feat(compile): add nvcc compiler

cinn/backends/compiler.cc

@@ -25,6 +25,7 @@
 #include "cinn/backends/nvrtc/nvrtc_util.h"
 #include "cinn/runtime/cuda/cuda_module.h"
 #include "cinn/runtime/cuda/cuda_util.h"
+#include "cinn/runtime/flags.h"
 #endif
 
 DECLARE_string(cinn_source_code_save_path);
@@ -123,16 +124,13 @@ void Compiler::CompileCudaModule(const Module& module, const std::string& code)
   SourceCodePrint::GetInstance()->write(source_code);
   using runtime::cuda::CUDAModule;
 
-  backends::nvrtc::Compiler compiler;
-
+  nvrtc::Compiler compiler;
   auto ptx = compiler(source_code);
   CHECK(!ptx.empty()) << "Compile PTX failed from source code:\n" << source_code;
-
   cuda_module_.reset(
       new CUDAModule(ptx, compiler.compile_to_cubin() ? CUDAModule::Kind::CUBIN : CUDAModule::Kind::PTX));
 
   RuntimeSymbols symbols;
-
   for (auto& fn : device_module.functions()) {
     std::string kernel_fn_name = fn->name;
     auto fn_kernel             = cuda_module_->GetFunction(0, kernel_fn_name);

cinn/backends/nvrtc/nvrtc_util.cc

@@ -17,19 +17,30 @@
 #include <cuda.h>
 #include <cuda_runtime.h>
 #include <nvrtc.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#include <fstream>
+#include <iostream>
 
 #include "cinn/backends/cuda_util.h"
 #include "cinn/backends/nvrtc/header_generator.h"
 #include "cinn/common/common.h"
+#include "cinn/runtime/flags.h"
 #include "cinn/utils/string.h"
 
+DECLARE_string(cinn_nvcc_cmd_path);
 DECLARE_bool(nvrtc_compile_to_cubin);
 
 namespace cinn {
 namespace backends {
 namespace nvrtc {
 
 std::string Compiler::operator()(const std::string& code, bool include_headers) {
+  if (runtime::CanUseNvccCompiler()) {
+    return CompileWithNvcc(code);
+  }
   return CompileCudaSource(code, include_headers);
 }
 
@@ -140,6 +151,89 @@ std::string Compiler::CompileCudaSource(const std::string& code, bool include_he
   return data;
 }
 
+std::string Compiler::CompileWithNvcc(const std::string& cuda_c) {
+  // read dir source
+  std::string dir = "./source";
+  if (access(dir.c_str(), 0) == -1) {
+    CHECK(mkdir(dir.c_str(), 7) != -1) << "Fail to mkdir " << dir;
+  }
+
+  // get unqiue prefix name
+  prefix_name_ = dir + "/" + common::UniqName("rtc_tmp");
+
+  auto cuda_c_file = prefix_name_ + ".cu";
+  std::ofstream ofs(cuda_c_file, std::ios::out);
+  CHECK(ofs.is_open()) << "Fail to open file " << cuda_c_file;
+  ofs << cuda_c;
+  ofs.close();
+
+  CompileToPtx();
+  CompileToCubin();
+
+  return prefix_name_ + ".cubin";
+}
+
+// std::string Compiler::GetPtx() { return ReadFile(prefix_name_ + ".ptx", std::ios::in); }
+
+void Compiler::CompileToPtx() {
+  auto include_dir            = common::Context::Global().runtime_include_dir();
+  std::string include_dir_str = "";
+  for (auto dir : include_dir) {
+    if (include_dir_str.empty()) {
+      include_dir_str = dir;
+    } else {
+      include_dir_str += ":" + dir;
+    }
+  }
+
+  std::string options = std::string("export PATH=") + FLAGS_cinn_nvcc_cmd_path +
+                        std::string(":$PATH && nvcc -std=c++14 --ptx -O3 -I ") + include_dir_str;
+  options += " -arch=" + GetDeviceArch();
+  options += " -o " + prefix_name_ + ".ptx";
+  options += " " + prefix_name_ + ".cu";
+
+  VLOG(2) << "Nvcc Compile Options : " << options;
+  CHECK(system(options.c_str()) == 0) << options;
+}
+
+void Compiler::CompileToCubin() {
+  std::string options =
+      std::string("export PATH=") + FLAGS_cinn_nvcc_cmd_path + std::string(":$PATH && nvcc --cubin -O3");
+  options += " -arch=" + GetDeviceArch();
+  options += " -o " + prefix_name_ + ".cubin";
+  options += " " + prefix_name_ + ".ptx";
+
+  VLOG(2) << "Nvcc Compile Options : " << options;
+  CHECK(system(options.c_str()) == 0) << options;
+}
+
+std::string Compiler::GetDeviceArch() {
+  int major = 0, minor = 0;
+  if (cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, 0) == cudaSuccess &&
+      cudaDeviceGetAttribute(&minor, cudaDevAttrComputeCapabilityMinor, 0) == cudaSuccess) {
+    return "sm_" + std::to_string(major) + std::to_string(minor);
+  } else {
+    LOG(WARNING) << "cannot detect compute capability from your device, "
+                 << "fall back to compute_30.";
+    return "sm_30";
+  }
+}
+
+std::string Compiler::ReadFile(const std::string& file_name, std::ios_base::openmode mode) {
+  // open cubin file
+  std::ifstream ifs(file_name, mode);
+  CHECK(ifs.is_open()) << "Fail to open file " << file_name;
+  ifs.seekg(std::ios::end);
+  auto len = ifs.tellg();
+  ifs.seekg(0);
+
+  // read cubin file
+  std::string file_data(len, ' ');
+  ifs.read(&file_data[0], len);
+  ifs.close();
+  return std::move(file_data);
+}
+
 }  // namespace nvrtc
 }  // namespace backends
 }  // namespace cinn

cinn/backends/nvrtc/nvrtc_util.h

@@ -70,6 +70,19 @@ class Compiler {
    * whether to compile the source code into cubin, only works with cuda version > 11.1
    */
   bool compile_to_cubin_{false};
+
+  // compile with nvcc
+  std::string CompileWithNvcc(const std::string&);
+
+  // compile to ptx
+  void CompileToPtx();
+  // compile to cubin
+  void CompileToCubin();
+  std::string GetDeviceArch();
+
+  std::string ReadFile(const std::string&, std::ios_base::openmode);
+
+  std::string prefix_name_{""};
 };
 
 }  // namespace nvrtc

cinn/hlir/framework/parallel_compiler.cc

@@ -28,6 +28,7 @@
 #include "cinn/common/context.h"
 #include "cinn/hlir/framework/pass.h"
 #include "cinn/ir/module.h"
+#include "cinn/runtime/flags.h"
 
 DECLARE_int32(cinn_parallel_compile_size);
 DECLARE_int32(cinn_parallel_compile_thread);
@@ -178,10 +179,9 @@ void ParallelCompiler::Task::CodegenAndJit() {
     backends::nvrtc::Compiler compiler;
     auto ptx = compiler(cuda_c);
     CHECK(!ptx.empty()) << "Compile PTX failed from source code:\n" << cuda_c;
-    graph->SavePTXCode(ptx);
-
     // load cumodule
     cumodule.reset(new CUDAModule(ptx, compiler.compile_to_cubin() ? CUDAModule::Kind::CUBIN : CUDAModule::Kind::PTX));
+
     // register kernel
     backends::RuntimeSymbols symbols;
     for (auto& fn : dmodule.functions()) {

cinn/runtime/cuda/cuda_module.cc

@@ -25,6 +25,7 @@
 
 #include "cinn/backends/cuda_util.h"
 #include "cinn/runtime/cuda/cuda_util.h"
+#include "cinn/runtime/flags.h"
 #include "cinn/utils/profiler.h"
 
 namespace cinn {
@@ -106,16 +107,11 @@ CUfunction CUDAModule::GetFunction(int device_id, const std::string& func_name)
     jit_options[4]  = CU_JIT_GENERATE_LINE_INFO;
     jit_opt_vals[4] = reinterpret_cast<void*>(value);
 
-    CUresult status = cuModuleLoadDataEx(
-        &module_per_card_[device_id], data_.c_str(), jit_num_options, jit_options.data(), jit_opt_vals.data());
-
-    if (CUDA_SUCCESS != status) {
-      RAW_LOG(ERROR, "PTX JIT ERROR LOG: %s\n.", log_buffer.data());
-      const char* name;
-      cuGetErrorName(status, &name);
-      const char* msg;
-      cuGetErrorString(status, &msg);
-      RAW_LOG(FATAL, "The error `%s` occurs while compiling the ptx! And its message is `%s`.", name, msg);
+    if (runtime::CanUseNvccCompiler()) {
+      CUDA_DRIVER_CALL(cuModuleLoad(&module_per_card_[device_id], data_.c_str()));
+    } else {
+      CUDA_DRIVER_CALL(cuModuleLoadDataEx(
+          &module_per_card_[device_id], data_.c_str(), jit_num_options, jit_options.data(), jit_opt_vals.data()));
     }
   }
 
@@ -127,11 +123,15 @@ CUfunction CUDAModule::GetFunction(int device_id, const std::string& func_name)
 CUdeviceptr CUDAModule::GetGlobal(int device_id, const std::string& name, size_t nbytes) {
   if (!module_per_card_[device_id]) {
     std::lock_guard<std::mutex> lock(mutex_);
-    CUDA_DRIVER_CALL(cuModuleLoadData(&module_per_card_[device_id], data_.c_str()));
+    if (runtime::CanUseNvccCompiler()) {
+      CUDA_DRIVER_CALL(cuModuleLoad(&module_per_card_[device_id], data_.c_str()));
+    } else {
+      CUDA_DRIVER_CALL(cuModuleLoadData(&module_per_card_[device_id], data_.c_str()));
+    }
   }
 
-  CUdeviceptr global;
   size_t _nbytes;
+  CUdeviceptr global;
   CUDA_DRIVER_CALL(cuModuleGetGlobal(&global, &_nbytes, module_per_card_[device_id], name.c_str()));
   return global;
 }

cinn/runtime/flags.cc

@@ -16,6 +16,9 @@
 
 #include <gflags/gflags.h>
 #include <glog/logging.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
 
 #include <unordered_set>
 
@@ -35,6 +38,9 @@ using ::GFLAGS_NAMESPACE::Int64FromEnv;
 using ::GFLAGS_NAMESPACE::StringFromEnv;
 
 DEFINE_string(cinn_x86_builtin_code_root, StringFromEnv("FLAGS_cinn_x86_builtin_code_root", ""), "");
+DEFINE_string(cinn_nvcc_cmd_path,
+              StringFromEnv("FLAGS_cinn_nvcc_cmd_path", "/usr/local/cuda/bin"),
+              "Setting nvcc default path!");
 
 DEFINE_int32(cinn_parallel_compile_size,
              Int32FromEnv("FLAGS_cinn_parallel_compile_size", 16),
@@ -82,9 +88,13 @@ DEFINE_bool(cinn_use_dense_merge_pass,
             "Whether use dense merge pass.");
 
 DEFINE_bool(nvrtc_compile_to_cubin,
-            BoolFromEnv("FLAGS_nvrtc_compile_to_cubin", false),
+            BoolFromEnv("FLAGS_nvrtc_compile_to_cubin", true),
             "Whether nvrtc compile cuda source into cubin instead of ptx (only works after cuda-11.1).");
 
+DEFINE_bool(cinn_compile_with_nvrtc,
+            BoolFromEnv("FLAGS_cinn_compile_with_nvrtc", true),
+            "Whether nvrtc compile cuda source with nvrtc(default nvcc).");
+
 // FLAGS for performance analysis and accuracy debug
 DEFINE_bool(cinn_sync_run,
             BoolFromEnv("FLAGS_cinn_sync_run", false),
@@ -180,6 +190,11 @@ unsigned long long RandomSeed::Clear() {
   return old_seed;
 }
 
+bool CanUseNvccCompiler() {
+  std::string nvcc_dir = FLAGS_cinn_nvcc_cmd_path + "/nvcc";
+  return (access(nvcc_dir.c_str(), 0) == -1 ? false : true) && (!FLAGS_cinn_compile_with_nvrtc);
+}
+
 bool IsCompiledWithCUDA() {
 #if !defined(CINN_WITH_CUDA)
   return false;

cinn/runtime/flags.h

@@ -27,6 +27,8 @@ bool CheckStringFlagFalse(const std::string &flag);
 void SetCinnCudnnDeterministic(bool state);
 bool GetCinnCudnnDeterministic();
 
+bool CanUseNvccCompiler();
+
 class RandomSeed {
  public:
   static unsigned long long GetOrSet(unsigned long long seed = 0);