From 2d6fd6f00c3b43744c30ec31a8581c97b763391e Mon Sep 17 00:00:00 2001
From: Martin Bauer <martin.bauer@fau.de>
Date: Sun, 6 Oct 2019 12:53:10 +0200
Subject: [PATCH] UniformGridGenerated: parallel initialization for NUMA first
touch
---
.../UniformGridGenerated.cpp | 29 ++++++++++++-------
.../UniformGridGenerated.py | 13 +++++----
2 files changed, 26 insertions(+), 16 deletions(-)
diff --git a/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.cpp b/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.cpp
index 8ccfa107d..bbc904de3 100644
--- a/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.cpp
+++ b/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.cpp
@@ -53,9 +53,18 @@ int main( int argc, char **argv )
uint_t timesteps = parameters.getParameter<uint_t>( "timesteps", uint_c( 60 ));
const real_t shearVelocityMagnitude = parameters.getParameter<real_t>("shearVelocityMagnitude", 0.08);
+
+ auto pdfFieldAdder = [](IBlock* const block, StructuredBlockStorage * const storage) {
+ return new PdfField_T(storage->getNumberOfXCells(*block),
+ storage->getNumberOfYCells(*block),
+ storage->getNumberOfZCells(*block),
+ uint_t(1),
+ field::fzyx,
+ make_shared<field::AllocateAligned<real_t, 64>>());
+ };
+
// Creating fields
- //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 pdfFieldId = blocks->addStructuredBlockData<PdfField_T>(pdfFieldAdder, "pdfs");
BlockDataID velFieldId = field::addToStorage< VelocityField_T >( blocks, "vel", real_t( 0 ), field::fzyx );
pystencils::GenMacroSetter setterKernel(pdfFieldId, velFieldId);
@@ -78,21 +87,21 @@ int main( int argc, char **argv )
if( timeStepMode == "twoField")
{
timeLoop.add() << BeforeFunction(twoFieldComm, "communication" )
- << Sweep( pystencils::GenLbKernel(pdfFieldId), "LB stream & collide1" );
+ << Sweep( pystencils::GenLbKernel(pdfFieldId, omega), "LB stream & collide1" );
timeLoop.add() << BeforeFunction(twoFieldComm, "communication" )
- << Sweep( pystencils::GenLbKernel(pdfFieldId), "LB stream & collide2" );
+ << Sweep( pystencils::GenLbKernel(pdfFieldId, omega), "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" );
+ timeLoop.add() << Sweep( pystencils::GenLbKernel(pdfFieldId, omega), "LB stream & collide1" );
+ timeLoop.add() << Sweep( pystencils::GenLbKernel(pdfFieldId, omega), "LB stream & collide2" );
} else if ( timeStepMode == "aa") {
- timeLoop.add() << Sweep( pystencils::GenLbKernelAAEven(pdfFieldId), "AA Even" );
+ timeLoop.add() << Sweep( pystencils::GenLbKernelAAEven(pdfFieldId, omega), "AA Even" );
timeLoop.add() << BeforeFunction( aaPullComm )
- << Sweep( pystencils::GenLbKernelAAOdd(pdfFieldId), "AA Odd")
+ << Sweep( pystencils::GenLbKernelAAOdd(pdfFieldId, omega), "AA Odd")
<< AfterFunction( aaPushComm );
} else if ( timeStepMode == "aaKernelOnly") {
- timeLoop.add() << Sweep( pystencils::GenLbKernelAAEven(pdfFieldId), "AA Even" );
- timeLoop.add() << Sweep( pystencils::GenLbKernelAAOdd(pdfFieldId), "AA Odd");
+ timeLoop.add() << Sweep( pystencils::GenLbKernelAAEven(pdfFieldId, omega), "AA Even" );
+ timeLoop.add() << Sweep( pystencils::GenLbKernelAAOdd(pdfFieldId, omega), "AA Odd");
} else {
WALBERLA_ABORT("Invalid value for timeStepMode ");
}
diff --git a/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.py b/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.py
index 2afbd8d18..74315df21 100644
--- a/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.py
+++ b/apps/benchmarks/UniformGridGenerated/UniformGridGenerated.py
@@ -1,12 +1,11 @@
import sympy as sp
import pystencils as ps
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 = 1.6#sp.symbols("omega")
+omega = sp.symbols("omega")
omega_fill = sp.symbols("omega_:10")
options_dict = {
@@ -116,15 +115,17 @@ with CodeGeneration() as ctx:
# Sweeps
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)
+ generate_sweep(ctx, 'GenLbKernelAAEven', update_rule_aa_even, cpu_vectorize_info={'assume_aligned': True},
+ cpu_openmp=True, ghost_layers=1)
+ generate_sweep(ctx, 'GenLbKernelAAOdd', update_rule_aa_odd, cpu_vectorize_info={'assume_aligned': True},
+ cpu_openmp=True, 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_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)
+ generate_sweep(ctx, 'GenMacroSetter', setter_assignments, cpu_openmp=True)
+ generate_sweep(ctx, 'GenMacroGetter', getter_assignments, cpu_openmp=True)
# Communication
generate_pack_info_from_kernel(ctx, 'GenPackInfo', update_rule_two_field,
--
GitLab