cxVNNclAlgorithm.cpp

Go to the documentation of this file.

/*=========================================================================
This file is part of CustusX, an Image Guided Therapy Application.

Copyright (c) 2008-2014, SINTEF Department of Medical Technology
All rights reserved.

Redistribution and use in source and binary forms, with or without 
modification, are permitted provided that the following conditions are met:

1. Redistributions of source code must retain the above copyright notice, 
   this list of conditions and the following disclaimer.

2. Redistributions in binary form must reproduce the above copyright notice, 
   this list of conditions and the following disclaimer in the documentation 
   and/or other materials provided with the distribution.

3. Neither the name of the copyright holder nor the names of its contributors 
   may be used to endorse or promote products derived from this software 
   without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" 
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE 
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER 
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
=========================================================================*/

#include "cxVNNclAlgorithm.h"

#include <QFileInfo>
#include "cxLogger.h"
#include "cxTypeConversions.h"
#include "HelperFunctions.hpp"
#include <vtkImageData.h>
#include <recConfig.h>
#include "cxVolumeHelpers.h"

namespace cx
{
VNNclAlgorithm::VNNclAlgorithm() :
                mRuntime(new oul::RuntimeMeasurementsManager()),
                mKernelMeasurementName("vnncl_execute_kernel")
{
        oul::DeviceCriteria criteria;
        criteria.setTypeCriteria(oul::DEVICE_TYPE_GPU);
        bool enableProfilling = true;
        mOulContex = oul::opencl()->createContextPtr(criteria, NULL, enableProfilling);
        if(enableProfilling)
                mRuntime->enable();
}

VNNclAlgorithm::~VNNclAlgorithm()
{}

bool VNNclAlgorithm::initCL(QString kernelFilePath, int nMaxPlanes, int nPlanes, int method, int planeMethod, int nStarts, float brightnessWeight, float newnessWeight)
{
        // READ KERNEL FILE
        report(QString("Kernel path: %1").arg(kernelFilePath));
        std::string source = oul::readFile(kernelFilePath.toStdString());

        QFileInfo path(kernelFilePath);
        path.absolutePath();

        // BUILD PROGRAM
        cl::Program clprogram = this->buildCLProgram(source, path.absolutePath().toStdString(), nMaxPlanes, nPlanes, method, planeMethod, nStarts,brightnessWeight, newnessWeight);

        // CREATE KERNEL
        mKernel = mOulContex->createKernel(clprogram, "voxel_methods");

        return true;
}

cl::Program VNNclAlgorithm::buildCLProgram(std::string program_src, std::string kernelPath, int nMaxPlanes, int nPlanes, int method, int planeMethod, int nStarts, float newnessWeight, float brightnessWeight)
{
        cl::Program retval;
        cl_int err = 0;
        VECTOR_CLASS<cl::Device> devices;
        try
        {
                QString define = "-D MAX_PLANES=%1 -D N_PLANES=%2 -D METHOD=%3 -D PLANE_METHOD=%4 -D MAX_MULTISTART_STARTS=%5 -D NEWNESS_FACTOR=%6 -D BRIGHTNESS_FACTOR=%7";
                define = define.arg(nMaxPlanes).arg(nPlanes).arg(method).arg(planeMethod).arg(nStarts).arg(newnessWeight).arg(brightnessWeight);

                int programID = mOulContex->createProgramFromString(program_src, "-I " + std::string(kernelPath) + " " + define.toStdString());
                retval = mOulContex->getProgram(programID);

        } catch (cl::Error &error)
        {
                reportError("Could not build a OpenCL program. Reason: "+QString(error.what()));
                reportError(qstring_cast(oul::getCLErrorString(error.err())));
        }
        return retval;
}

bool VNNclAlgorithm::initializeFrameBlocks(frameBlock_t* framePointers, int numBlocks, ProcessedUSInputDataPtr inputFrames)
{
        // Compute the size of each frame in bytes
        Eigen::Array3i dims = inputFrames->getDimensions();
        size_t frameSize = dims[0] * dims[1];
        size_t numFrames = dims[2];
        report(QString("Input dims: (%1, %2, %3)").arg(dims[0]).arg(dims[1]).arg(dims[2]));

        // Find out how many frames needs to be in each block
        size_t framesPerBlock = numFrames / numBlocks;
        report(QString("Frames: %1, Blocks: %2, Frames per block: %3").arg(numFrames).arg(numBlocks).arg(framesPerBlock));

        // Some blocks will need to contain one extra frame
        // (numFrames and numBlocks is probably not evenly divisible)
        size_t numBigBlocks = numFrames % numBlocks;

        // Allocate the big blocks
        report(QString("Allocating %1 big blocks outside of OpenCL").arg(numBigBlocks));
        for (unsigned int block = 0; block < numBigBlocks; block++)
        {
                framePointers[block].length = (1 + framesPerBlock) * frameSize;
                framePointers[block].data = new unsigned char[framePointers[block].length];
        }

        // Then the small ones
        report(QString("Allocating %1 small blocks outside of OpenCL").arg(numBlocks - numBigBlocks));
        for (int block = numBigBlocks; block < numBlocks; block++)
        {
                framePointers[block].length = (framesPerBlock) * frameSize;
                framePointers[block].data = new unsigned char[framePointers[block].length];
        }

        // Now fill them
        unsigned int frame = 0;
        for (int block = 0; block < numBlocks; block++)
        {
                for (unsigned int frameInThisBlock = 0; frameInThisBlock < framePointers[block].length / frameSize; frameInThisBlock++)
                {
                        memcpy(&(framePointers[block].data[frameInThisBlock * frameSize]), inputFrames->getFrame(frame), frameSize);
                        frame++;
                }
        }
        return true;
}

bool VNNclAlgorithm::reconstruct(ProcessedUSInputDataPtr input, vtkImageDataPtr outputData, float radius, int nClosePlanes)
{
        mMeasurementNames.clear();

        int numBlocks = 10; // FIXME? needs to be the same as the number of input bscans to the voxel_method kernel

        // Split input US into blocks
        // Splits and copies data from the processed input in the way the kernel will processes it, which is per frameBlock
        frameBlock_t* inputBlocks = new frameBlock_t[numBlocks];
        size_t nPlanes_numberOfInputImages = input->getDimensions()[2];
        this->initializeFrameBlocks(inputBlocks, numBlocks, input);

        // Allocate CL memory for each frame block
        VECTOR_CLASS<cl::Buffer> clBlocks;
        report("Allocating OpenCL input block buffers");
        for (int i = 0; i < numBlocks; i++)
        {
                //TODO why does the context suddenly contain a "dummy" device?
                cl::Buffer buffer = mOulContex->createBuffer(mOulContex->getContext(), CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, inputBlocks[i].length, inputBlocks[i].data, "block buffer "+QString::number(i).toStdString());
                clBlocks.push_back(buffer);
        }
        // Allocate output memory
        int *outputDims = outputData->GetDimensions();

        size_t outputVolumeSize = outputDims[0] * outputDims[1] * outputDims[2] * sizeof(unsigned char);

        report(QString("Allocating CL output buffer, size %1").arg(outputVolumeSize));

        cl_ulong globalMemUse = 10 * inputBlocks[0].length + outputVolumeSize + sizeof(float) * 16 * nPlanes_numberOfInputImages + sizeof(cl_uchar) * input->getDimensions()[0] * input->getDimensions()[1];
        if(isUsingTooMuchMemory(outputVolumeSize, inputBlocks[0].length, globalMemUse))
                return false;

        cl::Buffer outputBuffer = mOulContex->createBuffer(mOulContex->getContext(), CL_MEM_WRITE_ONLY, outputVolumeSize, NULL, "output volume buffer");

        // Fill the plane matrices
        float *planeMatrices = new float[16 * nPlanes_numberOfInputImages]; //4x4 (matrix) = 16
        this->fillPlaneMatrices(planeMatrices, input);

        cl::Buffer clPlaneMatrices = mOulContex->createBuffer(mOulContex->getContext(), CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, nPlanes_numberOfInputImages * sizeof(float) * 16, planeMatrices, "plane matrices buffer");

        // US Probe mask
        cl::Buffer clMask = mOulContex->createBuffer(mOulContex->getContext(), CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
                        sizeof(cl_uchar) * input->getMask()->GetDimensions()[0] * input->getMask()->GetDimensions()[1],
                        input->getMask()->GetScalarPointer(), "mask buffer");

        double *out_spacing = outputData->GetSpacing();
        float spacings[2];
        float f_out_spacings[3];
        f_out_spacings[0] = out_spacing[0];
        f_out_spacings[1] = out_spacing[1];
        f_out_spacings[2] = out_spacing[2];


        spacings[0] = input->getSpacing()[0];
        spacings[1] = input->getSpacing()[1];

        //TODO why 4? because float4 is used??
        size_t planes_eqs_size =  sizeof(cl_float)*4*nPlanes_numberOfInputImages;

        // Find the optimal local work size
        size_t local_work_size;
        unsigned int deviceNumber = 0;
        cl::Device device = mOulContex->getDevice(deviceNumber);
        mKernel.getWorkGroupInfo(device, CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE, &local_work_size);

        size_t close_planes_size = this->calculateSpaceNeededForClosePlanes(mKernel, device, local_work_size, nPlanes_numberOfInputImages, nClosePlanes);

        this->setKernelArguments(
                        mKernel,
                        outputDims[0],
                        outputDims[1],
                        outputDims[2],
                        f_out_spacings[0],
                        f_out_spacings[1],
                        f_out_spacings[2],
                        input->getDimensions()[0],
                        input->getDimensions()[1],
                        spacings[0],
                        spacings[1],
                        clBlocks,
                        outputBuffer,
                        clPlaneMatrices,
                        clMask,
                        planes_eqs_size,
                        close_planes_size,
                        radius);

        report(QString("Using %1 as local workgroup size").arg(local_work_size));

        // We will divide the work into cubes of CUBE_DIM^3 voxels. The global work size is the total number of voxels divided by that.
        int cube_dim = 4;
        int cube_dim_pow3 = cube_dim * cube_dim * cube_dim;
        // Global work items:
        size_t global_work_size = (((outputDims[0] + cube_dim) * (outputDims[1] + cube_dim) * (outputDims[2] + cube_dim)) / cube_dim_pow3); // = number of cubes = number of kernels to run

        // Round global_work_size up to nearest multiple of local_work_size
        if (global_work_size % local_work_size)
                global_work_size = ((global_work_size / local_work_size) + 1) * local_work_size; // ceil(...)

        unsigned int queueNumber = 0;
        cl::CommandQueue queue = mOulContex->getQueue(queueNumber);
        this->measureAndExecuteKernel(queue, mKernel, global_work_size, local_work_size, mKernelMeasurementName);
        this->measureAndReadBuffer(queue, outputBuffer, outputVolumeSize, outputData->GetScalarPointer(), "vnncl_read_buffer");
        setDeepModified(outputData);
        // Cleaning up
        report(QString("Done, freeing GPU memory"));
        this->freeFrameBlocks(inputBlocks, numBlocks);
        delete[] inputBlocks;

        inputBlocks = NULL;

        return true;
}

void VNNclAlgorithm::fillPlaneMatrices(float *planeMatrices, ProcessedUSInputDataPtr input)
{
        std::vector<TimedPosition> vecPosition = input->getFrames();

        // Sanity check on the number of frames
        if (input->getDimensions()[2] != vecPosition.end() - vecPosition.begin())
        {
                reportError(QString("Number of frames %1 != %2 dimension 2 of US input").arg(input->getDimensions()[2]).arg(vecPosition.end() - vecPosition.begin()));
                return;
        }

        int i = 0;
        for (std::vector<TimedPosition>::iterator it = vecPosition.begin(); it != vecPosition.end(); ++it)
        {
                Transform3D pos = it->mPos;

                // Now store the result in the output
                for (int j = 0; j < 16; j++)
                {
                        planeMatrices[i++] = pos(j / 4, j % 4);
                }
        }
}

void VNNclAlgorithm::setProfiling(bool on)
{
        if(on)
                mRuntime->enable();
        else
                mRuntime->disable();
}

double VNNclAlgorithm::getTotalExecutionTime()
{
        double totalExecutionTime = -1;
        std::set<std::string>::iterator it;
        for(it = mMeasurementNames.begin(); it != mMeasurementNames.end(); ++it)
        {
                oul::RuntimeMeasurement measurement =  mRuntime->getTiming(*it);
                totalExecutionTime += measurement.getSum();
        }
        return totalExecutionTime;
}

double VNNclAlgorithm::getKernelExecutionTime()
{
//      double kernelExecutionTime = -1;
        return mRuntime->getTiming(mKernelMeasurementName).getSum();

}

void VNNclAlgorithm::freeFrameBlocks(frameBlock_t *framePointers, int numBlocks)
{
        for (int i = 0; i < numBlocks; i++)
        {
                delete[] framePointers[i].data;
        }
}

void VNNclAlgorithm::setKernelArguments(
                cl::Kernel kernel,
                int volume_xsize,
        int volume_ysize,
        int volume_zsize,
        float volume_xspacing,
        float volume_yspacing,
        float volume_zspacing,
        int in_xsize,
        int in_ysize,
        float in_xspacing,
        float in_yspacing,
        std::vector<cl::Buffer>& blocks,
        cl::Buffer out_volume,
        cl::Buffer plane_matrices,
        cl::Buffer mask,
        size_t plane_eqs_size,
        size_t close_planes_size,
        float radius)
{
        int arg = 0;
        kernel.setArg(arg++, volume_xsize);
        kernel.setArg(arg++, volume_ysize);
        kernel.setArg(arg++, volume_zsize);
        kernel.setArg(arg++, volume_xspacing);
        kernel.setArg(arg++, volume_yspacing);
        kernel.setArg(arg++, volume_zspacing);
        kernel.setArg(arg++, in_xsize);
        kernel.setArg(arg++, in_ysize);
        kernel.setArg(arg++, in_xspacing);
        kernel.setArg(arg++, in_yspacing);
        for (int i = 0; i < blocks.size(); i++)
        {
                kernel.setArg(arg++, blocks[i]);
        }
        kernel.setArg(arg++, out_volume);
        kernel.setArg(arg++, plane_matrices);
        kernel.setArg(arg++, mask);
        kernel.setArg<cl::LocalSpaceArg>(arg++, cl::__local(plane_eqs_size));
        kernel.setArg<cl::LocalSpaceArg>(arg++, cl::__local(close_planes_size));
        kernel.setArg(arg++, radius);
}

size_t VNNclAlgorithm::calculateSpaceNeededForClosePlanes(cl::Kernel kernel, cl::Device device, size_t local_work_size, size_t nPlanes_numberOfInputImages, int nClosePlanes)
{
        // Find out how much local memory the device has
        size_t dev_local_mem_size;
        dev_local_mem_size = device.getInfo<CL_DEVICE_LOCAL_MEM_SIZE>();

        // Find the maximum work group size
        size_t max_work_size;
        kernel.getWorkGroupInfo(device, CL_KERNEL_WORK_GROUP_SIZE, &max_work_size);

        // Now find the largest multiple of the preferred work group size that will fit into local mem
        size_t constant_local_mem = sizeof(cl_float) * 4 * nPlanes_numberOfInputImages;

        size_t varying_local_mem = (sizeof(cl_float) + sizeof(cl_short) + sizeof(cl_uchar) + sizeof(cl_uchar)) * (nClosePlanes + 1);  //see _close_plane struct in kernels.cl
        report(QString("Device has %1 bytes of local memory").arg(dev_local_mem_size));
        dev_local_mem_size -= constant_local_mem + 128; //Hmmm? 128?

        // How many work items can the local mem support?
        size_t maxItems = dev_local_mem_size / varying_local_mem;
        // And what is the biggest multiple of local_work_size that fits into that?
        int multiple = maxItems / local_work_size;

        if(multiple == 0)
        {
                // If the maximum amount of work items is smaller than the preferred multiple, we end up here.
                // This means that the local memory use is so big we can't even fit into the preferred multiple, and
                // have use a sub-optimal local work size.
                local_work_size = std::min(max_work_size, maxItems);
        }
        else
        {
                // Otherwise, we make it fit into the local work size.
                local_work_size = std::min(max_work_size, multiple * local_work_size);
        }

        size_t close_planes_size = varying_local_mem*local_work_size;

        return close_planes_size;
}

bool VNNclAlgorithm::isUsingTooMuchMemory(size_t outputVolumeSize, size_t inputBlocksLength, cl_ulong globalMemUse)
{
        bool usingTooMuchMemory = false;

        unsigned int deviceNumber = 0;
        cl_ulong maxAllocSize = mOulContex->getDevice(deviceNumber).getInfo<CL_DEVICE_MAX_MEM_ALLOC_SIZE>();
        cl_ulong globalMemSize = mOulContex->getDevice(deviceNumber).getInfo<CL_DEVICE_GLOBAL_MEM_SIZE>();
        if (maxAllocSize < outputVolumeSize)
        {
                reportError(QString("Output volume size too large! %1 > %2\n").arg(outputVolumeSize).arg(maxAllocSize));
                usingTooMuchMemory = true;
        }

        if (maxAllocSize < inputBlocksLength)
        {
                reportError(QString("Input blocks too large! %1 > %2\n").arg(inputBlocksLength).arg(maxAllocSize));
                usingTooMuchMemory = true;
        }

        if (globalMemSize < globalMemUse)
        {
                reportError(QString("Using too much global memory! %1 > %2").arg(globalMemUse).arg(globalMemSize));
                usingTooMuchMemory = true;
        }

        report(QString("Using %1 of %2 global memory").arg(globalMemUse).arg(globalMemSize));
        return usingTooMuchMemory;
}

void VNNclAlgorithm::measureAndExecuteKernel(cl::CommandQueue queue, cl::Kernel kernel, size_t global_work_size, size_t local_work_size, std::string measurementName)
{
        this->startProfiling(measurementName, queue);
        mOulContex->executeKernel(queue, kernel, global_work_size, local_work_size);
        this->stopProfiling(measurementName, queue);
}

void VNNclAlgorithm::measureAndReadBuffer(cl::CommandQueue queue, cl::Buffer outputBuffer, size_t outputVolumeSize, void *outputData, std::string measurementName)
{
        this->startProfiling(measurementName, queue);
        mOulContex->readBuffer(queue, outputBuffer, outputVolumeSize, outputData);
        this->stopProfiling(measurementName, queue);
}

void VNNclAlgorithm::startProfiling(std::string name, cl::CommandQueue queue) {
        if(!mRuntime->isEnabled())
                return;

        mRuntime->startCLTimer(name, queue);
        mMeasurementNames.insert(name);
}

void VNNclAlgorithm::stopProfiling(std::string name, cl::CommandQueue queue) {
        if(!mRuntime->isEnabled())
                return;

        mRuntime->stopCLTimer(name, queue);
        mRuntime->printAll();
}



} /* namespace cx */