, std::array> 3-dim AoS vertical vectorized (Not there, yet, but would re-implement the current vc_vc plugin, while making the extra vc_array4 wrapper superfluous): storage::vector<3, scalar_t, Vc::SimdArray> 3-dim SoA of size N, std::array based: storage::vector<3, std::array<scalar_t, N>, std::array> 3-dim vetorized SoA (in this PR): storage::vector<3, Vc::Vector<scalar_t>, std::array> Also adds benchmarks to compare to the std::array and Eigen AoS plugins. The benchmarks contain a base class that holds the basic configuration (e.g. number of warmup samples) and a derived vector benchmark class that holds the samples of random inititialized vectors and is templated on the kind of vector operation to be benchmarked (unary and binary are possible). The benchmarks for the getter namespace also make use of the vector benchmark class. The benchmarks for single and double precision are then instantiated per plugin, including a prescription on how to produce a random vector.

krasznaa

Everything should be resolved now... 🤔

niermann999 marked this pull request as ready for review March 16, 2023 13:36

niermann999 force-pushed the vectorized-algebra branch 2 times, most recently from f4d13c2 to 0007b20 Compare March 16, 2023 14:23

niermann999 requested review from krasznaa, beomki-yeo and stephenswat March 16, 2023 14:23

niermann999 force-pushed the vectorized-algebra branch from 0007b20 to 709cc16 Compare March 16, 2023 15:05

krasznaa reviewed Mar 17, 2023

View reviewed changes

.gitignore Outdated Show resolved Hide resolved

benchmarks/common/include/benchmark/common/benchmark_getter.hpp Outdated Show resolved Hide resolved

math/vc_soa/include/algebra/math/impl/vc_soa_getter.hpp Show resolved Hide resolved

niermann999 force-pushed the vectorized-algebra branch 2 times, most recently from 8715bae to 4ad985f Compare March 25, 2023 15:45

niermann999 mentioned this pull request Mar 29, 2023

Benchmarks Fix, main branch (2023.03.29.) #99

Merged

niermann999 force-pushed the vectorized-algebra branch from 4ad985f to e8855fd Compare March 29, 2023 13:18

niermann999 added the enhancement New feature or request label Mar 29, 2023

niermann999 mentioned this pull request Mar 29, 2023

feat: add a Vc SoA implementation for transform3 #97

Merged

niermann999 force-pushed the vectorized-algebra branch from 5bf6c0a to 623307a Compare March 30, 2023 11:49

niermann999 requested a review from krasznaa March 30, 2023 11:50

niermann999 force-pushed the vectorized-algebra branch from 623307a to c8fa79b Compare March 30, 2023 12:39

krasznaa reviewed Mar 30, 2023

View reviewed changes

niermann999 force-pushed the vectorized-algebra branch from c8fa79b to b382a6b Compare March 31, 2023 11:53

wermos reviewed Mar 31, 2023

View reviewed changes

benchmarks/common/include/benchmark/common/benchmark_getter.hpp Outdated Show resolved Hide resolved

niermann999 force-pushed the vectorized-algebra branch 3 times, most recently from 998a7cb to c0b49af Compare September 21, 2023 12:08

krasznaa reviewed Oct 7, 2023

View reviewed changes

storage/common/include/algebra/storage/vector.hpp Outdated Show resolved Hide resolved

niermann999 force-pushed the vectorized-algebra branch 4 times, most recently from 25a6bde to 18be9e6 Compare October 9, 2023 11:05

krasznaa mentioned this pull request Oct 9, 2023

Silence Eigen Warnings From CUDA on Windows, main branch (2023.10.09.) #110

Merged

niermann999 force-pushed the vectorized-algebra branch from 18be9e6 to 58b46de Compare October 9, 2023 11:43

krasznaa reviewed Oct 9, 2023

View reviewed changes

storage/common/include/algebra/storage/vector.hpp Outdated Show resolved Hide resolved

This was referenced Oct 10, 2023

Windows tests acts-project/covfie#2

Closed

Add CMake Preset for ARM64, main branch (2023.10.14.) #111

Merged

krasznaa reviewed Oct 14, 2023

View reviewed changes

benchmarks/CMakeLists.txt Show resolved Hide resolved

niermann999 force-pushed the vectorized-algebra branch 2 times, most recently from 1881907 to d4c423d Compare October 14, 2023 10:10

stephenswat mentioned this pull request Oct 16, 2023

Disable floating point contraction at link time VcDevel/Vc#359

Open

beomki-yeo reviewed Apr 5, 2024

View reviewed changes

tests/vc_soa/vc_soa.cpp Show resolved Hide resolved

krasznaa requested changes Apr 8, 2024

View reviewed changes

benchmarks/CMakeLists.txt Show resolved Hide resolved

storage/vc/CMakeLists.txt Outdated Show resolved Hide resolved

tests/CMakeLists.txt Outdated Show resolved Hide resolved

tests/CMakeLists.txt Outdated Show resolved Hide resolved

niermann999 force-pushed the vectorized-algebra branch from d4c423d to a4eac49 Compare April 11, 2024 15:43

niermann999 requested a review from krasznaa April 11, 2024 16:03

niermann999 force-pushed the vectorized-algebra branch from 9d3ed3d to d26be44 Compare April 12, 2024 11:27

niermann999 added 2 commits April 12, 2024 13:29

Apply comments

d26be44

krasznaa approved these changes Apr 12, 2024

View reviewed changes

niermann999 merged commit 9e684de into acts-project:main Apr 12, 2024
24 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: SoA layout for vector #95

feat: SoA layout for vector #95

niermann999 commented Mar 13, 2023 •

edited

Loading

niermann999 commented Mar 13, 2023 •

edited

Loading

niermann999 commented Mar 16, 2023

krasznaa left a comment

niermann999 commented Mar 17, 2023

krasznaa left a comment

niermann999 commented Mar 31, 2023

krasznaa commented Oct 7, 2023

krasznaa commented Oct 9, 2023

niermann999 commented Oct 9, 2023

niermann999 commented Oct 9, 2023

krasznaa left a comment

krasznaa left a comment

krasznaa left a comment

feat: SoA layout for vector #95

feat: SoA layout for vector #95

Conversation

niermann999 commented Mar 13, 2023 • edited Loading

niermann999 commented Mar 13, 2023 • edited Loading

niermann999 commented Mar 16, 2023

krasznaa left a comment

Choose a reason for hiding this comment

niermann999 commented Mar 17, 2023

krasznaa left a comment

Choose a reason for hiding this comment

niermann999 commented Mar 31, 2023

krasznaa commented Oct 7, 2023

krasznaa commented Oct 9, 2023

niermann999 commented Oct 9, 2023

niermann999 commented Oct 9, 2023

krasznaa left a comment

Choose a reason for hiding this comment

krasznaa left a comment

Choose a reason for hiding this comment

krasznaa left a comment

Choose a reason for hiding this comment

niermann999 commented Mar 13, 2023 •

edited

Loading

niermann999 commented Mar 13, 2023 •

edited

Loading