SUSY HRDCOD=1 builds fail because sin/cos/atan are not constexpr (only CUDA fails, while C++ succeeds) #627

valassi · 2023-04-03T18:52:17Z

SUSY HRDCOD=1 builds fail because sin/cos/atan are not constexpr

See for instance susy gg to tt in PR #625, and in particular 586195d

The log says

The HRDCOD=1 build now fails with

ccache /usr/local/cuda-12.0/bin/nvcc  -O3  -lineinfo -I. -I../../src -I../../../../../tools -I/usr/local/cuda-12.0/include/ -DUSE_NVTX -gencode arch=compute_70,code=compute_70 -gencode arch=compute_70,code=sm_70 -use_fast_math -std=c++17  -ccbin /usr/lib64/ccache/g++ -DMGONGPU_FPTYPE_DOUBLE -DMGONGPU_FPTYPE2_DOUBLE -DMGONGPU_HARDCODE_PARAM -Xcompiler -fPIC -c gcheck_sa.cu -o gcheck_sa.o
../../src/Parameters_MSSM_SLHA2.h(310): error: expression must have a constant value
../../src/Parameters_MSSM_SLHA2.h(310): note #2703-D: cannot call non-constexpr function "atan(double) noexcept(true)"
/usr/local/cuda-12.0/include/crt/math_functions.h(4137): here
../../src/Parameters_MSSM_SLHA2.h(725): error: expression must have a constant value
../../src/Parameters_MSSM_SLHA2.h(725): note #2703-D: cannot call non-constexpr function "cos(double) noexcept(true)"
/usr/local/cuda-12.0/include/crt/math_functions.h(553): here
../../src/Parameters_MSSM_SLHA2.h(726): error: expression must have a constant value
../../src/Parameters_MSSM_SLHA2.h(726): note #2703-D: cannot call non-constexpr function "sin(double) noexcept(true)"
/usr/local/cuda-12.0/include/crt/math_functions.h(520): here
3 errors detected in the compilation of "gcheck_sa.cu".

The text was updated successfully, but these errors were encountered:

valassi · 2024-02-14T08:20:45Z

Update Feb 2024. Actually, it is only nvcc that complains that sin/cos/atan are not constexpr. The C++ only build (CUDA_HOME=none using the old setup) succeeds, and the runtime tests are ok.

Maybe we can try to hardcode the gcc implementation of constexpr sin/cos/atan into our nvcc code?

See #625 (comment)

…for mdl_G__exp__2 in FPTYPE=f 'make HRDCOD=0 FPTYPE=f -j' builds now succeed and the tests succeed too. 'make HRDCOD=0 FPTYPE=m -j' builds now also succeed and the tests succeed too. 'CUDA_HOME=none make HRDCOD=1' builds and tests also succeed with FPTYPE=d,f,m The only pending problem are CUDA HRDCOD=1 builds because of missing constexpr sin/cos/atan (madgraph5#627) This is probably ready to be backported to CODEGEN and eventually merged

…ons for HRDCOD=1 cuda builds (madgraph5#627) Builds still fail because assert is not constexpr (I need to extend the function to cover any value and not just 0 to pi/2)

valassi · 2024-02-28T12:59:03Z

I have made a first prototype in PR #625 here 46694af

Still need some fixes (extend the range beyond 0 to pi/2, add arctan).

The implementation is very simply based on a Taylor series expansion, which is easily done in constexpr.

…ons for HRDCOD=1 cuda builds (madgraph5#627) Builds still fail because assert is not constexpr (I need to extend the function to cover any value and not just 0 to pi/2). Also missing is atan.

…rig functions for HRDCOD=1 cuda builds (madgraph5#627) Extend coverage of sin/cos/tan to any value, and add an atan function. Now the build fails with the following: HRDCOD=1 make ccache /usr/local/cuda-12.0/bin/nvcc -Xcompiler -O3 -lineinfo -I. -I../../src -I/usr/local/cuda-12.0/include/ -DUSE_NVTX -gencode arch=compute_70,code=compute_70 -gencode arch=compute_70,code=sm_70 -use_fast_math -std=c++17 -ccbin /usr/lib64/ccache/g++ -DMGONGPU_FPTYPE_DOUBLE -DMGONGPU_FPTYPE2_DOUBLE -DMGONGPU_HARDCODE_PARAM -Xcompiler -fPIC -c -x cu CPPProcess.cc -o CPPProcess_cu.o ../../src/Parameters_MSSM_SLHA2.h(818): error: identifier "mg5amcGpu::Parameters_MSSM_SLHA2::mdl_I51x11" is undefined in device code

…builds (madgraph5#627): it was enought to define mdl_I51x11 as a "__device__" constexpr. NB: Now the build fully succeeds and runTest.exe also succeeds for HRDCOD=1!

valassi · 2024-02-28T16:19:36Z

The fix for HRDCOD=1 builds in SUSY is now complete

I implemented the relevant trig functions as constexpr using Taylor series
I added device to the relevant constexpr variables
I close this issue. The fix will be committed in PR fix susy gg_tt builds and add it to the repo #625.

(SUSY itself is almost completed, only some minor pieces missing to CODEGEN)

cc @oliviermattelaer @roiser

…mplementation of constexpr trig functions for HRDCOD=1 cuda builds (madgraph5#627) Also add it to output.py in CODEGEN

… from susyggtt.sa (madgraph5#627)

…onstexpr_math.h tests from susyggtt.sa (madgraph5#627)

… used madgraph5#627 Also add a comment about std::complex (replace is done twice) The differences in generated code are now the following: ______________________________________________________________________________ git diff /data/avalassi/GPU2023/madgraph4gpuBis/epochX/cudacpp/susy_gg_tt.sa/SubProcesses/P1_Sigma_MSSM_SLHA2_gg_ttx/CPPProcess.cc ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ epochX/cudacpp/susy_gg_tt.sa/SubProcesses/P1_Sigma_MSSM_SLHA2_gg_ttx/CPPProcess.cc /tmp/git-blob-dFUXLJ/CPPProcess.cc 0651b576632e9d50c3ed2bf9461bd41309c90821 100644 epochX/cudacpp/susy_gg_tt.sa/SubProcesses/P1_Sigma_MSSM_SLHA2_gg_ttx/CPPProcess.cc 0000000000000000000000000000000000000000 100644 79c79 < __device__ const fptype cIPD[3] = { (fptype)Parameters_MSSM_SLHA2::mdl_MT, (fptype)Parameters_MSSM_SLHA2::mdl_WT, (fptype)Parameters_MSSM_SLHA2::mdl_I51x11 }; --- > __device__ const fptype cIPD[2] = { (fptype)Parameters_MSSM_SLHA2::mdl_MT, (fptype)Parameters_MSSM_SLHA2::mdl_WT }; 83c83 < __device__ __constant__ fptype cIPD[3]; --- > __device__ __constant__ fptype cIPD[2]; 86c86 < static fptype cIPD[3]; --- > static fptype cIPD[2]; 505c505 < const fptype tIPD[3] = { (fptype)m_pars->mdl_MT, (fptype)m_pars->mdl_WT, (fptype)m_pars->mdl_I51x11 }; --- > const fptype tIPD[2] = { (fptype)m_pars->mdl_MT, (fptype)m_pars->mdl_WT }; 508c508 < gpuMemcpyToSymbol( cIPD, tIPD, 3 * sizeof( fptype ) ); --- > gpuMemcpyToSymbol( cIPD, tIPD, 2 * sizeof( fptype ) ); 511c511 < memcpy( cIPD, tIPD, 3 * sizeof( fptype ) ); --- > memcpy( cIPD, tIPD, 2 * sizeof( fptype ) ); 514c514 < //for ( i=0; i<3; i++ ) std::cout << std::setprecision(17) << "tIPD[i] = " << tIPD[i] << std::endl; --- > //for ( i=0; i<2; i++ ) std::cout << std::setprecision(17) << "tIPD[i] = " << tIPD[i] << std::endl; ______________________________________________________________________________ git diff /data/avalassi/GPU2023/madgraph4gpuBis/epochX/cudacpp/susy_gg_tt.sa/src/Parameters_MSSM_SLHA2.h ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ epochX/cudacpp/susy_gg_tt.sa/src/Parameters_MSSM_SLHA2.h /tmp/git-blob-fVhC4J/Parameters_MSSM_SLHA2.h 2805ef7d81bd8340b2a77b2ec16df7639165e30c 100644 epochX/cudacpp/susy_gg_tt.sa/src/Parameters_MSSM_SLHA2.h 0000000000000000000000000000000000000000 100644 23,24d22 < #include "constexpr_math.h" < 295c293 < __device__ constexpr double mdl_I51x11 = 1.; --- > constexpr double mdl_I51x11 = 1.; 370c368 < constexpr cxsmpl<double> mdl_bb = ( ( -mdl_mHd2 + mdl_mHu2 - mdl_MZ__exp__2 * constexpr_cos( 2. * mdl_beta ) ) * constexpr_tan( 2. * mdl_beta ) ) / 2.; --- > constexpr cxsmpl<double> mdl_bb = ( ( -mdl_mHd2 + mdl_mHu2 - mdl_MZ__exp__2 * constexpr_cos( 2. * mdl_beta ) ) * tan( 2. * mdl_beta ) ) / 2.; 801,802c799 < //const double mdl_I51x11 = Parameters_MSSM_SLHA2::getInstance()->mdl_I51x11; // fix HRDCOD=0 susy builds < const fptype mdl_I51x11 = cIPD[2]; // fix HRDCOD=0 susy builds --- > const fptype mdl_I51x11;

This was referenced Apr 3, 2023

Add an example of a calculation with nprocesses>1 #272

Closed

processes for the paper #344

Open

valassi mentioned this issue May 26, 2023

Updated workplan (May 2023) towards a first alpha release #671

Open

valassi mentioned this issue Feb 14, 2024

fix susy gg_tt builds and add it to the repo #625

Merged

valassi changed the title ~~SUSY HRDCOD=1 builds fail because sin/cos/atan are not constexpr~~ SUSY HRDCOD=1 builds fail because sin/cos/atan are not constexpr (only CUDA fails, while C++ succeeds) Feb 14, 2024

valassi self-assigned this Feb 28, 2024

valassi linked a pull request Feb 28, 2024 that will close this issue

fix susy gg_tt builds and add it to the repo #625

Merged

valassi closed this as completed Feb 28, 2024

valassi added a commit to valassi/madgraph4gpu that referenced this issue Feb 29, 2024

[susy2] in CODEGEN, add constexpr_math.h from susyggtt.sa, with the i…

e1e62eb

…mplementation of constexpr trig functions for HRDCOD=1 cuda builds (madgraph5#627) Also add it to output.py in CODEGEN

valassi added a commit to valassi/madgraph4gpu that referenced this issue Feb 29, 2024

[susy2] in CODEGEN, add to testmisc.cc the tests for constexpr_math.h…

4354adf

… from susyggtt.sa (madgraph5#627)

valassi added a commit to valassi/madgraph4gpu that referenced this issue Feb 29, 2024

[susy2] in CODEGEN cudacpp.mk, add optional quadmath dependency for c…

edc5005

…onstexpr_math.h tests from susyggtt.sa (madgraph5#627)

valassi mentioned this issue Mar 8, 2024

comprehensive set of fixes for SUSY (complete) and EFT (partial) - keep susy_gg_tt, susy_gg_t1t1, heft_gg_h in repo and test them in CI #824

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

SUSY HRDCOD=1 builds fail because sin/cos/atan are not constexpr (only CUDA fails, while C++ succeeds) #627

SUSY HRDCOD=1 builds fail because sin/cos/atan are not constexpr (only CUDA fails, while C++ succeeds) #627

valassi commented Apr 3, 2023

valassi commented Feb 14, 2024

valassi commented Feb 28, 2024

valassi commented Feb 28, 2024 •

edited

Loading

SUSY HRDCOD=1 builds fail because sin/cos/atan are not constexpr (only CUDA fails, while C++ succeeds) #627

SUSY HRDCOD=1 builds fail because sin/cos/atan are not constexpr (only CUDA fails, while C++ succeeds) #627

Comments

valassi commented Apr 3, 2023

valassi commented Feb 14, 2024

valassi commented Feb 28, 2024

valassi commented Feb 28, 2024 • edited Loading

valassi commented Feb 28, 2024 •

edited

Loading