Option to build without "-march=native"

[paoloczi]

The code builds with "-march=native" which is best for performance but reduces portability. There are use cases where portability is more important than performance, and therefore it would be nice to also have the option to build without "-march=native".

[rvaser]

Implemented in the newest version (3.0.0). You can remove -marc=native from flags with -Dspoa_optimize_for_native=OFF when running cmake.

Best regards,
Robert

[paoloczi]

Thank you for the quick response!

[Hikoyu]

-mssse3 or -msse4.1 instead of -march=native clearly affect the performance of recent CPU (Haswell or later).
I benchmarked the calculation time under each complilation condition: -mssse3, -msse4.1, -mavx2, and -march=native.
The number of sequences was 30.
This test was performed on my Linux server (Ubuntu 18.04, Xeon E3-1240L v3).

[rvaser]

Thanks for the evaluation, -march=native is the default build parameter. I was wondering how did you manage to compile with -mssse3? I am looking at the Intel Instrinsics Guide (https://software.intel.com/sites/landingpage/IntrinsicsGuide/) and the function _mm_max_epi32 is supported in SSE4.1 and onwards.

[Hikoyu]

I just edited line 20 of 'CMakeLists.txt' and executed cmake with following options -DCMAKE_CXX_FLAGS=-fPIC -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -Dspoa_build_executable=ON.
Compiler is gcc-7.4.0.
The result of reverse-compiling by objdump showed that specifying -mssse3 caused little vectorization. (Of course, pmaxsd or vpmaxsd instructions were not included.)

[rvaser]

Ah, I forgot that spoa also has SISD code, not specifying sse4.1 or avx2 will fallback to it :)

[paoloczi]

My request for an option to build without -march=native was motivated by situations where portability of compiled code is more important than performance (thank you @rvaser for adding that option!). Spoa correctly makes -march=native the default.

@Hikoyu it would be nice and useful if you could add to your plot a curve for compilation without -march and without any -m options (this is what happens when spoa is built with -Dspoa_optimize_for_native=OFF). That results in code with maximum portability (that is you can run the code on a machine different from the one that built it). That would give us a quantitative measure of the performance cost of portability for spoa. Extrapolating from your curves, it could be close to a factor of two, but it would be nice to know.

[Hikoyu]

Sorry guys, previous benchmark results were strange (too slow). I might have done something inappropriate compilation.
So, I recompiled spoa-v3.0.0 for each condition without editing 'CMakeLists.txt' and retried the benchmark tests. I also performed depth-dependence test.

Cmake options were shown below.
native: -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -Dspoa_build_executable=ON -DCMAKE_CXX_FLAGS='-fPIC'
without native: -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -Dspoa_build_executable=ON -Dspoa_optimize_for_native=OFF -DCMAKE_CXX_FLAGS='-fPIC'
ssse3: -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -Dspoa_build_executable=ON -Dspoa_optimize_for_native=OFF -DCMAKE_CXX_FLAGS='-fPIC -mssse3'
sse4.1: -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -Dspoa_build_executable=ON -Dspoa_optimize_for_native=OFF -DCMAKE_CXX_FLAGS='-fPIC -msse4.1'
avx2: -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -Dspoa_build_executable=ON -Dspoa_optimize_for_native=OFF -DCMAKE_CXX_FLAGS='-fPIC -mavx2'

The results of without native and native coincided with those of ssse3 and avx2 respectively.
So, you cannot see the green and purple plots.

@rvaser
I got it. Probably, without native also used the SISD code?

@paoloczi
I got your point, but without native caused more than 3 times performance degradation according to my recent data. I recommend adding at least -msse4.1, which is available in Penryn (2nd generation of Intel Core2) or later CPUs.

[paoloczi]

Thank you @Hikoyu for doing this. The performance ratio between native and portable is 3 to 5, much greater than I expected. I use spoa in the Shasta long read assembler, and in that case I also build with -march=native by default. However, most users of this type of software expect to be able to download a pre-built release and use it on whatever machine they choose. For that reason the release build is done without -march=native, and the documentation explains that for better performance you should rebuild the code on the run machine. Not an ideal situation, but a reasonable compromise.

Fortunately for the Shasta assembler spoa takes only a small fraction of the total assembly time (5% or so when using -march=native). However, it is likely/possible that the rest of the assembly code will also be impacted. I have not yet had the time to do a comparison with/without -march=native for assembly of a large genome.

Any additional thoughts/suggestions/perspective on this?

[Hikoyu]

As you pointed out, -march=native should be removed for binary distribution. But, little vectorized SISD code appeared to be used when -Dspoa_optimize_for_native=OFF was specified in cmake.
Well, how old CPUs would you like to support?
For example, recent (2013-) Intel processors (Haswell or later) support avx2. Semi-old (2008-) Intel processors (Penryn or later) support sse4.1.
If you want to support older processors, just remove -march=native.
The other cases, I supposed that specifying -msse4.1 or -mavx2 instead of -march=native maintain the portability without losing performance.

[paoloczi]

Thanks @Hikoyu for your perspective. I suspect the next step on this for me would be to quantify the cost of portability for my case like you did for spoa, and then make a decision based on that information and your considerations.

[rvaser]

@Hikoyu yes, when compiling without -march=native the SISD code is used. Either sse4.1 or avx2 need to be set in order to run SIMD code. Also, thanks for the follow-up evaluation!

@paoloczi I am agreeing with @Hikoyu that you should consider shipping with -msse4.1 as most of CPUs nowadays support it.

[paoloczi]

Thanks @rvaser, I will definitely consider your suggestion for my next release, after some performance experiments.

On the other hand, here are some TensorFlow users complaining that it does not run on their machine without sse 4.1. I think the answer for these people should be "if you want to do this type of computation, get a new computer"...

[Hikoyu]

@paoloczi I think they should get a new computer because their machines are probably over 10 years old. In addition, 32-bit Intel processors do not support SSE4.1, but almost all 64-bit Intel processors (except for 1st generation of Intel Core2) support.

[rvaser]

A quick followup, v3.0.2 now has a new flag (spoa_optimize_for_portability) which compiles with -msse4.1. This flag is used in Bioconda in order to avoid problems tied with -march=native.