From c933bf255788551f28740370ec0b1e1653896e90 Mon Sep 17 00:00:00 2001
From: Martin Bauer <martin.bauer@fau.de>
Date: Sat, 5 Oct 2019 10:50:32 +0200
Subject: [PATCH] Corrections in UniformGridGenerated Benchmark
---
.../UniformGridGenerated/CMakeLists.txt | 5 +-
.../UniformGridGenerated/UniformGrid.prm | 16 ++---
.../UniformGridGenerated.cpp | 64 +++++++++++++++----
.../UniformGridGenerated.py | 41 +++++++-----
4 files changed, 86 insertions(+), 40 deletions(-)
diff --git a/apps/benchmarks/UniformGridGenerated/CMakeLists.txt b/apps/benchmarks/UniformGridGenerated/CMakeLists.txt
index 1e8028d39..836baeefa 100644
--- a/apps/benchmarks/UniformGridGenerated/CMakeLists.txt
+++ b/apps/benchmarks/UniformGridGenerated/CMakeLists.txt
@@ -2,7 +2,10 @@ waLBerla_link_files_to_builddir( "*.prm" )
waLBerla_python_file_generates(UniformGridGenerated.py
- GenLbKernel.cpp GenMacroGetter.cpp GenMacroSetter.cpp GenPackInfo.cpp GenDefines.h)
+ GenMacroGetter.cpp GenMacroSetter.cpp
+ GenPackInfo.cpp GenPackInfoAAPush.cpp GenPackInfoAAPull.cpp
+ GenLbKernel.cpp GenLbKernelAAEven.cpp GenLbKernelAAOdd.cpp
+ GenDefines.h)
foreach(config trt )
diff --git a/apps/benchmarks/UniformGridGenerated/UniformGrid.prm b/apps/benchmarks/UniformGridGenerated/UniformGrid.prm
index 622515557..ae22ab6b5 100644
--- a/apps/benchmarks/UniformGridGenerated/UniformGrid.prm
+++ b/apps/benchmarks/UniformGridGenerated/UniformGrid.prm
@@ -1,26 +1,22 @@
DomainSetup
{
blocks < 1, 1, 1 >;
- cellsPerBlock < 64, 64, 128 >;
+ cellsPerBlock < 256, 128, 128 >;
periodic < 1, 1, 1 >;
}
Parameters
{
- timesteps 200; // time steps of one performance measurement
- warmupSteps 10; // number of steps to run before measurement starts
- outerIterations 15; // how many measurements to conduct
+ timesteps 400; // time steps of one performance measurement
+ warmupSteps 1; // number of steps to run before measurement starts
+ outerIterations 1; // how many measurements to conduct
- vtkWriteFrequency 200; // write a VTK file every n'th step, if zero VTK output is disabled
+ vtkWriteFrequency 0; // write a VTK file every n'th step, if zero VTK output is disabled
cudaEnabledMPI false; // switch on if you have a CUDA-enabled MPI implementation
-
- timeStepStrategy kernelOnly; // can be: noOverlap, simpleOverlap, complexOverlap, kernelOnly
- innerOuterSplit < 8, 1, 1>; // slice-thickness that 'outer'-kernels process when overlapping
-
+ timeStepMode aaKernelOnly; // can be: noOverlap, simpleOverlap, complexOverlap, kernelOnly
remainingTimeLoggerFrequency 0; // interval in seconds to log the estimated remaining time
omega 1.8;
- initShearFlow 1;
useGui 0;
}
diff --git a/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.cpp b/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.cpp
index 69db6180b..8ccfa107d 100644
--- a/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.cpp
+++ b/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.cpp
@@ -1,13 +1,11 @@
#include "core/Environment.h"
#include "core/logging/Initialization.h"
-#include "core/math/Random.h"
#include "python_coupling/CreateConfig.h"
#include "python_coupling/PythonCallback.h"
#include "python_coupling/DictWrapper.h"
#include "blockforest/Initialization.h"
#include "field/vtk/VTKWriter.h"
#include "field/AddToStorage.h"
-#include "field/communication/PackInfo.h"
#include "blockforest/communication/UniformBufferedScheme.h"
#include "timeloop/all.h"
#include "core/timing/TimingPool.h"
@@ -17,11 +15,18 @@
#include "InitShearVelocity.h"
#include "GenDefines.h"
-#include "GenPackInfo.h"
-#include "GenLbKernel.h"
#include "GenMacroGetter.h"
#include "GenMacroSetter.h"
+#include "GenLbKernel.h"
+#include "GenLbKernelAAEven.h"
+#include "GenLbKernelAAOdd.h"
+
+#include "GenPackInfo.h"
+#include "GenPackInfoAAPush.h"
+#include "GenPackInfoAAPull.h"
+
+
using namespace walberla;
using PdfField_T = GhostLayerField< real_t, Stencil_T::Q >;
@@ -43,13 +48,14 @@ int main( int argc, char **argv )
Vector3<uint_t> cellsPerBlock = config->getBlock( "DomainSetup" ).getParameter<Vector3<uint_t> >( "cellsPerBlock" );
// Reading parameters
auto parameters = config->getOneBlock( "Parameters" );
- const std::string timeStepStrategy = parameters.getParameter<std::string>( "timeStepStrategy", "normal");
+ const std::string timeStepMode = parameters.getParameter<std::string>( "timeStepMode", "twoField");
const real_t omega = parameters.getParameter<real_t>( "omega", real_c( 1.4 ));
- const uint_t timesteps = parameters.getParameter<uint_t>( "timesteps", uint_c( 50 ));
+ uint_t timesteps = parameters.getParameter<uint_t>( "timesteps", uint_c( 60 ));
const real_t shearVelocityMagnitude = parameters.getParameter<real_t>("shearVelocityMagnitude", 0.08);
// Creating fields
- BlockDataID pdfFieldId = field::addToStorage< PdfField_T >( blocks, "pdfs", real_t( std::nan("") ), field::fzyx );
+ //BlockDataID pdfFieldId = field::addToStorage< PdfField_T >( blocks, "pdfs", real_t( std::nan("") ), field::fzyx );
+ BlockDataID pdfFieldId = field::addToStorage< PdfField_T >( blocks, "pdfs", 0.0, field::fzyx );
BlockDataID velFieldId = field::addToStorage< VelocityField_T >( blocks, "vel", real_t( 0 ), field::fzyx );
pystencils::GenMacroSetter setterKernel(pdfFieldId, velFieldId);
@@ -59,12 +65,38 @@ int main( int argc, char **argv )
for( auto & b : *blocks)
setterKernel(&b);
- SweepTimeloop timeLoop( blocks->getBlockStorage(), timesteps );
- blockforest::communication::UniformBufferedScheme< Stencil_T > communication( blocks );
- communication.addPackInfo( make_shared< pystencils::GenPackInfo >( pdfFieldId ) );
+ blockforest::communication::UniformBufferedScheme< Stencil_T > twoFieldComm(blocks );
+ twoFieldComm.addPackInfo(make_shared< pystencils::GenPackInfo >(pdfFieldId ) );
+
+ blockforest::communication::UniformBufferedScheme< Stencil_T > aaPullComm(blocks);
+ aaPullComm.addPackInfo(make_shared< pystencils::GenPackInfoAAPull>(pdfFieldId));
+
+ blockforest::communication::UniformBufferedScheme< Stencil_T > aaPushComm(blocks);
+ aaPushComm.addPackInfo(make_shared< pystencils::GenPackInfoAAPush>(pdfFieldId));
+
+ SweepTimeloop timeLoop( blocks->getBlockStorage(), timesteps / 2 );
+ if( timeStepMode == "twoField")
+ {
+ timeLoop.add() << BeforeFunction(twoFieldComm, "communication" )
+ << Sweep( pystencils::GenLbKernel(pdfFieldId), "LB stream & collide1" );
+ timeLoop.add() << BeforeFunction(twoFieldComm, "communication" )
+ << Sweep( pystencils::GenLbKernel(pdfFieldId), "LB stream & collide2" );
+
+ } else if ( timeStepMode == "twoFieldKernelOnly") {
+ timeLoop.add() << Sweep( pystencils::GenLbKernel(pdfFieldId), "LB stream & collide1" );
+ timeLoop.add() << Sweep( pystencils::GenLbKernel(pdfFieldId), "LB stream & collide2" );
+ } else if ( timeStepMode == "aa") {
+ timeLoop.add() << Sweep( pystencils::GenLbKernelAAEven(pdfFieldId), "AA Even" );
+ timeLoop.add() << BeforeFunction( aaPullComm )
+ << Sweep( pystencils::GenLbKernelAAOdd(pdfFieldId), "AA Odd")
+ << AfterFunction( aaPushComm );
+ } else if ( timeStepMode == "aaKernelOnly") {
+ timeLoop.add() << Sweep( pystencils::GenLbKernelAAEven(pdfFieldId), "AA Even" );
+ timeLoop.add() << Sweep( pystencils::GenLbKernelAAOdd(pdfFieldId), "AA Odd");
+ } else {
+ WALBERLA_ABORT("Invalid value for timeStepMode ");
+ }
- timeLoop.add() << BeforeFunction( communication, "communication" )
- << Sweep( pystencils::GenLbKernel(pdfFieldId, omega), "LB stream & collide" );
int warmupSteps = parameters.getParameter<int>( "warmupSteps", 2 );
int outerIterations = parameters.getParameter<int>( "outerIterations", 1 );
@@ -108,6 +140,14 @@ int main( int argc, char **argv )
WALBERLA_LOG_INFO_ON_ROOT( "Starting simulation with " << timesteps << " time steps" );
simTimer.start();
timeLoop.run();
+ /*
+ pystencils::GenLbKernelAAEven k1(pdfFieldId, omega);
+ pystencils::GenLbKernelAAOdd k2(pdfFieldId, omega);
+ for(int t=0; t < timesteps / 2; ++t)
+ { for( auto & b : *blocks) {
+ k1(&b);
+ k2(&b);
+ }}*/
simTimer.end();
WALBERLA_LOG_INFO_ON_ROOT( "Simulation finished" );
auto time = simTimer.last();
diff --git a/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.py b/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.py
index 7027cf7ee..2afbd8d18 100644
--- a/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.py
+++ b/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.py
@@ -4,8 +4,9 @@ from lbmpy.creationfunctions import create_lb_update_rule
from lbmpy.fieldaccess import StreamPullTwoFieldsAccessor
from pystencils_walberla import CodeGeneration, generate_pack_info_from_kernel, generate_sweep
from lbmpy.macroscopic_value_kernels import macroscopic_values_getter, macroscopic_values_setter
+from lbmpy.fieldaccess import AAEvenTimeStepAccessor, AAOddTimeStepAccessor
-omega = sp.symbols("omega")
+omega = 1.6#sp.symbols("omega")
omega_fill = sp.symbols("omega_:10")
options_dict = {
@@ -18,6 +19,7 @@ options_dict = {
'trt': {
'method': 'trt',
'stencil': 'D3Q19',
+ 'compressible': False,
'relaxation_rate': omega,
},
'mrt': {
@@ -74,16 +76,12 @@ const bool infoCsePdfs = {cse_pdfs};
with CodeGeneration() as ctx:
- accessor = StreamPullTwoFieldsAccessor()
- assert not accessor.is_inplace, "This app does not work for inplace accessors"
-
common_options = {
'field_name': 'pdfs',
'temporary_field_name': 'pdfs_tmp',
- 'kernel_type': accessor,
'optimization': {'cse_global': False,
- 'cse_pdfs': True,
- 'split': True}
+ 'cse_pdfs': False,
+ 'split': False}
}
config_name = ctx.config
noopt = False
@@ -101,9 +99,6 @@ with CodeGeneration() as ctx:
options.update(common_options)
options = options.copy()
- if noopt:
- options['optimization']['cse_global'] = False
- options['optimization']['cse_pdfs'] = False
if d3q27:
options['stencil'] = 'D3Q27'
@@ -112,20 +107,32 @@ with CodeGeneration() as ctx:
pdfs, velocity_field = ps.fields("pdfs({q}), velocity(3) : double[3D]".format(q=q), layout='fzyx')
options['optimization']['symbolic_field'] = pdfs
- update_rule = create_lb_update_rule(**options)
- vec = {'nontemporal': True, 'assume_aligned': True, 'assume_inner_stride_one': True}
+ update_rule_two_field = create_lb_update_rule(**options)
+ update_rule_aa_even = create_lb_update_rule(kernel_type=AAEvenTimeStepAccessor(), **options)
+ options['optimization']['split'] = True
+ update_rule_aa_odd = create_lb_update_rule(kernel_type=AAOddTimeStepAccessor(), **options)
+
+ vec = {'nontemporal': False, 'assume_aligned': True, 'assume_inner_stride_one': True}
# Sweeps
- generate_sweep(ctx, 'GenLbKernel', update_rule, field_swaps=[('pdfs', 'pdfs_tmp')])
- setter_assignments = macroscopic_values_setter(update_rule.method, velocity=velocity_field.center_vector,
+ generate_sweep(ctx, 'GenLbKernel', update_rule_two_field, field_swaps=[('pdfs', 'pdfs_tmp')])
+ generate_sweep(ctx, 'GenLbKernelAAEven', update_rule_aa_even, cpu_vectorize_info={'assume_aligned': True}, cpu_openmp=6, ghost_layers=1)
+ generate_sweep(ctx, 'GenLbKernelAAOdd', update_rule_aa_odd, cpu_vectorize_info={'assume_aligned': True}, cpu_openmp=6, ghost_layers=1)
+
+ setter_assignments = macroscopic_values_setter(update_rule_two_field.method, velocity=velocity_field.center_vector,
pdfs=pdfs.center_vector, density=1)
- getter_assignments = macroscopic_values_getter(update_rule.method, velocity=velocity_field.center_vector,
- pdfs=pdfs.center_vector, density=None)
+ getter_assignments = macroscopic_values_getter(update_rule_two_field.method, velocity=velocity_field.center_vector,
+ pdfs=pdfs.center_vector, density=None)
generate_sweep(ctx, 'GenMacroSetter', setter_assignments)
generate_sweep(ctx, 'GenMacroGetter', getter_assignments)
# Communication
- generate_pack_info_from_kernel(ctx, 'GenPackInfo', update_rule, cpu_vectorize_info={'instruction_set': None})
+ generate_pack_info_from_kernel(ctx, 'GenPackInfo', update_rule_two_field,
+ cpu_vectorize_info={'instruction_set': None})
+ generate_pack_info_from_kernel(ctx, 'GenPackInfoAAPull', update_rule_aa_odd, kind='pull',
+ cpu_vectorize_info={'instruction_set': None})
+ generate_pack_info_from_kernel(ctx, 'GenPackInfoAAPush', update_rule_aa_odd, kind='push',
+ cpu_vectorize_info={'instruction_set': None})
# Info Header
infoHeaderParams = {
--
GitLab