Merge pull request apache#3 from DongBaiYue/sycl

[SYCL] Support auto_scheduler

CMakeLists.txt

@@ -29,6 +29,7 @@ tvm_option(USE_OPENCL "Build with OpenCL" OFF)
 tvm_option(USE_OPENCL_GTEST "Path to OpenCL specific gtest version for runtime cpp tests." /path/to/opencl/gtest)
 tvm_option(USE_VULKAN "Build with Vulkan" OFF)
 tvm_option(USE_SYCL "Build with SYCL" OFF)
+tvm_option(SYCL_TEMP_FOLDER "Build with SYCL" "/tmp/tvm_sycl")
 
 
 # Whether to use spirv-tools.and SPIRV-Headers from Khronos github or gitlab.
@@ -525,6 +526,9 @@ add_library(tvm_objs OBJECT ${COMPILER_SRCS})
 add_library(tvm_runtime_objs OBJECT ${RUNTIME_SRCS})
 add_library(tvm_libinfo_objs OBJECT ${LIBINFO_FILE})
 if(USE_SYCL)
+ set(CMAKE_CXX_COMPILER "${USE_SYCL}/bin/clang++")
+ add_definitions(-DSYCL_CXX_COMPILER="${USE_SYCL}/bin/clang++")
+ add_definitions(-DSYCL_TEMP_FOLDER="${SYCL_TEMP_FOLDER}")
  set(CMAKE_CXX_FLAGS "-fsycl -fsycl-targets=nvptx64-nvidia-cuda ${CMAKE_CXX_FLAGS}")
  add_library(tvm_sycl_runtime_objs OBJECT ${SYCL_RUNTIME_SRCS})
  add_library(tvm SHARED $<TARGET_OBJECTS:tvm_objs> $<TARGET_OBJECTS:tvm_runtime_objs> $<TARGET_OBJECTS:tvm_sycl_runtime_objs> $<TARGET_OBJECTS:tvm_libinfo_objs>)

cmake/config.cmake

@@ -65,10 +65,12 @@ set(USE_AOCL OFF)
 # Whether enable SYCL runtime
 #
 # Possible values:
-# - ON: enable SYCL with cmake's auto search
+# - ON: enable SYCL with cmake's auto search, not support now.
 # - OFF: disable SYCL
 # - /path/to/sycl: use specific path to sycl
 set(USE_SYCL OFF)
+# if enable SYCL runtime, please set storage path for sycl temporary files. The default path is "/tmp/tvm_sycl/".
+set(SYCL_TEMP_FOLDER "/tmp/tvm_sycl")
 
 # Whether enable OpenCL runtime
 #

cmake/modules/SYCL.cmake

@@ -20,8 +20,8 @@ if(USE_SYCL)
  tvm_file_glob(GLOB RUNTIME_SYCL_SRCS src/runtime/sycl/*.cc)
  list(APPEND SYCL_RUNTIME_SRCS ${RUNTIME_SYCL_SRCS})
 
- include_directories(BEFORE SYSTEM $ENV{SYCL_ROOT_DIR}/include/sycl/)
- list(APPEND TVM_RUNTIME_LINKER_LIBS $ENV{SYCL_ROOT_DIR}/lib/)
+ include_directories(BEFORE SYSTEM ${USE_SYCL}/include/sycl/)
+ list(APPEND TVM_RUNTIME_LINKER_LIBS ${USE_SYCL}/lib/)
 else()
  list(APPEND COMPILER_SRCS src/target/opt/build_sycl_off.cc)
 endif(USE_SYCL)

python/tvm/auto_scheduler/utils.py

@@ -293,6 +293,15 @@ def call_func_with_timeout(
  worker.send(func, args, kwargs, timeout)
  try:
  res = worker.recv()
+ except ChildProcessError:
+ # when a sycl runtime error occurs in the child process used for measure, the child process is terminated.
+ # Restart measure the previous schedule.
+ print("Restart measure the previous schedule.")
+ worker.send(func, args, kwargs, timeout)
+ try:
+ res = worker.recv()
+ except Exception: # pylint: disable=broad-except
+ res = Exception(make_traceback_info())
  except Exception: # pylint: disable=broad-except
  res = Exception(make_traceback_info())
 

python/tvm/autotvm/tophub.py

@@ -108,7 +108,11 @@ def context(target, extra_files=None):
  device = tgt.attrs.get("device", "")
  if device != "":
  possible_names.append(_alias(device))
- possible_names.append(tgt.kind.name)
+ # use opencl pre-tuned log when target=sycl, because lack of sycl pre-tuned log
+ if(tgt.kind.name == "sycl"):
+ possible_names.append("opencl")
+ else:
+ possible_names.append(tgt.kind.name)
 
  all_packages = list(PACKAGE_VERSION.keys())
  for name in possible_names:

src/auto_scheduler/search_policy/utils.h

@@ -58,7 +58,8 @@ inline bool IsGPUTask(const SearchTask& task) {
  (task)->target->kind->device_type == kDLVulkan ||
  (task)->target->kind->device_type == kDLMetal ||
  (task)->target->kind->device_type == kDLROCM ||
- (task)->target->kind->device_type == kOpenGL;
+ (task)->target->kind->device_type == kOpenGL ||
+ (task)->target->kind->device_type == kDLSYCL;
 }
 
 /*! \brief Return whether the search task is targeting a CUDA GPU. */
@@ -70,6 +71,10 @@ inline bool IsCUDATask(const SearchTask& task) {
 inline bool IsOpenCLTask(const SearchTask& task) {
  return (task)->target->kind->device_type == kDLOpenCL;
 }
+/*! \brief Return whether the search task is targeting a SYCL GPU. */
+inline bool IsSYCLTask(const SearchTask& task) {
+ return (task)->target->kind->device_type == kDLSYCL;
+}
 
 /*! \brief Argsort. Order: largest to smallest */
 template <typename T>

src/auto_scheduler/search_task.cc

@@ -154,6 +154,33 @@ HardwareParams HardwareParamsNode::GetDefaultHardwareParams(const Target& target
 
  return HardwareParams(-1, 16, 64, max_shared_memory_per_block, max_local_memory_per_block,
  max_threads_per_block, max_vthread_extent, warp_size);
+ } else if (device_type == kDLSYCL) {
+ auto dev = Device{static_cast<DLDeviceType>(device_type), 0};
+ auto device_name = "device_api.sycl";
+ auto func = tvm::runtime::Registry::Get(device_name);
+ ICHECK(func != nullptr) << "Cannot find SYCL device_api in registry";
+ auto device_api = static_cast<tvm::runtime::DeviceAPI*>(((*func)()).operator void*());
+
+ tvm::runtime::TVMRetValue ret;
+
+ //device_api->GetAttr(dev, tvm::runtime::DeviceAttrKind::kCacheLineSize, &ret);
+ //int cache_line_size = ret;
+
+ device_api->GetAttr(dev, tvm::runtime::DeviceAttrKind::kMaxSharedMemoryPerBlock, &ret);
+ int max_shared_memory_per_block = ret;
+
+ int max_local_memory_per_block = INT32_MAX;
+
+ device_api->GetAttr(dev, tvm::runtime::DeviceAttrKind::kMaxThreadsPerBlock, &ret);
+ int max_threads_per_block = ret;
+
+ device_api->GetAttr(dev, tvm::runtime::DeviceAttrKind::kWarpSize, &ret);
+ int warp_size = ret;
+
+ int max_vthread_extent = std::max(1, warp_size / 4);
+ return HardwareParams(-1, 16, 64, max_shared_memory_per_block, max_local_memory_per_block,
+ max_threads_per_block, max_vthread_extent, warp_size);
+
  } else {
  LOG(FATAL) << "No default hardware parameters for target: " << target;
  }

src/runtime/opencl/opencl_module.cc

@@ -212,13 +212,6 @@ void OpenCLModuleNode::Init() {
  // zero initialize cl_program pointers for each device kernel
  for (auto& kv : parsed_kernels_) {
  programs_.insert({kv.first, std::vector<cl_program>(workspace_->devices.size(), nullptr)});
- std::string kernel_name = kv.first;
- std::string kernel_code = kv.second;
- std::string filepath = "/tmp/tvm_sycl/"+kernel_name+"_opencl.cpp";
- std::ofstream myfile;
- myfile.open(filepath);
- myfile << kernel_code;
- myfile.close();
  }
 }