[driver] fixed some bugs

include/driver/platform.h

@@ -40,7 +40,7 @@ class Platform
 {
 public:
   //Accessors
-  std::string name() const;
+  std::string name() const { return "CUDA"; }
   std::string version() const;
   std::vector<Device> devices() const;
 private:

lib/driver/dispatch.cpp

@@ -85,18 +85,6 @@ namespace driver
 #define CUDA_DEFINE10(ret, fname, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10) DEFINE10(cuinit, cuda_, ret, fname, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10)
 #define CUDA_DEFINE11(ret, fname, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11) DEFINE11(cuinit, cuda_, ret, fname, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11)
 
-#define NVRTC_DEFINE1(ret, fname, t1) DEFINE1(nvrtcinit, nvrtc_, ret, fname, t1)
-#define NVRTC_DEFINE2(ret, fname, t1, t2) DEFINE2(nvrtcinit, nvrtc_, ret, fname, t1, t2)
-#define NVRTC_DEFINE3(ret, fname, t1, t2, t3) DEFINE3(nvrtcinit, nvrtc_, ret, fname, t1, t2, t3)
-#define NVRTC_DEFINE4(ret, fname, t1, t2, t3, t4) DEFINE4(nvrtcinit, nvrtc_, ret, fname, t1, t2, t3, t4)
-#define NVRTC_DEFINE5(ret, fname, t1, t2, t3, t4, t5) DEFINE5(nvrtcinit, nvrtc_, ret, fname, t1, t2, t3, t4, t5)
-#define NVRTC_DEFINE6(ret, fname, t1, t2, t3, t4, t5, t6) DEFINE6(nvrtcinit, nvrtc_, ret, fname, t1, t2, t3, t4, t5, t6)
-#define NVRTC_DEFINE7(ret, fname, t1, t2, t3, t4, t5, t6, t7) DEFINE7(nvrtcinit, nvrtc_, ret, fname, t1, t2, t3, t4, t5, t6, t7)
-#define NVRTC_DEFINE8(ret, fname, t1, t2, t3, t4, t5, t6, t7, t8) DEFINE8(nvrtcinit, nvrtc_, ret, fname, t1, t2, t3, t4, t5, t6, t7, t8)
-#define NVRTC_DEFINE9(ret, fname, t1, t2, t3, t4, t5, t6, t7, t8, t9) DEFINE9(nvrtcinit, nvrtc_, ret, fname, t1, t2, t3, t4, t5, t6, t7, t8, t9)
-#define NVRTC_DEFINE10(ret, fname, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10) DEFINE10(nvrtcinit, nvrtc_, ret, fname, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10)
-#define NVRTC_DEFINE11(ret, fname, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11) DEFINE11(nvrtcinit, nvrtc_, ret, fname, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11)
-
 #define NVML_DEFINE0(ret, fname) DEFINE0(nvmlinit, nvml_, ret, fname)
 #define NVML_DEFINE1(ret, fname, t1) DEFINE1(nvmlinit, nvml_, ret, fname, t1)
 #define NVML_DEFINE2(ret, fname, t1, t2) DEFINE2(nvmlinit, nvml_, ret, fname, t1, t2)
@@ -127,12 +115,6 @@ bool dispatch::cuinit(){
   return cuda_ != nullptr;
 }
 
-bool dispatch::nvrtcinit(){
-  if(nvrtc_==nullptr)
-    nvrtc_ = dlopen("libnvrtc.so", RTLD_LAZY);
-  return nvrtc_ != nullptr;
-}
-
 bool dispatch::nvmlinit(){
   if(nvml_==nullptr)
     nvml_ = dlopen("libnvidia-ml.so", RTLD_LAZY);
@@ -194,13 +176,6 @@ CUDA_DEFINE4(CUresult, cuMemsetD8Async, CUdeviceptr, unsigned char, size_t, CUst
 CUDA_DEFINE1(CUresult, cuCtxPushCurrent_v2, CUcontext)
 CUDA_DEFINE1(CUresult, cuCtxPopCurrent_v2, CUcontext*)
 
-NVRTC_DEFINE3(nvrtcResult, nvrtcCompileProgram, nvrtcProgram, int, const char **)
-NVRTC_DEFINE2(nvrtcResult, nvrtcGetProgramLogSize, nvrtcProgram, size_t *)
-NVRTC_DEFINE2(nvrtcResult, nvrtcGetPTX, nvrtcProgram, char *)
-NVRTC_DEFINE2(nvrtcResult, nvrtcGetPTXSize, nvrtcProgram, size_t *)
-NVRTC_DEFINE6(nvrtcResult, nvrtcCreateProgram, nvrtcProgram *, const char *, const char *, int, const char **, const char **)
-NVRTC_DEFINE2(nvrtcResult, nvrtcGetProgramLog, nvrtcProgram, char *)
-
 NVML_DEFINE2(nvmlReturn_t, nvmlDeviceGetHandleByPciBusId_v2, const char *, nvmlDevice_t*)
 NVML_DEFINE3(nvmlReturn_t, nvmlDeviceGetClockInfo, nvmlDevice_t, nvmlClockType_t, unsigned int*)
 NVML_DEFINE3(nvmlReturn_t, nvmlDeviceGetMaxClockInfo, nvmlDevice_t, nvmlClockType_t, unsigned int*)

lib/driver/error.cpp

@@ -94,23 +94,6 @@ void check(CUresult err)
   }
 }
 
-void check(nvrtcResult err){
-  using namespace exception::nvrtc;
-
-  switch(err)
-  {
-  case NVRTC_SUCCESS:                         break;
-  case NVRTC_ERROR_OUT_OF_MEMORY:             throw out_of_memory();
-  case NVRTC_ERROR_PROGRAM_CREATION_FAILURE:  throw program_creation_failure();
-  case NVRTC_ERROR_INVALID_INPUT:             throw invalid_input();
-  case NVRTC_ERROR_INVALID_PROGRAM:           throw invalid_program();
-  case NVRTC_ERROR_INVALID_OPTION:            throw invalid_option();
-  case NVRTC_ERROR_COMPILATION:               throw compilation();
-  case NVRTC_ERROR_BUILTIN_OPERATION_FAILURE: throw builtin_operation_failure();
-  default: throw unknown_error();
-  }
-}
-
 void check(cublasStatus_t err){
   using namespace exception::cublas;
   switch(err)

lib/driver/module.cpp

@@ -34,68 +34,19 @@ namespace tdl
 namespace driver
 {
 
-CUjit_target_enum cutarget(Device::Architecture arch){
+Module::Module(Context const & context, std::string const & source) : context_(context), source_(source){
-  switch(arch){
-    case Device::Architecture::SM_2_0: return CU_TARGET_COMPUTE_20;
-    case Device::Architecture::SM_2_1: return CU_TARGET_COMPUTE_21;
-    case Device::Architecture::SM_3_0: return CU_TARGET_COMPUTE_30;
-    case Device::Architecture::SM_3_5: return CU_TARGET_COMPUTE_35;
-    case Device::Architecture::SM_3_7: return CU_TARGET_COMPUTE_37;
-    case Device::Architecture::SM_5_0: return CU_TARGET_COMPUTE_50;
-    case Device::Architecture::SM_5_2: return CU_TARGET_COMPUTE_52;
-    case Device::Architecture::SM_6_0: return CU_TARGET_COMPUTE_60;
-    case Device::Architecture::SM_6_1: return CU_TARGET_COMPUTE_61;
-    default: throw;
-  }
-}
-
-inline std::pair<int, int> ptx(std::pair<int, int> sm){
-  if(sm.first == 7) return {6, 0};
-  if(sm.first == 6) return {5, 0};
-  if(sm.first == 5) return {4, 3};
-  throw;
-}
-
-std::string Module::header(Device const & device){
-  auto cc = device.compute_capability();
-  auto vptx = ptx(cc);
-  std::string header;
-  header += ".version " + std::to_string(vptx.first) + "." + std::to_string(vptx.second) + "\n";
-  header += ".target sm_" + std::to_string(cc.first) + std::to_string(cc.second) + "\n";
-  header += ".address_size 64\n";
-  return header;
-}
-
-Module::Module(Context const & context, std::string const & source) : context_(context), source_(header(context.device()) + source){
   ContextSwitcher ctx_switch(context_);
-
+  // JIT compile source-code
-  //Path to custom PTX compiler
+  CUjit_option opt[] = {CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES, CU_JIT_ERROR_LOG_BUFFER};
-  std::string compiler = tools::getenv("ISAAC_PTXAS");
+  unsigned int errbufsize = 8096;
-  if(compiler.size()){
+  std::string errbuf(errbufsize, 0);
-    auto cc = context.device().compute_capability();
+  void* optval[] = {(void*)(uintptr_t)errbufsize, (void*)errbuf.data()};
-    std::string out = context.cache_path() + "tmp.o";
+  try{
-    std::string opt = " --gpu-name sm_" + std::to_string(cc.first) + std::to_string(cc.second)
+    dispatch::cuModuleLoadDataEx(&*cu_, source_.data(), 2, opt, optval);
-                    + "  -o " + out
+  }catch(exception::cuda::base const &){
-                    + "  -ias \"" + source_ + "\"";
+    std::cerr << "Compilation Failed! Log: " << std::endl;
-    std::string cmd = compiler + opt;
+    std::cerr << errbuf << std::endl;
-    if(std::system(cmd.c_str()) != 0)
+    throw;
-      throw;
-    dispatch::cuModuleLoad(&*cu_, out.c_str());
-  }
-  //JIT Compilation
-  else{
-    CUjit_option opt[] = {CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES, CU_JIT_ERROR_LOG_BUFFER};
-    unsigned int errbufsize = 8096;
-    std::string errbuf(errbufsize, 0);
-    //CUjit_target_enum target = cutarget(context.device().architecture());
-    void* optval[] = {(void*)(uintptr_t)errbufsize, (void*)errbuf.data()};
-    try{
-      dispatch::cuModuleLoadDataEx(&*cu_, source_.data(), 2, opt, optval);
-    }catch(exception::cuda::base const &){
-      std::cerr << "Compilation Failed! Log: " << std::endl;
-      std::cerr << errbuf << std::endl;
-      throw;
-    }
   }
 }
 

lib/driver/platform.cpp

@@ -37,9 +37,6 @@ std::string Platform::version() const{
   return std::to_string(version);
 }
 
-std::string Platform::name() const
-{ return (std::string)"CUDA"; }
-
 std::vector<Device> Platform::devices() const{
   std::vector<Device> devices;
   int N;