[MODULE] Enable OpenCL and CUDA Modules

include/tvm/runtime/c_runtime_api.h

@@ -373,34 +373,6 @@ TVM_DLL int TVMFuncListGlobalNames(int *out_size,
                                    const char*** out_array);
 
 // Array related apis for quick proptying
-/*!
- * \brief Initialize certain type of devices, this may
- *  not be necessary for all device types. But is needed for OpenCL.
- *
- * \param dev_mask The device mask of device type to be initialized
- * \param option_keys Additional option  keys to pass.
- * \param option_vals Additional option values to pass
- * \param num_options Number of options to be passed into it.
- * \param out_code 1: success, 0: already initialized
- * \return 0 when success, -1 when failure happens
- */
-TVM_DLL int TVMDeviceInit(int dev_mask,
-                          const char** option_keys,
-                          const char** option_vals,
-                          int num_options,
-                          int *out_code);
-
-
-/*!
- * \brief Whether the specified context is enabled.
- *
- * \param ctx The context to be checked.
- * \param out_enabled whether the ctx is enabled.
- * \return Whether the function is successful.
- */
-TVM_DLL int TVMContextEnabled(TVMContext ctx,
-                              int* out_enabled);
-
 /*!
  * \brief Allocate a nd-array's memory,
  *  including space of shape, of given spec.

include/tvm/runtime/packed_func.h

@@ -535,8 +535,9 @@ inline const char* TypeCode2Str(int type_code) {
 }
 
 inline std::ostream& operator<<(std::ostream& os, TVMType t) {  // NOLINT(*)
-  os << TypeCode2Str(t.code)
-     << static_cast<int>(t.bits);
+  os << TypeCode2Str(t.code);
+  if (t.code == kHandle) return os;
+  os << static_cast<int>(t.bits);
   if (t.lanes != 1) {
     os << 'x' << static_cast<int>(t.lanes);
   }
@@ -559,7 +560,7 @@ inline TVMType String2TVMType(std::string s) {
     t.code = kUInt; scan = s.c_str() + 4;
   } else if (s.substr(0, 5) == "float") {
     t.code = kFloat; scan = s.c_str() + 5;
-  } else if (s == "handle") {
+  } else if (s.substr(0, 6) == "handle") {
     t.code = kHandle;
     t.bits = 64;  // handle uses 64 bit by default.
     scan = s.c_str() + 6;

python/tvm/__init__.py

@@ -15,7 +15,7 @@
 from . import module
 
 from . import ndarray as nd
-from .ndarray import cpu, gpu, opencl, init_opencl, cl
+from .ndarray import cpu, gpu, opencl, cl
 
 from ._base import TVMError
 from .api import *

python/tvm/_ctypes/_ndarray.py

@@ -7,10 +7,9 @@
 import numpy as np
 
 from .._base import _LIB, check_call
-from .._base import c_array, c_str
+from .._base import c_array
 from ._types import TVMType, tvm_index_t
 
-
 class TVMContext(ctypes.Structure):
     """TVM context strucure."""
     _fields_ = [("dev_mask", ctypes.c_int),
@@ -29,12 +28,6 @@ def __repr__(self):
         return "%s(%d)" % (
             TVMContext.MASK2STR[self.dev_mask], self.dev_id)
 
-    @property
-    def enabled(self):
-        ret = ctypes.c_int()
-        check_call(_LIB.TVMContextEnabled(self, ctypes.byref(ret)))
-        return ret.value != 0
-
 
 class TVMArray(ctypes.Structure):
     """TVMValue in C API"""
@@ -141,30 +134,6 @@ def sync(ctx):
     check_call(_LIB.TVMSynchronize(ctx, None))
 
 
-def init_opencl(**kwargs):
-    """Initialize the opencl with the options.
-
-    Parameters
-    ----------
-    kwargs : dict
-        The options
-    """
-    keys = []
-    vals = []
-    for k, v in kwargs.items():
-        keys.append(c_str(k))
-        vals.append(c_str(v))
-    dev_mask = ctypes.c_int(4)
-    out_code = ctypes.c_int()
-    check_call(_LIB.TVMDeviceInit(
-        dev_mask,
-        c_array(ctypes.c_char_p, keys),
-        c_array(ctypes.c_char_p, vals),
-        ctypes.c_int(len(keys)),
-        ctypes.byref(out_code)))
-    return out_code.value != 0
-
-
 class NDArrayBase(object):
     """A simple Device/CPU Array object in runtime."""
     __slots__ = ["handle"]

python/tvm/addon/testing.py

@@ -0,0 +1,34 @@
+"""Utilities to make tempdir"""
+from __future__ import absolute_import as _abs
+import os
+import tempfile
+import shutil
+
+class TempDirectory(object):
+    """Helper object to manage temp directory during testing"""
+    def __init__(self):
+        self.temp_dir = tempfile.mkdtemp()
+
+    def __del__(self):
+        shutil.rmtree(self.temp_dir)
+
+    def relpath(self, name):
+        """Relative path in temp dir
+
+        Parameters
+        ----------
+        name : str
+            The name of the file.
+        """
+        return os.path.join(self.temp_dir, name)
+
+
+def tempdir():
+    """Return a new temp dir which deletes the contents when exit
+
+    Returns
+    -------
+    temp : TempDirectory
+        The temp directory object
+    """
+    return TempDirectory()

python/tvm/libinfo.py

@@ -1,9 +1,11 @@
 # coding: utf-8
 """Information about nnvm."""
 from __future__ import absolute_import
+import sys
 import os
 import platform
 
+
 def find_lib_path():
     """Find dynamic library files.
 
@@ -12,6 +14,7 @@ def find_lib_path():
     lib_path : list(string)
         List of all found path to the libraries
     """
+    use_runtime = os.environ.get("TVM_USE_RUNTIME_LIB", False)
     curr_path = os.path.dirname(os.path.abspath(os.path.expanduser(__file__)))
     api_path = os.path.join(curr_path, '../../lib/')
     cmake_build_path = os.path.join(curr_path, '../../build/Release/')
@@ -26,15 +29,24 @@ def find_lib_path():
             dll_path.append(os.path.join(curr_path, '../../windows', vs_configuration))
     elif os.name == "posix" and os.environ.get('LD_LIBRARY_PATH', None):
         dll_path.extend([p.strip() for p in os.environ['LD_LIBRARY_PATH'].split(":")])
+
     if os.name == 'nt':
-        dll_path = [os.path.join(p, 'libtvm.dll') for p in dll_path]
+        lib_dll_path = [os.path.join(p, 'libtvm.dll') for p in dll_path]
+        runtime_dll_path = [os.path.join(p, 'libtvm_runtime.dll') for p in dll_path]
     else:
-        dll_path = [os.path.join(p, 'libtvm.so') for p in dll_path]
-    lib_path = [p for p in dll_path if os.path.exists(p) and os.path.isfile(p)]
-    if len(lib_path) == 0:
+        lib_dll_path = [os.path.join(p, 'libtvm.so') for p in dll_path]
+        runtime_dll_path = [os.path.join(p, 'libtvm_runtime.so') for p in dll_path]
+
+    dll_path = runtime_dll_path if use_runtime else lib_dll_path
+    lib_found = [p for p in dll_path if os.path.exists(p) and os.path.isfile(p)]
+
+    if len(lib_found) == 0:
         raise RuntimeError('Cannot find the files.\n' +
                            'List of candidates:\n' + str('\n'.join(dll_path)))
-    return lib_path
+    if use_runtime:
+        sys.stderr.write("Loading runtime library... this is execution only\n")
+        sys.stderr.flush()
+    return lib_found
 
 
 # current version

python/tvm/ndarray.py

@@ -9,7 +9,6 @@
 from ._ctypes._ndarray import TVMContext, TVMType, NDArrayBase
 from ._ctypes._ndarray import cpu, gpu, opencl, empty, sync
 from ._ctypes._ndarray import _init_ndarray_module
-from ._ctypes._ndarray import init_opencl
 from ._ctypes._function import Function
 
 cl = opencl

src/api/api_codegen.cc

@@ -21,7 +21,7 @@ TVM_REGISTER_API(_codegen_build)
     }
   });
 
-TVM_REGISTER_API(_codegen_target_enabled)
+TVM_REGISTER_API(_codegen_enabled)
 .set_body([](TVMArgs args, TVMRetValue *ret) {
     *ret = TargetEnabled(args[0]);
   });

src/codegen/build_cuda.cc

@@ -61,10 +61,13 @@ runtime::Module BuildCUDA(Array<LoweredFunc> funcs) {
   if (const auto* f = Registry::Get("tvm_callback_cuda_postproc")) {
     code = (*f)(code).operator std::string();
   }
-
+  std::string fmt = "ptx";
   std::string ptx;
   if (const auto* f = Registry::Get("tvm_callback_cuda_compile")) {
     ptx = (*f)(code).operator std::string();
+    // Dirty matching to check PTX vs cubin.
+    // TODO(tqchen) more reliable checks
+    if (ptx[0] != '/') fmt = "cubin";
   } else {
     ptx = NVRTCCompile(code);
   }
@@ -80,7 +83,7 @@ runtime::Module BuildCUDA(Array<LoweredFunc> funcs) {
     }
     fmap[f->name] = info;
   }
-  return CUDAModuleCreate(ptx, "ptx", fmap, code);
+  return CUDAModuleCreate(ptx, fmt, fmap, code);
 }
 
 TVM_REGISTER_API(_codegen_build_cuda)

src/runtime/c_runtime_api.cc

@@ -200,38 +200,6 @@ int TVMFuncCreateFromCFunc(TVMPackedCFunc func,
   API_END();
 }
 
-int TVMDeviceInit(int dev_mask,
-                  const char** option_keys,
-                  const char** option_vals,
-                  int num_options,
-                  int* out_code) {
-  API_BEGIN();
-  *out_code = 1;
-  switch (dev_mask) {
-    case kOpenCL: {
-      *out_code = DeviceInit<kOpenCL>(option_keys, option_vals, num_options);
-      break;
-    }
-    default: break;
-  }
-  API_END();
-}
-
-int TVMContextEnabled(TVMContext ctx,
-                      int* out_enabled) {
-  API_BEGIN();
-  if (ctx.dev_mask == kGPU && TVM_CUDA_RUNTIME == 0) {
-    *out_enabled = 0;
-  } else if (ctx.dev_mask == kOpenCL && TVM_OPENCL_RUNTIME == 0) {
-    *out_enabled = 0;
-  } else {
-    TVM_DEVICE_SWITCH(ctx, {
-        *out_enabled = CheckEnabled<xpu>(ctx);
-      });
-  }
-  API_END();
-}
-
 int TVMArrayAlloc(const tvm_index_t* shape,
                   tvm_index_t ndim,
                   TVMType dtype,

src/runtime/cuda/cuda_module.cc

@@ -5,6 +5,7 @@
 #include "./cuda_module.h"
 
 #if TVM_CUDA_RUNTIME
+
 #include <tvm/runtime/registry.h>
 #include <cuda.h>
 #include <cuda_runtime.h>
@@ -60,7 +61,12 @@ class CUDAModuleNode : public runtime::ModuleNode {
 
   void SaveToFile(const std::string& file_name,
                   const std::string& format) final {
-    LOG(FATAL) << "Not implemented";
+    std::string fmt = GetFileFormat(file_name, format);
+    CHECK_EQ(fmt, fmt_)
+        << "Can only save to format=" << fmt_;
+    std::string meta_file = GetMetaFilePath(file_name);
+    SaveMetaDataToFile(meta_file, fmap_);
+    SaveBinaryToFile(file_name, data_);
   }
 
   std::string GetSource(const std::string& format) final {
@@ -212,9 +218,13 @@ Module CUDAModuleCreate(
 // Load module from module.
 Module CUDAModuleLoad(const std::string& file_name,
                       const std::string& format) {
+  std::string data;
+  std::unordered_map<std::string, FunctionInfo> fmap;
   std::string fmt = GetFileFormat(file_name, format);
-  std::string data = LoadBinaryFile(file_name);
-  return CUDAModuleCreate(data, fmt, {{}}, std::string());
+  std::string meta_file = GetMetaFilePath(file_name);
+  LoadBinaryFromFile(file_name, &data);
+  LoadMetaDataFromFile(meta_file, &fmap);
+  return CUDAModuleCreate(data, fmt, fmap, std::string());
 }
 
 TVM_REGISTER_GLOBAL(_module_loadfile_cubin)

src/runtime/device_api.h

@@ -11,31 +11,6 @@
 
 namespace tvm {
 namespace runtime {
-/*!
- * \brief Initialize the device.
- * \param option_keys Additional option  keys to pass.
- * \param option_vals Additional option values to pass
- * \param num_options Number of options to be passed into it.
- * \return 0 if success, 1: if already initialized
- * \tparam xpu The device mask.
- */
-template<TVMDeviceMask xpu>
-inline bool DeviceInit(const char** option_keys,
-                       const char** option_vals,
-                       int num_options) {
-  return true;
-}
-
-/*!
- * \brief Whether ctx is enabled.
- * \param ctx The device context to perform operation.
- * \tparam xpu The device mask.
- */
-template<TVMDeviceMask xpu>
-inline bool CheckEnabled(TVMContext ctx) {
-  return true;
-}
-
 /*!
  * \brief Allocate a data space on device.
  * \param ctx The device context to perform operation.