Click here to Skip to main content
15,867,943 members
Articles / OpenCL
Article

Part 8: Heterogeneous workflows using OpenCL

Rate me:
Please Sign up or sign in to vote.
4.50/5 (2 votes)
13 Feb 2012CPOL12 min read 29.4K   3  
This article will demonstrate how to incorporate OpenCL into heterogeneous workflows via a general-purpose “click together tools” framework that can stream arbitrary messages within a single workstation, across a network of machines, or within a cloud computing framework.

The previous article, part 7, in this series on portable parallelism with OpenCL™ demonstrated how to create C/C++ plugins that can be dynamically loaded at runtime to add massively parallel OpenCL™ capabilities to an already running application. Developers who understand how to use OpenCL in a dynamically loaded runtime environment have the ability to create plugins that accelerate the performance of existing applications by an order of magnitude or more – simply by writing a new plugin that uses OpenCL.

This article will demonstrate how to incorporate OpenCL into heterogeneous workflows via a general-purpose "click together tools" framework that can stream arbitrary messages (vectors, arrays, and arbitrary, complex nested structures) within a single workstation, across a network of machines, or within a cloud computing framework. The ability to create scalable workflows is important because data handling and transformation can be as complex and time consuming as the computational problem used to generate a desired result. My production version of the framework described in this tutorial has successfully integrated multiple supercomputers and numerous computation nodes into a single unified workflow in both commercial and research environments.

For generality, this tutorial uses the freely downloadable Google protobufs (Protocol Buffers) package so readers can easily extend the example codes to operate on their own data structures. Google protobufs provides binary interoperability across machines and the ability to incorporate applications written in C/C++, Python, Java, R and many other languages into their workflows. Used in everyday computing at Google, protobufs are production hardened. Google also claims this binary format provides a 20 to 100–times increase in performance over XML.

Google protobufs in a click-together framework

For many applications, preprocessing the data can be as complicated and time consuming as the actual computation that generates the desired results. "Click together tools" is a common design pattern that enables flexible and efficient data workflows by creating a pipeline of applications to process information. Each element in the pipeline reads data from an input (usually stdin), performs some filtering or transformation operation, and writes the result to the output (usually stdout). Information flows through these pipelines as packets of information comprised of streams of bytes. All elements in a "click-together" pipeline have the ability to read a message (sometimes called a packet of information) and write it to an output. Based on the type of the packet of information, elements in the pipeline can decide to operate on the data or just pass it along to other elements in the pipeline.

A "click together" framework naturally exploits the parallelism of multi-core processors because each element in the pipeline is a separate application. The operating system scheduler ensures that any applications that have the data necessary to perform work will run – generally on separate processor cores. Buffering between applications allows very large data sets to be processed on the fly. Similarly, the parallelism of multiple machines can be exploited by piping data across machines with ssh, socat, or equivalent socket-based applications or libraries.

Scalable, high-performance workflows can be constructed by tying together multiple machines across a network or in the cloud. Component and plugin reuse reduces errors as most workflows can be constructed from existing "known working" applications. The flexibility of dynamic runtime loading allows OpenCL to be used as a high-performance, massively parallel scripting language to create and implement new workflows. Highly complex multi-stream work flows can be easily constructed using a load-balancing split operation based on simple socket based programming techniques like select() or poll() to determine when a stream is ready for more data. One possible workflow showing the use of multi-core processors and multiple GPUs within a system and across a network is represented in the following figure.

image001.jpg

Figure 1: Example workflow

Keep in mind that each stream of information can be written out to disk as an archive of the work performed, to checkpoint results, or to use as the input deck for a later application.

Experience gained from decades of using this click-together framework has shown that each packet of information needs to be preceded by a header similar to the one shown below:

Version number Size (in bytes) of packet Packet Type ID Size (in bytes) of packet
Figure 2: Example header

It is important that the header include a version number as it allows libraries to transparently select the correct formatting and serialization methods. For example, data streams that I saved to disk in the early 1980s at Los Alamos National Laboratory are still usable today. For robustness, it is necessary to duplicate the size of the packet to detect "silent" transmission errors. Otherwise, bit-errors in the packet size can cause bizarre failures as an application may suddenly attempt to allocate 232 or 264 bytes of memory (depending on the number of bits used to store the size). I have seen such errors when preprocessing data sets on clusters of machines when the processing takes many weeks to complete. Machine failures can cause successful transmission of bogus information across a TCP network.

The redundancy in the size information provides a high likelihood that bit rot in persistent streams will be found. Some disk subsystems, especially inexpensive ones, are susceptible to bit rot caused by multi-bit data errors. So long as the size is correctly known, the packet information can be properly loaded into memory where other more extensive checks or error recovery can occur. Even if that particular packet is corrupt, the remaining packets in the stream can be correctly loaded so all is not lost.

Following is a simple definition of a header that can be adapted to a variety of languages

C++
struct simpleHeader {
   uint64_t version, size1, packetID, size2;
}
Example 1: Header structure

Each application, regardless of language, must be able to read and understand this header. The application programmer can decide what to do with each packet of information. At the very least, the programmer can just pass on the packet of information without affecting the packet contents, or to discard the packet and remove it from the data stream. Regardless, all header data is transferred between applications in binary format using network standard byte order so that arbitrary machine architectures can be used. The type of streaming protocol has run successfully on most machines sold since the mid-1980s.

The following pseudo-code describes how to read and write one or more packets of information. For clarity, this pseudo code does check that every I/O operation was successful. Actual production code needs to be very strict about checking every operation.

C++
while ( read the binary header information == SUCCESS)
{
    •    Compare header sizes (a mismatch flags an unrecoverable error)
    •    Allocate size bytes (after converting from network standard byte order)
    •    Binary read of size bytes into the allocated memory.

// perform the write
    •    Binary write the header in network standard byte order
    •    Binary write the packet information
}
Example 2: pseudo code showing how to use protobufs in a stream

Most users will utilize a common data interchange format for the packet data. By appropriately specifying the type of the packet in the header, proprietary and special high-performance formats can also be mixed into any data stream. Whenever possible, binary data should be used for performance reasons. As previously mentioned, this tutorial uses Google Protocol Buffers (protobufs) because they are a well-supported free binary data interchange format that is fast, well-tested, and robust. Google uses protobufs for most of their internal RPC protocols and file formats. Code for a variety of destination languages can be generated from a common protobuf description. Generators exist for many common languages including C, C++, Python, Java, Ruby, PHP, Matlab, Visual Basic and others. I have used protobufs on workstations and supercomputers.

The following protobuf specification demonstrates how to describe messages containing vectors of various types. For simplicity only floating-point and double precision vectors are defined.

C++
package tutorial;

enum Packet {
  UNKNOWN=0;
  PB_VEC_FLOAT=1;
  PB_VEC_DOUBLE=2;
  PB_VEC_INT=3;
}
message FloatVector {
  repeated float values    = 1 [packed = true];
  optional string name = 2;
}
message DoubleVector {
  repeated float values    = 1 [packed = true];
  optional string name = 2;
}
Example 3: tutorial.proto

The .proto file is compiled to a destination source language with the protoc source language generator. Following is the command to generate a C++ source package using tutorial.proto. The protoc compiler will also generate Java and Python source packages. Consult the protobuf website for links to source generators for other languages.

protoc --cpp_out=. tutorial.proto

Example 4: protoc command to generate C++ code

Linux users can install protobufs from the application manager such as "apt-get" under Ubuntu. Cygwin and Windows users will need to download and install protobufs from code.google.com. Google provides Visual Studio solutions to help with building the code generator and libraries.

The following file packetheader.h contains the methods to read and write the header and packet information in a stream containing multiple protobuf messages. For generality, note that the message type is defined via the enum in the .proto file. Your own message packets can be utilized by adding to these definitions.

For brevity, many essential checks have been left out of packetheader.h. C++ purists will note that cin and cout are changed to support binary information in the method setPacket_binaryIO(). This was done for convenience as it allows the use of operating system pipes (denoted with ‘|’) to easily "click together" applications. While not part of the C++ standard, most C++ runtime systems support binary I/O on std::cin and std::cout. Those C++ programmers who object to this practice can either (1) change the scripts to manually specify the FIFOs (First-In First-Out queues) and network connections so they can use binary I/O according to the C++ standard, or (2) use the C language. Windows programmers will note that packetheader.h uses the Microsoft provided _setmode() method to perform binary IO.

C++
#ifndef PACKET_HEADER_H
#define PACKET_HEADER_H
 
#ifdef _WIN32
#include <stdio.h>
#include <fcntl.h>
#include <io.h>
#include <stdint.h>
#include <Winsock2.h>
#else
#include <arpa/inet.h>
#endif
 
// a simple version identifier
static const uint32_t version=1;
 
// change cin and cout so C++ can use binary
inline bool setPacket_binaryIO()
{
#ifdef _WIN32
  if(_setmode( _fileno(stdin), _O_BINARY) == -1)
     return false;
  if(_setmode( _fileno(stdout), _O_BINARY) == -1) 
     return false;
#endif
  return true;
}
 
inline bool writePacketHdr (uint32_t size, uint32_t type, std::ostream *out)
{
  size = htonl(size);
  type = htonl(type);
  out->write((const char *)&version, sizeof(uint32_t));
  out->write((const char *)&size, sizeof(uint32_t));
  out->write((const char *)&type, sizeof(uint32_t));
  out->write((const char *)&size, sizeof(uint32_t));
  return true;
}
 
template <typename T>
bool writeProtobuf(T &pb, uint32_t type, std::ostream *out)
{
  writePacketHdr(pb.ByteSize(), type, out);
  pb.SerializeToOstream(out);
  return true;
}
 
inline bool readPacketHdr (uint32_t *size, uint32_t *type, std::istream *in) 
{
  uint32_t size2, myversion;
 
  in->read((char *)&myversion, sizeof(uint32_t)); myversion = ntohl(myversion);
  if(!in->good()) return(false);
  in->read((char *)size, sizeof(uint32_t)); *size = ntohl(*size);
  if(!in->good()) return(false);
  in->read((char *)type, sizeof(uint32_t)); *type = ntohl(*type);
  if(!in->good()) return(false);
  in->read((char *)&size2, sizeof(uint32_t)); size2 = ntohl(size2);
  if(!in->good()) return(false);
 
  if(*size != size2) return(false);
  return(true);
}
 
template <typename T>
bool readProtobuf(T *pb, uint32_t size, std::istream *in)
{
  char *blob = new char[size];
  in->read(blob,size);
  bool ret = pb->ParseFromArray(blob,size);
  delete [] blob;
  return ret;
}
#endif

The program testWrite.cc demonstrates how to create and write both double and float vector messages. The default vector length is 100 elements. Larger messages can be created by specifying a size on the command-line.

C++
// Rob Farber
#include <iostream>
using namespace std;
#include "tutorial.pb.h"
#include "packetheader.h"
 
int main(int argc, char *argv[])
{
  GOOGLE_PROTOBUF_VERIFY_VERSION;
 
  int vec_len = 100;
  // allow user to change the size of the data if they wish
  if(argc > 1) vec_len = atoi(argv[1]);
  
  // change cin and cout to binary mode
  // NOTE: this is not part of the C++ standard
  if(!setPacket_binaryIO()) return -1;
 
  tutorial::FloatVector vec;
  for(int i=0; i < vec_len; i++) vec.add_values(i);
 
  tutorial::DoubleVector vec_d;
  for(int i=0; i < 2*vec_len; i++) vec_d.add_values(i);
  
  vec.set_name("A");
  writeProtobuf<tutorial::FloatVector>(vec, tutorial::PB_VEC_FLOAT,
                                  &std::cout);
  vec_d.set_name("B");
  writeProtobuf<tutorial::DoubleVector>(vec_d, tutorial::PB_VEC_DOUBLE,
                                  &std::cout);
  return(0);
}
Example 6: testWrite.cc

The program testRead.cc demonstrates how to read the header and messages via a stream. The string associated with the optional name in the protobuf message is printed when provided.

C++
// Rob Farber
#include <iostream>
#include "packetheader.h"
#include "tutorial.pb.h"
using namespace std;
 
int main(int argc, char *argv[])
{
  GOOGLE_PROTOBUF_VERIFY_VERSION;
 
  // Change cin and cout to binary mode
  // NOTE: this is not part of the C++ standard
  if(!setPacket_binaryIO()) return -1;
 
  uint32_t size, type;
  while(readPacketHdr(&size, &type, &std::cin)) {
    switch(type) {
    case tutorial::PB_VEC_FLOAT: {
      tutorial::FloatVector vec;
      if(!readProtobuf<tutorial::FloatVector>(&vec, size, &std::cin))
       break;
      if(vec.has_name() == true) cerr << "vec_float " << vec.name() << endl;
      cerr << vec.values_size() << " elements" << endl;
    } break;
    case tutorial::PB_VEC_DOUBLE: {
      tutorial::DoubleVector vec;
      if(!readProtobuf<tutorial::DoubleVector>(&vec, size, &std::cin))
       break;
      if(vec.has_name() == true) cerr << "vec_double " << vec.name() << endl;
      cerr << vec.values_size() << " elements" << endl;
    } break;
    default:
      cerr << "Unknown packet type" << endl;
    }
  }
  return(0);
}
Example 7: testRead.cc

These applications can be built and tested under Linux with the following commands:

g++ -I . testWrite.cc tutorial.pb.cc -l protobuf -o testWrite -lpthread
g++ -I . testRead.cc tutorial.pb.cc -l protobuf -o testRead -lpthread
 
echo "----------- simple test -----------------"
./testWrite | ./testRead
Example 8: Linux build and test commands
C++
bda$ sh BUILD.linux 
----------- simple test -----------------
vec_float A
100 elements
vec_double B
200 elements
Example 9: Output of Linux commands

A Click-Together Framework

Combining the previous dynamic compile/link and protobuf examples yields the powerful "click together" framework discussed earlier in this article.

Following is the complete source for dynFunc.cc, which combines the streaming of protobuf messages and the dynamic compilation of C/C++ methods.

For flexibility, the init(), func(), and fini() methods have the ability to modify or create new messages that can be passed onto other applications in the click-together framework. It is the responsibility of the plugin author to create a char array that will hold the modified protobuf message. A dynFree() method was added so the plugin author can free the memory region. This makes the plugin framework language agnostic. For example, C source code would use malloc()/free() while C++ source code would use new/delete .

Each method can return a pointer to a character array that contains a modified message that is to be written out. Returning NULL implies no message need be written. The plugin author can pass on the original message by returning the message pointer. In this case, dynFree() is not called because the plugin framework performed the allocation.

C++
//Rob Farber
#include <cstdlib>
#include <sys/types.h>
#include <dlfcn.h>
#include <string>
#include <iostream>
#include "packetheader.h"
 
using namespace std;
 
void *lib_handle;
 
typedef char* (*initFini_t)(const char*, const char*, uint32_t*, uint32_t*);
typedef char* (*func_t)(const char*, const char*, uint32_t*, uint32_t*, char*);
typedef void (*dynFree_t)(char*);
 
int main(int argc, char **argv) 
{
  if(argc < 2) {
    cerr << "Use: sourcefilename" << endl;
    return -1;
  }
  string base_filename(argv[1]);
  base_filename = base_filename.substr(0,base_filename.find_last_of("."));
  
  // build the shared object or dll
  string buildCommand("g++ -fPIC -shared ");
  buildCommand += string(argv[1]) 
    + string(" -o ") + base_filename + string(".so ");
 
  cerr << "Compiling with \"" << buildCommand << "\"" << endl;
  if(system(buildCommand.c_str())) {
    cerr << "compile command failed!" << endl;
    cerr << "Build command " << buildCommand << endl;
    return -1;
  }
  
  // load the library -------------------------------------------------
  string nameOfLibToLoad("./");
  nameOfLibToLoad += base_filename;
  
  nameOfLibToLoad += ".so";
  lib_handle = dlopen(nameOfLibToLoad.c_str(), RTLD_LAZY);
  if (!lib_handle) {
    cerr << "Cannot load library: " << dlerror() << endl;
    return -1;
  }
  
  // load the symbols -------------------------------------------------
  initFini_t dynamicInit= NULL;
  func_t dynamicFunc= NULL;
  initFini_t dynamicFini= NULL;
  dynFree_t dynamicFree= NULL;
 
  // reset errors
  dlerror();
  
  // load the function pointers
  dynamicFunc= (func_t) dlsym(lib_handle, "func");
  const char* dlsym_error = dlerror();
  if (dlsym_error) { cerr << "sym load: " << dlsym_error << endl; return -1;}
  dynamicInit= (initFini_t) dlsym(lib_handle, "init");
  dlsym_error = dlerror();
  if (dlsym_error) { cerr << "sym load: " << dlsym_error << endl; return -1;}
  dynamicFini= (initFini_t) dlsym(lib_handle, "fini");
  dlsym_error = dlerror();
  if (dlsym_error) { cerr << "sym load: " << dlsym_error << endl; return -1;}
  dlsym_error = dlerror();
  if (dlsym_error) { cerr << "sym load: " << dlsym_error << endl; return -1;}
  dynamicFree= (dynFree_t) dlsym(lib_handle, "dynFree");
 
  // 
  // work with protobufs
  // 
 
  //enable C++ binary cin and cout
  if (!setPacket_binaryIO()) {
     cerr << "Cannot set binary mode for cin and cout!" << endl;
     return -1;
     }
 
  uint32_t size, type;
  char *retBlob;
 
  // handle initialization and put information on output stream when told
  if( (retBlob=(*dynamicInit)(argv[0], base_filename.c_str(),&size, &type)) ) {
    writePacketHdr(size, type, &std::cout);
    cout.write(retBlob, size);
    (dynamicFree)(retBlob);
  }
 
  // read stream from cin and put information on output stream when told
  while(readPacketHdr(&size, &type, &std::cin)) {
    char *blob = new char[size];
    cin.read(blob, size);
    retBlob =(*dynamicFunc)(argv[0], base_filename.c_str(), &size, &type, blob);
    if(retBlob) {
      writePacketHdr(size, type, &std::cout);
      cout.write(retBlob, size);
      // optimization: if retBlob == blob then allocated was by this program
      if(retBlob != blob) (dynamicFree)(retBlob);
    }
    delete [] blob;
  }
 
  // handle finalization (fini) and put information on output stream when told
  if( retBlob = (*dynamicFini)(argv[0], base_filename.c_str(),&size, &type) ) {
    writePacketHdr(size, type, &std::cout);
    cout.write(retBlob, size);
    (dynamicFree)(retBlob);
  }
  
  // unload the library -----------------------------------------------
  dlclose(lib_handle);
  return 0;
}
Example 10: dynFunc.cc

The source for passthrough.cc is included below. This plugin simply passes all messages along to the next application in the pipeline. Linux users will note the g++ command includes use of the –rdynamic option, which tells the linker to check the executable for any unresolved symbols. (While Visual Studio code is not provided in this tutorial, it is important that Visual Studio users specify a #pragma comment() to inform the linker about needed libraries. In this way the protobuf methods can be linked with the plugin.)

C++
//passthrough.cc (Rob Farber)
#include <stdlib.h>
#include <stdint.h>
#include <iostream>
#include "tutorial.pb.h"
using namespace std;
 
extern "C" char* init(const char* progname, const char* sourcename,
          uint32_t *size, uint32_t *type) {
  return(NULL); 
}
 
extern "C" char* func(const char* progname, const char* sourcename, 
          uint32_t *size, uint32_t *type, char *blob)
{
  return(blob); //Note: this is a special case that will not invoke dynFree
}
 
extern "C" char* fini(const char* progname, const char* sourcename,
          uint32_t *size, uint32_t *type) {
  return(NULL); 
}
 
extern "C" void dynFree(char* pt) {
  cerr << "dynFree" << endl;
  if(pt) delete [] pt;
}
Example 11: passthrough.cc

The following source file, reduction.cc, demonstrates how to use init() and fini() to calculate the sum of either a float or double protobuf vector message on the host.

C++
//reduction.cc (Rob Farber)
#include <stdlib.h>
#include <stdint.h>
#include <iostream>
#include "tutorial.pb.h"
using namespace std;
 
extern "C" char* init(const char* progname, const char* sourcename, 
                    uint32_t *size, uint32_t *type) {
  return(NULL); 
}
 
extern "C" char* func(const char* progname, const char* sourcename, 
                    uint32_t *size, uint32_t *type, char *blob)
{
  switch(*type) {
  case tutorial::PB_VEC_FLOAT: {
    tutorial::FloatVector vec;
    if(!vec.ParseFromArray(blob,*size)) {
      cerr << progname << "," << sourcename << "Illegal packet" << endl;
    } else {
      if(vec.has_name() == true) cerr << "vec_float " << vec.name() << " ";
      float sum=0.f;
      for(int i=0; i < vec.values_size(); i++) sum += vec.values(i);
      cerr << "sum of vector " << sum << endl;
      cerr << "\tlast value in vector is " << vec.values(vec.values_size()-1)
          << endl;
      cerr << "\tvector size is " << vec.values_size() << endl;
    }
  } break;
  case tutorial::PB_VEC_DOUBLE: {
    tutorial::DoubleVector vec;
    if(!vec.ParseFromArray(blob,*size)) {
      cerr << progname << "," << sourcename << "Illegal packet" << endl;
    } else {
      if(vec.has_name() == true) cerr << "vec_double " << vec.name() << " ";
      double sum=0.;
      for(int i=0; i < vec.values_size(); i++) sum += vec.values(i);
      cerr << "sum of vector " << sum << endl;
      cerr << "\tlast value in vector is " << vec.values(vec.values_size()-1)
          << endl;
      cerr << "\tvector size is " << vec.values_size() << endl;
    }
  } break;
  default:
    cerr << "Unknown packet type" << endl;
  }
  return(NULL);
}
 
extern "C" char* fini(const char* progname, const char* sourcename, 
                    uint32_t *size, uint32_t *type) {
  return(NULL); 
}
 
extern "C" void dynFree(char* pt) {
  if(pt) delete [] pt;
}
Example 12: reduction.cc

Using OpenCL in a Click Together Framework

The source code for dynFunc.cc is easily modified to include the code needed to parse the device type (either CPU or GPU) and to add a call to a new method, oclSetupFunc(), which passes the OpenCL context and name of the kernel source file to the plugin. The plugin can then build the OpenCL source code and call the OpenCL kernels in the init(),

func()
, and fini() methods. Changes to the code are highlighted in bold font in source for oclFunc.cc below.

C++
//Rob Farber (dynOCL.cc)
#include <cstdlib>
#include <sys/types.h>
#include <dlfcn.h>
#include <string>
#include <iostream>
#include "packetheader.h"
 
#define PROFILING // Define to see the time the kernel takes
#define __NO_STD_VECTOR // Use cl::vector instead of STL version
#define __CL_ENABLE_EXCEPTIONS // needed for exceptions
#include <CL/cl.hpp>
#include <fstream>
 
using namespace std;
void *lib_handle;
 
typedef char* (*initFini_t)(const char*, const char*, uint32_t*, uint32_t*);
typedef char* (*func_t)(const char*, const char*, uint32_t*, uint32_t*, char*);
typedef void (*dynFree_t)(char*);
typedef void (*oclSetup_t)(const char*, cl::CommandQueue*);
 
int main(int argc, char **argv) 
{
  if(argc < 3) {
    cerr << "Use: sourcefilename cpu|gpu oclSource" << endl;
    return -1;
  }
  string base_filename(argv[1]);
  base_filename = base_filename.substr(0,base_filename.find_last_of("."));
  
  // build the shared object or dll
  string buildCommand("g++ -fPIC -shared -I $AMDAPPSDKROOT/include ");
  buildCommand += string(argv[1]) 
    + string(" -o ") + base_filename + string(".so ");
 
  cerr << "Compiling with \"" << buildCommand << "\"" << endl;
  if(system(buildCommand.c_str())) {
    cerr << "compile command failed!" << endl;
    cerr << "Build command " << buildCommand << endl;
    return -1;
  }
  
  // load the library -------------------------------------------------
  string nameOfLibToLoad("./");
  nameOfLibToLoad += base_filename;
  
  nameOfLibToLoad += ".so";
  lib_handle = dlopen(nameOfLibToLoad.c_str(), RTLD_LAZY);
  if (!lib_handle) {
    cerr << "Cannot load library: " << dlerror() << endl;
    return -1;
  }
  
  // load the symbols -------------------------------------------------
  initFini_t dynamicInit= NULL;
  func_t dynamicFunc= NULL;
  initFini_t dynamicFini= NULL;
  dynFree_t dynamicFree= NULL;
 
  // reset errors
  dlerror();
  
  // load the function pointers
  dynamicFunc= (func_t) dlsym(lib_handle, "func");
  const char* dlsym_error = dlerror();
  if (dlsym_error) { cerr << "sym load: " << dlsym_error << endl; return -1;}
  dynamicInit= (initFini_t) dlsym(lib_handle, "init");
  dlsym_error = dlerror();
  if (dlsym_error) { cerr << "sym load: " << dlsym_error << endl; return -1;}
  dynamicFini= (initFini_t) dlsym(lib_handle, "fini");
  dlsym_error = dlerror();
  if (dlsym_error) { cerr << "sym load: " << dlsym_error << endl; return -1;}
  dynamicFree= (dynFree_t) dlsym(lib_handle, "dynFree");
  dlsym_error = dlerror();
  if (dlsym_error) { cerr << "sym load: " << dlsym_error << endl; return -1;}
  // add a function to specify the ocl context and kernel file
  oclSetup_t oclSetupFunc;
  oclSetupFunc = (oclSetup_t) dlsym(lib_handle, "oclSetup");
  dlsym_error = dlerror();
  if (dlsym_error) { cerr << "sym load: " << dlsym_error << endl; return -1;}
 
  // -------------------------------------------------------------- 
  // Setup OCL context
  //
  const string platformName(argv[2]);
  const char* oclKernelFile = argv[3];
  int ret= -1;
 
  cl::vector<int> deviceType;
  cl::vector< cl::CommandQueue > contextQueues;
 
  // crudely parse the command line arguments. 
  if(platformName.compare("cpu")==0)
    deviceType.push_back(CL_DEVICE_TYPE_CPU);
  else if(platformName.compare("gpu")==0) 
    deviceType.push_back(CL_DEVICE_TYPE_GPU);
  else { cerr << "Invalid device type!" << endl; return(1); }
 
  // create the context and queues
  try {
    cl::vector< cl::Platform > platformList;
    cl::Platform::get(&platformList);
 
    // Get all the appropriate devices for the platform the
    // implementation thinks we should be using.
    // find the user-specified devices
    cl::vector<cl::Device> devices;
    for(int i=0; i < deviceType.size(); i++) {
      cl::vector<cl::Device> dev;
      platformList[0].getDevices(deviceType[i], &dev);
      for(int j=0; j < dev.size(); j++) devices.push_back(dev[j]);
    }
 
    // set a single context
    cl_context_properties cprops[] = {CL_CONTEXT_PLATFORM, NULL, 0};
    cl::Context context(devices, cprops);
    cerr << "Using the following device(s) in one context" << endl;
    for(int i=0; i < devices.size(); i++)  {
      cerr << "  " << devices[i].getInfo<CL_DEVICE_NAME>() << endl;
    }
 
    // Create the separate command queues to perform work
    for(int i=0; i < devices.size(); i++)  {
#ifdef PROFILING
      cl::CommandQueue queue(context, devices[i],CL_QUEUE_PROFILING_ENABLE);
#else
      cl::CommandQueue queue(context, devices[i],0);
#endif
      contextQueues.push_back( queue );
    }
  } catch (cl::Error error) {
    cerr << "caught exception: " << error.what() 
        << '(' << error.err() << ')' << endl;
    return(-1);
  }
  oclSetupFunc(oclKernelFile, &contextQueues[0]);
 
  // -------------------------------------------------------------- 
  // work with protobufs
  // 
 
  //enable C++ binary cin and cout
  if (!setPacket_binaryIO()) {
     cerr << "Cannot set binary mode for cin and cout!" << endl;
     return -1;
     }
 
  uint32_t size, type;
  char *retBlob;
 
  // handle initialization and put information on output stream when told
  if( (retBlob=(*dynamicInit)(argv[0], base_filename.c_str(),&size, &type)) ) {
    writePacketHdr(size, type, &std::cout);
    cout.write(retBlob, size);
    (dynamicFree)(retBlob);
  }
 
  // read stream from cin and put information on output stream when told
  while(readPacketHdr(&size, &type, &std::cin)) {
    char *blob = new char[size];
    cin.read(blob, size);
    retBlob =(*dynamicFunc)(argv[0], base_filename.c_str(), &size, &type, blob);
    if(retBlob) {
      writePacketHdr(size, type, &std::cout);
      cout.write(retBlob, size);
      // optimization: if retBlob == blob then allocated was by this program
      if(retBlob != blob) (dynamicFree)(retBlob);
    }
    delete [] blob;
  }
 
  // handle finalization (fini) and put information on output stream when told
  if( retBlob = (*dynamicFini)(argv[0], base_filename.c_str(),&size, &type) ) {
    writePacketHdr(size, type, &std::cout);
    cout.write(retBlob, size);
    (dynamicFree)(retBlob);
  }
  
  // unload the library -----------------------------------------------
  dlclose(lib_handle);
  return 0;
}

The following OpenCL kernel adds each element in the vector to itself on the device. The resulting vector is then passed on to the next application in the pipeline. Note, the number of copies was not optimized in this example.

C++
inline __kernel void init(int veclen, __global TYPE1* c, int offset)
{
}
 
inline __kernel void func(int veclen, __global TYPE1* c, int offset)
{
  // get the index of the test we are performing
  int index = get_global_id(0);
 
  c[index + offset*veclen] += c[index + offset*veclen];
}
 
inline __kernel void fini(int veclen, __global TYPE1* c, int offset)
{
}
Example 14: simpleAdd.cl source that adds the elements in a vector to itself

The following commands are used to build and run some tests under Linux:

C++
gcc -rdynamic -o dynFunc dynFunc.cc tutorial.pb.cc -l protobuf -ldl -lpthread
gcc -I $AMDAPPSDKROOT/include -rdynamic -o dynOCL dynOCL.cc tutorial.pb.cc -L $AMDAPPSDKROOT/lib/x86_64 -lOpenCL -lprotobuf -ldl -lpthread  

echo "--------------- showing a dynamic reduction ----------------------------"
./testWrite | ./dynFunc reduction.cc
 
echo "---------------- Pass through demo -------------------------------"
./testWrite | ./dynFunc passthrough.cc \
   | ./dynFunc reduction.cc
 
 
echo "------------- the float vector contains values*4 -----------------------"
# increase the float vector by a factor of four
./testWrite \
       | ./dynOCL oclFunc.cc cpu simpleAdd.cl \
       | ./dynOCL oclFunc.cc gpu simpleAdd.cl \
       | ./dynFunc reduction.cc
Example 15: Linux commands to build and test the examples

These commands generated the following output under Linux. Note the OpenCL plugin that uses simpleAdd.cl is called twice in the last test to increase the value of the floating-point vector values by a factor of four as shown in bold font below. This example demonstrates that OpenCL plugins can be chained together. In addition, the first use of simpleAdd.cl runs on the CPU and the second runs on a GPU.

C++
$: sh BUILD.linux 
--------------- showing a dynamic reduction ----------------------------
Compiling with "g++ -fPIC -shared reduction.cc -o reduction.so "
vec_float A sum of vector 4950
       last value in vector is 99
       vector size is 100
vec_double B sum of vector 19900
       last value in vector is 199
       vector size is 200
---------------- Pass through demo -------------------------------
Compiling with "g++ -fPIC -shared passthrough.cc -o passthrough.so "
Compiling with "g++ -fPIC -shared reduction.cc -o reduction.so "
vec_float A sum of vector 4950
       last value in vector is 99
       vector size is 100
vec_double B sum of vector 19900
       last value in vector is 199
       vector size is 200
------------- the float vector contains values*4 -----------------------
Compiling with "gcc -fPIC -shared -I $AMDAPPSDKROOT/include oclFunc.cc -o oclFunc.so "
Compiling with "g++ -fPIC -shared reduction.cc -o reduction.so "
Compiling with "gcc -fPIC -shared -I $AMDAPPSDKROOT/include oclFunc.cc -o oclFunc.so "
Using the following device(s) in one context
  AMD Phenom(tm) II X6 1055T Processor
building OCL source (simpleAdd.cl)
   buildOptions -D TYPE1=float 
Using the following device(s) in one context
  Cypress
building OCL source (simpleAdd.cl)
   buildOptions -D TYPE1=float 
dynFree
dynFree
vec_float A sum of vector 19800
       last value in vector is 396
       vector size is 100
vec_double B sum of vector 19900
       last value in vector is 199
       vector size is 200
Example 16: Output of the Linux commands and tests

Summary

With the ability to create OpenCL plugins, application programmers have the ability to write and support generic applications that will deliver accelerated performance when a GPU is present and CPU-based performance when a GPU is not available. These plugin architectures are well-understood and a convenient way to leverage existing applications and code bases. They also help preserve existing software investments.

The ability to dynamically compile OpenCL plugins opens a host of opportunities for optimizing code generators and transparently running a single OpenCL application on multiple device back ends. With this technique, OpenCL applications can deliver optimized code for specific problems and achieve very high performance – far beyond what a single generic code can deliver. As OpenCL matures, these applications will also transparently benefit from any compiler and other performance improvements.

The workflow example in this article demonstrates how to use OpenCL to exploit hybrid CPU/GPU computation. Incorporating OpenCL into the flexibility, robustness, scalability, and performance of "click together" pipeline workflows gives programmers the ability to capitalize on GPU acceleration and CPU capabilities in their production workflows.

License

This article, along with any associated source code and files, is licensed under The Code Project Open License (CPOL)


Written By
United States United States
Rob Farber is a senior scientist and research consultant at the Irish Center for High-End Computing in Dublin, Ireland and Chief Scientist for BlackDog Endeavors, LLC in the US.

Rob has been on staff at several US national laboratories including Los Alamos National Laboratory, Lawrence Berkeley National Laboratory, and at Pacific Northwest National Laboratory. He also served as an external faculty member at the Santa Fe Institute, co-founded two successful start-ups, and has been a consultant to Fortune 100 companies. His articles have appeared in Dr. Dobb's Journal and Scientific Computing, among others. He recently completed a book teaching massive parallel computing.

Comments and Discussions

 
-- There are no messages in this forum --