[GitHub] piiswrong closed pull request #8398: fix topk with big shape

[GitBox <gi...@apache.org>]

piiswrong closed pull request #8398: fix topk with big shape
URL: https://github.com/apache/incubator-mxnet/pull/8398
 
 
   

This is a PR merged from a forked repository.
As GitHub hides the original diff on merge, it is displayed below for
the sake of provenance:

As this is a foreign pull request (from a fork), the diff is supplied
below (as it won't show otherwise due to GitHub magic):

diff --git a/src/operator/tensor/indexing_op.h b/src/operator/tensor/indexing_op.h
index 1c731088ee..4eb0e46908 100644
--- a/src/operator/tensor/indexing_op.h
+++ b/src/operator/tensor/indexing_op.h
@@ -277,7 +277,8 @@ void AddTakeGradLargeBatchCaller(const OpContext& ctx, mshadow::Tensor<xpu, 2, D
   Tensor<xpu, 1, char> temp_storage(&workspace[pos], Shape1(temp_storage_size), s);
   Kernel<tcast_clip, xpu>::Launch(s, index.shape_.Size(), sorted_data.dptr_, index.dptr_,
     static_cast<int>(dst.shape_[0]));
-  original_index = range<int>(0, index.shape_.Size());
+  Kernel<range_fwd, xpu>::Launch(s, index.shape_.Size(),
+    1, 0, 1, kWriteTo, original_index.dptr_);
   int num_bits = ilog2((dst.shape_[0] - 1));
   mxnet::op::SortByKey(sorted_data, original_index, true, &temp_storage, 0, num_bits);
   mxnet::op::AddTakeGradLargeBatch(dst, sorted_data, original_index, src, &temp_storage);
diff --git a/src/operator/tensor/ordering_op-inl.h b/src/operator/tensor/ordering_op-inl.h
index 560554151a..b491b64cc9 100644
--- a/src/operator/tensor/ordering_op-inl.h
+++ b/src/operator/tensor/ordering_op-inl.h
@@ -213,7 +213,9 @@ void TopKImpl(RunContext ctx,
   } else {
     sorted_dat = reshape(dat, Shape1(src.Size()));
   }
-  indices = range<int>(0, batch_size * element_num);
+  mxnet_op::Kernel<range_fwd, xpu>::Launch(s, batch_size * element_num, 1, 0, 1,
+    kWriteTo, indices.dptr_);
+
   CHECK_EQ(sorted_dat.CheckContiguous(), true);
   CHECK_EQ(indices.CheckContiguous(), true);
   if (param.ret_typ == topk_enum::kReturnMask) {
@@ -383,7 +385,8 @@ void TopKBackward_(const nnvm::NodeAttrs& attrs,
     inputs[0].get_with_shape<xpu, 2, real_t>(Shape2(inputs[0].shape_.Size(), 1), s);
   Tensor<xpu, 2, real_t> in_grad =
     outputs[0].get_with_shape<xpu, 2, real_t>(Shape2(outputs[0].shape_.Size(), 1), s);
-  batch_shift = range<real_t>(0, batch_size, 1) * element_num;
+  mxnet_op::Kernel<range_fwd, xpu>::Launch(s, batch_size, 1, 0.0f,
+    static_cast<real_t>(element_num), kWriteTo, batch_shift.dptr_);
   if (do_transpose) {
     Tensor<xpu, 1, real_t> indices = inputs[2].FlatTo1D<xpu, real_t>(s);
     TShape src_shape = outputs[0].shape_.FlatTo3D(axis);
@@ -411,7 +414,8 @@ void TopKBackward_(const nnvm::NodeAttrs& attrs,
   } else if (kAddTo == req[0]) {
     // TODO(sxjscience) We can use AddTakeGrad in the future.
     // However, the current implementation of AddTakeGrad is not so efficient.
-    dummy_index = range<real_t>(0, sel_indices.shape_.Size());
+    mxnet_op::Kernel<range_fwd, xpu>::Launch(s, sel_indices.shape_.Size(), 1, 0.0f,
+      1.0f, kWriteTo, dummy_index.dptr_);
     mxnet::op::AddTakeGradLargeBatch(in_grad, sel_indices, dummy_index, out_grad);
   } else if (kNullOp == req[0]) {
     return;
diff --git a/tests/python/unittest/test_ndarray.py b/tests/python/unittest/test_ndarray.py
index fc8c350bbb..28aab7a4cb 100644
--- a/tests/python/unittest/test_ndarray.py
+++ b/tests/python/unittest/test_ndarray.py
@@ -593,6 +593,10 @@ def gt_topk(dat, axis, ret_typ, k, is_ascend):
     nd_ret_argsort = mx.nd.argsort(a_nd, axis=None, is_ascend=False).asnumpy()
     gt = gt_topk(a_npy, axis=None, ret_typ="indices", k=5*5*5*5, is_ascend=False)
     assert_almost_equal(nd_ret_argsort, gt)
+    
+    # test topk with a big shape
+    a = mx.nd.arange(0, 54686454, step=1, repeat=1)
+    assert_almost_equal(a.topk(k=54686454).asnumpy(), a.asnumpy()[::-1])
 
 def test_ndarray_equal():
     x = mx.nd.zeros((2, 3))


 

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services