public GPU_func()
{
km = CudafyTranslator.Cudafy();
gpu = CudafyHost.GetDevice(CudafyModes.Target, CudafyModes.DeviceId);
gpu.LoadModule(km);
GPU_prop = gpu.GetDeviceProperties();
}
///// <summary>
///// Initializes a new instance of the <see cref="KernelMethodInfo"/> class.
///// </summary>
///// <param name="type">The type.</param>
///// <param name="method">The method.</param>
///// <param name="gpuMethodType">Type of the gpu method.</param>
///// <param name="noDummyInclude"></param>
///// <param name="parentModule"></param>
//public KernelMethodInfo(Type type, MethodInfo method, eKernelMethodType gpuMethodType, bool noDummyInclude, CudafyModule parentModule)
// : this(type, method, gpuMethodType, false, false, parentModule)
//{
//}
/// <summary>
/// Initializes a new instance of the <see cref="KernelMethodInfo"/> class.
/// </summary>
/// <param name="type">The type.</param>
/// <param name="method">The method.</param>
/// <param name="gpuMethodType">Type of the gpu method.</param>
/// <param name="isDummy">if set to <c>true</c> is dummy.</param>
/// <param name="behaviour"></param>
/// <param name="parentModule">Module of which this is a part.</param>
public KernelMethodInfo(Type type, MethodInfo method, eKernelMethodType gpuMethodType, bool isDummy, eCudafyDummyBehaviour behaviour, CudafyModule parentModule)
{
Type = type;
Method = method;
MethodType = gpuMethodType;
DeserializedChecksum = 0;
IsDummy = isDummy;
Behaviour = behaviour;
ParentModule = parentModule;
}
private void miOpen_Click(object sender, EventArgs e)
{
try
{
if (openFileDialog.ShowDialog() == System.Windows.Forms.DialogResult.OK)
{
_module = CudafyModule.Deserialize(openFileDialog.FileName);
FillForm();
}
}
catch (Exception ex)
{
HandleException(ex);
}
}
public GpuRenderer()
{
var availableOpenCLDevices = CudafyHost.GetDeviceProperties(eGPUType.OpenCL);
if (availableOpenCLDevices.Any() == false)
{
throw new Exception("No OpenCL devices found...");
}
var device = availableOpenCLDevices.First();
Module = CudafyTranslator.Cudafy(eArchitecture.OpenCL12);
var blockSide =
Enumerable
.Range(1, 15)
.Reverse()
.First(count => count * count <= device.MaxThreadsPerBlock);
BlockSize = new dim3(blockSide, blockSide);
// Initialize gpu and load the module (avoids reloading every time)
gpu = CudafyHost.GetDevice(eGPUType.OpenCL);
gpu.LoadModule(Module);
}
public static void cudaTransposeAndMultiply(ref MathNet.Numerics.LinearAlgebra.Double.DenseMatrix dm)
{
Cudafy.CudafyModule km = Cudafy.Translator.CudafyTranslator.Cudafy();
km.Serialize();
GPGPU gpu = CudafyHost.GetDevice(eGPUType.Cuda);
int cols = dm.ColumnCount, rows = dm.RowCount;
dm.Storage.ToColumnMajorArray();
double[] a = dm.ToColumnWiseArray();
dm = new MathNet.Numerics.LinearAlgebra.Double.DenseMatrix(1, 1);
double[] dev_a = gpu.Allocate <double>(a.Length);
GPGPUBLAS blas = GPGPUBLAS.Create(gpu);
double[] a_d = gpu.CopyToDevice <double>(a);
double[] c_d = gpu.Allocate <double>(cols * cols);
gpu.StartTimer();
blas.GEMM(cols, rows, cols, 1, a_d, a_d, 0, c_d, Cudafy.Maths.BLAS.Types.cublasOperation.T);
a = new double[cols * cols];
gpu.CopyFromDevice <double>(c_d, a);
gpu.FreeAll();
dm = new MathNet.Numerics.LinearAlgebra.Double.DenseMatrix(cols, cols, a);
}
public static double DTW(double[] _x, double[] _y)
{
// gpu.EnableMultithreading();
// gpu.EnableSmartCopy();
CudafyModes.Target = eGPUType.Cuda;
CudafyTranslator.AllowClasses = true;
CudafyTranslator.GenerateDebug = true;
gpu = CudafyHost.GetDevice(CudafyModes.Target, 0);
arch = gpu.GetArchitecture();
km = CudafyTranslator.Cudafy(arch);
// km = CudafyTranslator.Cudafy(arch);
// gpu.EnableMultithreading();
// gpu.SetCurrentContext();
gpu.LoadModule(km);
return CalcDTW(_x, _y);
}
private static CudafyModule Deserialize(CudafyModule km, XDocument doc, string path)
{
XElement root = doc.Element(csCUDAFYMODULE);
if (root == null)
throw new XmlException(string.Format(GES.csELEMENT_X_NOT_FOUND, csCUDAFYMODULE));
string vStr = root.GetAttributeValue(csVERSION);
Version version = new Version(vStr);
Version curVers = typeof(CudafyModule).Assembly.GetName().Version;
if (version < new Version(1, 25))
throw new CudafyException(CudafyException.csVERSION_MISMATCH_EXPECTED_X_GOT_X, "1.25 or later", version);
if (version.Major != curVers.Major || version.Minor > curVers.Minor)
throw new CudafyException(CudafyException.csVERSION_MISMATCH_EXPECTED_X_GOT_X, curVers, version);
string name = root.TryGetAttributeValue(csNAME);
km.Name = name != null ? name : km.Name;
// Cuda Source
bool? hasCudaSrc = root.TryGetAttributeBoolValue(csHASCUDASOURCECODE);
XElement scse = root.Element(csSOURCECODES);
if (hasCudaSrc == true && scse == null) // Legacy support
{
km.SourceCode = root.TryGetElementBase64(csCUDASOURCECODE);
}
// else if sourcecodes node
else if (scse != null)
{
foreach (var sce in scse.Elements(csSOURCECODEFILE))
{
SourceCodeFile scf = new SourceCodeFile();
scf.ID = sce.TryGetAttributeValue(csID);
scf.Language = sce.TryGetAttributeEnum<eLanguage>(csLANGUAGE);
scf.Source = UnicodeEncoding.ASCII.GetString(Convert.FromBase64String(sce.Value));
scf.Architecture = sce.TryGetAttributeEnum<eArchitecture>(csARCH);
km._sourceCodes.Add(scf);
}
}
else
{
km.SourceCode = string.Empty;
}
bool? hasDebug = root.TryGetAttributeBoolValue(csDEBUGINFO);
km.GenerateDebug = hasDebug.HasValue && hasDebug.Value;
// PTX
bool? hasPtx = root.TryGetAttributeBoolValue(csHASPTX);
if (hasPtx == true && root.Element(csPTX) != null) // legacy support V0.3 or less
{
string ptx = root.Element(csPTX).Value;
byte[] ba = Convert.FromBase64String(ptx);
km._PTXModules.Add(new PTXModule() { PTX = UnicodeEncoding.ASCII.GetString(ba), Platform = km.CurrentPlatform, Architecture = eArchitecture.Unknown });
}
else if (hasPtx == true)
{
XElement ptxsxe = root.Element(csPTXMODULES);
foreach (XElement xe in ptxsxe.Elements(csPTXMODULE))
{
string ptx = xe.Value; //xe.Element(csPTX).Value;
string platformStr = xe.GetAttributeValue(csPLATFORM);
ePlatform platform = (ePlatform)Enum.Parse(typeof(ePlatform), platformStr);
string archStr = xe.TryGetAttributeValue(csARCH);
eArchitecture arch = eArchitecture.Unknown;
string sourcecodeId = xe.TryGetAttributeValue(csSOURCECODEFILE);
if (archStr != null)
arch = (eArchitecture)Enum.Parse(typeof(eArchitecture), archStr);
byte[] ba = Convert.FromBase64String(ptx);
km._PTXModules.Add(new PTXModule() { PTX = UnicodeEncoding.ASCII.GetString(ba), Platform = platform, Architecture = arch, SourceCodeID = sourcecodeId });
}
}
// Binary
bool? hasBinary = root.TryGetAttributeBoolValue(csHASBINARY);
if (hasBinary == true)
{
XElement binsxe = root.Element(csBINARYMODULES);
foreach (XElement xe in binsxe.Elements(csBINARYMODULE))
{
string bin = xe.Value; //xe.Element(csBINARY).Value;
string platformStr = xe.GetAttributeValue(csPLATFORM);
string archStr = xe.GetAttributeValue(csARCH);
ePlatform platform = (ePlatform)Enum.Parse(typeof(ePlatform), platformStr);
eArchitecture arch = (eArchitecture)Enum.Parse(typeof(eArchitecture), archStr);
string sourcecodeId = xe.TryGetAttributeValue(csSOURCECODEFILE);
byte[] ba = Convert.FromBase64String(bin);
km._BinaryModules.Add(new BinaryModule() { Binary = ba, Platform = platform, Architecture = arch, SourceCodeID = sourcecodeId });
}
}
// Functions
XElement funcs = root.Element(csFUNCTIONS);
if (funcs != null)
{
foreach (var xe in funcs.Elements(KernelMethodInfo.csCUDAFYKERNELMETHOD))
{
KernelMethodInfo kmi = KernelMethodInfo.Deserialize(xe, km, path);
km.Functions.Add(kmi.Method.Name, kmi);
}
}
// Constants
XElement constants = root.Element(csCONSTANTS);
if (constants != null)
{
foreach (var xe in constants.Elements(KernelConstantInfo.csCUDAFYCONSTANTINFO))
{
KernelConstantInfo kci = KernelConstantInfo.Deserialize(xe, path);
km.Constants.Add(kci.Name, kci);
}
}
// Types
XElement types = root.Element(csTYPES);
if (constants != null)
{
foreach (var xe in types.Elements(KernelTypeInfo.csCUDAFYTYPE))
{
KernelTypeInfo kti = KernelTypeInfo.Deserialize(xe, path);
km.Types.Add(kti.Name, kti);
}
}
return km;
}
public static CudafyModule Deserialize(Stream stream)
{
CudafyModule km = new CudafyModule();
#if NET35
XmlReader reader = XmlReader.Create(stream);
XDocument doc = XDocument.Load(reader);
#else
XDocument doc = XDocument.Load(stream);
#endif
return Deserialize(km, doc, null);
}
/// <summary>
/// Deserializes the specified file.
/// </summary>
/// <param name="filename">The filename.</param>
/// <returns>Cudafy module.</returns>
public static CudafyModule Deserialize(string filename)
{
CudafyModule km = new CudafyModule();
if (!File.Exists(filename))
filename += csFILEEXT;
XDocument doc = XDocument.Load(filename);
string path = Path.GetDirectoryName(filename);
return Deserialize(km, doc, path);
}
internal static KernelMethodInfo Deserialize(XElement xe, CudafyModule parentModule, string directory = null)
{
string methodName = xe.GetAttributeValue(csNAME);
string methodTypeName = xe.GetAttributeValue(csTYPE);
eKernelMethodType methodType = (eKernelMethodType)Enum.Parse(typeof(eKernelMethodType), methodTypeName);
bool? isDummy = xe.TryGetAttributeBoolValue(csISDUMMY);
string behaviourStr = xe.TryGetAttributeValue(csDUMMYBEHAVIOUR);
string typeName = xe.Element(csTYPE).Value;
string assemblyFullName = xe.Element(csASSEMBLY).Value;
string assemblyName = xe.Element(csASSEMBLYNAME).Value;
string assemblyPath = xe.TryGetElementValue(csASSEMBLYPATH);
long checksum = XmlConvert.ToInt64(xe.Element(csCHECKSUM).Value);
eCudafyDummyBehaviour behaviour = string.IsNullOrEmpty(behaviourStr) ? eCudafyDummyBehaviour.Default : (eCudafyDummyBehaviour)Enum.Parse(typeof(eCudafyDummyBehaviour), behaviourStr);
MethodInfo mi = null;
KernelMethodInfo kmi = null;
if(!string.IsNullOrEmpty(typeName) && !string.IsNullOrEmpty(assemblyFullName))
{
Assembly assembly = null;
try
{
assembly = Assembly.Load(assemblyFullName);
}
catch (FileNotFoundException)
{
directory = directory != null ? directory : string.Empty;
assemblyName = directory + Path.DirectorySeparatorChar + assemblyName;
if (File.Exists(assemblyName + ".dll"))
{
assembly = Assembly.LoadFrom(assemblyName + ".dll");
}
else if (File.Exists(assemblyName + ".exe"))
{
assembly = Assembly.LoadFrom(assemblyName + ".exe");
}
else if (!string.IsNullOrEmpty(assemblyPath))
{
assembly = Assembly.LoadFrom(assemblyPath);
}
else
throw;
}
if (assembly == null)
throw new CudafyException(CudafyException.csCOULD_NOT_LOAD_ASSEMBLY_X, assemblyFullName);
Type type = assembly.GetType(typeName);
if (type == null)
throw new CudafyException(CudafyException.csCOULD_NOT_FIND_TYPE_X_IN_ASSEMBLY_X, typeName, assemblyFullName);
mi = type.GetMethod(methodName, BindingFlags.Static | BindingFlags.NonPublic | BindingFlags.Public | BindingFlags.Instance);
if (mi == null)
throw new CudafyException(CudafyException.csCOULD_NOT_FIND_METHOD_X_IN_TYPE_X_IN_ASSEMBLY_X, methodName, typeName, assemblyFullName);
kmi = new KernelMethodInfo(type, mi, methodType, isDummy == true ? true : false, behaviour, parentModule);
}
kmi.DeserializedChecksum = checksum;
return kmi;
}
private static CudafyModule DoCudafy(CudafyModule cm, params Type[] types)
{
MemoryStream output = new MemoryStream();
var outputSw = new StreamWriter(output);
MemoryStream structs = new MemoryStream();
var structsSw = new StreamWriter(structs);
var structsPto = new PlainTextOutput(structsSw);
MemoryStream declarations = new MemoryStream();
var declarationsSw = new StreamWriter(declarations);
var declarationsPto = new PlainTextOutput(declarationsSw);
MemoryStream code = new MemoryStream();
var codeSw = new StreamWriter(code);
var codePto = new PlainTextOutput(codeSw);
bool isDummy = false;
eCudafyDummyBehaviour behaviour = eCudafyDummyBehaviour.Default;
Dictionary<string, ModuleDefinition> modules = new Dictionary<string,ModuleDefinition>();
var compOpts = new DecompilationOptions { FullDecompilation = true };
CUDALanguage.Reset();
bool firstPass = true;
if(cm == null)
cm = new CudafyModule();// #######!!!
else
firstPass = false;
// Test structs
//foreach (var strct in types.Where(t => !t.IsClass))
// if (strct.GetCustomAttributes(typeof(CudafyAttribute), false).Length == 0)
// throw new CudafyLanguageException(CudafyLanguageException.csCUDAFY_ATTRIBUTE_IS_MISSING_ON_X, strct.Name);
IEnumerable<Type> typeList = GetWithNestedTypes(types);
foreach (var type in typeList)
{
if(!modules.ContainsKey(type.Assembly.Location))
modules.Add(type.Assembly.Location, ModuleDefinition.ReadModule(type.Assembly.Location));
}
// Additional loop to compile in order
foreach (var requestedType in typeList)
{
foreach (var kvp in modules)
{
foreach (var td in kvp.Value.Types)
{
List<TypeDefinition> tdList = new List<TypeDefinition>();
tdList.Add(td);
tdList.AddRange(td.NestedTypes);
Type type = null;
foreach (var t in tdList)
{
//type = typeList.Where(tt => tt.FullName.Replace("+", "") == t.FullName.Replace("/", "")).FirstOrDefault();
// Only select type if this matches the requested type (to ensure order is maintained).
type = requestedType.FullName.Replace("+", "") == t.FullName.Replace("/", "") ? requestedType : null;
if (type == null)
continue;
Debug.WriteLine(t.FullName);
// Types
var attr = t.GetCudafyType(out isDummy, out behaviour);
if (attr != null)
{
_cl.DecompileType(t, structsPto, compOpts);
if (firstPass)
cm.Types.Add(type.FullName.Replace("+", ""), new KernelTypeInfo(type, isDummy, behaviour));// #######!!!
}
else if (t.Name == td.Name)
{
// Fields
foreach (var fi in td.Fields)
{
attr = fi.GetCudafyType(out isDummy, out behaviour);
if (attr != null)
{
VerifyMemberName(fi.Name);
System.Reflection.FieldInfo fieldInfo = type.GetField(fi.Name, BindingFlags.Static | BindingFlags.Public | BindingFlags.NonPublic);
if (fieldInfo == null)
throw new CudafyLanguageException(CudafyLanguageException.csX_ARE_NOT_SUPPORTED, "Non-static fields");
int[] dims = _cl.GetFieldInfoDimensions(fieldInfo);
_cl.DecompileCUDAConstantField(fi, dims, codePto, compOpts);
var kci = new KernelConstantInfo(fi.Name, fieldInfo, isDummy);
if (firstPass)
cm.Constants.Add(fi.Name, kci);// #######!!!
CUDALanguage.AddConstant(kci);
}
}
#warning TODO Only Global Methods can be called from host
#warning TODO For OpenCL may need to do Methods once all Constants have been handled
// Methods
foreach (var med in td.Methods)
{
attr = med.GetCudafyType(out isDummy, out behaviour);
if (attr != null)
{
if (!med.IsStatic)
throw new CudafyLanguageException(CudafyLanguageException.csX_ARE_NOT_SUPPORTED, "Non-static methods");
_cl.DecompileMethodDeclaration(med, declarationsPto, new DecompilationOptions { FullDecompilation = false });
_cl.DecompileMethod(med, codePto, compOpts);
MethodInfo mi = type.GetMethod(med.Name, BindingFlags.Static | BindingFlags.Public | BindingFlags.NonPublic);
if (mi == null)
continue;
VerifyMemberName(med.Name);
eKernelMethodType kmt = eKernelMethodType.Device;
kmt = GetKernelMethodType(attr, mi);
if (firstPass)
cm.Functions.Add(med.Name, new KernelMethodInfo(type, mi, kmt, isDummy, behaviour, cm));// #######!!!
}
}
}
}
}
}
}
codeSw.Flush();
if (CudafyTranslator.Language == eLanguage.OpenCL)
{
outputSw.WriteLine("#if defined(cl_khr_fp64)");
outputSw.WriteLine("#pragma OPENCL EXTENSION cl_khr_fp64: enable");
outputSw.WriteLine("#elif defined(cl_amd_fp64)");
outputSw.WriteLine("#pragma OPENCL EXTENSION cl_amd_fp64: enable");
outputSw.WriteLine("#endif");
}
foreach (var oh in CUDALanguage.OptionalHeaders)
{
if (oh.Used && !oh.AsResource)
outputSw.WriteLine(oh.IncludeLine);
else if (oh.Used)
outputSw.WriteLine(GetResourceString(oh.IncludeLine));
}
foreach (var oh in CUDALanguage.OptionalFunctions)
{
if (oh.Used)
outputSw.WriteLine(oh.Code);
}
declarationsSw.WriteLine();
declarationsSw.Flush();
structsSw.WriteLine();
structsSw.Flush();
foreach (var def in cm.GetDummyDefines())
outputSw.WriteLine(def);
foreach (var inc in cm.GetDummyStructIncludes())
outputSw.WriteLine(inc);
foreach (var inc in cm.GetDummyIncludes())
outputSw.WriteLine(inc);
outputSw.Flush();
output.Write(structs.GetBuffer(), 0, (int)structs.Length);
output.Write(declarations.GetBuffer(), 0, (int)declarations.Length);
output.Write(code.GetBuffer(), 0, (int)code.Length);
outputSw.Flush();
#if DEBUG
using (FileStream fs = new FileStream("output.cu", FileMode.Create))
{
fs.Write(output.GetBuffer(), 0, (int)output.Length);
}
#endif
String s = Encoding.UTF8.GetString(output.GetBuffer(), 0, (int)output.Length);
//cm.SourceCode = s;// #######!!!
var scf = new SourceCodeFile(s, Language, _architecture);
cm.AddSourceCodeFile(scf);
return cm;
}
private void btnCompile_Click(object sender, EventArgs e)
{
try
{
if (_module == null)
{
_module = new CudafyModule();
}
if (_module != null)
{
eArchitecture arch = (eArchitecture)Enum.Parse(typeof(eArchitecture), cbArch.SelectedItem as string);
if(arch == eArchitecture.OpenCL)
{
MessageBox.Show(this, "OpenCL modules are not compiled.", "Information", MessageBoxButtons.OK, MessageBoxIcon.Warning);
return;
}
if (!cb32bit.Checked && !cb64bit.Checked)
{
MessageBox.Show(this, "Select a platform.", "Information", MessageBoxButtons.OK, MessageBoxIcon.Information);
return;
}
_module.SourceCode = tbSource.Text;
NvccCompilerOptions opt = null;
_module.CompilerOptionsList.Clear();
_module.RemovePTXModules();
if (cb64bit.Checked)
{
opt = NvccCompilerOptions.Createx64(arch);
_module.CompilerOptionsList.Add(opt);
}
if (cb32bit.Checked)
{
opt = NvccCompilerOptions.Createx86(arch);
_module.CompilerOptionsList.Add(opt);
}
_module.Compile(eGPUCompiler.CudaNvcc, false);
FillForm();
}
}
catch (Exception ex)
{
HandleException(ex);
}
}