[v1.7.x] backport mixed type binary ops to v1.7.x (apache#18649)

* Fix Windows GPU CI (apache#17962)

Update Windows CI to use VS 2019 and enable x64 bit toolchain. Previously we are using an older 32 bit toolchain causing OOM errors during linking. Switching to x64 bit toolchain on the older VS version previously used by the CI was attempted in apache#17912 and did not work. Update to Cuda 10.2 as it is required by VS 2019. Switch to ninja-build on Windows to speed up build as ninja-build is now preinstalled. Remove logic to install cmake 3.16 on every PR as cmake 3.17 is now preinstalled. Add build retrials due to cuda thrust + VS2019 flakyness.

Co-authored-by: vexilligera <vexilligera@gmail.com>

* backport mixed type

Co-authored-by: Leonard Lausen <lausen@amazon.com>
Co-authored-by: vexilligera <vexilligera@gmail.com>

.gitignore

@@ -121,6 +121,10 @@ cmake_install.cmake
 # Mac OS X
 .DS_Store
 
+# Windows
+windows_package.7z
+windows_package
+
 #Notebook Automated Test
 !tests/nightly/test_tutorial_config.txt
 !tests/nightly/TestNotebook

ci/build_windows.py

@@ -33,13 +33,15 @@
 import zipfile
 from distutils.dir_util import copy_tree
 from enum import Enum
-from subprocess import check_call
+from subprocess import check_call, call
 
 from util import *
 
 KNOWN_VCVARS = {
+    # https://gitlab.kitware.com/cmake/cmake/issues/18920
     'VS 2015': r'C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\bin\x86_amd64\vcvarsx86_amd64.bat',
-    'VS 2017': r'C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Auxiliary\Build\vcvarsx86_amd64.bat'
+    'VS 2017': r'C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Auxiliary\Build\vcvarsx86_amd64.bat',
+    'VS 2019': r'C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Auxiliary\Build\vcvars64.bat',
 }
 
 
@@ -54,6 +56,8 @@ class BuildFlavour(Enum):
 
 CMAKE_FLAGS = {
     'WIN_CPU': (
+        '-DCMAKE_C_COMPILER=cl '
+        '-DCMAKE_CXX_COMPILER=cl '
         '-DUSE_CUDA=OFF '
         '-DUSE_CUDNN=OFF '
         '-DENABLE_CUDA_RTC=OFF '
@@ -67,6 +71,8 @@ class BuildFlavour(Enum):
         '-DCMAKE_BUILD_TYPE=Release')
 
     , 'WIN_CPU_MKLDNN': (
+        '-DCMAKE_C_COMPILER=cl '
+        '-DCMAKE_CXX_COMPILER=cl '
         '-DUSE_CUDA=OFF '
         '-DUSE_CUDNN=OFF '
         '-DENABLE_CUDA_RTC=OFF '
@@ -80,6 +86,8 @@ class BuildFlavour(Enum):
         '-DCMAKE_BUILD_TYPE=Release')
 
     , 'WIN_CPU_MKLDNN_MKL': (
+        '-DCMAKE_C_COMPILER=cl '
+        '-DCMAKE_CXX_COMPILER=cl '
         '-DUSE_CUDA=OFF '
         '-DUSE_CUDNN=OFF '
         '-DENABLE_CUDA_RTC=OFF '
@@ -93,6 +101,8 @@ class BuildFlavour(Enum):
         '-DCMAKE_BUILD_TYPE=Release')
 
     , 'WIN_CPU_MKL': (
+        '-DCMAKE_C_COMPILER=cl '
+        '-DCMAKE_CXX_COMPILER=cl '
         '-DUSE_CUDA=OFF '
         '-DUSE_CUDNN=OFF '
         '-DENABLE_CUDA_RTC=OFF '
@@ -106,6 +116,8 @@ class BuildFlavour(Enum):
         '-DCMAKE_BUILD_TYPE=Release')
 
     , 'WIN_GPU': (
+        '-DCMAKE_C_COMPILER=cl '
+        '-DCMAKE_CXX_COMPILER=cl '
         '-DUSE_CUDA=ON '
         '-DUSE_CUDNN=ON '
         '-DENABLE_CUDA_RTC=ON '
@@ -120,6 +132,8 @@ class BuildFlavour(Enum):
         '-DCMAKE_BUILD_TYPE=Release')
 
     , 'WIN_GPU_MKLDNN': (
+        '-DCMAKE_C_COMPILER=cl '
+        '-DCMAKE_CXX_COMPILER=cl '
         '-DUSE_CUDA=ON '
         '-DUSE_CUDNN=ON '
         '-DENABLE_CUDA_RTC=ON '
@@ -140,38 +154,40 @@ def windows_build(args):
     logging.info("Using vcvars environment:\n{}".format(args.vcvars))
 
     path = args.output
-    os.makedirs(path, exist_ok=True)
 
-    mxnet_root = get_mxnet_root()
-    logging.info("Found MXNet root: {}".format(mxnet_root))
+    # cuda thrust + VS 2019 is flaky: try multiple times if fail
+    MAXIMUM_TRY = 5
+    build_try = 0
+
+    while build_try < MAXIMUM_TRY:
+        if os.path.exists(path):
+            shutil.rmtree(path)
+        os.makedirs(path, exist_ok=True)
 
-    url = 'https://github.com/Kitware/CMake/releases/download/v3.16.1/cmake-3.16.1-win64-x64.zip'
-    with tempfile.TemporaryDirectory() as tmpdir:
-        cmake_file_path = download_file(url, tmpdir)
-        with zipfile.ZipFile(cmake_file_path, 'r') as zip_ref:
-            # Create $tmpdir\cmake-3.16.1-win64-x64\bin\cmake.exe
-            zip_ref.extractall(tmpdir)
+        mxnet_root = get_mxnet_root()
+        logging.info("Found MXNet root: {}".format(mxnet_root))
 
         with remember_cwd():
             os.chdir(path)
-            cmd = "\"{}\" && {} -G \"NMake Makefiles JOM\" {} {}".format(
-                args.vcvars,
-                os.path.join(tmpdir, 'cmake-3.16.1-win64-x64', 'bin', 'cmake.exe'),
-                CMAKE_FLAGS[args.flavour], mxnet_root)
+            cmd = "\"{}\" && cmake -GNinja {} {}".format(args.vcvars,
+                                                         CMAKE_FLAGS[args.flavour],
+                                                         mxnet_root)
             logging.info("Generating project with CMake:\n{}".format(cmd))
             check_call(cmd, shell=True)
 
-            cmd = "\"{}\" && jom".format(args.vcvars)
-            logging.info("Building with jom:\n{}".format(cmd))
+            cmd = "\"{}\" && ninja".format(args.vcvars)
+            logging.info("Building:\n{}".format(cmd))
 
             t0 = int(time.time())
-            check_call(cmd, shell=True)
-
-            logging.info(
-                "Build flavour: {} complete in directory: \"{}\"".format(
-                    args.flavour, os.path.abspath(path)))
-            logging.info("Build took {}".format(
-                datetime.timedelta(seconds=int(time.time() - t0))))
+            ret = call(cmd, shell=True)
+
+            if ret != 0:
+                build_try += 1
+                logging.info("{} build(s) have failed".format(build_try))
+            else:
+                logging.info("Build flavour: {} complete in directory: \"{}\"".format(args.flavour, os.path.abspath(path)))
+                logging.info("Build took {}".format(datetime.timedelta(seconds=int(time.time() - t0))))
+                break
     windows_package(args)
 
 
@@ -233,7 +249,7 @@ def main():
 
     parser.add_argument("--vcvars",
         help="vcvars batch file location, typically inside vs studio install dir",
-        default=KNOWN_VCVARS['VS 2015'],
+        default=KNOWN_VCVARS['VS 2019'],
         type=str)
 
     parser.add_argument("--arch",
@@ -258,7 +274,7 @@ def main():
         if 'OpenCV_DIR' not in os.environ:
             os.environ["OpenCV_DIR"] = "C:\\Program Files\\OpenCV-v3.4.1\\build"
         if 'CUDA_PATH' not in os.environ:
-            os.environ["CUDA_PATH"] = "C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v9.2"
+            os.environ["CUDA_PATH"] = "C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.2"
         if 'MKL_ROOT' not in os.environ:
             os.environ["MKL_ROOT"] = "C:\\Program Files (x86)\\IntelSWTools\\compilers_and_libraries\\windows\\mkl"
         windows_build(args)

include/mxnet/imperative.h

@@ -117,6 +117,16 @@ class Imperative {
     }
     return is_np_shape_thread_local_ ? 1 : 0;
   }
+
+  /*! \brief return current numpy default dtype compatibility status.
+  * */
+  bool is_np_default_dtype() const {
+    if (is_np_default_dtype_global_) {
+      return true;
+    }
+    return false;
+  }
+
   /*! \brief specify numpy compatibility off, thread local on or global on. */
   bool set_is_np_shape(int is_np_shape) {
     NumpyShape flag = static_cast<NumpyShape>(is_np_shape);
@@ -215,6 +225,7 @@ class Imperative {
   static MX_THREAD_LOCAL bool is_np_shape_thread_local_;
 #endif
   bool is_np_shape_global_{false};
+  bool is_np_default_dtype_global_{false};
   /*! \brief node count used for naming */
   std::atomic<uint64_t> node_count_{0};
   /*! \brief variable count used for naming */

python/mxnet/symbol/numpy/_symbol.py

@@ -1529,13 +1529,15 @@ def _ufunc_helper(lhs, rhs, fn_array, fn_scalar, lfn_scalar, rfn_scalar=None, ou
         if isinstance(rhs, numeric_types):
             return fn_scalar(lhs, rhs, out=out)
         else:
+            is_int = isinstance(rhs, integer_types)
             if rfn_scalar is None:
                 # commutative function
-                return lfn_scalar(rhs, float(lhs), out=out)
+                return lfn_scalar(rhs, scalar=float(lhs), is_int=is_int, out=out)
             else:
-                return rfn_scalar(rhs, float(lhs), out=out)
+                return rfn_scalar(rhs, scalar=float(lhs), is_int=is_int, out=out)
     elif isinstance(rhs, numeric_types):
-        return lfn_scalar(lhs, float(rhs), out=out)
+        is_int = isinstance(rhs, integer_types)
+        return lfn_scalar(lhs, scalar=float(rhs), is_int=is_int, out=out)
     elif isinstance(rhs, Symbol):
         return fn_array(lhs, rhs, out=out)
     else:

src/common/utils.h

@@ -31,6 +31,7 @@
 #include <nnvm/node.h>
 #include <mxnet/engine.h>
 #include <mxnet/ndarray.h>
+#include <mxnet/imperative.h>
 #include <mxnet/op_attr_types.h>
 #include <mxnet/graph_attr_types.h>
 #include <nnvm/graph_attr_types.h>
@@ -874,6 +875,11 @@ inline bool is_float(const int dtype) {
   return dtype == mshadow::kFloat32 || dtype == mshadow::kFloat64 || dtype == mshadow::kFloat16;
 }
 
+inline bool is_int(const int dtype) {
+  return dtype == mshadow::kUint8 || dtype == mshadow::kInt8 ||
+         dtype == mshadow::kInt32 || dtype == mshadow::kInt64;
+}
+
 inline int get_more_precise_type(const int type1, const int type2) {
   if (type1 == type2) return type1;
   if (is_float(type1) && is_float(type2)) {
@@ -910,6 +916,19 @@ inline int np_binary_out_infer_type(const int type1, const int type2) {
   return get_more_precise_type(type1, type2);
 }
 
+inline int GetDefaultDtype() {
+  return Imperative::Get()->is_np_default_dtype() ?
+         mshadow::kFloat64 :
+         mshadow::kFloat32;
+}
+
+inline int GetDefaultDtype(int dtype) {
+  if (dtype != -1) return dtype;
+  return Imperative::Get()->is_np_default_dtype() ?
+         mshadow::kFloat64 :
+         mshadow::kFloat32;
+}
+
 }  // namespace common
 }  // namespace mxnet
 #endif  // MXNET_COMMON_UTILS_H_

src/operator/contrib/gradient_multiplier_op.cc

@@ -77,9 +77,7 @@ In forward pass it acts as an identity transform. During backpropagation it
 multiplies the gradient from the subsequent level by a scalar factor lambda and passes it to
 the preceding layer.
 )code" ADD_FILELINE)
-.set_attr_parser([](NodeAttrs* attrs) {
-    attrs->parsed = dmlc::stod(attrs->dict["scalar"]);
-  })
+.set_attr_parser(ParamParser<NumpyBinaryScalarParam>)
 .set_attr<FInferStorageType>("FInferStorageType", ElemwiseStorageType<1, 1, false, true, true>)
 .set_attr<FCompute>("FCompute<cpu>", UnaryOp::IdentityCompute<cpu>)
 .set_attr<FComputeEx>("FComputeEx<cpu>", UnaryOp::IdentityComputeEx<cpu>)
@@ -88,7 +86,7 @@ the preceding layer.
   [](const NodeAttrs& attrs){
     return std::vector<bool>{true};
   })
-.add_argument("scalar", "float", "lambda multiplier");
+.add_arguments(NumpyBinaryScalarParam::__FIELDS__());
 
 MXNET_OPERATOR_REGISTER_BINARY_SCALAR(_contrib_backward_gradientmultiplier)
 .set_attr<nnvm::TIsBackward>("TIsBackward", true)