diff --git a/RTCP/Cobalt/CobaltTest/test/tManyPartTABOutput.cc b/RTCP/Cobalt/CobaltTest/test/tManyPartTABOutput.cc
index b3dbc3a8aaa00206857740876aefdf68911a2526..7f2ee0a62ad72f1d86d7ec53b2a4879f23846fb0 100644
--- a/RTCP/Cobalt/CobaltTest/test/tManyPartTABOutput.cc
+++ b/RTCP/Cobalt/CobaltTest/test/tManyPartTABOutput.cc
@@ -30,7 +30,7 @@
#include <Common/LofarLogger.h>
#include <CoInterface/Parset.h>
#include <GPUProc/Pipelines/BeamFormerPipeline.h>
-#include <GPUProc/SubbandProcs/BeamFormerSubbandProc.h>
+#include <GPUProc/cuda/SubbandProcs/SubbandProcOutputData.h>
#include <GPUProc/Station/StationInput.h>
#include <GPUProc/Storage/StorageProcesses.h>
@@ -44,8 +44,7 @@ using boost::str;
SmartPtr<SubbandProcOutputData> getTestSbCohData(const Parset& ps, gpu::Context& ctx,
unsigned blockIdx, unsigned sbIdx)
{
- // BeamFormedData is a sub-class of SubbandProcOutputData.
- BeamFormedData *bfData = new BeamFormedData(ps, ctx);
+ SubbandProcOutputData *bfData = new SubbandProcOutputData(ps, ctx);
bfData->blockID.block = blockIdx;
bfData->blockID.globalSubbandIdx = sbIdx;
diff --git a/RTCP/Cobalt/CobaltTest/test/tMultiPartTABOutput.cc b/RTCP/Cobalt/CobaltTest/test/tMultiPartTABOutput.cc
index 2f3e6c3719def70091ce26701c2b4ecdd576c449..e8473b121be5579b504461ada83a8652eef1d752 100644
--- a/RTCP/Cobalt/CobaltTest/test/tMultiPartTABOutput.cc
+++ b/RTCP/Cobalt/CobaltTest/test/tMultiPartTABOutput.cc
@@ -30,7 +30,7 @@
#include <Common/LofarLogger.h>
#include <CoInterface/Parset.h>
#include <GPUProc/Pipelines/BeamFormerPipeline.h>
-#include <GPUProc/SubbandProcs/BeamFormerSubbandProc.h>
+#include <GPUProc/cuda/SubbandProcs/SubbandProcOutputData.h>
#include <GPUProc/Station/StationInput.h>
#include <GPUProc/Storage/StorageProcesses.h>
@@ -44,8 +44,7 @@ using boost::str;
SmartPtr<SubbandProcOutputData> getTestSbIncohData(const Parset& ps, gpu::Context& ctx,
unsigned blockIdx, unsigned sbIdx)
{
- // BeamFormedData is a sub-class of SubbandProcOutputData.
- BeamFormedData *bfData = new BeamFormedData(ps, ctx);
+ SubbandProcOutputData *bfData = new SubbandProcOutputData(ps, ctx);
bfData->blockID.block = blockIdx;
bfData->blockID.globalSubbandIdx = sbIdx;
diff --git a/RTCP/Cobalt/GPUProc/src/CMakeLists.txt b/RTCP/Cobalt/GPUProc/src/CMakeLists.txt
index 68663c1f5faa564ad6d3e6e1e0a2f3aa2929a09f..44ee8edd9987e5ffffb66e5ff393ce6a0b24d604 100644
--- a/RTCP/Cobalt/GPUProc/src/CMakeLists.txt
+++ b/RTCP/Cobalt/GPUProc/src/CMakeLists.txt
@@ -56,7 +56,8 @@ if(USE_CUDA)
cuda/Pipelines/BeamFormerPipeline.cc
# cuda/Pipelines/UHEP_Pipeline.cc
cuda/SubbandProcs/SubbandProc.cc
- cuda/SubbandProcs/BeamFormerSubbandProc.cc
+ cuda/SubbandProcs/SubbandProcInputData.cc
+ cuda/SubbandProcs/SubbandProcOutputData.cc
cuda/SubbandProcs/BeamFormerFactories.cc
cuda/SubbandProcs/CorrelatorStep.cc
cuda/SubbandProcs/BeamFormerPreprocessingStep.cc
diff --git a/RTCP/Cobalt/GPUProc/src/SubbandProcs/BeamFormerSubbandProc.h b/RTCP/Cobalt/GPUProc/src/SubbandProcs/BeamFormerSubbandProc.h
deleted file mode 100644
index dac9468ff602722838a57ec33fae7203a1c43546..0000000000000000000000000000000000000000
--- a/RTCP/Cobalt/GPUProc/src/SubbandProcs/BeamFormerSubbandProc.h
+++ /dev/null
@@ -1,41 +0,0 @@
-//# BeamFormerSubbandProc.h
-//#
-//# Copyright (C) 2013 ASTRON (Netherlands Institute for Radio Astronomy)
-//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
-//#
-//# This file is part of the LOFAR software suite.
-//# The LOFAR software suite is free software: you can redistribute it and/or
-//# modify it under the terms of the GNU General Public License as published
-//# by the Free Software Foundation, either version 3 of the License, or
-//# (at your option) any later version.
-//#
-//# The LOFAR software suite is distributed in the hope that it will be useful,
-//# but WITHOUT ANY WARRANTY; without even the implied warranty of
-//# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
-//# GNU General Public License for more details.
-//#
-//# You should have received a copy of the GNU General Public License along
-//# with the LOFAR software suite. If not, see <http://www.gnu.org/licenses/>.
-//#
-//# $Id$
-
-// \file
-// Include the right GPU API include with our options.
-
-#ifndef LOFAR_GPUPROC_BEAM_FORMER_SUBBAND_PROC_H
-#define LOFAR_GPUPROC_BEAM_FORMER_SUBBAND_PROC_H
-
-#if defined (USE_CUDA) && defined (USE_OPENCL)
-# error "Either CUDA or OpenCL must be enabled, not both"
-#endif
-
-#if defined (USE_CUDA)
-# include <GPUProc/cuda/SubbandProcs/BeamFormerSubbandProc.h>
-#elif defined (USE_OPENCL)
-# include <GPUProc/opencl/SubbandProcs/BeamFormerSubbandProc.h>
-#else
-# error "Either CUDA or OpenCL must be enabled, not neither"
-#endif
-
-#endif
-
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/BeamFormerPipeline.cc b/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/BeamFormerPipeline.cc
index 70feebaf0b4a4361709a5d74809d0611a70a5503..0e8128425de5cca7f5a896df9a3f48688e9ef3a6 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/BeamFormerPipeline.cc
+++ b/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/BeamFormerPipeline.cc
@@ -37,7 +37,7 @@
#include <CoInterface/SmartPtr.h>
#include <CoInterface/Stream.h>
#include <CoInterface/TimeFuncs.h>
-#include <GPUProc/SubbandProcs/BeamFormerSubbandProc.h>
+#include <GPUProc/SubbandProcs/SubbandProc.h>
#include <GPUProc/gpu_wrapper.h>
#include <GPUProc/gpu_utils.h>
#include <GPUProc/global_defines.h>
@@ -151,7 +151,7 @@ namespace LOFAR
for (size_t i = 0; i < workQueues.size(); ++i) {
gpu::Context context(devices[i % devices.size()]);
- workQueues[i] = new BeamFormerSubbandProc(ps, context, factories, nrSubbandsPerSubbandProc);
+ workQueues[i] = new SubbandProc(ps, context, factories, nrSubbandsPerSubbandProc);
}
}
@@ -227,8 +227,6 @@ namespace LOFAR
// Process pool elements until end-of-output
while ((data = inputQueue.remove()) != NULL)
{
- BeamFormedData &beamFormedData = dynamic_cast<BeamFormedData&>(*data);
-
const struct BlockID id = data->blockID;
ASSERT( globalSubbandIdx == id.globalSubbandIdx );
ASSERT( id.block >= 0 ); // Negative blocks should not reach storage
@@ -267,7 +265,7 @@ namespace LOFAR
const size_t nrSamples = stokes.nrSamples;
// Our data has the shape
- // beamFormedData.(in)coherentData[tab][stokes][sample][channel]
+ // data->(in)coherentData[tab][stokes][sample][channel]
//
// To transpose our data, we copy a slice representing
// slice[sample][channel]
@@ -291,8 +289,8 @@ namespace LOFAR
MultiDimArray<float, 2> srcData(
boost::extents[nrSamples][nrChannels],
file.coherent
- ? beamFormedData.coherentData[file.coherentIdxInSAP][file.stokesNr].origin()
- : beamFormedData.incoherentData[file.incoherentIdxInSAP][file.stokesNr].origin(),
+ ? data->coherentData[file.coherentIdxInSAP][file.stokesNr].origin()
+ : data->incoherentData[file.incoherentIdxInSAP][file.stokesNr].origin(),
false);
// Copy data to block
@@ -366,13 +364,12 @@ namespace LOFAR
// Process pool elements until end-of-output
while ((data = inputQueue.remove()) != NULL) {
- BeamFormedData &beamFormedData = dynamic_cast<BeamFormedData&>(*data);
- CorrelatedData &correlatedData = beamFormedData.correlatedData;
+ CorrelatedData &correlatedData = data->correlatedData;
const struct BlockID id = data->blockID;
ASSERT( globalSubbandIdx == id.globalSubbandIdx );
- if (beamFormedData.emit_correlatedData) {
+ if (data->emit_correlatedData) {
ASSERT(ps.settings.correlator.enabled);
ASSERT(outputStream.get());
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/BeamFormerPipeline.h b/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/BeamFormerPipeline.h
index d330e07a84570fa2b9eef32ad167122d0c66445d..7ad287c2403c7ae4431bfe8e7d1f83ed7fc3c422 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/BeamFormerPipeline.h
+++ b/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/BeamFormerPipeline.h
@@ -28,6 +28,8 @@
#include "Pipeline.h"
+#include <MACIO/RTmetadata.h>
+
#include <GPUProc/SubbandProcs/BeamFormerFactories.h>
namespace LOFAR
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/Pipeline.cc b/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/Pipeline.cc
index d1af82ea14e71d5680f6923e0d9ed7a8805b06cf..bd9abc776ab41f4cdd8c1702e465fce10b51b569 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/Pipeline.cc
+++ b/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/Pipeline.cc
@@ -39,7 +39,6 @@
#include <GPUProc/gpu_utils.h>
#include <GPUProc/global_defines.h>
#include <GPUProc/Kernels/Kernel.h>
-#include <GPUProc/SubbandProcs/SubbandProc.h>
#include <InputProc/SampleType.h>
#include <InputProc/RSPTimeStamp.h>
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerCoherentStep.cc b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerCoherentStep.cc
index b3fa3e46d47c3fcf551c2d11c8bab64a515ddf3b..9f0ed9b8d7c431f070ba3573a83b5cc3d9d8bde4 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerCoherentStep.cc
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerCoherentStep.cc
@@ -32,8 +32,6 @@
#include <CoInterface/Parset.h>
#include "SubbandProc.h"
-
-#include "BeamFormerSubbandProc.h"
#include "BeamFormerCoherentStep.h"
#include <iomanip>
@@ -184,7 +182,7 @@ void BeamFormerCoherentStep::process(const SubbandProcInputData &input)
}
-void BeamFormerCoherentStep::readOutput(BeamFormedData &output)
+void BeamFormerCoherentStep::readOutput(SubbandProcOutputData &output)
{
if (nrCoherent(output.blockID) == 0)
return;
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerCoherentStep.h b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerCoherentStep.h
index e1bdbc3677456d7e53ca3cd1991f073af95513c6..12572f7953f7248dbc8fee579c92a5dde59514e4 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerCoherentStep.h
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerCoherentStep.h
@@ -33,7 +33,8 @@
#include <GPUProc/MultiDimArrayHostBuffer.h>
#include <CoInterface/BlockID.h>
-#include "SubbandProc.h"
+#include "SubbandProcInputData.h"
+#include "SubbandProcOutputData.h"
#include "ProcessStep.h"
#include <GPUProc/Kernels/BeamFormerKernel.h>
@@ -50,7 +51,6 @@ namespace LOFAR
{
//# Forward declarations
struct BeamFormerFactories;
- class BeamFormedData;
class BeamFormerCoherentStep: public ProcessStep
{
@@ -83,7 +83,7 @@ namespace LOFAR
void process(const SubbandProcInputData &input);
- void readOutput(BeamFormedData &output);
+ void readOutput(SubbandProcOutputData &output);
private:
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerFactories.cc b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerFactories.cc
index 2ec4a06d5703191cbfa0c92c44325c7adbdc8520..1b90590e4233c2bd52bed8a3c0754547646556a6 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerFactories.cc
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerFactories.cc
@@ -22,7 +22,6 @@
#include <lofar_config.h>
#include "BeamFormerFactories.h"
-#include "BeamFormerSubbandProc.h"
namespace LOFAR
{
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerIncoherentStep.cc b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerIncoherentStep.cc
index 2e3b13512974b1bcba5d75eec7bbe17436daa569..bc31c3a4a788311a7362943304cad93008c595da 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerIncoherentStep.cc
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerIncoherentStep.cc
@@ -32,7 +32,6 @@
#include <CoInterface/Parset.h>
#include "SubbandProc.h"
-#include "BeamFormerSubbandProc.h"
#include "BeamFormerIncoherentStep.h"
@@ -171,7 +170,7 @@ namespace LOFAR
}
- void BeamFormerIncoherentStep::readOutput(BeamFormedData &output)
+ void BeamFormerIncoherentStep::readOutput(SubbandProcOutputData &output)
{
if (nrIncoherent(output.blockID) == 0)
return;
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerIncoherentStep.h b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerIncoherentStep.h
index 33a4fe5b2ac19a5ec7001260989bced60331bf08..12587f3d1ef8fd81270f7056b4de6178544cc194 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerIncoherentStep.h
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerIncoherentStep.h
@@ -33,7 +33,8 @@
#include <GPUProc/MultiDimArrayHostBuffer.h>
#include <CoInterface/BlockID.h>
-#include "SubbandProc.h"
+#include "SubbandProcInputData.h"
+#include "SubbandProcOutputData.h"
#include "ProcessStep.h"
#include <GPUProc/Kernels/FFT_Kernel.h>
@@ -47,9 +48,6 @@ namespace LOFAR
{
namespace Cobalt
{
- //# Forward declarations
- class BeamFormedData;
-
class BeamFormerIncoherentStep : public ProcessStep
{
public:
@@ -81,7 +79,7 @@ namespace LOFAR
void process(const SubbandProcInputData &input);
- void readOutput(BeamFormedData &output);
+ void readOutput(SubbandProcOutputData &output);
private:
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerPreprocessingStep.h b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerPreprocessingStep.h
index edead2f3e7eb9a2bb9569e953967fb4ca8fc3818..5789e36452d810198f17dd333d24f3c54aa67ef1 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerPreprocessingStep.h
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerPreprocessingStep.h
@@ -32,7 +32,8 @@
#include <GPUProc/MultiDimArrayHostBuffer.h>
#include <CoInterface/BlockID.h>
-#include "SubbandProc.h"
+#include "SubbandProcInputData.h"
+#include "SubbandProcOutputData.h"
#include "ProcessStep.h"
#include <GPUProc/KernelFactory.h>
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerSubbandProc.cc b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerSubbandProc.cc
deleted file mode 100644
index d7fe3df399f2078ddcdb6b79359f89d94c6ef7e0..0000000000000000000000000000000000000000
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerSubbandProc.cc
+++ /dev/null
@@ -1,238 +0,0 @@
-//# BeamFormerSubbandProc.cc
-//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
-//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
-//#
-//# This file is part of the LOFAR software suite.
-//# The LOFAR software suite is free software: you can redistribute it and/or
-//# modify it under the terms of the GNU General Public License as published
-//# by the Free Software Foundation, either version 3 of the License, or
-//# (at your option) any later version.
-//#
-//# The LOFAR software suite is distributed in the hope that it will be useful,
-//# but WITHOUT ANY WARRANTY; without even the implied warranty of
-//# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
-//# GNU General Public License for more details.
-//#
-//# You should have received a copy of the GNU General Public License along
-//# with the LOFAR software suite. If not, see <http://www.gnu.org/licenses/>.
-//#
-//# $Id$
-
-#include <lofar_config.h>
-
-#include "BeamFormerSubbandProc.h"
-#include "BeamFormerFactories.h"
-
-#include <GPUProc/global_defines.h>
-#include <GPUProc/gpu_wrapper.h>
-
-#include <CoInterface/Parset.h>
-#include <ApplCommon/PosixTime.h>
-#include <Common/LofarLogger.h>
-
-#include <iomanip>
-
-
-namespace LOFAR
-{
- namespace Cobalt
- {
- BeamFormedData::BeamFormedData(
- const Parset &ps,
- gpu::Context &context) :
- coherentData(ps.settings.beamFormer.anyCoherentTABs()
- ? boost::extents[ps.settings.beamFormer.maxNrCoherentTABsPerSAP()]
- [ps.settings.beamFormer.coherentSettings.nrStokes]
- [ps.settings.beamFormer.coherentSettings.nrSamples]
- [ps.settings.beamFormer.coherentSettings.nrChannels]
- : boost::extents[0][0][0][0],
- context, 0),
-
- incoherentData(ps.settings.beamFormer.anyIncoherentTABs()
- ? boost::extents[ps.settings.beamFormer.maxNrIncoherentTABsPerSAP()]
- [ps.settings.beamFormer.incoherentSettings.nrStokes]
- [ps.settings.beamFormer.incoherentSettings.nrSamples]
- [ps.settings.beamFormer.incoherentSettings.nrChannels]
- : boost::extents[0][0][0][0],
- context, 0),
-
- correlatedData(ps.settings.correlator.enabled ? ps.settings.antennaFields.size() : 0,
- ps.settings.correlator.enabled ? ps.settings.correlator.nrChannels : 0,
- ps.settings.correlator.enabled ? ps.settings.correlator.nrSamplesPerChannel : 0,
- context),
- emit_correlatedData(false)
- {
- }
-
-
- BeamFormerSubbandProc::BeamFormerSubbandProc(
- const Parset &parset,
- gpu::Context &context,
- BeamFormerFactories &factories,
- size_t nrSubbandsPerSubbandProc)
- :
- SubbandProc(parset, context, nrSubbandsPerSubbandProc),
- prevBlock(-1),
- prevSAP(-1),
- inputCounter(context, "input")
- {
- // See doc/bf-pipeline.txt
- size_t devA_size = 0;
- size_t devB_size = 0;
-
- if (factories.correlator) {
- CorrelatorStep::Factories &cf = *factories.correlator;
-
- devA_size = std::max(devA_size,
- cf.firFilter ? cf.firFilter->bufferSize(FIR_FilterKernel::INPUT_DATA)
- : cf.delayAndBandPass.bufferSize(DelayAndBandPassKernel::INPUT_DATA));
- devB_size = std::max(devB_size,
- cf.correlator.bufferSize(CorrelatorKernel::INPUT_DATA));
- }
-
- if (factories.preprocessing) {
- devA_size = std::max(devA_size,
- factories.preprocessing->intToFloat.bufferSize(IntToFloatKernel::OUTPUT_DATA));
- devB_size = std::max(devB_size,
- factories.preprocessing->intToFloat.bufferSize(IntToFloatKernel::OUTPUT_DATA));
- }
-
- if (factories.incoherentStokes) {
- ASSERT(factories.preprocessing);
-
- /* incoherentStokes uses devA and devB, but the sizes provided b the preprocessing
- pipeline are already sufficient. */
- }
-
- // NOTE: For an explanation of the different buffers being used, please refer
- // to the document bf-pipeline.txt in the GPUProc/doc directory.
- devA.reset(new gpu::DeviceMemory(context, devA_size));
- devB.reset(new gpu::DeviceMemory(context, devB_size));
-
- //################################################
- // Create objects containing the kernel and device buffers
-
- if (factories.correlator) {
- correlatorStep = std::auto_ptr<CorrelatorStep>(
- new CorrelatorStep(parset, queue, context, *factories.correlator,
- devA, devB, nrSubbandsPerSubbandProc));
- }
-
- if (factories.preprocessing) {
- preprocessingStep = std::auto_ptr<BeamFormerPreprocessingStep>(
- new BeamFormerPreprocessingStep(parset, queue, context, *factories.preprocessing,
- devA, devB));
- }
-
- if (factories.coherentStokes) {
- coherentStep = std::auto_ptr<BeamFormerCoherentStep>(
- new BeamFormerCoherentStep(parset, queue, context, *factories.coherentStokes,
- devB));
- }
-
- if (factories.incoherentStokes) {
- incoherentStep = std::auto_ptr<BeamFormerIncoherentStep>(
- new BeamFormerIncoherentStep(parset, queue, context, *factories.incoherentStokes,
- devA, devB));
- }
-
-
- LOG_INFO_STR("Pipeline configuration: "
- << (correlatorStep.get() ? "[correlator] " : "")
- << (preprocessingStep.get() ? "[bf preproc] " : "")
- << (coherentStep.get() ? "[coh stokes] " : "")
- << (incoherentStep.get() ? "[incoh stokes] " : "")
- );
-
- // put enough objects in the outputPool to operate
- for (size_t i = 0; i < nrOutputElements(); ++i)
- {
- outputPool.free.append(new BeamFormedData(ps, context));
- }
- }
-
-
- void BeamFormerSubbandProc::processSubband( SubbandProcInputData &input,
- SubbandProcOutputData &_output)
- {
- BeamFormedData &output = dynamic_cast<BeamFormedData&>(_output);
-
- //*******************************************************************
- // calculate some variables depending on the input subband
- size_t block = input.blockID.block;
- unsigned SAP = ps.settings.subbands[input.blockID.globalSubbandIdx].SAP;
-
- //****************************************
- // Send inputs to GPU
- queue.writeBuffer(*devA, input.inputSamples, inputCounter, true);
-
- // Some additional buffers
- // Only upload delays if they changed w.r.t. the previous subband.
- if ((int)SAP != prevSAP || (ssize_t)block != prevBlock) {
- if (correlatorStep.get()) {
- correlatorStep->writeInput(input);
- }
-
- if (preprocessingStep.get()) {
- preprocessingStep->writeInput(input);
- }
-
- if (coherentStep.get()) {
- coherentStep->writeInput(input);
- }
-
- prevSAP = SAP;
- prevBlock = block;
- }
-
- // ************************************************
- // Start the GPU processing
-
- if (correlatorStep.get()) {
- output.correlatedData.blockID = input.blockID;
-
- correlatorStep->process(input);
- correlatorStep->readOutput(output.correlatedData);
- }
-
- if (preprocessingStep.get()) {
- preprocessingStep->process(input);
- }
-
- if (coherentStep.get())
- {
- coherentStep->process(input);
- coherentStep->readOutput(output);
- }
-
- if (incoherentStep.get())
- {
- incoherentStep->process(input);
- incoherentStep->readOutput(output);
- }
-
- // ************************************************
- // Do CPU computations while the GPU is working
-
- if (correlatorStep.get()) {
- correlatorStep->processCPU(input, output.correlatedData);
- }
-
- // Synchronise to assure that all the work in the data is done
- queue.synchronize();
- }
-
- void BeamFormerSubbandProc::postprocessSubband(SubbandProcOutputData &_output)
- {
- BeamFormedData &output = dynamic_cast<BeamFormedData&>(_output);
-
- if (correlatorStep.get()) {
- output.emit_correlatedData = correlatorStep->postprocessSubband(output.correlatedData);
- } else {
- output.emit_correlatedData = false;
- }
- }
-
- }
-}
-
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerSubbandProc.h b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerSubbandProc.h
deleted file mode 100644
index a5772ed0b83cbf439c48b7767b223b995cea4aba..0000000000000000000000000000000000000000
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/BeamFormerSubbandProc.h
+++ /dev/null
@@ -1,107 +0,0 @@
-//# BeamFormerSubbandProc.h
-//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
-//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
-//#
-//# This file is part of the LOFAR software suite.
-//# The LOFAR software suite is free software: you can redistribute it and/or
-//# modify it under the terms of the GNU General Public License as published
-//# by the Free Software Foundation, either version 3 of the License, or
-//# (at your option) any later version.
-//#
-//# The LOFAR software suite is distributed in the hope that it will be useful,
-//# but WITHOUT ANY WARRANTY; without even the implied warranty of
-//# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
-//# GNU General Public License for more details.
-//#
-//# You should have received a copy of the GNU General Public License along
-//# with the LOFAR software suite. If not, see <http://www.gnu.org/licenses/>.
-//#
-//# $Id$
-
-#ifndef LOFAR_GPUPROC_CUDA_BEAM_FORMER_SUBBAND_PROC_H
-#define LOFAR_GPUPROC_CUDA_BEAM_FORMER_SUBBAND_PROC_H
-
-#include <complex>
-#include <memory>
-
-#include <boost/shared_ptr.hpp>
-#include <Common/LofarLogger.h>
-#include <CoInterface/CorrelatedData.h>
-#include <CoInterface/Parset.h>
-
-#include <GPUProc/gpu_wrapper.h>
-
-#include <GPUProc/MultiDimArrayHostBuffer.h>
-#include <GPUProc/Pipelines/BeamFormerPipeline.h>
-
-#include "CorrelatorStep.h"
-#include "BeamFormerPreprocessingStep.h"
-#include "BeamFormerCoherentStep.h"
-#include "BeamFormerIncoherentStep.h"
-
-#include "SubbandProc.h"
-
-namespace LOFAR
-{
- namespace Cobalt
- {
- //# Forward declarations
- struct BeamFormerFactories;
-
- // Our output data type
- class BeamFormedData: public SubbandProcOutputData
- {
- public:
-
- MultiDimArrayHostBuffer<float, 4> coherentData;
- MultiDimArrayHostBuffer<float, 4> incoherentData;
-
- CorrelatorStep::CorrelatedData correlatedData;
- bool emit_correlatedData;
-
- BeamFormedData(const Parset &ps,
- gpu::Context &context);
- };
-
- class BeamFormerSubbandProc : public SubbandProc
- {
- public:
- BeamFormerSubbandProc(const Parset &parset, gpu::Context &context,
- BeamFormerFactories &factories,
- size_t nrSubbandsPerSubbandProc = 1);
-
- // Beam form the data found in the input data buffer
- virtual void processSubband(SubbandProcInputData &input,
- SubbandProcOutputData &output);
-
- // Do post processing on the CPU
- virtual void postprocessSubband(SubbandProcOutputData &output);
-
- private:
- // The previously processed SAP/block, or -1 if nothing has been
- // processed yet. Used in order to determine if new delays have
- // to be uploaded.
- ssize_t prevBlock;
- signed int prevSAP;
-
- // @{
- // Device memory buffers. These buffers are used interleaved. For details,
- // please refer to the document bf-pipeline.txt in the directory
- // GPUProc/doc.
- boost::shared_ptr<gpu::DeviceMemory> devA;
- boost::shared_ptr<gpu::DeviceMemory> devB;
- // @}
-
- PerformanceCounter inputCounter;
-
- std::auto_ptr<CorrelatorStep> correlatorStep;
- std::auto_ptr<BeamFormerPreprocessingStep> preprocessingStep;
- std::auto_ptr<BeamFormerCoherentStep> coherentStep;
- std::auto_ptr<BeamFormerIncoherentStep> incoherentStep;
- };
-
- }
-}
-
-#endif
-
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/CorrelatorStep.cc b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/CorrelatorStep.cc
index 88a5bd224a45d304fecd938d94556551a14e30a6..99b0a1bbbb7d2e913b7c6772bf9722de744e5429 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/CorrelatorStep.cc
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/CorrelatorStep.cc
@@ -21,6 +21,7 @@
#include <lofar_config.h>
#include "CorrelatorStep.h"
+#include "SubbandProc.h"
#include <GPUProc/global_defines.h>
#include <GPUProc/gpu_wrapper.h>
@@ -35,21 +36,6 @@ namespace LOFAR
{
namespace Cobalt
{
- CorrelatorStep::CorrelatedData::CorrelatedData(
- unsigned nrStations, unsigned nrChannels,
- unsigned maxNrValidSamples, gpu::Context &context)
- :
- MultiDimArrayHostBuffer<fcomplex, 4>(
- boost::extents
- [nrStations * (nrStations + 1) / 2]
- [nrChannels][NR_POLARIZATIONS]
- [NR_POLARIZATIONS],
- context, 0),
- LOFAR::Cobalt::CorrelatedData(nrStations, nrChannels,
- maxNrValidSamples, this->origin(),
- this->num_elements(), heapAllocator, 1)
- {
- }
CorrelatorStep::Factories::Factories(const Parset &ps, size_t nrSubbandsPerSubbandProc) :
@@ -95,7 +81,7 @@ namespace LOFAR
// First transform the flags to channel flags: taking in account
// reduced resolution in time and the size of the filter
- convertFlagsToChannelFlags(parset, inputFlags, flagsPerChannel);
+ SubbandProc::Flagger::convertFlagsToChannelFlags(parset, inputFlags, flagsPerChannel);
// Calculate the number of flags per baseline and assign to
// output object.
@@ -313,14 +299,14 @@ namespace LOFAR
}
- void CorrelatorStep::readOutput(CorrelatedData &output)
+ void CorrelatorStep::readOutput(SubbandProcOutputData &output)
{
// Read data back from the kernel
- queue.readBuffer(output, devE, outputCounter, false);
+ queue.readBuffer(output.correlatedData, devE, outputCounter, false);
}
- void CorrelatorStep::processCPU(const SubbandProcInputData &input, CorrelatedData &output)
+ void CorrelatorStep::processCPU(const SubbandProcInputData &input, SubbandProcOutputData &output)
{
// Propagate the flags.
MultiDimArray<LOFAR::SparseSet<unsigned>, 1> flags = input.inputFlags;
@@ -332,30 +318,30 @@ namespace LOFAR
flags, input.blockID.subbandProcSubbandIdx);
}
- Flagger::propagateFlags(ps, flags, output);
+ Flagger::propagateFlags(ps, flags, output.correlatedData);
}
- bool CorrelatorStep::integrate(CorrelatedData &output)
+ bool CorrelatorStep::integrate(SubbandProcOutputData &output)
{
const size_t idx = output.blockID.subbandProcSubbandIdx;
const size_t nblock = ps.settings.correlator.nrBlocksPerIntegration;
// We don't want to copy the data if we don't need to integrate.
if (nblock == 1) {
- output.setSequenceNumber(output.blockID.block);
+ output.correlatedData.setSequenceNumber(output.blockID.block);
return true;
}
integratedData[idx].first++;
if (integratedData[idx].first < nblock) {
- *integratedData[idx].second += output;
+ *integratedData[idx].second += output.correlatedData;
return false;
}
else {
- output += *integratedData[idx].second;
- output.setSequenceNumber(output.blockID.block / nblock);
+ output.correlatedData += *integratedData[idx].second;
+ output.correlatedData.setSequenceNumber(output.blockID.block / nblock);
integratedData[idx].first = 0;
integratedData[idx].second->reset();
return true;
@@ -363,7 +349,7 @@ namespace LOFAR
}
- bool CorrelatorStep::postprocessSubband(CorrelatedData &output)
+ bool CorrelatorStep::postprocessSubband(SubbandProcOutputData &output)
{
if (!integrate(output)) {
// Not yet done constructing output block
@@ -372,7 +358,7 @@ namespace LOFAR
// The flags are already copied to the correct location
// now the flagged amount should be applied to the visibilities
- Flagger::applyWeights(ps, output);
+ Flagger::applyWeights(ps, output.correlatedData);
return true;
}
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/CorrelatorStep.h b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/CorrelatorStep.h
index 3fe345122d95b01eba8d38529b22902532c6f474..5075f0888866cd4fd13512485fd7a76e0d27c721 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/CorrelatorStep.h
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/CorrelatorStep.h
@@ -25,8 +25,8 @@
#include <vector>
#include <utility> // for std::pair
-#include <Common/LofarLogger.h>
#include <CoInterface/Parset.h>
+#include <CoInterface/SmartPtr.h>
#include <boost/shared_ptr.hpp>
#include <GPUProc/gpu_wrapper.h>
@@ -35,7 +35,8 @@
#include <CoInterface/BlockID.h>
#include <CoInterface/CorrelatedData.h>
-#include "SubbandProc.h"
+#include "SubbandProcInputData.h"
+#include "SubbandProcOutputData.h"
#include "ProcessStep.h"
#include <GPUProc/PerformanceCounter.h>
@@ -53,17 +54,6 @@ namespace LOFAR
class CorrelatorStep: public ProcessStep
{
public:
- struct CorrelatedData:
- public MultiDimArrayHostBuffer<fcomplex,4>,
- public LOFAR::Cobalt::CorrelatedData,
- public SubbandProcOutputData
- {
- CorrelatedData(unsigned nrStations,
- unsigned nrChannels,
- unsigned maxNrValidSamples,
- gpu::Context &context);
- };
-
struct Factories {
Factories(const Parset &ps, size_t nrSubbandsPerSubbandProc);
@@ -89,17 +79,17 @@ namespace LOFAR
void writeInput(const SubbandProcInputData &input);
void process(const SubbandProcInputData &input);
- void processCPU(const SubbandProcInputData &input, CorrelatedData &output);
+ void processCPU(const SubbandProcInputData &input, SubbandProcOutputData &output);
- void readOutput(CorrelatedData &output);
+ void readOutput(SubbandProcOutputData &output);
- bool postprocessSubband(CorrelatedData &output);
+ bool postprocessSubband(SubbandProcOutputData &output);
// Collection of functions to tranfer the input flags to the output.
// \c propagateFlags can be called parallel to the kernels.
// After the data is copied from the the shared buffer
// \c applyWeights can be used to weight the visibilities
- class Flagger: public SubbandProc::Flagger
+ class Flagger
{
public:
// 1. Convert input flags to channel flags, calculate the amount flagged
@@ -169,7 +159,7 @@ namespace LOFAR
std::vector< std::pair< size_t, SmartPtr<LOFAR::Cobalt::CorrelatedData> > >
integratedData;
- bool integrate(CorrelatedData &output);
+ bool integrate(SubbandProcOutputData &output);
};
}
}
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/ProcessStep.h b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/ProcessStep.h
index fedbbf5589f7c55180035e3724089d1d63cea292..63ed74a484c0e26e788636737c52be38afe17630 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/ProcessStep.h
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/ProcessStep.h
@@ -25,9 +25,7 @@
#include <GPUProc/gpu_wrapper.h>
-#include <GPUProc/SubbandProcs/SubbandProc.h>
-
-#include "SubbandProc.h"
+#include "SubbandProcInputData.h"
namespace LOFAR
{
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProc.cc b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProc.cc
index f0abf75dd323b64a507b37cb4a7e92d09fab2222..ea4a4018017db7ae1e723a886db27e09ad7087ff 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProc.cc
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProc.cc
@@ -21,18 +21,28 @@
#include <lofar_config.h>
#include "SubbandProc.h"
+#include "BeamFormerFactories.h"
-#include <Common/LofarLogger.h>
+#include <GPUProc/global_defines.h>
+#include <GPUProc/gpu_wrapper.h>
+#include <CoInterface/Parset.h>
#include <CoInterface/Align.h>
-#include <GPUProc/global_defines.h>
+#include <ApplCommon/PosixTime.h>
+#include <Common/LofarLogger.h>
+
+#include <iomanip>
namespace LOFAR
{
namespace Cobalt
{
- SubbandProc::SubbandProc(const Parset &ps, gpu::Context &context, size_t nrSubbandsPerSubbandProc)
+ SubbandProc::SubbandProc(
+ const Parset &ps,
+ gpu::Context &context,
+ BeamFormerFactories &factories,
+ size_t nrSubbandsPerSubbandProc)
:
inputPool("SubbandProc::inputPool"),
processPool("SubbandProc::processPool"),
@@ -40,8 +50,79 @@ namespace LOFAR
ps(ps),
nrSubbandsPerSubbandProc(nrSubbandsPerSubbandProc),
- queue(gpu::Stream(context))
+ queue(gpu::Stream(context)),
+ prevBlock(-1),
+ prevSAP(-1),
+ inputCounter(context, "input")
{
+ // See doc/bf-pipeline.txt
+ size_t devA_size = 0;
+ size_t devB_size = 0;
+
+ if (factories.correlator) {
+ CorrelatorStep::Factories &cf = *factories.correlator;
+
+ devA_size = std::max(devA_size,
+ cf.firFilter ? cf.firFilter->bufferSize(FIR_FilterKernel::INPUT_DATA)
+ : cf.delayAndBandPass.bufferSize(DelayAndBandPassKernel::INPUT_DATA));
+ devB_size = std::max(devB_size,
+ cf.correlator.bufferSize(CorrelatorKernel::INPUT_DATA));
+ }
+
+ if (factories.preprocessing) {
+ devA_size = std::max(devA_size,
+ factories.preprocessing->intToFloat.bufferSize(IntToFloatKernel::OUTPUT_DATA));
+ devB_size = std::max(devB_size,
+ factories.preprocessing->intToFloat.bufferSize(IntToFloatKernel::OUTPUT_DATA));
+ }
+
+ if (factories.incoherentStokes) {
+ ASSERT(factories.preprocessing);
+
+ /* incoherentStokes uses devA and devB, but the sizes provided b the preprocessing
+ pipeline are already sufficient. */
+ }
+
+ // NOTE: For an explanation of the different buffers being used, please refer
+ // to the document bf-pipeline.txt in the GPUProc/doc directory.
+ devA.reset(new gpu::DeviceMemory(context, devA_size));
+ devB.reset(new gpu::DeviceMemory(context, devB_size));
+
+ //################################################
+ // Create objects containing the kernel and device buffers
+
+ if (factories.correlator) {
+ correlatorStep = std::auto_ptr<CorrelatorStep>(
+ new CorrelatorStep(ps, queue, context, *factories.correlator,
+ devA, devB, nrSubbandsPerSubbandProc));
+ }
+
+ if (factories.preprocessing) {
+ preprocessingStep = std::auto_ptr<BeamFormerPreprocessingStep>(
+ new BeamFormerPreprocessingStep(ps, queue, context, *factories.preprocessing,
+ devA, devB));
+ }
+
+ if (factories.coherentStokes) {
+ coherentStep = std::auto_ptr<BeamFormerCoherentStep>(
+ new BeamFormerCoherentStep(ps, queue, context, *factories.coherentStokes,
+ devB));
+ }
+
+ if (factories.incoherentStokes) {
+ incoherentStep = std::auto_ptr<BeamFormerIncoherentStep>(
+ new BeamFormerIncoherentStep(ps, queue, context, *factories.incoherentStokes,
+ devA, devB));
+ }
+
+
+ LOG_INFO_STR("Pipeline configuration: "
+ << (correlatorStep.get() ? "[correlator] " : "")
+ << (preprocessingStep.get() ? "[bf preproc] " : "")
+ << (coherentStep.get() ? "[coh stokes] " : "")
+ << (incoherentStep.get() ? "[incoh stokes] " : "")
+ );
+
// put enough objects in the inputPool to operate
//
// At least 3 items are needed for a smooth Pool operation.
@@ -49,11 +130,14 @@ namespace LOFAR
for (size_t i = 0; i < nrInputDatas; ++i) {
inputPool.free.append(new SubbandProcInputData(ps, context), false);
}
+
+ // put enough objects in the outputPool to operate
+ for (size_t i = 0; i < nrOutputElements(); ++i)
+ {
+ outputPool.free.append(new SubbandProcOutputData(ps, context));
+ }
}
- SubbandProc::~SubbandProc()
- {
- }
size_t SubbandProc::nrOutputElements() const
{
@@ -70,84 +154,14 @@ namespace LOFAR
return 7 * nrSubbandsPerSubbandProc;
}
-
- void SubbandProcInputData::applyMetaData(const Parset &ps,
- unsigned station, unsigned SAP,
- const SubbandMetaData &metaData)
- {
- // extract and apply the flags
- inputFlags[station] = metaData.flags;
-
- flagInputSamples(station, metaData);
-
- // extract and assign the delays for the station beams
-
- // X polarisation
- delaysAtBegin[SAP][station][0] = ps.settings.antennaFields[station].delay.x + metaData.stationBeam.delayAtBegin;
- delaysAfterEnd[SAP][station][0] = ps.settings.antennaFields[station].delay.x + metaData.stationBeam.delayAfterEnd;
- phase0s[station][0] = ps.settings.antennaFields[station].phase0.x;
-
- // Y polarisation
- delaysAtBegin[SAP][station][1] = ps.settings.antennaFields[station].delay.y + metaData.stationBeam.delayAtBegin;
- delaysAfterEnd[SAP][station][1] = ps.settings.antennaFields[station].delay.y + metaData.stationBeam.delayAfterEnd;
- phase0s[station][1] = ps.settings.antennaFields[station].phase0.y;
-
- if (ps.settings.beamFormer.enabled)
- {
- // we already compensated for the delay for the first beam
- double compensatedDelay = (metaData.stationBeam.delayAfterEnd +
- metaData.stationBeam.delayAtBegin) * 0.5;
-
- size_t nrTABs = ps.settings.beamFormer.SAPs[SAP].nrCoherent;
-
- ASSERTSTR(metaData.TABs.size() == nrTABs, "Need delays for " << nrTABs << " coherent TABs, but got delays for " << metaData.TABs.size() << " TABs");
-
- // Note: We only get delays for the coherent TABs
- for (unsigned tab = 0; tab < nrTABs; tab++)
- {
- // subtract the delay that was already compensated for
- tabDelays[SAP][station][tab] = (metaData.TABs[tab].delayAtBegin +
- metaData.TABs[tab].delayAfterEnd) * 0.5 -
- compensatedDelay;
- }
-
- // Zero padding entries that exist because we always produce maxNrCoherentTABsPerSAP for any subband
- for (unsigned tab = nrTABs; tab < ps.settings.beamFormer.maxNrCoherentTABsPerSAP(); tab++)
- tabDelays[SAP][station][tab] = 0.0;
- }
- }
-
-
- // flag the input samples.
- void SubbandProcInputData::flagInputSamples(unsigned station,
- const SubbandMetaData& metaData)
- {
-
- // Get the size of a sample in bytes.
- size_t sizeof_sample = sizeof *inputSamples.origin();
-
- // Calculate the number elements to skip when striding over the second
- // dimension of inputSamples.
- size_t stride = inputSamples[station][0].num_elements();
-
- // Zero the bytes in the input data for the flagged ranges.
- for(SparseSet<unsigned>::const_iterator it = metaData.flags.getRanges().begin();
- it != metaData.flags.getRanges().end(); ++it)
- {
- void *offset = inputSamples[station][it->begin].origin();
- size_t size = stride * (it->end - it->begin) * sizeof_sample;
- memset(offset, 0, size);
- }
- }
-
- void SubbandProc::Flagger::convertFlagsToChannelFlags(Parset const &parset,
+ void SubbandProc::Flagger::convertFlagsToChannelFlags(Parset const &ps,
MultiDimArray<LOFAR::SparseSet<unsigned>, 1>const &inputFlags,
MultiDimArray<SparseSet<unsigned>, 2>& flagsPerChannel)
{
- unsigned numberOfChannels = parset.nrChannelsPerSubband();
+ unsigned numberOfChannels = ps.nrChannelsPerSubband();
unsigned log2NrChannels = log2(numberOfChannels);
//Convert the flags per sample to flags per channel
- for (unsigned station = 0; station < parset.nrStations(); station ++)
+ for (unsigned station = 0; station < ps.nrStations(); station ++)
{
// get the flag ranges
const SparseSet<unsigned>::Ranges &ranges = inputFlags[station].getRanges();
@@ -160,7 +174,7 @@ namespace LOFAR
{
// do nothing, just take the ranges as supplied
begin_idx = it->begin;
- end_idx = std::min(parset.nrSamplesPerChannel(), it->end );
+ end_idx = std::min(ps.nrSamplesPerChannel(), it->end );
}
else
{
@@ -181,7 +195,7 @@ namespace LOFAR
// The min is needed, because flagging the last input
// samples would cause NR_TAPS subsequent samples to
// be flagged, which aren't necessarily part of this block.
- end_idx = std::min(parset.nrSamplesPerChannel() + 1,
+ end_idx = std::min(ps.nrSamplesPerChannel() + 1,
((it->end - 1) >> log2NrChannels) + 1);
}
@@ -192,6 +206,84 @@ namespace LOFAR
}
}
}
+
+
+ void SubbandProc::processSubband( SubbandProcInputData &input,
+ SubbandProcOutputData &output)
+ {
+ //*******************************************************************
+ // calculate some variables depending on the input subband
+ size_t block = input.blockID.block;
+ unsigned SAP = ps.settings.subbands[input.blockID.globalSubbandIdx].SAP;
+
+ //****************************************
+ // Send inputs to GPU
+ queue.writeBuffer(*devA, input.inputSamples, inputCounter, true);
+
+ // Some additional buffers
+ // Only upload delays if they changed w.r.t. the previous subband.
+ if ((int)SAP != prevSAP || (ssize_t)block != prevBlock) {
+ if (correlatorStep.get()) {
+ correlatorStep->writeInput(input);
+ }
+
+ if (preprocessingStep.get()) {
+ preprocessingStep->writeInput(input);
+ }
+
+ if (coherentStep.get()) {
+ coherentStep->writeInput(input);
+ }
+
+ prevSAP = SAP;
+ prevBlock = block;
+ }
+
+ // ************************************************
+ // Start the GPU processing
+
+ if (correlatorStep.get()) {
+ correlatorStep->process(input);
+ correlatorStep->readOutput(output);
+ }
+
+ if (preprocessingStep.get()) {
+ preprocessingStep->process(input);
+ }
+
+ if (coherentStep.get())
+ {
+ coherentStep->process(input);
+ coherentStep->readOutput(output);
+ }
+
+ if (incoherentStep.get())
+ {
+ incoherentStep->process(input);
+ incoherentStep->readOutput(output);
+ }
+
+ // ************************************************
+ // Do CPU computations while the GPU is working
+
+ if (correlatorStep.get()) {
+ correlatorStep->processCPU(input, output);
+ }
+
+ // Synchronise to assure that all the work in the data is done
+ queue.synchronize();
+ }
+
+
+ void SubbandProc::postprocessSubband(SubbandProcOutputData &output)
+ {
+ if (correlatorStep.get()) {
+ output.emit_correlatedData = correlatorStep->postprocessSubband(output);
+ } else {
+ output.emit_correlatedData = false;
+ }
+ }
}
}
+
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProc.h b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProc.h
index 5597ddbada71d708d2ee807e73919db91a08af7d..dd9c81129972c7d64eda642c506175db711e3a0c 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProc.h
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProc.h
@@ -23,7 +23,12 @@
#include <string>
#include <map>
+#include <complex>
+#include <memory>
+#include <boost/shared_ptr.hpp>
+#include <Common/LofarLogger.h>
+#include <CoInterface/CorrelatedData.h>
#include <CoInterface/Parset.h>
#include <CoInterface/Pool.h>
#include <CoInterface/SmartPtr.h>
@@ -34,6 +39,13 @@
#include <GPUProc/gpu_wrapper.h>
#include <GPUProc/MultiDimArrayHostBuffer.h>
+#include "SubbandProcInputData.h"
+#include "SubbandProcOutputData.h"
+#include "CorrelatorStep.h"
+#include "BeamFormerPreprocessingStep.h"
+#include "BeamFormerCoherentStep.h"
+#include "BeamFormerIncoherentStep.h"
+
// \file
// TODO: Update documentation
@@ -41,101 +53,8 @@ namespace LOFAR
{
namespace Cobalt
{
- // Collect all inputData for the correlatorSubbandProc item:
- // \arg inputsamples
- // \arg delays
- // \arg phaseOffSets
- // \arg flags
- // It also contains a read function parsing all this data from an input stream.
- class SubbandProcInputData
- {
- public:
- // Which block this InputData represents
- struct BlockID blockID;
-
- // Delays are computed and applied in double precision,
- // otherwise the to be computed phase shifts become too inprecise.
-
- //!< Whole sample delays at the start of the workitem
- MultiDimArrayHostBuffer<double, 3> delaysAtBegin;
-
- //!< Whole sample delays at the end of the workitem
- MultiDimArrayHostBuffer<double, 3> delaysAfterEnd;
-
- //!< Remainder of delays
- MultiDimArrayHostBuffer<double, 2> phase0s;
-
- //!< Delays for TABs (aka pencil beams) after station beam correction
- MultiDimArrayHostBuffer<double, 3> tabDelays;
-
- // inputdata with flagged data set to zero
- MultiDimArrayHostBuffer<char, 4> inputSamples;
-
- // The input flags
- MultiDimArray<SparseSet<unsigned>, 1> inputFlags;
-
- // CPU-side holder for the Meta Data
- std::vector<SubbandMetaData> metaData; // [station]
-
- // Create the inputData object we need shared host/device memory on the
- // supplied devicequeue
- SubbandProcInputData(size_t n_beams, size_t n_stations,
- size_t n_polarizations, size_t n_coherent_tabs,
- size_t n_samples, size_t bytes_per_complex_sample,
- gpu::Context &context,
- unsigned int hostBufferFlags = 0)
- :
- delaysAtBegin(boost::extents[n_beams][n_stations][n_polarizations],
- context, hostBufferFlags),
- delaysAfterEnd(boost::extents[n_beams][n_stations][n_polarizations],
- context, hostBufferFlags),
- phase0s(boost::extents[n_stations][n_polarizations],
- context, hostBufferFlags),
- tabDelays(boost::extents[n_beams][n_stations][n_coherent_tabs],
- context, hostBufferFlags),
- inputSamples(boost::extents[n_stations][n_samples][n_polarizations][bytes_per_complex_sample],
- context, hostBufferFlags), // TODO: The size of the buffer is NOT validated
- inputFlags(boost::extents[n_stations]),
- metaData(n_stations)
- {
- }
-
- // Short-hand constructor pulling all relevant values from a Parset
- SubbandProcInputData(const Parset &ps,
- gpu::Context &context,
- unsigned int hostBufferFlags = 0)
- :
- delaysAtBegin(boost::extents[ps.settings.SAPs.size()][ps.settings.antennaFields.size()][NR_POLARIZATIONS],
- context, hostBufferFlags),
- delaysAfterEnd(boost::extents[ps.settings.SAPs.size()][ps.settings.antennaFields.size()][NR_POLARIZATIONS],
- context, hostBufferFlags),
- phase0s(boost::extents[ps.settings.antennaFields.size()][NR_POLARIZATIONS],
- context, hostBufferFlags),
- tabDelays(boost::extents[ps.settings.SAPs.size()][ps.settings.antennaFields.size()][ps.settings.beamFormer.maxNrCoherentTABsPerSAP()],
- context, hostBufferFlags),
- inputSamples(boost::extents[ps.settings.antennaFields.size()][ps.settings.blockSize][NR_POLARIZATIONS][ps.nrBytesPerComplexSample()],
- context, hostBufferFlags), // TODO: The size of the buffer is NOT validated
- inputFlags(boost::extents[ps.settings.antennaFields.size()]),
- metaData(ps.settings.antennaFields.size())
- {
- }
-
- // process the given meta data
- void applyMetaData(const Parset &ps, unsigned station,
- unsigned SAP, const SubbandMetaData &metaData);
-
- // set all flagged inputSamples to zero.
- void flagInputSamples(unsigned station, const SubbandMetaData& metaData);
- };
-
- class SubbandProcOutputData
- {
- public:
- struct BlockID blockID;
-
- // Need a virtual destructor to make type polymorphic
- virtual ~SubbandProcOutputData() {}
- };
+ //# Forward declarations
+ struct BeamFormerFactories;
/*
* The SubbandProc does the following transformation:
@@ -179,15 +98,15 @@ namespace LOFAR
class SubbandProc {
public:
SubbandProc(const Parset &ps, gpu::Context &context,
+ BeamFormerFactories &factories,
size_t nrSubbandsPerSubbandProc = 1);
- virtual ~SubbandProc();
class Flagger
{
public:
// 1.1 Convert the flags per station to channel flags, change time scale
// if nchannel > 1
- static void convertFlagsToChannelFlags(Parset const &parset,
+ static void convertFlagsToChannelFlags(Parset const &ps,
MultiDimArray<SparseSet<unsigned>, 1> const &inputFlags,
MultiDimArray<SparseSet<unsigned>, 2> &flagsPerChannel);
};
@@ -204,10 +123,10 @@ namespace LOFAR
Pool<SubbandProcOutputData> outputPool;
// Correlate the data found in the input data buffer
- virtual void processSubband(SubbandProcInputData &input, SubbandProcOutputData &output) = 0;
+ void processSubband(SubbandProcInputData &input, SubbandProcOutputData &output);
// Do post processing on the CPU.
- virtual void postprocessSubband(SubbandProcOutputData &output) = 0;
+ void postprocessSubband(SubbandProcOutputData &output);
protected:
const Parset &ps;
@@ -215,6 +134,27 @@ namespace LOFAR
gpu::Stream queue;
+ // The previously processed SAP/block, or -1 if nothing has been
+ // processed yet. Used in order to determine if new delays have
+ // to be uploaded.
+ ssize_t prevBlock;
+ signed int prevSAP;
+
+ // @{
+ // Device memory buffers. These buffers are used interleaved. For details,
+ // please refer to the document bf-pipeline.txt in the directory
+ // GPUProc/doc.
+ boost::shared_ptr<gpu::DeviceMemory> devA;
+ boost::shared_ptr<gpu::DeviceMemory> devB;
+ // @}
+
+ PerformanceCounter inputCounter;
+
+ std::auto_ptr<CorrelatorStep> correlatorStep;
+ std::auto_ptr<BeamFormerPreprocessingStep> preprocessingStep;
+ std::auto_ptr<BeamFormerCoherentStep> coherentStep;
+ std::auto_ptr<BeamFormerIncoherentStep> incoherentStep;
+
// Returns the number of output elements to create to get a smooth
// running pipeline.
size_t nrOutputElements() const;
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProcInputData.cc b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProcInputData.cc
new file mode 100644
index 0000000000000000000000000000000000000000..fa4e3c752050e3c59c80392b68b914ce385a1f03
--- /dev/null
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProcInputData.cc
@@ -0,0 +1,144 @@
+//# SubbandProcInputData.cc
+//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
+//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
+//#
+//# This file is part of the LOFAR software suite.
+//# The LOFAR software suite is free software: you can redistribute it and/or
+//# modify it under the terms of the GNU General Public License as published
+//# by the Free Software Foundation, either version 3 of the License, or
+//# (at your option) any later version.
+//#
+//# The LOFAR software suite is distributed in the hope that it will be useful,
+//# but WITHOUT ANY WARRANTY; without even the implied warranty of
+//# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+//# GNU General Public License for more details.
+//#
+//# You should have received a copy of the GNU General Public License along
+//# with the LOFAR software suite. If not, see <http://www.gnu.org/licenses/>.
+//#
+//# $Id$
+
+#include <lofar_config.h>
+
+#include "SubbandProcInputData.h"
+
+#include <CoInterface/Config.h>
+
+namespace LOFAR
+{
+ namespace Cobalt
+ {
+ SubbandProcInputData::SubbandProcInputData(size_t n_beams, size_t n_stations,
+ size_t n_polarizations, size_t n_coherent_tabs,
+ size_t n_samples, size_t bytes_per_complex_sample,
+ gpu::Context &context,
+ unsigned int hostBufferFlags)
+ :
+ delaysAtBegin(boost::extents[n_beams][n_stations][n_polarizations],
+ context, hostBufferFlags),
+ delaysAfterEnd(boost::extents[n_beams][n_stations][n_polarizations],
+ context, hostBufferFlags),
+ phase0s(boost::extents[n_stations][n_polarizations],
+ context, hostBufferFlags),
+ tabDelays(boost::extents[n_beams][n_stations][n_coherent_tabs],
+ context, hostBufferFlags),
+ inputSamples(boost::extents[n_stations][n_samples][n_polarizations][bytes_per_complex_sample],
+ context, hostBufferFlags), // TODO: The size of the buffer is NOT validated
+ inputFlags(boost::extents[n_stations]),
+ metaData(n_stations)
+ {
+ }
+
+ // Short-hand constructor pulling all relevant values from a Parset
+ SubbandProcInputData::SubbandProcInputData(const Parset &ps,
+ gpu::Context &context,
+ unsigned int hostBufferFlags)
+ :
+ delaysAtBegin(boost::extents[ps.settings.SAPs.size()][ps.settings.antennaFields.size()][NR_POLARIZATIONS],
+ context, hostBufferFlags),
+ delaysAfterEnd(boost::extents[ps.settings.SAPs.size()][ps.settings.antennaFields.size()][NR_POLARIZATIONS],
+ context, hostBufferFlags),
+ phase0s(boost::extents[ps.settings.antennaFields.size()][NR_POLARIZATIONS],
+ context, hostBufferFlags),
+ tabDelays(boost::extents[ps.settings.SAPs.size()][ps.settings.antennaFields.size()][ps.settings.beamFormer.maxNrCoherentTABsPerSAP()],
+ context, hostBufferFlags),
+ inputSamples(boost::extents[ps.settings.antennaFields.size()][ps.settings.blockSize][NR_POLARIZATIONS][ps.nrBytesPerComplexSample()],
+ context, hostBufferFlags), // TODO: The size of the buffer is NOT validated
+ inputFlags(boost::extents[ps.settings.antennaFields.size()]),
+ metaData(ps.settings.antennaFields.size())
+ {
+ }
+
+
+ void SubbandProcInputData::applyMetaData(const Parset &ps,
+ unsigned station, unsigned SAP,
+ const SubbandMetaData &metaData)
+ {
+ // extract and apply the flags
+ inputFlags[station] = metaData.flags;
+
+ flagInputSamples(station, metaData);
+
+ // extract and assign the delays for the station beams
+
+ // X polarisation
+ delaysAtBegin[SAP][station][0] = ps.settings.antennaFields[station].delay.x + metaData.stationBeam.delayAtBegin;
+ delaysAfterEnd[SAP][station][0] = ps.settings.antennaFields[station].delay.x + metaData.stationBeam.delayAfterEnd;
+ phase0s[station][0] = ps.settings.antennaFields[station].phase0.x;
+
+ // Y polarisation
+ delaysAtBegin[SAP][station][1] = ps.settings.antennaFields[station].delay.y + metaData.stationBeam.delayAtBegin;
+ delaysAfterEnd[SAP][station][1] = ps.settings.antennaFields[station].delay.y + metaData.stationBeam.delayAfterEnd;
+ phase0s[station][1] = ps.settings.antennaFields[station].phase0.y;
+
+ if (ps.settings.beamFormer.enabled)
+ {
+ // we already compensated for the delay for the first beam
+ double compensatedDelay = (metaData.stationBeam.delayAfterEnd +
+ metaData.stationBeam.delayAtBegin) * 0.5;
+
+ size_t nrTABs = ps.settings.beamFormer.SAPs[SAP].nrCoherent;
+
+ ASSERTSTR(metaData.TABs.size() == nrTABs, "Need delays for " << nrTABs << " coherent TABs, but got delays for " << metaData.TABs.size() << " TABs");
+
+ // Note: We only get delays for the coherent TABs
+ for (unsigned tab = 0; tab < nrTABs; tab++)
+ {
+ // subtract the delay that was already compensated for
+ tabDelays[SAP][station][tab] = (metaData.TABs[tab].delayAtBegin +
+ metaData.TABs[tab].delayAfterEnd) * 0.5 -
+ compensatedDelay;
+ }
+
+ // Zero padding entries that exist because we always produce maxNrCoherentTABsPerSAP for any subband
+ for (unsigned tab = nrTABs; tab < ps.settings.beamFormer.maxNrCoherentTABsPerSAP(); tab++)
+ tabDelays[SAP][station][tab] = 0.0;
+ }
+ }
+
+
+ // flag the input samples.
+ void SubbandProcInputData::flagInputSamples(unsigned station,
+ const SubbandMetaData& metaData)
+ {
+
+ // Get the size of a sample in bytes.
+ size_t sizeof_sample = sizeof *inputSamples.origin();
+
+ // Calculate the number elements to skip when striding over the second
+ // dimension of inputSamples.
+ size_t stride = inputSamples[station][0].num_elements();
+
+ // Zero the bytes in the input data for the flagged ranges.
+ for(SparseSet<unsigned>::const_iterator it = metaData.flags.getRanges().begin();
+ it != metaData.flags.getRanges().end(); ++it)
+ {
+ void *offset = inputSamples[station][it->begin].origin();
+ size_t size = stride * (it->end - it->begin) * sizeof_sample;
+ memset(offset, 0, size);
+ }
+ }
+ }
+}
+
+
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProcInputData.h b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProcInputData.h
new file mode 100644
index 0000000000000000000000000000000000000000..a002bf5b75111b862dfdba844e2e234454fb0270
--- /dev/null
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProcInputData.h
@@ -0,0 +1,99 @@
+//# SubbandProcInputData.h
+//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
+//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
+//#
+//# This file is part of the LOFAR software suite.
+//# The LOFAR software suite is free software: you can redistribute it and/or
+//# modify it under the terms of the GNU General Public License as published
+//# by the Free Software Foundation, either version 3 of the License, or
+//# (at your option) any later version.
+//#
+//# The LOFAR software suite is distributed in the hope that it will be useful,
+//# but WITHOUT ANY WARRANTY; without even the implied warranty of
+//# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+//# GNU General Public License for more details.
+//#
+//# You should have received a copy of the GNU General Public License along
+//# with the LOFAR software suite. If not, see <http://www.gnu.org/licenses/>.
+//#
+//# $Id$
+
+#ifndef LOFAR_GPUPROC_CUDA_SUBBAND_PROC_INPUT_DATA_H
+#define LOFAR_GPUPROC_CUDA_SUBBAND_PROC_INPUT_DATA_H
+
+#include <vector>
+
+#include <CoInterface/BlockID.h>
+#include <CoInterface/Parset.h>
+#include <CoInterface/SubbandMetaData.h>
+#include <GPUProc/MultiDimArrayHostBuffer.h>
+#include <GPUProc/gpu_wrapper.h>
+
+// \file
+// TODO: Update documentation
+
+namespace LOFAR
+{
+ namespace Cobalt
+ {
+ // Collect all inputData for the correlatorSubbandProc item:
+ // \arg inputsamples
+ // \arg delays
+ // \arg phaseOffSets
+ // \arg flags
+ // It also contains a read function parsing all this data from an input stream.
+ class SubbandProcInputData
+ {
+ public:
+ // Which block this InputData represents
+ struct BlockID blockID;
+
+ // Delays are computed and applied in double precision,
+ // otherwise the to be computed phase shifts become too inprecise.
+
+ //!< Whole sample delays at the start of the workitem
+ MultiDimArrayHostBuffer<double, 3> delaysAtBegin;
+
+ //!< Whole sample delays at the end of the workitem
+ MultiDimArrayHostBuffer<double, 3> delaysAfterEnd;
+
+ //!< Remainder of delays
+ MultiDimArrayHostBuffer<double, 2> phase0s;
+
+ //!< Delays for TABs (aka pencil beams) after station beam correction
+ MultiDimArrayHostBuffer<double, 3> tabDelays;
+
+ // inputdata with flagged data set to zero
+ MultiDimArrayHostBuffer<char, 4> inputSamples;
+
+ // The input flags
+ MultiDimArray<SparseSet<unsigned>, 1> inputFlags;
+
+ // CPU-side holder for the Meta Data
+ std::vector<SubbandMetaData> metaData; // [station]
+
+ // Create the inputData object we need shared host/device memory on the
+ // supplied devicequeue
+ SubbandProcInputData(size_t n_beams, size_t n_stations,
+ size_t n_polarizations, size_t n_coherent_tabs,
+ size_t n_samples, size_t bytes_per_complex_sample,
+ gpu::Context &context,
+ unsigned int hostBufferFlags = 0);
+
+ // Short-hand constructor pulling all relevant values from a Parset
+ SubbandProcInputData(const Parset &ps,
+ gpu::Context &context,
+ unsigned int hostBufferFlags = 0);
+
+ // process the given meta data
+ void applyMetaData(const Parset &ps, unsigned station,
+ unsigned SAP, const SubbandMetaData &metaData);
+
+ // set all flagged inputSamples to zero.
+ void flagInputSamples(unsigned station, const SubbandMetaData& metaData);
+ };
+ }
+}
+
+#endif
+
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProcOutputData.cc b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProcOutputData.cc
new file mode 100644
index 0000000000000000000000000000000000000000..228434f83180fce035cfaed2c67bc9d2ab8a754d
--- /dev/null
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProcOutputData.cc
@@ -0,0 +1,75 @@
+//# SubbandProcOutputData.cc
+//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
+//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
+//#
+//# This file is part of the LOFAR software suite.
+//# The LOFAR software suite is free software: you can redistribute it and/or
+//# modify it under the terms of the GNU General Public License as published
+//# by the Free Software Foundation, either version 3 of the License, or
+//# (at your option) any later version.
+//#
+//# The LOFAR software suite is distributed in the hope that it will be useful,
+//# but WITHOUT ANY WARRANTY; without even the implied warranty of
+//# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+//# GNU General Public License for more details.
+//#
+//# You should have received a copy of the GNU General Public License along
+//# with the LOFAR software suite. If not, see <http://www.gnu.org/licenses/>.
+//#
+//# $Id$
+
+#include <lofar_config.h>
+
+#include "SubbandProcOutputData.h"
+
+namespace LOFAR
+{
+ namespace Cobalt
+ {
+ SubbandProcOutputData::SubbandProcOutputData(
+ const Parset &ps,
+ gpu::Context &context) :
+ coherentData(ps.settings.beamFormer.anyCoherentTABs()
+ ? boost::extents[ps.settings.beamFormer.maxNrCoherentTABsPerSAP()]
+ [ps.settings.beamFormer.coherentSettings.nrStokes]
+ [ps.settings.beamFormer.coherentSettings.nrSamples]
+ [ps.settings.beamFormer.coherentSettings.nrChannels]
+ : boost::extents[0][0][0][0],
+ context, 0),
+
+ incoherentData(ps.settings.beamFormer.anyIncoherentTABs()
+ ? boost::extents[ps.settings.beamFormer.maxNrIncoherentTABsPerSAP()]
+ [ps.settings.beamFormer.incoherentSettings.nrStokes]
+ [ps.settings.beamFormer.incoherentSettings.nrSamples]
+ [ps.settings.beamFormer.incoherentSettings.nrChannels]
+ : boost::extents[0][0][0][0],
+ context, 0),
+
+ correlatedData(ps.settings.correlator.enabled ? ps.settings.antennaFields.size() : 0,
+ ps.settings.correlator.enabled ? ps.settings.correlator.nrChannels : 0,
+ ps.settings.correlator.enabled ? ps.settings.correlator.nrSamplesPerChannel : 0,
+ context),
+ emit_correlatedData(false)
+ {
+ }
+
+
+ SubbandProcOutputData::CorrelatedData::CorrelatedData(
+ unsigned nrStations, unsigned nrChannels,
+ unsigned maxNrValidSamples, gpu::Context &context)
+ :
+ MultiDimArrayHostBuffer<fcomplex, 4>(
+ boost::extents
+ [nrStations * (nrStations + 1) / 2]
+ [nrChannels][NR_POLARIZATIONS]
+ [NR_POLARIZATIONS],
+ context, 0),
+ LOFAR::Cobalt::CorrelatedData(nrStations, nrChannels,
+ maxNrValidSamples, this->origin(),
+ this->num_elements(), heapAllocator, 1)
+ {
+ }
+ }
+}
+
+
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProcOutputData.h b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProcOutputData.h
new file mode 100644
index 0000000000000000000000000000000000000000..58743cee0c3ceee310a5560a084478b930eac6fc
--- /dev/null
+++ b/RTCP/Cobalt/GPUProc/src/cuda/SubbandProcs/SubbandProcOutputData.h
@@ -0,0 +1,65 @@
+//# SubbandProcOutputData.h
+//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
+//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
+//#
+//# This file is part of the LOFAR software suite.
+//# The LOFAR software suite is free software: you can redistribute it and/or
+//# modify it under the terms of the GNU General Public License as published
+//# by the Free Software Foundation, either version 3 of the License, or
+//# (at your option) any later version.
+//#
+//# The LOFAR software suite is distributed in the hope that it will be useful,
+//# but WITHOUT ANY WARRANTY; without even the implied warranty of
+//# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+//# GNU General Public License for more details.
+//#
+//# You should have received a copy of the GNU General Public License along
+//# with the LOFAR software suite. If not, see <http://www.gnu.org/licenses/>.
+//#
+//# $Id$
+
+#ifndef LOFAR_GPUPROC_CUDA_SUBBAND_PROC_OUTPUT_DATA_H
+#define LOFAR_GPUPROC_CUDA_SUBBAND_PROC_OUTPUT_DATA_H
+
+#include <CoInterface/BlockID.h>
+#include <CoInterface/Parset.h>
+#include <CoInterface/CorrelatedData.h>
+#include <GPUProc/gpu_wrapper.h>
+#include <GPUProc/MultiDimArrayHostBuffer.h>
+
+// \file
+// TODO: Update documentation
+
+namespace LOFAR
+{
+ namespace Cobalt
+ {
+ // Our output data type
+ class SubbandProcOutputData
+ {
+ public:
+ struct BlockID blockID;
+
+ MultiDimArrayHostBuffer<float, 4> coherentData;
+ MultiDimArrayHostBuffer<float, 4> incoherentData;
+
+ struct CorrelatedData:
+ public MultiDimArrayHostBuffer<fcomplex,4>,
+ public LOFAR::Cobalt::CorrelatedData
+ {
+ CorrelatedData(unsigned nrStations,
+ unsigned nrChannels,
+ unsigned maxNrValidSamples,
+ gpu::Context &context);
+ };
+
+ CorrelatedData correlatedData;
+ bool emit_correlatedData;
+
+ SubbandProcOutputData(const Parset &ps, gpu::Context &context);
+ };
+ }
+}
+
+#endif
+
diff --git a/RTCP/Cobalt/GPUProc/test/SubbandProcs/tBeamFormerSubbandProcProcessSb.cc b/RTCP/Cobalt/GPUProc/test/SubbandProcs/tBeamFormerSubbandProcProcessSb.cc
index 6ea4fa512575b4331601ab60ca2fd7566202c6a3..920fdcff6c5902387a0986a3886af7a5decfaf1e 100644
--- a/RTCP/Cobalt/GPUProc/test/SubbandProcs/tBeamFormerSubbandProcProcessSb.cc
+++ b/RTCP/Cobalt/GPUProc/test/SubbandProcs/tBeamFormerSubbandProcProcessSb.cc
@@ -29,7 +29,7 @@
#include <CoInterface/Parset.h>
#include <CoInterface/fpequals.h>
#include <GPUProc/gpu_utils.h>
-#include <GPUProc/SubbandProcs/BeamFormerSubbandProc.h>
+#include <GPUProc/SubbandProcs/SubbandProc.h>
#include <GPUProc/SubbandProcs/BeamFormerFactories.h>
using namespace std;
@@ -113,7 +113,7 @@ int main() {
// transform the data order).
BeamFormerFactories factories(ps);
- BeamFormerSubbandProc bwq(ps, ctx, factories);
+ SubbandProc bwq(ps, ctx, factories);
SubbandProcInputData in(ps, ctx);
@@ -162,7 +162,7 @@ int main() {
for (size_t i = 0; i < in.tabDelays.num_elements(); i++)
in.tabDelays.get<float>()[i] = 0.0f;
- BeamFormedData out(ps, ctx);
+ SubbandProcOutputData out(ps, ctx);
for (size_t i = 0; i < out.coherentData.num_elements(); i++)
out.coherentData.get<float>()[i] = 42.0f;
diff --git a/RTCP/Cobalt/GPUProc/test/SubbandProcs/tCoherentStokesBeamFormerSubbandProcProcessSb.cc b/RTCP/Cobalt/GPUProc/test/SubbandProcs/tCoherentStokesBeamFormerSubbandProcProcessSb.cc
index 5b1551e6ede0ba70a78aeffdeee50da0aa9b59a4..e7934ad830ec2cb4e04854758e1346389b980605 100644
--- a/RTCP/Cobalt/GPUProc/test/SubbandProcs/tCoherentStokesBeamFormerSubbandProcProcessSb.cc
+++ b/RTCP/Cobalt/GPUProc/test/SubbandProcs/tCoherentStokesBeamFormerSubbandProcProcessSb.cc
@@ -29,7 +29,7 @@
#include <CoInterface/Parset.h>
#include <CoInterface/fpequals.h>
#include <GPUProc/gpu_utils.h>
-#include <GPUProc/SubbandProcs/BeamFormerSubbandProc.h>
+#include <GPUProc/SubbandProcs/SubbandProc.h>
#include <GPUProc/SubbandProcs/BeamFormerFactories.h>
#include "../Kernels/KernelTestHelpers.h"
@@ -142,7 +142,7 @@ int main(/*int argc, char *argv[]*/) {
// transform the data order).
BeamFormerFactories factories(ps);
- BeamFormerSubbandProc bwq(ps, ctx, factories);
+ SubbandProc bwq(ps, ctx, factories);
SubbandProcInputData in(
nrBeams, nrStations, nrPolarisations, maxNrTABsPerSAP,
@@ -193,7 +193,7 @@ int main(/*int argc, char *argv[]*/) {
for (size_t i = 0; i < in.tabDelays.num_elements(); i++)
in.tabDelays.get<float>()[i] = 0.0f;
- BeamFormedData out(ps, ctx);
+ SubbandProcOutputData out(ps, ctx);
for (size_t i = 0; i < out.coherentData.num_elements(); i++)
out.coherentData.get<float>()[i] = 42.0f;
diff --git a/RTCP/Cobalt/GPUProc/test/SubbandProcs/tCorrelatorSubbandProcProcessSb.cc b/RTCP/Cobalt/GPUProc/test/SubbandProcs/tCorrelatorSubbandProcProcessSb.cc
index 3961c18d83b6fc4e673567508f72644d900ac442..bc9f88f9f003e16f04b4bf8cb3afe0837716a69d 100644
--- a/RTCP/Cobalt/GPUProc/test/SubbandProcs/tCorrelatorSubbandProcProcessSb.cc
+++ b/RTCP/Cobalt/GPUProc/test/SubbandProcs/tCorrelatorSubbandProcProcessSb.cc
@@ -26,7 +26,8 @@
#include <CoInterface/fpequals.h>
#include <CoInterface/Parset.h>
#include <GPUProc/gpu_utils.h>
-#include <GPUProc/SubbandProcs/BeamFormerSubbandProc.h>
+#include <GPUProc/SubbandProcs/BeamFormerFactories.h>
+#include <GPUProc/SubbandProcs/SubbandProc.h>
using namespace std;
using namespace LOFAR;
@@ -83,13 +84,13 @@ int main() {
// transform the data order).
BeamFormerFactories factories(ps, 1);
- BeamFormerSubbandProc cwq(ps, ctx, factories);
+ SubbandProc cwq(ps, ctx, factories);
SubbandProcInputData in(
nrBeams, nrStations, nrPolarisations, maxNrTABsPerSAP,
nrSamplesPerSubband, nrBytesPerComplexSample, ctx);
- BeamFormedData out(ps, ctx);
+ SubbandProcOutputData out(ps, ctx);
LOG_INFO_STR(
"\nInput info:" <<
diff --git a/RTCP/Cobalt/GPUProc/test/SubbandProcs/tFlysEyeBeamFormerSubbandProcProcessSb.cc b/RTCP/Cobalt/GPUProc/test/SubbandProcs/tFlysEyeBeamFormerSubbandProcProcessSb.cc
index 5ed0f53c618952569dd439218dff38f854cb99dd..7fd3b2d02f1e34e7eb06dff7863cec70d7e50a76 100644
--- a/RTCP/Cobalt/GPUProc/test/SubbandProcs/tFlysEyeBeamFormerSubbandProcProcessSb.cc
+++ b/RTCP/Cobalt/GPUProc/test/SubbandProcs/tFlysEyeBeamFormerSubbandProcProcessSb.cc
@@ -30,7 +30,7 @@
#include <CoInterface/Parset.h>
#include <CoInterface/fpequals.h>
#include <GPUProc/gpu_utils.h>
-#include <GPUProc/SubbandProcs/BeamFormerSubbandProc.h>
+#include <GPUProc/SubbandProcs/SubbandProc.h>
#include <GPUProc/SubbandProcs/BeamFormerFactories.h>
using namespace std;
@@ -130,7 +130,7 @@ int main() {
// transform the data order).
BeamFormerFactories factories(ps);
- BeamFormerSubbandProc bwq(ps, ctx, factories);
+ SubbandProc bwq(ps, ctx, factories);
SubbandProcInputData in(
nrBeams, nrStations, nrPolarisations, nrTABs,
@@ -182,7 +182,7 @@ int main() {
in.tabDelays.get<float>()[i] = 0.0f;
// Allocate buffer for output signal
- BeamFormedData out(ps, ctx);
+ SubbandProcOutputData out(ps, ctx);
for (size_t i = 0; i < out.coherentData.num_elements(); i++)
out.coherentData.get<float>()[i] = 42.0f;