From f73e08e3e4720044b38f2921127dd4c70dacca5d Mon Sep 17 00:00:00 2001
From: Alexander van Amesfoort <amesfoort@astron.nl>
Date: Tue, 4 Jun 2013 17:33:28 +0000
Subject: [PATCH] Task #4338: MultiDimArrayHostBuffer.h, 1st version,
CorrelatorWorkQueue partially adapted
---
RTCP/Cobalt/GPUProc/src/BandPass.cc | 9 +-
RTCP/Cobalt/GPUProc/src/FilterBank.h | 4 -
.../{Buffers.h => MultiDimArrayHostBuffer.h} | 12 +--
.../src/cuda/MultiDimArrayHostBuffer.h | 62 ++++++++++++
.../GPUProc/src/cuda/Pipelines/Pipeline.cc | 1 -
.../src/cuda/WorkQueues/BeamFormerWorkQueue.h | 2 +-
.../cuda/WorkQueues/CorrelatorWorkQueue.cc | 99 ++++++++-----------
.../src/cuda/WorkQueues/CorrelatorWorkQueue.h | 87 +++++++++-------
.../src/cuda/WorkQueues/UHEP_WorkQueue.h | 2 +-
.../{Buffers.h => MultiDimArrayHostBuffer.h} | 6 +-
.../GPUProc/src/opencl/Pipelines/Pipeline.cc | 5 -
.../opencl/WorkQueues/BeamFormerWorkQueue.h | 2 +-
.../opencl/WorkQueues/CorrelatorWorkQueue.h | 2 +-
.../src/opencl/WorkQueues/UHEP_WorkQueue.h | 2 +-
14 files changed, 170 insertions(+), 125 deletions(-)
rename RTCP/Cobalt/GPUProc/src/{Buffers.h => MultiDimArrayHostBuffer.h} (80%)
create mode 100644 RTCP/Cobalt/GPUProc/src/cuda/MultiDimArrayHostBuffer.h
rename RTCP/Cobalt/GPUProc/src/opencl/{Buffers.h => MultiDimArrayHostBuffer.h} (97%)
diff --git a/RTCP/Cobalt/GPUProc/src/BandPass.cc b/RTCP/Cobalt/GPUProc/src/BandPass.cc
index db79484649a..c42e4b74983 100644
--- a/RTCP/Cobalt/GPUProc/src/BandPass.cc
+++ b/RTCP/Cobalt/GPUProc/src/BandPass.cc
@@ -2130,7 +2130,7 @@ namespace LOFAR
const std::complex<float> l = out[(i - 3 * nrChannels / 2) % fftSize];
const std::complex<float> r = out[i + nrChannels / 2];
- factors[i] = pow(2, 25) / sqrt(abs(m * m + l * l + r * r));
+ factors[i] = std::pow(2, 25) / std::sqrt(std::abs(m * m + l * l + r * r));
}
}
@@ -2140,11 +2140,12 @@ namespace LOFAR
} // namespace LOFAR
-#if 0
+#ifdef TEST_COMPUTE_BANDPASS_CORRECTION_FACTORS
int main()
{
- std::vector<float> factors(4096);
- BandPass::computeCorrectionFactors(&factors[0], 4096);
+ unsigned nrChannelsPerSb = 4096;
+ std::vector<float> factors(nrChannelsPerSb);
+ BandPass::computeCorrectionFactors(&factors[0], nrChannelsPerSb);
return 0;
}
#endif
diff --git a/RTCP/Cobalt/GPUProc/src/FilterBank.h b/RTCP/Cobalt/GPUProc/src/FilterBank.h
index 4e9b057b219..b4f186ff863 100644
--- a/RTCP/Cobalt/GPUProc/src/FilterBank.h
+++ b/RTCP/Cobalt/GPUProc/src/FilterBank.h
@@ -23,10 +23,6 @@
#define USE_ORIGINAL_FILTER 0
-#if 0 || !defined HAVE_BGP
-#define FIR_C_IMPLEMENTATION
-#endif
-
#include <boost/multi_array.hpp>
namespace LOFAR
diff --git a/RTCP/Cobalt/GPUProc/src/Buffers.h b/RTCP/Cobalt/GPUProc/src/MultiDimArrayHostBuffer.h
similarity index 80%
rename from RTCP/Cobalt/GPUProc/src/Buffers.h
rename to RTCP/Cobalt/GPUProc/src/MultiDimArrayHostBuffer.h
index 315e490f93b..dee99d41fa5 100644
--- a/RTCP/Cobalt/GPUProc/src/Buffers.h
+++ b/RTCP/Cobalt/GPUProc/src/MultiDimArrayHostBuffer.h
@@ -1,4 +1,4 @@
-//# Buffers.h
+//# MultiDimArrayHostBuffer.h
//#
//# Copyright (C) 2013 ASTRON (Netherlands Institute for Radio Astronomy)
//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
@@ -20,19 +20,19 @@
//# $Id$
// \file
-// Support for our GPU processing buffer types.
+// Support for our multi-dim array-ed GPU host buffer.
-#ifndef LOFAR_GPUPROC_BUFFERS_H
-#define LOFAR_GPUPROC_BUFFERS_H
+#ifndef LOFAR_GPUPROC_MULTI_DIM_ARRAY_HOST_BUFFER_H
+#define LOFAR_GPUPROC_MULTI_DIM_ARRAY_HOST_BUFFER_H
#if defined (USE_CUDA) && defined (USE_OPENCL)
# error "Either CUDA or OpenCL must be enabled, not both"
#endif
#if defined (USE_CUDA)
-# include "cuda/Buffers.h"
+# include "cuda/MultiDimArrayHostBuffer.h"
#elif defined (USE_OPENCL)
-# include "opencl/Buffers.h"
+# include "opencl/MultiDimArrayHostBuffer.h"
#else
# error "Either CUDA or OpenCL must be enabled, not neither"
#endif
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/MultiDimArrayHostBuffer.h b/RTCP/Cobalt/GPUProc/src/cuda/MultiDimArrayHostBuffer.h
new file mode 100644
index 00000000000..c2fa3409c35
--- /dev/null
+++ b/RTCP/Cobalt/GPUProc/src/cuda/MultiDimArrayHostBuffer.h
@@ -0,0 +1,62 @@
+//# MultiDimArrayHostBuffer.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_MULTI_DIM_ARRAY_HOST_BUFFER_H
+#define LOFAR_GPUPROC_CUDA_MULTI_DIM_ARRAY_HOST_BUFFER_H
+
+#include <CoInterface/MultiDimArray.h>
+
+#include "gpu_wrapper.h"
+
+namespace LOFAR
+{
+ namespace Cobalt
+ {
+
+ // A MultiDimArray allocated as a HostBuffer
+ // Note: Elements are not constructed/destructed.
+ template <typename T, unsigned DIM>
+ class MultiDimArrayHostBuffer : public gpu::HostMemory,
+ public MultiDimArray<T, DIM>
+ {
+ public:
+ template <typename ExtentList>
+ MultiDimArrayHostBuffer(const ExtentList &extents, gpu::Context &context,
+ unsigned int flags)
+ :
+ HostMemory(context, size(), flags),
+ MultiDimArray<T, DIM>(extents, gpu::HostMemory::get<T>(), false)
+ {
+ }
+
+ using HostMemory::size;
+
+ private:
+ MultiDimArrayHostBuffer(); // don't use
+ MultiDimArrayHostBuffer(const MultiDimArrayHostBuffer<T, DIM> &rhs); // don't use
+ MultiDimArrayHostBuffer<T, DIM> &operator=(const MultiDimArrayHostBuffer<T, DIM> &rhs); // don't use
+ using MultiDimArray<T, DIM>::resize; // don't use
+ };
+
+ } // namespace Cobalt
+} // namespace LOFAR
+
+#endif
+
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/Pipeline.cc b/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/Pipeline.cc
index 0a6d1bfe490..5065e76f421 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/Pipeline.cc
+++ b/RTCP/Cobalt/GPUProc/src/cuda/Pipelines/Pipeline.cc
@@ -25,7 +25,6 @@
#include <Common/LofarLogger.h>
#include <Common/lofar_iomanip.h>
-#include <GPUProc/Buffers.h>
#include <GPUProc/gpu_utils.h>
namespace LOFAR
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/BeamFormerWorkQueue.h b/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/BeamFormerWorkQueue.h
index cc8d02bc4b0..8f117daac2b 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/BeamFormerWorkQueue.h
+++ b/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/BeamFormerWorkQueue.h
@@ -26,7 +26,7 @@
#include <Common/LofarLogger.h>
#include <CoInterface/Parset.h>
-#include <GPUProc/Buffers.h>
+#include <GPUProc/MultiDimArrayHostBuffer.h>
#include <GPUProc/BandPass.h>
#include <GPUProc/Pipelines/BeamFormerPipeline.h>
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/CorrelatorWorkQueue.cc b/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/CorrelatorWorkQueue.cc
index 9aee86578bb..4f217501aa5 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/CorrelatorWorkQueue.cc
+++ b/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/CorrelatorWorkQueue.cc
@@ -55,65 +55,40 @@ namespace LOFAR
WorkQueue( parset, context ),
prevBlock(-1),
prevSAP(-1),
- devInput(ps.nrBeams(),
- ps.nrStations(),
- NR_POLARIZATIONS,
- ps.nrHistorySamples() + ps.nrSamplesPerSubband(),
- ps.nrBytesPerComplexSample(),
- queue,
-
- // reserve enough space in inputSamples for the output of
- // the delayAndBandPassKernel.
- ps.nrStations() * NR_POLARIZATIONS * ps.nrSamplesPerSubband() * sizeof(std::complex<float>)),
- devFilteredData(queue,
- CL_MEM_READ_WRITE,
-
- // reserve enough space for the output of the
- // firFilterKernel,
- std::max(ps.nrStations() * NR_POLARIZATIONS * ps.nrSamplesPerSubband() * sizeof(std::complex<float>),
- // and the correlatorKernel.
- ps.nrBaselines() * ps.nrChannelsPerSubband() * NR_POLARIZATIONS * NR_POLARIZATIONS * sizeof(std::complex<float>))),
- devFIRweights(queue,
- CL_MEM_READ_ONLY,
- ps.nrChannelsPerSubband() * NR_TAPS * sizeof(float)),
- firFilterKernel(ps,
- queue,
- programs.firFilterProgram,
- devFilteredData,
- devInput.inputSamples,
- devFIRweights),
- fftKernel(ps,
- context,
- devFilteredData),
- bandPassCorrectionWeights(boost::extents[ps.nrChannelsPerSubband()],
- queue,
- CL_MEM_WRITE_ONLY,
- CL_MEM_READ_ONLY),
- delayAndBandPassKernel(ps,
- programs.delayAndBandPassProgram,
+ // TODO: have the Kernel classes be able to provide input and output buffer sizes to use below
+ devInput(ps.nrBeams(), ps.nrStations(), NR_POLARIZATIONS,
+ ps.nrHistorySamples() + ps.nrSamplesPerSubband(),
+ ps.nrBytesPerComplexSample(), context,
+
+ // reserve enough space in inputSamples for the output of the delayAndBandPassKernel.
+ ps.nrStations() * NR_POLARIZATIONS * ps.nrSamplesPerSubband() * sizeof(std::complex<float>)),
+ // reserve enough space for the output of the firFilterKernel,
+ devFilteredData(context, std::max(
+ ps.nrStations() * NR_POLARIZATIONS * ps.nrSamplesPerSubband() * sizeof(std::complex<float>),
+ // and the correlatorKernel.
+ ps.nrBaselines() * ps.nrChannelsPerSubband() * NR_POLARIZATIONS * NR_POLARIZATIONS * sizeof(std::complex<float>)
+ )),
+ devFIRweights(context, ps.nrChannelsPerSubband() * NR_TAPS * sizeof(float)),
+ devBandPassCorrectionWeights(context, ps.nrChannelsPerSubband() * sizeof(float)),
+
+ firFilterKernel(ps, queue, programs.firFilterProgram,
+ devFilteredData, devInput.inputSamples, devFIRweights),
+ fftKernel(ps, context, devFilteredData),
+ delayAndBandPassKernel(ps, programs.delayAndBandPassProgram,
devInput.inputSamples,
devFilteredData,
devInput.delaysAtBegin,
devInput.delaysAfterEnd,
devInput.phaseOffsets,
- bandPassCorrectionWeights),
+ devBandPassCorrectionWeights),
#if defined USE_NEW_CORRELATOR
- correlateTriangleKernel(ps,
- queue,
- programs.correlatorProgram,
- devFilteredData,
- devInput.inputSamples),
- correlateRectangleKernel(ps,
- queue,
- programs.correlatorProgram,
- devFilteredData,
- devInput.inputSamples)
+ correlateTriangleKernel(ps, queue, programs.correlatorProgram,
+ devFilteredData, devInput.inputSamples),
+ correlateRectangleKernel(ps, queue, programs.correlatorProgram,
+ devFilteredData, devInput.inputSamples)
#else
- correlatorKernel(ps,
- queue,
- programs.correlatorProgram,
- devFilteredData,
- devInput.inputSamples)
+ correlatorKernel(ps, queue, programs.correlatorProgram,
+ devFilteredData, devInput.inputSamples)
#endif
{
// put enough objects in the inputPool to operate
@@ -141,7 +116,6 @@ namespace LOFAR
ps.nrStations(),
ps.nrChannelsPerSubband(),
ps.integrationSteps(),
- devFilteredData,
*this));
}
@@ -172,16 +146,21 @@ namespace LOFAR
addTimer("GPU - compute");
addTimer("GPU - wait");
- // Copy the FIR filter weights to the device in two steps (TODO: make FilterBank supply the right buffer, or do like BandPassCorrectionWeights below)
- size_t fbBytes = filterBank.getWeights().num_elements() * sizeof(float);
- gpu::HostMemory fbBuffer(context, fbBytes);
- std::memcpy(fbBuffer.get<void>(), filterBank.getWeights().origin(), fbBytes);
- queue.writeBuffer(devFIRweights, fbBuffer, true);
+ // Copy the FIR filter and bandpass weights to the device.
+ // Note that these constant weights are now (unnecessarily) stored on the
+ // device for every workqueue. A single copy per device could be used, but
+ // first verify that the device platform still allows workqueue overlap.
+ size_t firWeightsSize = filterBank.getWeights().num_elements() * sizeof(float);
+ gpu::HostMemory firWeights(context, firWeightsSize);
+ std::memcpy(firWeights.get<void>(), filterBank.getWeights().origin(), fbBytes);
+ queue.writeBuffer(devFIRweights, firWeights, true);
if (ps.correctBandPass())
{
- BandPass::computeCorrectionFactors(bandPassCorrectionWeights.origin(), ps.nrChannelsPerSubband());
- bandPassCorrectionWeights.hostToDevice(true);
+ gpu::HostMemory bpWeights(context, ps.nrChannelsPerSubband() * sizeof(float));
+ BandPass::computeCorrectionFactors(bpWeights.origin(),
+ ps.nrChannelsPerSubband());
+ queue.writeBuffer(devBandPassCorrectionWeights, bpWeights, true);
}
}
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/CorrelatorWorkQueue.h b/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/CorrelatorWorkQueue.h
index b7c9b54b191..af86ee52a39 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/CorrelatorWorkQueue.h
+++ b/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/CorrelatorWorkQueue.h
@@ -33,7 +33,7 @@
#include <CoInterface/SubbandMetaData.h>
#include <GPUProc/global_defines.h>
-#include <GPUProc/Buffers.h>
+#include <GPUProc/MultiDimArrayHostBuffer.h>
#include <GPUProc/FilterBank.h>
#include <GPUProc/Pipelines/CorrelatorPipelinePrograms.h>
#include <GPUProc/Kernels/FIR_FilterKernel.h>
@@ -99,24 +99,29 @@ namespace LOFAR
};
// A CorrelatedData object tied to a HostBuffer and WorkQueue. Such links
- // are needed for performance -- the visibilities are stored in a buffer
- // directly linked to the GPU output buffer.
- class CorrelatedDataHostBuffer: public MultiArrayHostBuffer<fcomplex, 4>, public CorrelatedData
+ // After the visibilities have been written to storage, we need remember
+ // the queue to recycle the buffer.
+ class CorrelatedDataHostBuffer: public CorrelatedData,
+ public MultiDimArrayHostBuffer<fcomplex, 4>
{
public:
- CorrelatedDataHostBuffer(unsigned nrStations, unsigned nrChannels, unsigned maxNrValidSamples, DeviceBuffer &deviceBuffer, CorrelatorWorkQueue &queue)
+ CorrelatedDataHostBuffer(unsigned nrStations, unsigned nrChannels,
+ unsigned maxNrValidSamples, CorrelatorWorkQueue &workQueue)
:
- MultiArrayHostBuffer<fcomplex, 4>(boost::extents[nrStations * (nrStations + 1) / 2][nrChannels][NR_POLARIZATIONS][NR_POLARIZATIONS], CL_MEM_WRITE_ONLY, deviceBuffer),
- CorrelatedData(nrStations, nrChannels, maxNrValidSamples, this->origin(), this->num_elements(), heapAllocator, 1),
- queue(queue)
+ CorrelatedData(nrStations, nrChannels, maxNrValidSamples, this->origin(),
+ this->num_elements(), heapAllocator, 1),
+ MultiDimArrayHostBuffer<fcomplex, 4>(boost::extents[nrStations * (nrStations + 1) / 2]
+ [nrChannels][NR_POLARIZATIONS]
+ [NR_POLARIZATIONS], 0),
+ workQueue(workQueue)
{
}
- // Annotation required, as we'll loose track of the exact order
+ // Annotation required, as we'll lose track of the exact order
size_t block;
unsigned subband;
- CorrelatorWorkQueue &queue;
+ CorrelatorWorkQueue &workQueue;
private:
CorrelatedDataHostBuffer();
@@ -134,24 +139,26 @@ namespace LOFAR
{
public:
- // The set of GPU buffers to link our HostBuffers to.
+ // The set of GPU buffers to link our host buffers to.
+ // Device buffers may be reused between different pairs of kernels,
+ // since device memory size is a concern. Use inputSamplesMinSize
+ // to specify a minimum derived from other uses apart from input.
struct DeviceBuffers
{
- DeviceBuffer delaysAtBegin;
- DeviceBuffer delaysAfterEnd;
- DeviceBuffer phaseOffsets;
- DeviceBuffer inputSamples;
+ gpu::DeviceMemory delaysAtBegin;
+ gpu::DeviceMemory delaysAfterEnd;
+ gpu::DeviceMemory phaseOffsets;
+ gpu::DeviceMemory inputSamples;
DeviceBuffers(size_t n_beams, size_t n_stations, size_t n_polarizations,
- size_t n_samples, size_t bytes_per_complex_sample,
- gpu::Stream &queue,
- size_t inputSamplesMinSize = 0,
- cl_mem_flags deviceBufferFlags = CL_MEM_READ_ONLY)
+ size_t n_samples, size_t bytes_per_complex_sample,
+ gpu::Context &context, size_t inputSamplesMinSize = 0)
:
- delaysAtBegin(queue, deviceBufferFlags, n_beams * n_stations * n_polarizations * sizeof(float)),
- delaysAfterEnd(queue, deviceBufferFlags, n_beams * n_stations * n_polarizations * sizeof(float)),
- phaseOffsets(queue, deviceBufferFlags, n_stations * n_polarizations * sizeof(float)),
- inputSamples(queue, CL_MEM_READ_WRITE, std::max(inputSamplesMinSize, n_stations * n_samples * n_polarizations * bytes_per_complex_sample))
+ delaysAtBegin (context, n_beams * n_stations * n_polarizations * sizeof(float)),
+ delaysAfterEnd(context, n_beams * n_stations * n_polarizations * sizeof(float)),
+ phaseOffsets (context, n_stations * n_polarizations * sizeof(float)),
+ inputSamples (context, std::max(inputSamplesMinSize,
+ n_samples * n_stations * n_polarizations * bytes_per_complex_sample))
{
}
};
@@ -162,26 +169,30 @@ namespace LOFAR
// Relevant subband
unsigned subband;
- MultiArrayHostBuffer<float, 3> delaysAtBegin; //!< Whole sample delays at the start of the workitem
- MultiArrayHostBuffer<float, 3> delaysAfterEnd;//!< Whole sample delays at the end of the workitem
- MultiArrayHostBuffer<float, 2> phaseOffsets; //!< Remainder of delays
+ //!< Whole sample delays at the start of the workitem
+ MultiDimArrayHostBuffer<float, 3> delaysAtBegin;
+
+ //!< Whole sample delays at the end of the workitem
+ MultiDimArrayHostBuffer<float, 3> delaysAfterEnd;
+
+ //!< Remainder of delays
+ MultiDimArrayHostBuffer<float, 2> phaseOffsets;
// inputdata with flagged data set to zero
- MultiArrayHostBuffer<char, 4> inputSamples;
+ MultiDimArrayHostBuffer<char, 4> inputSamples;
// The input flags
- MultiDimArray<SparseSet<unsigned>,1> inputFlags;
+ MultiDimArray<SparseSet<unsigned>, 1> inputFlags;
// Create the inputData object we need shared host/device memory on the supplied devicequeue
WorkQueueInputData(size_t n_beams, size_t n_stations, size_t n_polarizations,
size_t n_samples, size_t bytes_per_complex_sample,
- DeviceBuffers &deviceBuffers,
- cl_mem_flags hostBufferFlags = CL_MEM_WRITE_ONLY)
+ unsigned int hostBufferFlags = 0)
:
- delaysAtBegin(boost::extents[n_beams][n_stations][n_polarizations], hostBufferFlags, deviceBuffers.delaysAtBegin),
- delaysAfterEnd(boost::extents[n_beams][n_stations][n_polarizations], hostBufferFlags, deviceBuffers.delaysAfterEnd),
- phaseOffsets(boost::extents[n_stations][n_polarizations], hostBufferFlags, deviceBuffers.phaseOffsets),
- inputSamples(boost::extents[n_stations][n_samples][n_polarizations][bytes_per_complex_sample], hostBufferFlags, deviceBuffers.inputSamples), // TODO: The size of the buffer is NOT validated
+ delaysAtBegin(boost::extents[n_beams][n_stations][n_polarizations], hostBufferFlags),
+ delaysAfterEnd(boost::extents[n_beams][n_stations][n_polarizations], hostBufferFlags),
+ phaseOffsets(boost::extents[n_stations][n_polarizations], hostBufferFlags),
+ inputSamples(boost::extents[n_stations][n_samples][n_polarizations][bytes_per_complex_sample], hostBufferFlags), // TODO: The size of the buffer is NOT validated
inputFlags(boost::extents[n_stations])
{
}
@@ -251,7 +262,7 @@ namespace LOFAR
// in the InputData class
WorkQueueInputData::DeviceBuffers devInput;
- DeviceBuffer devFilteredData;
+ gpu::DeviceMemory devFilteredData;
public:
// A pool of input data, to allow items to be filled and
@@ -263,11 +274,13 @@ namespace LOFAR
Pool<CorrelatedDataHostBuffer> outputPool;
private:
+ // Constant input buffers for the kernels
+ gpu::DeviceMemory devFIRweights;
+ gpu::DeviceMemory devBandPassCorrectionWeights;
+
// Compiled kernels
- DeviceBuffer devFIRweights;
FIR_FilterKernel firFilterKernel;
Filter_FFT_Kernel fftKernel;
- MultiArraySharedBuffer<float, 1> bandPassCorrectionWeights;
DelayAndBandPassKernel delayAndBandPassKernel;
#if defined USE_NEW_CORRELATOR
CorrelateTriangleKernel correlateTriangleKernel;
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/UHEP_WorkQueue.h b/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/UHEP_WorkQueue.h
index ec9e95809d2..2abeb81523d 100644
--- a/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/UHEP_WorkQueue.h
+++ b/RTCP/Cobalt/GPUProc/src/cuda/WorkQueues/UHEP_WorkQueue.h
@@ -27,7 +27,7 @@
#include <CoInterface/Parset.h>
#include <GPUProc/global_defines.h>
-#include <GPUProc/Buffers.h>
+#include <GPUProc/MultiDimArrayHostBuffer.h>
#include <GPUProc/Pipelines/UHEP_Pipeline.h>
#include <GPUProc/Kernels/UHEP_TriggerKernel.h>
#include "WorkQueue.h"
diff --git a/RTCP/Cobalt/GPUProc/src/opencl/Buffers.h b/RTCP/Cobalt/GPUProc/src/opencl/MultiDimArrayHostBuffer.h
similarity index 97%
rename from RTCP/Cobalt/GPUProc/src/opencl/Buffers.h
rename to RTCP/Cobalt/GPUProc/src/opencl/MultiDimArrayHostBuffer.h
index 191340906ac..41d48185c8c 100644
--- a/RTCP/Cobalt/GPUProc/src/opencl/Buffers.h
+++ b/RTCP/Cobalt/GPUProc/src/opencl/MultiDimArrayHostBuffer.h
@@ -1,4 +1,4 @@
-//# Buffers.h
+//# MultiDimArrayHostBuffer.h
//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
//#
@@ -18,8 +18,8 @@
//#
//# $Id$
-#ifndef LOFAR_GPUPROC_OPENCL_BUFFERS_H
-#define LOFAR_GPUPROC_OPENCL_BUFFERS_H
+#ifndef LOFAR_GPUPROC_OPENCL_MULTI_DIM_ARRAY_HOST_BUFFER_H
+#define LOFAR_GPUPROC_OPENCL_MULTI_DIM_ARRAY_HOST_BUFFER_H
#include <CoInterface/Allocator.h>
#include <CoInterface/MultiDimArray.h>
diff --git a/RTCP/Cobalt/GPUProc/src/opencl/Pipelines/Pipeline.cc b/RTCP/Cobalt/GPUProc/src/opencl/Pipelines/Pipeline.cc
index 376287aed29..dadb9d7483d 100644
--- a/RTCP/Cobalt/GPUProc/src/opencl/Pipelines/Pipeline.cc
+++ b/RTCP/Cobalt/GPUProc/src/opencl/Pipelines/Pipeline.cc
@@ -27,11 +27,6 @@
#include <GPUProc/gpu_utils.h>
-#if 0
-#include <boost/format.hpp>
-using boost::format;
-#endif
-
namespace LOFAR
{
namespace Cobalt
diff --git a/RTCP/Cobalt/GPUProc/src/opencl/WorkQueues/BeamFormerWorkQueue.h b/RTCP/Cobalt/GPUProc/src/opencl/WorkQueues/BeamFormerWorkQueue.h
index 6c07062aea9..aa0dbe0a5ae 100644
--- a/RTCP/Cobalt/GPUProc/src/opencl/WorkQueues/BeamFormerWorkQueue.h
+++ b/RTCP/Cobalt/GPUProc/src/opencl/WorkQueues/BeamFormerWorkQueue.h
@@ -26,7 +26,7 @@
#include <Common/LofarLogger.h>
#include <CoInterface/Parset.h>
-#include <GPUProc/Buffers.h>
+#include <GPUProc/MultiDimArrayHostBuffer.h>
#include <GPUProc/BandPass.h>
#include <GPUProc/Pipelines/BeamFormerPipeline.h>
diff --git a/RTCP/Cobalt/GPUProc/src/opencl/WorkQueues/CorrelatorWorkQueue.h b/RTCP/Cobalt/GPUProc/src/opencl/WorkQueues/CorrelatorWorkQueue.h
index d034bf4ebf9..3c507d4ec5c 100644
--- a/RTCP/Cobalt/GPUProc/src/opencl/WorkQueues/CorrelatorWorkQueue.h
+++ b/RTCP/Cobalt/GPUProc/src/opencl/WorkQueues/CorrelatorWorkQueue.h
@@ -33,7 +33,7 @@
#include <CoInterface/SubbandMetaData.h>
#include <GPUProc/global_defines.h>
-#include <GPUProc/Buffers.h>
+#include <GPUProc/MultiDimArrayHostBuffer.h>
#include <GPUProc/FilterBank.h>
#include <GPUProc/Pipelines/CorrelatorPipelinePrograms.h>
#include <GPUProc/Kernels/FIR_FilterKernel.h>
diff --git a/RTCP/Cobalt/GPUProc/src/opencl/WorkQueues/UHEP_WorkQueue.h b/RTCP/Cobalt/GPUProc/src/opencl/WorkQueues/UHEP_WorkQueue.h
index 11298670e2c..46c00cde1c3 100644
--- a/RTCP/Cobalt/GPUProc/src/opencl/WorkQueues/UHEP_WorkQueue.h
+++ b/RTCP/Cobalt/GPUProc/src/opencl/WorkQueues/UHEP_WorkQueue.h
@@ -27,7 +27,7 @@
#include <CoInterface/Parset.h>
#include <GPUProc/global_defines.h>
-#include <GPUProc/Buffers.h>
+#include <GPUProc/MultiDimArrayHostBuffer.h>
#include <GPUProc/Pipelines/UHEP_Pipeline.h>
#include <GPUProc/Kernels/UHEP_TriggerKernel.h>
#include "WorkQueue.h"
--
GitLab