Fix binary broadcast with zero-size tensors

src/operator/tensor/elemwise_binary_broadcast_op.h

@@ -61,16 +61,13 @@ inline bool BinaryBroadcastShape(const nnvm::NodeAttrs& attrs,
     int l = 1, r = 1;
     if (i >= bl) l = lhs[i-bl];
     if (i >= br) r = rhs[i-br];
+    if (!mxnet::dim_size_is_known(l) || !mxnet::dim_size_is_known(r)) continue;
     if (l != r) {
-      if (l == 0 || r == 0) {
-        // TODO(junwu): here is not compatible with NumPy.
-        // For example, (2, 3) cannot broadcast to (2, 0, 3).
-        out[i] = 0;
-      } else {
-        CHECK(l == 1 || r == 1)
-          << "operands could not be broadcast together with shapes " << lhs << " " << rhs;
-        out[i] = std::max(l, r);
-      }
+      // Make it compatible with NumPy.
+      // For example, (2, 3) cannot broadcast to (2, 0, 3), but (1, 3) can broadcast to (2, 0, 3).
+      CHECK(l == 1 || r == 1)
+        << "operands could not be broadcast together with shapes " << lhs << " " << rhs;
+      out[i] = (l == 1 ? r : l);
     } else {
       out[i] = l;
     }

tests/python/gpu/test_operator_gpu.py

@@ -1963,19 +1963,21 @@ def check_proposal_consistency(op, batch_size, with_nms=False):
 # The following 2 functions launch 0-thread kernels, an error that should be caught and signaled.
 def kernel_error_check_imperative():
     os.environ['MXNET_ENGINE_TYPE'] = 'NaiveEngine'
-    a = mx.nd.array([1,2,3],ctx=mx.gpu(0))
-    b = mx.nd.array([],ctx=mx.gpu(0))
-    c = (a / b).asnumpy()
+    with mx.enable_np_comp():
+        a = mx.nd.array([1,2,3],ctx=mx.gpu(0))
+        b = mx.nd.array([],ctx=mx.gpu(0))
+        c = (a / b).asnumpy()
 
 def kernel_error_check_symbolic():
     os.environ['MXNET_ENGINE_TYPE'] = 'NaiveEngine'
-    a = mx.sym.Variable('a')
-    b = mx.sym.Variable('b')
-    c = a / b
-    f = c.bind(mx.gpu(0), { 'a':mx.nd.array([1,2,3],ctx=mx.gpu(0)),
-                            'b':mx.nd.array([],ctx=mx.gpu(0))})
-    f.forward()
-    g = f.outputs[0].asnumpy()
+    with mx.enable_np_comp():
+        a = mx.sym.Variable('a')
+        b = mx.sym.Variable('b')
+        c = a / b
+        f = c.bind(mx.gpu(0), { 'a':mx.nd.array([1,2,3],ctx=mx.gpu(0)),
+                                'b':mx.nd.array([],ctx=mx.gpu(0))})
+        f.forward()
+        g = f.outputs[0].asnumpy()
 
 def test_kernel_error_checking():
     # Running tests that may throw exceptions out of worker threads will stop CI testing