!8006 [MS][GPU][CUDA][API] Adding new Ops - UnsortedSegmentMax GPU

Merge pull request !8006 from danishnxt/SegMax

mindspore/ccsrc/backend/kernel_compiler/gpu/arrays/unsorted_segment_max_gpu_kernel.cc

@@ -0,0 +1,36 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "backend/kernel_compiler/gpu/arrays/unsorted_segment_max_gpu_kernel.h"
+
+namespace mindspore {
+namespace kernel {
+MS_REG_GPU_KERNEL_ONE(
+  UnsortedSegmentMax,
+  KernelAttr().AddInputAttr(kNumberTypeFloat32).AddInputAttr(kNumberTypeInt32).AddOutputAttr(kNumberTypeFloat32),
+  UnsortedSegmentMaxGpuKernel, float)
+
+MS_REG_GPU_KERNEL_ONE(
+  UnsortedSegmentMax,
+  KernelAttr().AddInputAttr(kNumberTypeFloat16).AddInputAttr(kNumberTypeInt32).AddOutputAttr(kNumberTypeFloat16),
+  UnsortedSegmentMaxGpuKernel, half)
+
+MS_REG_GPU_KERNEL_ONE(
+  UnsortedSegmentMax,
+  KernelAttr().AddInputAttr(kNumberTypeInt32).AddInputAttr(kNumberTypeInt32).AddOutputAttr(kNumberTypeInt32),
+  UnsortedSegmentMaxGpuKernel, int)
+}  // namespace kernel
+}  // namespace mindspore

mindspore/ccsrc/backend/kernel_compiler/gpu/arrays/unsorted_segment_max_gpu_kernel.h

@@ -0,0 +1,123 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_UNSORTED_SEGMENT_MAX_H_
+#define MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_UNSORTED_SEGMENT_MAX_H_
+
+#include <vector>
+#include <limits>
+#include "backend/kernel_compiler/gpu/gpu_kernel.h"
+#include "backend/kernel_compiler/gpu/gpu_kernel_factory.h"
+#include "backend/kernel_compiler/gpu/cuda_impl/unsorted_segment_max.cuh"
+
+namespace mindspore {
+namespace kernel {
+template <typename T>
+class UnsortedSegmentMaxGpuKernel : public GpuKernel {
+ public:
+  UnsortedSegmentMaxGpuKernel()
+      : num_segments_(1),
+        inner_size_(1),
+        outer_size_(1),
+        input_size_(1),
+        segment_ids_size_(1),
+        output_size_(1),
+        is_null_input_(false) {}
+  ~UnsortedSegmentMaxGpuKernel() override = default;
+
+  const std::vector<size_t> &GetInputSizeList() const override { return input_size_list_; }
+  const std::vector<size_t> &GetOutputSizeList() const override { return output_size_list_; }
+  const std::vector<size_t> &GetWorkspaceSizeList() const override { return workspace_size_list_; }
+
+  bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &,
+              const std::vector<AddressPtr> &outputs, void *stream_ptr) override {
+    if (is_null_input_) {
+      return true;
+    }
+    T *input_addr = GetDeviceAddress<T>(inputs, 0);
+    int *indices_addr = GetDeviceAddress<int>(inputs, 1);
+    T *output_addr = GetDeviceAddress<T>(outputs, 0);
+
+    CHECK_CUDA_RET_WITH_EXCEPT(cudaMemsetAsync(output_addr, std::numeric_limits<T>::min(), outputs[0]->size,
+                                               reinterpret_cast<cudaStream_t>(stream_ptr)),
+                               "cudaMemSet Failed");
+    CalUnsortedSegmentMax(input_addr, indices_addr, num_segments_, outer_size_, inner_size_, output_addr,
+                          reinterpret_cast<cudaStream_t>(stream_ptr));
+    return true;
+  }
+
+  bool Init(const CNodePtr &kernel_node) override {
+    auto input_shapes = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 0);
+    is_null_input_ = CHECK_NULL_INPUT(input_shapes);
+    if (is_null_input_) {
+      MS_LOG(WARNING) << "UnsortedSegmentMax input is null";
+      InitSizeLists();
+      return true;
+    }
+    auto segment_ids_shapes = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 1);
+    auto output_shapes = AnfAlgo::GetOutputInferShape(kernel_node, 0);  // we get that from computation
+
+    num_segments_ = output_shapes[0];
+
+    input_size_ = 1;
+    for (size_t i = 0; i < input_shapes.size(); i++) {
+      input_size_ *= input_shapes[i];
+    }
+
+    segment_ids_size_ = 1;
+    for (size_t i = 0; i < segment_ids_shapes.size(); i++) {
+      segment_ids_size_ *= segment_ids_shapes[i];
+    }
+
+    output_size_ = 1;
+    for (size_t i = 0; i < output_shapes.size(); i++) {
+      output_size_ *= output_shapes[i];
+    }
+
+    outer_size_ = input_shapes[0];
+    inner_size_ = 1;
+    for (size_t i = 1; i < input_shapes.size(); i++) {
+      inner_size_ *= input_shapes[i];
+    }
+
+    InitSizeLists();
+    return true;
+  }
+
+ protected:
+  void InitSizeLists() override {
+    input_size_list_.push_back(input_size_ * sizeof(T));
+    input_size_list_.push_back(segment_ids_size_ * sizeof(int));
+    output_size_list_.push_back(output_size_ * sizeof(T));
+  }
+
+ private:
+  int num_segments_;
+  size_t inner_size_;
+  size_t outer_size_;
+  size_t input_size_;
+  size_t segment_ids_size_;
+  size_t output_size_;
+  bool is_null_input_;
+
+  std::vector<size_t> input_size_list_;
+  std::vector<size_t> output_size_list_;
+  std::vector<size_t> workspace_size_list_;
+};
+}  // namespace kernel
+}  // namespace mindspore
+
+#endif  // MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_UNSORTED_SEGMENT_MAX_H_

mindspore/ccsrc/backend/kernel_compiler/gpu/cuda_impl/unsorted_segment_max.cu

@@ -0,0 +1,79 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "backend/kernel_compiler/gpu/cuda_impl/unsorted_segment_max.cuh"
+#include <limits>
+
+template <typename T>
+__global__ void UnsortedSegmentMax(const T *input, const int *segment_ids, const int num_segments, size_t outer_size,
+                                   size_t inner_size, bool fp16_flag, T init_K, T *output) {
+  if (fp16_flag) {
+    init_K = __int2half_rd(-65504);  // min value representable by float16
+  }
+
+  for (int t_idx = blockIdx.x * blockDim.x + threadIdx.x; t_idx < KWARPSIZE * num_segments * inner_size;
+       t_idx += blockDim.x * gridDim.x) {
+    int segment_id = t_idx / KWARPSIZE / inner_size;
+    int inner_id = t_idx / KWARPSIZE % inner_size;
+    int lane_id = threadIdx.x % KWARPSIZE;
+    T threadK = init_K;
+
+    for (int i = lane_id; i < outer_size; i += KWARPSIZE) {
+      if (segment_ids[i] != segment_id) continue;
+      T other_K = input[i * inner_size + inner_id];
+      if (threadK < other_K) {
+        threadK = other_K;
+      }
+    }
+    __syncwarp();
+
+    for (int offset = KWARPSIZE / 2; offset > 0; offset /= 2) {
+      T other_K = __shfl_down_sync(0xffffffff, threadK, offset);
+      if (threadK < other_K) {
+        threadK = other_K;
+      }
+    }
+
+    __syncwarp();
+
+    if (lane_id == 0) {
+      output[segment_id * inner_size + inner_id] = threadK;
+    }
+    __syncthreads();
+  }
+}
+
+template <typename T>
+void CalUnsortedSegmentMax(const T *input, const int *segment_ids, const int num_segments, size_t outer_size,
+                           size_t inner_size, T *output, cudaStream_t stream) {
+  int size = (inner_size * KWARPSIZE * num_segments);
+  bool fp16_flag = false;
+  // handle fp16 min value
+  if (std::is_same<T, half>::value) {
+    fp16_flag = true;
+  }
+  T init_K = std::numeric_limits<T>::lowest();
+  UnsortedSegmentMax<<<GET_BLOCKS(size), GET_THREADS, 0, stream>>>(input, segment_ids, num_segments, outer_size,
+                                                                   inner_size, fp16_flag, init_K, output);
+  return;
+}
+
+template void CalUnsortedSegmentMax<float>(const float *input, const int *segment_ids, const int num_segments,
+                                           size_t outer_size, size_t inner_size, float *output, cudaStream_t stream);
+template void CalUnsortedSegmentMax<half>(const half *input, const int *segment_ids, const int num_segments,
+                                          size_t outer_size, size_t inner_size, half *output, cudaStream_t stream);
+template void CalUnsortedSegmentMax<int>(const int *input, const int *segment_ids, const int num_segments,
+                                         size_t outer_size, size_t inner_size, int *output, cudaStream_t stream);

mindspore/ccsrc/backend/kernel_compiler/gpu/cuda_impl/unsorted_segment_max.cuh

@@ -0,0 +1,30 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_CCSRC_KERNEL_GPU_CUDA_IMPL_UNSORTED_SEGMENT_MAX_H_
+#define MINDSPORE_CCSRC_KERNEL_GPU_CUDA_IMPL_UNSORTED_SEGMENT_MAX_H_
+
+#include <cuda_runtime.h>
+#include "runtime/device/gpu/cuda_common.h"
+
+// Setting warp size to sync data across threads
+#define KWARPSIZE 32
+
+template <typename T>
+void CalUnsortedSegmentMax(const T *input, const int *segment_ids, const int num_segments, size_t outer_size,
+                           size_t inner_size, T *output, cudaStream_t stream);
+
+#endif  // MINDSPORE_CCSRC_KERNEL_GPU_CUDA_IMPL_UNSORT_SEGMENT_MAX_H_

mindspore/ccsrc/backend/optimizer/pass/const_input_to_attr_registry.cc

@@ -49,6 +49,7 @@ ConstInputToAttrInfoRegistry::ConstInputToAttrInfoRegistry() {
   Register(prim::kPrimReduceAll->name(), {1});
   Register(prim::kPrimReduceAny->name(), {1});
   Register(prim::kPrimUnsortedSegmentMin->name(), {2});
+  Register(prim::kPrimUnsortedSegmentMax->name(), {2});
   Register(kSparseGatherV2, {2});
   Register(kUnsortedSegmentProdOpName, {2});
   Register(kSimpleMeanGradOpName, {1});

mindspore/core/base/core_ops.h

@@ -92,6 +92,7 @@ inline const PrimitivePtr kPrimEmbeddingLookupCommGrad = std::make_shared<Primit
 inline const PrimitivePtr kPrimSize = std::make_shared<Primitive>("Size");
 inline const PrimitivePtr kPrimArgMax = std::make_shared<Primitive>("Argmax");
 inline const PrimitivePtr kPrimPack = std::make_shared<Primitive>("Pack");
+inline const PrimitivePtr kPrimUnsortedSegmentMax = std::make_shared<Primitive>("UnsortedSegmentMax");
 inline const PrimitivePtr kPrimUnsortedSegmentSum = std::make_shared<Primitive>("UnsortedSegmentSum");
 inline const PrimitivePtr kPrimUnsortedSegmentMin = std::make_shared<Primitive>("UnsortedSegmentMin");
 inline const PrimitivePtr kPrimConcatOffset = std::make_shared<Primitive>("ConcatOffset");

mindspore/ops/operations/__init__.py

@@ -30,8 +30,8 @@ from .array_ops import (Argmax, Argmin, Cast, Concat, Pack, Unpack,
                         Shape, DynamicShape, Size, Slice, Split, TransShape, ParallelConcat, Padding, UniqueWithPad,
                         ScatterNdAdd, ScatterNdSub, ScatterNonAliasingAdd, ReverseV2, Rint,
                         Squeeze, StridedSlice, Tile, TensorScatterUpdate, EditDistance, Sort,
-                        Transpose, TruncatedNormal, TupleToArray, UnsortedSegmentMin, UnsortedSegmentProd,
-                        UnsortedSegmentSum, SpaceToDepth, DepthToSpace, SpaceToBatch, BatchToSpace,
+                        Transpose, TruncatedNormal, TupleToArray, UnsortedSegmentMin, UnsortedSegmentMax,
+                        UnsortedSegmentProd, UnsortedSegmentSum, SpaceToDepth, DepthToSpace, SpaceToBatch, BatchToSpace,
                         SpaceToBatchND, BatchToSpaceND, BroadcastTo, InplaceUpdate, ReverseSequence, EmbeddingLookup,
                         Unique, GatherD, Identity, RepeatElements)
 from .comm_ops import (AllGather, AllReduce, _AlltoAll, ReduceScatter, Broadcast,

mindspore/ops/operations/array_ops.py

@@ -1640,6 +1640,56 @@ class UnsortedSegmentMin(PrimitiveWithInfer):
         return out
 
 
+class UnsortedSegmentMax(PrimitiveWithInfer):
+    """
+    Computes the maximum along segments of a tensor.
+
+    Inputs:
+        - **input_x** (Tensor) - The shape is :math:`(x_1, x_2, ..., x_R)`.
+          The data type must be float16, float32 or int32.
+        - **segment_ids** (Tensor) - A `1-D` tensor whose shape is :math:`(x_1)`, the value must be >= 0.
+          The data type must be int32.
+        - **num_segments** (int) - The value spcifies the number of distinct `segment_ids`.
+
+    Outputs:
+        Tensor, set the number of `num_segments` as `N`, the shape is :math:`(N, x_2, ..., x_R)`.
+
+    Examples:
+        >>> input_x = Tensor(np.array([[1, 2, 3], [4, 5, 6], [4, 2, 1]]).astype(np.float32))
+        >>> segment_ids = Tensor(np.array([0, 1, 1]).astype(np.int32))
+        >>> num_segments = 2
+        >>> unsorted_segment_max = P.UnsortedSegmentMax()
+        >>> unsorted_segment_max(input_x, segment_ids, num_segments)
+        [[1., 2., 3.], [4., 5., 6.]]
+    """
+
+    @prim_attr_register
+    def __init__(self):
+        """Initialize UnsortedSegmentMax"""
+        self.init_prim_io_names(inputs=['x', 'segment_ids', 'num_segments'], outputs=['y'])
+
+    def __infer__(self, x, segment_ids, num_segments):
+        x_type = x['dtype']
+        x_shape = x['shape']
+        segment_ids_shape = segment_ids['shape']
+        valid_type = [mstype.float16, mstype.float32, mstype.int32]
+        validator.check_tensor_type_same({"x": x['dtype']}, valid_type, self.name)
+        validator.check_tensor_type_same({"segment_ids": segment_ids['dtype']}, [mstype.int32], self.name)
+        validator.check_equal_int(len(segment_ids_shape), 1, "rank of segment_ids_shape", self.name)
+        validator.check(f'first shape of input_x', x_shape[0],
+                        'length of segments_id', segment_ids_shape[0], Rel.EQ, self.name)
+        num_segments_v = num_segments['value']
+        validator.check_value_type('num_segments', num_segments_v, [int], self.name)
+        validator.check_positive_int(num_segments_v, "num_segments", self.name)
+        segment_ids_shape_len = len(segment_ids_shape)
+        out_shape = [num_segments_v]
+        out_shape += x_shape[segment_ids_shape_len:]
+        out = {'shape': out_shape,
+               'dtype': x_type,
+               'value': None}
+        return out
+
+
 class UnsortedSegmentProd(PrimitiveWithInfer):
     """
     Computes the product along segments of a tensor.

tests/st/ops/gpu/test_unsorted_segment_max.py

@@ -0,0 +1,204 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+import numpy as np
+import pytest
+
+import mindspore
+import mindspore.context as context
+import mindspore.nn as nn
+from mindspore import Tensor
+from mindspore.common import dtype as mstype
+from mindspore.ops import operations as P
+
+
+class UnsortedSegmentMaxNet(nn.Cell):
+    def __init__(self, num_segments):
+        super(UnsortedSegmentMaxNet, self).__init__()
+        self.unsorted_segment_max = P.UnsortedSegmentMax()
+        self.num_segments = num_segments
+
+    def construct(self, data, ids):
+        return self.unsorted_segment_max(data, ids, self.num_segments)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_1d_int32():
+    context.set_context(mode=context.PYNATIVE_MODE, device_target='GPU')
+    input_x = Tensor([1, 2, 3, 4], mstype.int32)
+    segment_ids = Tensor([0, 0, 1, 2], mstype.int32)
+    num_segments = 4
+    net = UnsortedSegmentMaxNet(num_segments)
+    output = net(input_x, segment_ids)
+    expect = [2, 3, 4, -2147483648]
+    assert (output.asnumpy() == expect).all()
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_2d_int32():
+    context.set_context(mode=context.GRAPH_MODE, device_target='GPU')
+    input_x = Tensor([[1, 2, 3, 4],
+                      [5, 6, 7, 8],
+                      [9, 10, 11, 12]], mstype.int32)
+    segment_ids = Tensor([2, 1, 1], mstype.int32)
+    num_segments = 4
+    net = UnsortedSegmentMaxNet(num_segments)
+    output = net(input_x, segment_ids)
+    expect = [[[-2147483648, -2147483648, -2147483648, -2147483648],
+               [9, 10, 11, 12],
+               [1, 2, 3, 4],
+               [-2147483648, -2147483648, -2147483648, -2147483648]]]
+    assert (output.asnumpy() == expect).all()
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_3d_float16():
+    context.set_context(mode=context.PYNATIVE_MODE, device_target='GPU')
+    input_x = Tensor(np.arange(
+        4 * 5 * 3, dtype=np.float16).reshape(4, 5, 3), dtype=mindspore.float16)
+    segment_ids = Tensor([2, 1, 1, -1], mstype.int32)
+    num_segments = 5
+    net = UnsortedSegmentMaxNet(num_segments)
+    output = net(input_x, segment_ids).asnumpy()
+    expect = np.array([[[-6.55e+04, -6.55e+04, -6.55e+04],
+                        [-6.55e+04, -6.55e+04, -6.55e+04],
+                        [-6.55e+04, -6.55e+04, -6.55e+04],
+                        [-6.55e+04, -6.55e+04, -6.55e+04],
+                        [-6.55e+04, -6.55e+04, -6.55e+04]],
+                       [[3.00e+01, 3.10e+01, 3.20e+01],
+                        [3.30e+01, 3.40e+01, 3.50e+01],
+                        [3.60e+01, 3.70e+01, 3.80e+01],
+                        [3.90e+01, 4.00e+01, 4.10e+01],
+                        [4.20e+01, 4.30e+01, 4.40e+01]],
+                       [[0.00e+00, 1.00e+00, 2.00e+00],
+                        [3.00e+00, 4.00e+00, 5.00e+00],
+                        [6.00e+00, 7.00e+00, 8.00e+00],
+                        [9.00e+00, 1.00e+01, 1.10e+01],
+                        [1.20e+01, 1.30e+01, 1.40e+01]],
+                       [[-6.55e+04, -6.55e+04, -6.55e+04],
+                        [-6.55e+04, -6.55e+04, -6.55e+04],
+                        [-6.55e+04, -6.55e+04, -6.55e+04],
+                        [-6.55e+04, -6.55e+04, -6.55e+04],
+                        [-6.55e+04, -6.55e+04, -6.55e+04]],
+                       [[-6.55e+04, -6.55e+04, -6.55e+04],
+                        [-6.55e+04, -6.55e+04, -6.55e+04],
+                        [-6.55e+04, -6.55e+04, -6.55e+04],
+                        [-6.55e+04, -6.55e+04, -6.55e+04],
+                        [-6.55e+04, -6.55e+04, -6.55e+04]]]).astype(np.float16)
+    np.testing.assert_array_almost_equal(output, expect)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_3d_float32():
+    context.set_context(mode=context.GRAPH_MODE, device_target='GPU')
+    input_x = Tensor(np.arange(
+        4 * 5 * 3, dtype=np.float32).reshape(4, 5, 3), dtype=mindspore.float32)
+    segment_ids = Tensor([2, 1, 1, -1], mstype.int32)
+    num_segments = 3
+    net = UnsortedSegmentMaxNet(num_segments)
+    output = net(input_x, segment_ids).asnumpy()
+    expect = np.array([[[-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38]],
+                       [[3.0000000e+01, 3.1000000e+01, 3.2000000e+01],
+                        [3.3000000e+01, 3.4000000e+01, 3.5000000e+01],
+                        [3.6000000e+01, 3.7000000e+01, 3.8000000e+01],
+                        [3.9000000e+01, 4.0000000e+01, 4.1000000e+01],
+                        [4.2000000e+01, 4.3000000e+01, 4.4000000e+01]],
+                       [[0.0000000e+00, 1.0000000e+00, 2.0000000e+00],
+                        [3.0000000e+00, 4.0000000e+00, 5.0000000e+00],
+                        [6.0000000e+00, 7.0000000e+00, 8.0000000e+00],
+                        [9.0000000e+00, 1.0000000e+01, 1.1000000e+01],
+                        [1.2000000e+01, 1.3000000e+01, 1.4000000e+01]]]).astype(np.float32)
+    np.testing.assert_array_almost_equal(output, expect)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_3d_single_init():
+    context.set_context(device_target='GPU')
+    input_x = Tensor(np.arange(
+        4 * 5 * 3, dtype=np.float32).reshape(4, 5, 3), dtype=mindspore.float32)
+    segment_ids = Tensor([3, 0, 1, -1], mstype.int32)
+    net = P.UnsortedSegmentMax()
+
+    num_segments = 4
+    output = net(input_x, segment_ids, num_segments).asnumpy()
+    expect = np.array([[[1.5000000e+01, 1.6000000e+01, 1.7000000e+01],
+                        [1.8000000e+01, 1.9000000e+01, 2.0000000e+01],
+                        [2.1000000e+01, 2.2000000e+01, 2.3000000e+01],
+                        [2.4000000e+01, 2.5000000e+01, 2.6000000e+01],
+                        [2.7000000e+01, 2.8000000e+01, 2.9000000e+01]],
+                       [[3.0000000e+01, 3.1000000e+01, 3.2000000e+01],
+                        [3.3000000e+01, 3.4000000e+01, 3.5000000e+01],
+                        [3.6000000e+01, 3.7000000e+01, 3.8000000e+01],
+                        [3.9000000e+01, 4.0000000e+01, 4.1000000e+01],
+                        [4.2000000e+01, 4.3000000e+01, 4.4000000e+01]],
+                       [[-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38]],
+                       [[0.0000000e+00, 1.0000000e+00, 2.0000000e+00],
+                        [3.0000000e+00, 4.0000000e+00, 5.0000000e+00],
+                        [6.0000000e+00, 7.0000000e+00, 8.0000000e+00],
+                        [9.0000000e+00, 1.0000000e+01, 1.1000000e+01],
+                        [1.2000000e+01, 1.3000000e+01, 1.4000000e+01]]]).astype(np.float32)
+    np.testing.assert_array_almost_equal(output, expect)
+
+    num_segments = 6
+    output = net(input_x, segment_ids, num_segments).asnumpy()
+    expect = np.array([[[1.5000000e+01, 1.6000000e+01, 1.7000000e+01],
+                        [1.8000000e+01, 1.9000000e+01, 2.0000000e+01],
+                        [2.1000000e+01, 2.2000000e+01, 2.3000000e+01],
+                        [2.4000000e+01, 2.5000000e+01, 2.6000000e+01],
+                        [2.7000000e+01, 2.8000000e+01, 2.9000000e+01]],
+                       [[3.0000000e+01, 3.1000000e+01, 3.2000000e+01],
+                        [3.3000000e+01, 3.4000000e+01, 3.5000000e+01],
+                        [3.6000000e+01, 3.7000000e+01, 3.8000000e+01],
+                        [3.9000000e+01, 4.0000000e+01, 4.1000000e+01],
+                        [4.2000000e+01, 4.3000000e+01, 4.4000000e+01]],
+                       [[-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38]],
+                       [[0.0000000e+00, 1.0000000e+00, 2.0000000e+00],
+                        [3.0000000e+00, 4.0000000e+00, 5.0000000e+00],
+                        [6.0000000e+00, 7.0000000e+00, 8.0000000e+00],
+                        [9.0000000e+00, 1.0000000e+01, 1.1000000e+01],
+                        [1.2000000e+01, 1.3000000e+01, 1.4000000e+01]],
+                       [[-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38]],
+                       [[-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38],
+                        [-3.4028235e+38, -3.4028235e+38, -3.4028235e+38]]]).astype(np.float32)
+    np.testing.assert_array_almost_equal(output, expect)