fractorium/Source/EmberCL/OpenCLWrapper.h
mfeemster 018ba26b5f --User changes
-Add support for multiple GPU devices.
  --These options are present in the command line and in Fractorium.
 -Change scheme of specifying devices from platform,device to just total device index.
  --Single number on the command line.
  --Change from combo boxes for device selection to a table of all devices in Fractorium.
 -Temporal samples defaults to 100 instead of 1000 which was needless overkill.

--Bug fixes
 -EmberAnimate, EmberRender, FractoriumSettings, FinalRenderDialog: Fix wrong order of arguments to Clamp() when assigning thread priority.
 -VariationsDC.h: Fix NVidia OpenCL compilation error in DCTriangleVariation.
 -FractoriumXformsColor.cpp: Checking for null pixmap pointer is not enough, must also check if the underlying buffer is null via call to QPixmap::isNull().

--Code changes
 -Ember.h: Add case for FLAME_MOTION_NONE and default in ApplyFlameMotion().
 -EmberMotion.h: Call base constructor.
 -EmberPch.h: #pragma once only on Windows.
 -EmberToXml.h:
  --Handle different types of exceptions.
  --Add default cases to ToString().
 -Isaac.h: Remove unused variable in constructor.
 -Point.h: Call base constructor in Color().
 -Renderer.h/cpp:
  --Add bool to Alloc() to only allocate memory for the histogram. Needed for multi-GPU.
  --Make CoordMap() return a const ref, not a pointer.
 -SheepTools.h:
  --Use 64-bit types like the rest of the code already does.
  --Fix some comment misspellings.
 -Timing.h: Make BeginTime(), EndTime(), ElapsedTime() and Format() be const functions.
 -Utils.h:
  --Add new functions Equal() and Split().
  --Handle more exception types in ReadFile().
  --Get rid of most legacy blending of C and C++ argument parsing.
 -XmlToEmber.h:
  --Get rid of most legacy blending of C and C++ code from flam3.
  --Remove some unused variables.
 -EmberAnimate:
  --Support multi-GPU processing that alternates full frames between devices.
  --Use OpenCLInfo instead of OpenCLWrapper for --openclinfo option.
  --Remove bucketT template parameter, and hard code float in its place.
  --If a render fails, exit since there is no point in continuing an animation with a missing frame.
  --Pass variables to threaded save better, which most likely fixes a very subtle bug that existed before.
  --Remove some unused variables.
 -EmberGenome, EmberRender:
  --Support multi-GPU processing that alternates full frames between devices.
  --Use OpenCLInfo instead of OpenCLWrapper for --openclinfo option.
  --Remove bucketT template parameter, and hard code float in its place.
 -EmberRender:
  --Support multi-GPU processing that alternates full frames between devices.
  --Use OpenCLInfo instead of OpenCLWrapper for --openclinfo option.
  --Remove bucketT template parameter, and hard code float in its place.
  --Only print values when not rendering with OpenCL, since they're always 0 in that case.
 -EmberCLPch.h:
  --#pragma once only on Windows.
  --#include <atomic>.
 -IterOpenCLKernelCreator.h: Add new kernel for summing two histograms. This is needed for multi-GPU.
 -OpenCLWrapper.h:
  --Move all OpenCL info related code into its own class OpenCLInfo.
  --Add members to cache the values of global memory size and max allocation size.
 -RendererCL.h/cpp:
  --Redesign to accomodate multi-GPU.
  --Constructor now takes a vector of devices.
  --Remove DumpErrorReport() function, it's handled in the base.
  --ClearBuffer(), ReadPoints(), WritePoints(), ReadHist() and WriteHist() now optionally take a device index as a parameter.
  --MakeDmap() override and m_DmapCL member removed because it no longer applies since the histogram is always float since the last commit.
  --Add new function SumDeviceHist() to sum histograms from two devices by first copying to a temporary on the host, then a temporary on the device, then summing.
  --m_Calls member removed, as it's now per-device.
  --OpenCLWrapper removed.
  --m_Seeds member is now a vector of vector of seeds, to accomodate a separate and different array of seeds for each device.
  --Added member m_Devices, a vector of unique_ptr of RendererCLDevice.
 -EmberCommon.h
  --Added Devices() function to convert from a vector of device indices to a vector of platform,device indices.
  --Changed CreateRenderer() to accept a vector of devices to create a single RendererCL which will split work across multiple devices.
  --Added CreateRenderers() function to accept a vector of devices to create multiple RendererCL, each which will render on a single device.
  --Add more comments to some existing functions.
 -EmberCommonPch.h: #pragma once only on Windows.
 -EmberOptions.h:
  --Remove --platform option, it's just sequential device number now with the --device option.
  --Make --out be OPT_USE_RENDER instead of OPT_RENDER_ANIM since it's an error condition when animating. It makes no sense to write all frames to a single image.
  --Add Devices() function to parse comma separated --device option string and return a vector of device indices.
  --Make int and uint types be 64-bit, so intmax_t and size_t.
  --Make better use of macros.
 -JpegUtils.h: Make string parameters to WriteJpeg() and WritePng() be const ref.
 -All project files: Turn off buffer security check option in Visual Studio (/Gs-)
 -deployment.pri: Remove the line OTHER_FILES +=, it's pointless and was causing problems.
 -Ember.pro, EmberCL.pro: Add CONFIG += plugin, otherwise it wouldn't link.
 -EmberCL.pro: Add new files for multi-GPU support.
 -build_all.sh: use -j4 and QMAKE=${QMAKE:/usr/bin/qmake}
 -shared_settings.pri:
  -Add version string.
  -Remove old DESTDIR definitions.
  -Add the following lines or else nothing would build:
   CONFIG(release, debug|release) {
    CONFIG += warn_off
    DESTDIR = ../../../Bin/release
   }

   CONFIG(debug, debug|release) {
    DESTDIR = ../../../Bin/debug
   }

   QMAKE_POST_LINK += $$quote(cp --update ../../../Data/flam3-palettes.xml $${DESTDIR}$$escape_expand(\n\t))
   LIBS += -L/usr/lib -lpthread
 -AboutDialog.ui: Another futile attempt to make it look correct on Linux.
 -FinalRenderDialog.h/cpp:
  --Add support for multi-GPU.
  --Change from combo boxes for device selection to a table of all devices.
  --Ensure device selection makes sense.
  --Remove "FinalRender" prefix of various function names, it's implied given the context.
 -FinalRenderEmberController.h/cpp:
  --Add support for multi-GPU.
  --Change m_FinishedImageCount to be atomic.
  --Move CancelRender() from the base to FinalRenderEmberController<T>.
  --Refactor RenderComplete() to omit any progress related functionality or image saving since it can be potentially ran in a thread.
  --Consolidate setting various renderer fields into SyncGuiToRenderer().
 -Fractorium.cpp: Allow for resizing of the options dialog to show the entire device table.
 -FractoriumCommon.h: Add various functions to handle a table showing the available OpenCL devices on the system.
 -FractoriumEmberController.h/cpp: Remove m_FinalImageIndex, it's no longer needed.
 -FractoriumRender.cpp: Scale the interactive sub batch count and quality by the number of devices used.
 -FractoriumSettings.h/cpp:
  --Temporal samples defaults to 100 instead of 1000 which was needless overkill.
  --Add multi-GPU support, remove old device,platform pair.
 -FractoriumToolbar.cpp: Disable OpenCL toolbar button if there are no devices present on the system.
 -FractoriumOptionsDialog.h/cpp:
  --Add support for multi-GPU.
  --Consolidate more assignments in DataToGui().
  --Enable/disable CPU/OpenCL items in response to OpenCL checkbox event.
 -Misc: Convert almost everything to size_t for unsigned, intmax_t for signed.
2015-09-12 18:33:45 -07:00

212 lines
6.9 KiB
C++

#pragma once
#include "EmberCLPch.h"
#include "OpenCLInfo.h"
/// <summary>
/// OpenCLWrapper, Spk, NamedBuffer, NamedImage2D, NamedImage2DGL classes.
/// </summary>
namespace EmberCLns
{
#if CL_VERSION_1_2
#define IMAGEGL2D cl::ImageGL
#else
#define IMAGEGL2D cl::Image2DGL
#endif
/// <summary>
/// Class to contain all of the things needed to store an OpenCL program.
/// The name of it, the source, the compiled program object and the kernel.
/// </summary>
class EMBERCL_API Spk
{
public:
string m_Name;
cl::Program::Sources m_Source;
cl::Program m_Program;
cl::Kernel m_Kernel;
};
/// <summary>
/// Class to hold an OpenCL buffer with a name to identify it by.
/// </summary>
class EMBERCL_API NamedBuffer
{
public:
NamedBuffer()
{
}
NamedBuffer(const cl::Buffer& buff, const string& name)
{
m_Buffer = buff;
m_Name = name;
}
cl::Buffer m_Buffer;
string m_Name;
};
/// <summary>
/// Class to hold a 2D image with a name to identify it by.
/// </summary>
class EMBERCL_API NamedImage2D
{
public:
NamedImage2D()
{
}
NamedImage2D(const cl::Image2D& image, const string& name)
{
m_Image = image;
m_Name = name;
}
cl::Image2D m_Image;
string m_Name;
};
/// <summary>
/// Class to hold a 2D image that is mapped to an OpenGL texture
/// and a name to identify it by.
/// </summary>
class EMBERCL_API NamedImage2DGL
{
public:
NamedImage2DGL()
{
}
NamedImage2DGL(const IMAGEGL2D& image, const string& name)
{
m_Image = image;
m_Name = name;
}
IMAGEGL2D m_Image;
string m_Name;
};
/// <summary>
/// Running kernels in OpenCL can require quite a bit of setup, tear down and
/// general housekeeping. This class helps shield the user from such hassles.
/// Its main utility is in holding collections of programs, buffers and images
/// all identified by names. That way, a user can access them as needed without
/// having to pollute their code.
/// In addition, writing to an existing object by name determines if the object
/// can be overwritten, or if it needs to be deleted and replaced by the new one.
/// This class derives from EmberReport, so the caller is able
/// to retrieve a text dump of error information if any errors occur.
/// </summary>
class EMBERCL_API OpenCLWrapper : public EmberReport
{
public:
OpenCLWrapper();
bool Init(size_t platformIndex, size_t deviceIndex, bool shared = false);
//Programs.
bool AddProgram(const string& name, const string& program, const string& entryPoint, bool doublePrecision);
void ClearPrograms();
//Buffers.
bool AddBuffer(const string& name, size_t size, cl_mem_flags flags = CL_MEM_READ_WRITE);
bool AddAndWriteBuffer(const string& name, void* data, size_t size, cl_mem_flags flags = CL_MEM_READ_WRITE);
bool WriteBuffer(const string& name, void* data, size_t size);
bool WriteBuffer(size_t bufferIndex, void* data, size_t size);
bool ReadBuffer(const string& name, void* data, size_t size);
bool ReadBuffer(size_t bufferIndex, void* data, size_t size);
int FindBufferIndex(const string& name);
size_t GetBufferSize(const string& name);
size_t GetBufferSize(size_t bufferIndex);
void ClearBuffers();
//Images.
bool AddAndWriteImage(const string& name, cl_mem_flags flags, const cl::ImageFormat& format, ::size_t width, ::size_t height, ::size_t row_pitch, void* data = NULL, bool shared = false, GLuint texName = 0);
bool WriteImage2D(size_t index, bool shared, size_t width, size_t height, size_t row_pitch, void* data);
bool ReadImage(const string& name, ::size_t width, ::size_t height, ::size_t row_pitch, bool shared, void* data);
bool ReadImage(size_t imageIndex, ::size_t width, ::size_t height, ::size_t row_pitch, bool shared, void* data);
int FindImageIndex(const string& name, bool shared);
size_t GetImageSize(const string& name, bool shared);
size_t GetImageSize(size_t imageIndex, bool shared);
bool CompareImageParams(cl::Image& image, cl_mem_flags flags, const cl::ImageFormat& format, ::size_t width, ::size_t height, ::size_t row_pitch);
void ClearImages(bool shared);
bool CreateImage2D(cl::Image2D& image2D, cl_mem_flags flags, cl::ImageFormat format, ::size_t width, ::size_t height, ::size_t row_pitch = 0, void* data = NULL);
bool CreateImage2DGL(IMAGEGL2D& image2DGL, cl_mem_flags flags, GLenum target, GLint miplevel, GLuint texobj);
bool EnqueueAcquireGLObjects(const string& name);
bool EnqueueAcquireGLObjects(IMAGEGL2D& image);
bool EnqueueReleaseGLObjects(const string& name);
bool EnqueueReleaseGLObjects(IMAGEGL2D& image);
bool EnqueueAcquireGLObjects(const VECTOR_CLASS<cl::Memory>* memObjects = NULL);
bool EnqueueReleaseGLObjects(const VECTOR_CLASS<cl::Memory>* memObjects = NULL);
bool CreateSampler(cl::Sampler& sampler, cl_bool normalizedCoords, cl_addressing_mode addressingMode, cl_filter_mode filterMode);
//Arguments.
bool SetBufferArg(size_t kernelIndex, cl_uint argIndex, const string& name);
bool SetBufferArg(size_t kernelIndex, cl_uint argIndex, size_t bufferIndex);
bool SetImageArg(size_t kernelIndex, cl_uint argIndex, bool shared, const string& name);
bool SetImageArg(size_t kernelIndex, cl_uint argIndex, bool shared, size_t imageIndex);
/// <summary>
/// Set an argument in the specified kernel, at the specified argument index.
/// Must keep this here in the .h because it's templated.
/// </summary>
/// <param name="kernelIndex">Index of the kernel whose argument will be set</param>
/// <param name="argIndex">Index of the argument to set</param>
/// <param name="arg">The argument value to set</param>
/// <returns>True if success, else false</returns>
template <typename T>
bool SetArg(size_t kernelIndex, cl_uint argIndex, T arg)
{
if (m_Init && kernelIndex < m_Programs.size())
{
cl_int err = m_Programs[kernelIndex].m_Kernel.setArg(argIndex, arg);
return m_Info.CheckCL(err, "cl::Kernel::setArg()");
}
return false;
}
//Kernels.
int FindKernelIndex(const string& name);
bool RunKernel(size_t kernelIndex, size_t totalGridWidth, size_t totalGridHeight, size_t totalGridDepth, size_t blockWidth, size_t blockHeight, size_t blockDepth);
//Accessors.
bool Ok() const;
bool Shared() const;
const cl::Context& Context() const;
size_t PlatformIndex() const;
size_t DeviceIndex() const;
const string& DeviceName() const;
size_t TotalDeviceIndex() const;
size_t LocalMemSize() const;
size_t GlobalMemSize() const;
size_t MaxAllocSize() const;
static void MakeEvenGridDims(size_t blockW, size_t blockH, size_t& gridW, size_t& gridH);
private:
bool CreateSPK(const string& name, const string& program, const string& entryPoint, Spk& spk, bool doublePrecision);
bool m_Init;
bool m_Shared;
size_t m_PlatformIndex;
size_t m_DeviceIndex;
size_t m_LocalMemSize;
size_t m_GlobalMemSize;
size_t m_MaxAllocSize;
cl::Platform m_Platform;
cl::Context m_Context;
cl::Device m_Device;
cl::CommandQueue m_Queue;
OpenCLInfo& m_Info;
std::vector<cl::Device> m_DeviceVec;
std::vector<Spk> m_Programs;
std::vector<NamedBuffer> m_Buffers;
std::vector<NamedImage2D> m_Images;
std::vector<NamedImage2DGL> m_GLImages;
};
}