enable for windows

Author: haojin2

src/operator/numpy/np_broadcast_reduce_op.h

@@ -222,15 +222,15 @@ void NumpyReduceAxesCompute(const nnvm::NodeAttrs& attrs,
                             const std::vector<TBlob>& inputs,
                             const std::vector<OpReqType>& req,
                             const std::vector<TBlob>& outputs) {
+  using namespace mshadow;
   if (req[0] == kNullOp) return;
   const NumpyReduceAxesParam& param = nnvm::get<NumpyReduceAxesParam>(attrs.parsed);
   if (param.initial.has_value()) {
     LOG(FATAL) << "initial is not supported yet";
   }
+  Stream<xpu>* s = ctx.get_stream<xpu>();
   if (inputs[0].shape_.Size() == 0 && outputs[0].shape_.Size() != 0) {
     using namespace mxnet_op;
-    using namespace mshadow;
-    Stream<xpu>* s = ctx.get_stream<xpu>();
     MSHADOW_TYPE_SWITCH(outputs[0].type_flag_, DType, {
       Kernel<set_zero, xpu>::Launch(s, outputs[0].shape_.Size(), outputs[0].dptr<DType>());
     });
@@ -247,6 +247,13 @@ void NumpyReduceAxesCompute(const nnvm::NodeAttrs& attrs,
       LOG(FATAL) << "Only reduce op: `sum` is supported for boolean ndarrays";
     }
     TVMOpReduce(ctx, inputs[0], param.axis, outputs[0], req[0], reducer_name);
+    if (normalize) {
+      using namespace mshadow::expr;
+      MSHADOW_TYPE_SWITCH(outputs[0].type_flag_, OType, {
+        auto out = outputs[0].FlatTo2D<xpu, OType>(s);
+        out /= scalar<OType>(inputs[0].Size()/outputs[0].Size());
+      });
+    }
     return;
   }
 #endif

src/operator/tensor/broadcast_reduce-inl.cuh

@@ -619,8 +619,6 @@ void Reduce(Stream<gpu> *s, const TBlob& small, const OpReqType req,
   ReduceImplConfig<ndim> config =
     ConfigureReduceImpl<ndim, DType>(small.shape_, big.shape_, NULL, NULL);
   if (safe_acc) {
-    // TODO(haojin2): Use real-only type swtich for windows temporarily due to CI issues.
-#ifndef _WIN32
     MXNET_ACC_TYPE_SWITCH(mshadow::DataType<DType>::kFlag, DataType, AType, {
       typedef typename std::conditional<safe_acc, AType, DataType>::type AccType;
       MSHADOW_TYPE_SWITCH(small.type_flag_, OType, {
@@ -630,17 +628,6 @@ void Reduce(Stream<gpu> *s, const TBlob& small, const OpReqType req,
           stream, small, req, big, workspace, config);
       });
     });
-#else
-    MXNET_REAL_ACC_TYPE_SWITCH(mshadow::DataType<DType>::kFlag, DataType, AType, {
-      typedef typename std::conditional<safe_acc, AType, DataType>::type AccType;
-      MSHADOW_TYPE_SWITCH(small.type_flag_, OType, {
-        typedef typename std::conditional<safe_acc, OType, DataType>::type OutType;
-        config = ConfigureReduceImpl<ndim, AccType>(small.shape_, big.shape_, NULL, NULL);
-        ReduceImpl<Reducer, ndim, AccType, DataType, OutType, OP>(
-          stream, small, req, big, workspace, config);
-      });
-    });
-#endif
   } else {
     ReduceImpl<Reducer, ndim, DType, DType, DType, OP>(stream, small, req, big, workspace, config);
   }

src/operator/tensor/broadcast_reduce-inl.h

@@ -241,8 +241,6 @@ void Reduce(Stream<cpu>* s, const TBlob& small, const OpReqType req,
       N, M, req == kAddTo, big.dptr<DType>(), small.dptr<DType>(),
       big.shape_.get<ndim>(), small.shape_.get<ndim>(), rshape, rstride);
   } else {
-    // TODO(haojin2): Use real-only type swtich for windows temporarily due to CI issues.
-#ifndef _WIN32
     MXNET_ACC_TYPE_SWITCH(mshadow::DataType<DType>::kFlag, DataType, AType, {
       typedef typename std::conditional<safe_acc, AType, DataType>::type AccType;
       MSHADOW_TYPE_SWITCH_WITH_BOOL(small.type_flag_, OType, {
@@ -252,17 +250,6 @@ void Reduce(Stream<cpu>* s, const TBlob& small, const OpReqType req,
           big.shape_.get<ndim>(), small.shape_.get<ndim>(), rshape, rstride);
       });
     });
-#else
-    MXNET_REAL_ACC_TYPE_SWITCH(mshadow::DataType<DType>::kFlag, DataType, AType, {
-      typedef typename std::conditional<safe_acc, AType, DataType>::type AccType;
-      MSHADOW_TYPE_SWITCH_WITH_BOOL(small.type_flag_, OType, {
-        typedef typename std::conditional<safe_acc, OType, DataType>::type OutType;
-        seq_reduce_compute<Reducer, ndim, AccType, DataType, OutType, OP>(
-          N, M, req == kAddTo, big.dptr<DataType>(), small.dptr<OutType>(),
-          big.shape_.get<ndim>(), small.shape_.get<ndim>(), rshape, rstride);
-      });
-    });
-#endif
   }
 }
 

tests/python/unittest/test_numpy_op.py

@@ -614,6 +614,7 @@ def hybrid_forward(self, F, a, *args, **kwargs):
     def is_int(dtype):
         return 'int' in dtype
 
+    is_windows = sys.platform.startswith('win')
     in_data_dim = random.choice([2, 3, 4])
     shape = rand_shape_nd(in_data_dim, dim=3)
     acc_type = {'float16': 'float32', 'float32': 'float64', 'float64': 'float64',
@@ -667,11 +668,20 @@ def is_int(dtype):
                         test_mean.hybridize()
 
                     if itype == 'bool':
-                        x = np.random.uniform(size=shape) > 0.5
+                        x = np.array(_np.random.uniform(size=shape) > 0.5)
                     else:
                         x = np.random.uniform(-128, 127, size=shape).astype(itype)
 
                     expected_ret = _np.mean(x.asnumpy(), axis=axis, dtype=dtype, keepdims=keepdims)
+
+                    if itype == 'bool':
+                        if is_op_runnable() and (not is_windows) and dtype not in ['float16', 'int8']:  # special handling of boolean ndarray
+                            y = test_mean(x)
+                            assert y.shape == expected_ret.shape
+                            assert_almost_equal(y.asnumpy(), expected_ret, rtol=1e-3 if dtype == 'float16' else 1e-3,
+                                                atol=1e-5 if dtype == 'float16' else 1e-5)
+                        continue
+
                     y = test_mean(x)
                     assert y.shape == expected_ret.shape
                     assert_almost_equal(y.asnumpy(), expected_ret, rtol=1e-3 if dtype == 'float16' else 1e-3,