!8861 [MS][GPU][DynamicShapeUpdate] Converting UnsortedSegmentMax from normal to dynamic shape op

From: @danishnxt
Reviewed-by: 
Signed-off-by:

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

@@ -22,15 +22,35 @@ 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)
+// Dynamic Mode
+MS_REG_GPU_KERNEL_ONE(UnsortedSegmentMax,
+                      KernelAttr()
+                        .AddInputAttr(kNumberTypeFloat32)
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddInputAttr(kNumberTypeInt64)
+                        .AddOutputAttr(kNumberTypeFloat32),
+                      UnsortedSegmentMaxGpuKernel, float)
+MS_REG_GPU_KERNEL_ONE(UnsortedSegmentMax,
+                      KernelAttr()
+                        .AddInputAttr(kNumberTypeFloat16)
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddInputAttr(kNumberTypeInt64)
+                        .AddOutputAttr(kNumberTypeFloat16),
+                      UnsortedSegmentMaxGpuKernel, half)
+MS_REG_GPU_KERNEL_ONE(UnsortedSegmentMax,
+                      KernelAttr()
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddInputAttr(kNumberTypeInt64)
+                        .AddOutputAttr(kNumberTypeInt32),
+                      UnsortedSegmentMaxGpuKernel, int)
 }  // namespace kernel
 }  // namespace mindspore

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

@@ -28,14 +28,7 @@ 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() { ResetResource(); }
   ~UnsortedSegmentMaxGpuKernel() override = default;
 
   const std::vector<size_t> &GetInputSizeList() const override { return input_size_list_; }
@@ -60,18 +53,24 @@ class UnsortedSegmentMaxGpuKernel : public GpuKernel {
   }
 
   bool Init(const CNodePtr &kernel_node) override {
-    auto input_shapes = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 0);
+    auto input_shapes = AnfAlgo::GetInputRealDeviceShapeIfExist(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
+    auto segment_ids_shapes = AnfAlgo::GetInputRealDeviceShapeIfExist(kernel_node, 1);
+    auto output_shapes = AnfAlgo::GetOutputRealDeviceShapeIfExist(kernel_node, 0);
+
+    size_t input_num = AnfAlgo::GetInputTensorNum(kernel_node);
+    if (input_num == 3) {
+      MS_LOG(INFO) << "UnsortedSegmentMax Kernel Input count is 3 - dynamic mode";
+    } else {
+      MS_LOG(INFO) << "UnsortedSegmentMax Kernel Input count is 2";
+    }
 
     num_segments_ = output_shapes[0];
-
     input_size_ = 1;
     for (size_t i = 0; i < input_shapes.size(); i++) {
       input_size_ *= input_shapes[i];
@@ -97,6 +96,19 @@ class UnsortedSegmentMaxGpuKernel : public GpuKernel {
     return true;
   }
 
+  void ResetResource() noexcept override {
+    num_segments_ = 1;
+    inner_size_ = 1;
+    outer_size_ = 1;
+    input_size_ = 1;
+    segment_ids_size_ = 1;
+    output_size_ = 1;
+    is_null_input_ = false;
+    input_size_list_.clear();
+    output_size_list_.clear();
+    workspace_size_list_.clear();
+  }
+
  protected:
   void InitSizeLists() override {
     input_size_list_.push_back(input_size_ * sizeof(T));

mindspore/core/abstract/infer_functions.h

@@ -113,6 +113,8 @@ AbstractBasePtr InferImplMakeDict(const AnalysisEnginePtr &, const PrimitivePtr
                                   const AbstractBasePtrList &args_spec_list);
 AbstractBasePtr InferImplUnsortedSegmentSum(const AnalysisEnginePtr &, const PrimitivePtr &primitive,
                                             const AbstractBasePtrList &args_spec_list);
+AbstractBasePtr InferImplUnsortedSegmentMax(const AnalysisEnginePtr &, const PrimitivePtr &primitive,
+                                            const AbstractBasePtrList &args_spec_list);
 AbstractBasePtr InferImplMakeSlice(const AnalysisEnginePtr &, const PrimitivePtr &primitive,
                                    const AbstractBasePtrList &args_spec_list);
 AbstractBasePtr InferImplMakeKwarg(const AnalysisEnginePtr &, const PrimitivePtr &primitive,

mindspore/core/abstract/prim_arrays.cc

@@ -273,6 +273,74 @@ AbstractBasePtr InferImplUnsortedSegmentSum(const AnalysisEnginePtr &, const Pri
   return std::make_shared<AbstractTensor>(x->element(), std::make_shared<Shape>(shape, min_shape, max_shape));
 }
 
+AbstractBasePtr InferImplUnsortedSegmentMax(const AnalysisEnginePtr &, const PrimitivePtr &primitive,
+                                            const AbstractBasePtrList &args_spec_list) {
+  const std::string op_name = primitive->name();
+  CheckArgsSize(op_name, args_spec_list, 3);
+  auto x = CheckArg<AbstractTensor>(op_name, args_spec_list, 0);
+  MS_EXCEPTION_IF_NULL(x);
+  MS_EXCEPTION_IF_NULL(x->shape());
+  auto segment_ids = CheckArg<AbstractTensor>(op_name, args_spec_list, 1);
+  MS_EXCEPTION_IF_NULL(segment_ids);
+  MS_EXCEPTION_IF_NULL(segment_ids->shape());
+  auto segment_ids_shape = segment_ids->shape()->shape();
+  (void)CheckTensorDType(x, {kFloat16, kFloat32, kInt32}, "Input 0 (x) for UnsortedSegmentMax should be %s");
+  (void)CheckTensorDType(segment_ids, {kInt32}, "Input 1 (segment_ids) for UnsortedSegmentMax should be %s");
+  // check if dynamic shape
+  bool x_is_dyn = (!x->shape()->min_shape().empty() && !x->shape()->max_shape().empty());
+  bool ids_is_dyn = (!segment_ids->shape()->min_shape().empty() && !segment_ids->shape()->max_shape().empty());
+  bool op_is_dynamic = x_is_dyn && ids_is_dyn;
+  auto x_shape = x->shape()->shape();
+  ShapeVector shape;
+  int64_t num_segments_value = 0;
+  if (args_spec_list[2]->isa<AbstractTensor>()) {  // num_segments is Tensor
+    auto num_segments = args_spec_list[2]->cast<AbstractTensorPtr>();
+    MS_EXCEPTION_IF_NULL(num_segments);
+    auto num_segments_value_ptr = num_segments->BuildValue();
+    MS_EXCEPTION_IF_NULL(num_segments_value_ptr);
+    auto num_segments_tensor = num_segments_value_ptr->cast<tensor::TensorPtr>();
+    MS_EXCEPTION_IF_NULL(num_segments_tensor);
+    num_segments_value = *static_cast<int64_t *>(num_segments_tensor->data_c());
+  } else if (args_spec_list[2]->isa<AbstractScalar>()) {  // num_segments is Scalar
+    auto num_segments = CheckArg<AbstractScalar>(op_name, args_spec_list, 2);
+    num_segments_value = GetValue<int64_t>(num_segments->BuildValue());
+  } else {
+    MS_LOG(EXCEPTION) << "num_segments incorrect type in UnsortedSegmentMax";
+  }
+  if (num_segments_value <= 0) {
+    MS_LOG(EXCEPTION) << "num_segments must be > 0 in UnsortedSegmentMax";
+  }
+  shape.emplace_back(num_segments_value);
+  shape.insert(shape.end(), x_shape.begin() + segment_ids_shape.size(), x_shape.end());
+  if (!op_is_dynamic) {
+    if (x_shape[0] != segment_ids_shape[0]) {
+      MS_LOG(EXCEPTION) << "Length of segment_ids must match first value of x shape UnsortedSegmentMax";
+    }
+    return std::make_shared<AbstractTensor>(x->element(), std::make_shared<Shape>(shape));
+  }
+  // is dynamic
+  ShapeVector min_shape;
+  ShapeVector max_shape;
+  min_shape.emplace_back(num_segments_value);
+  max_shape.emplace_back(num_segments_value);
+  // only run validation if shape values are known
+  bool x_any_shape = std::any_of(x_shape.begin(), x_shape.end(), [](int64_t dim) { return dim == Shape::SHP_ANY; });
+  bool ids_any_shape =
+    std::any_of(segment_ids_shape.begin(), segment_ids_shape.end(), [](int64_t dim) { return dim == Shape::SHP_ANY; });
+  if (!x_any_shape && !ids_any_shape) {
+    if (x_shape[0] != segment_ids_shape[0]) {
+      MS_LOG(EXCEPTION) << "Length of segment_ids must match first value of x shape UnsortedSegmentMax";
+    }
+  }
+  ShapeVector x_shape_min;
+  ShapeVector x_shape_max;
+  x_shape_min = (x_is_dyn) ? x->shape()->min_shape() : x->shape()->shape();
+  x_shape_max = (x_is_dyn) ? x->shape()->max_shape() : x->shape()->shape();
+  min_shape.insert(min_shape.end(), x_shape_min.begin() + segment_ids_shape.size(), x_shape_min.end());
+  max_shape.insert(max_shape.end(), x_shape_max.begin() + segment_ids_shape.size(), x_shape_max.end());
+  return std::make_shared<AbstractTensor>(x->element(), std::make_shared<Shape>(shape, min_shape, max_shape));
+}
+
 AbstractBasePtr InferImplScatterAdd(const AnalysisEnginePtr &, const PrimitivePtr &primitive,
                                     const AbstractBasePtrList &args_spec_list) {
   const std::string op_name = primitive->name();

mindspore/core/abstract/primitive_infer_map.cc

@@ -58,6 +58,7 @@ PrimitiveEvalImplMap &GetPrimitiveToEvalImplMap() {
     {prim::kPrimSparseGatherV2, {InferImplGatherV2, true}},
     {prim::kPrimEmbeddingLookup, {InferImplEmbeddingLookup, true}},
     {prim::kPrimUnsortedSegmentSum, {InferImplUnsortedSegmentSum, true}},
+    {prim::kPrimUnsortedSegmentMax, {InferImplUnsortedSegmentMax, true}},
     {prim::kPrimScatterAdd, {InferImplScatterAdd, true}},
     {prim::kPrimScatterUpdate, {InferImplScatterUpdate, true}},
     {prim::kPrimMapCacheIdx, {InferImplMapCacheIdx, true}},

mindspore/ops/operations/array_ops.py

@@ -1981,7 +1981,7 @@ class UnsortedSegmentMin(PrimitiveWithInfer):
         return out
 
 
-class UnsortedSegmentMax(PrimitiveWithInfer):
+class UnsortedSegmentMax(PrimitiveWithCheck):
     """
     Computes the maximum along segments of a tensor.
 
@@ -2017,27 +2017,21 @@ class UnsortedSegmentMax(PrimitiveWithInfer):
     def __init__(self):
         """Initialize UnsortedSegmentMax"""
         self.init_prim_io_names(inputs=['x', 'segment_ids', 'num_segments'], outputs=['y'])
+        self.add_prim_attr("dynamic_shape_depends", [2])
 
-    def __infer__(self, x, segment_ids, num_segments):
-        x_type = x['dtype']
-        x_shape = x['shape']
+    def __check__(self, x, segment_ids, num_segments):
         segment_ids_shape = segment_ids['shape']
         valid_type = [mstype.float16, mstype.float32, mstype.int32]
         validator.check_tensor_dtype_valid("x", x['dtype'], valid_type, self.name)
         validator.check_tensors_dtypes_same_and_valid({"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
+        num_segments_type = num_segments['dtype']
+        validator.check_subclass("num_segments", num_segments_type, [mstype.tensor, mstype.number], self.name)
+        if isinstance(num_segments_type, type(mstype.tensor)):
+            validator.check_tensor_dtype_valid("num_segments", num_segments_type, [mstype.int64],
+                                               self.name)
+        else:
+            validator.check_value_type('num_segments', num_segments['value'], [int], self.name)
 
 
 class UnsortedSegmentProd(PrimitiveWithInfer):

tests/st/ops/gpu/test_unsorted_segment_max.py

@@ -20,6 +20,7 @@ import mindspore
 import mindspore.context as context
 import mindspore.nn as nn
 from mindspore import Tensor
+from mindspore.ops.operations import _inner_ops as inner
 from mindspore.common import dtype as mstype
 from mindspore.ops import operations as P
 
@@ -139,7 +140,7 @@ def test_3d_float32():
 @pytest.mark.platform_x86_gpu_training
 @pytest.mark.env_onecard
 def test_3d_single_init():
-    context.set_context(device_target='GPU')
+    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([3, 0, 1, -1], mstype.int32)
@@ -202,3 +203,112 @@ def test_3d_single_init():
                         [-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)
+
+# For testing Dynamic Shape operation
+class UnsortedSegmentMaxDynNet(nn.Cell):
+    def __init__(self, num_segments):
+        super(UnsortedSegmentMaxDynNet, self).__init__()
+        self.unsorted_segment_max = P.UnsortedSegmentMax()
+        self.gpu_convert_to_dynamic_shape = inner.GpuConvertToDynamicShape()
+        self.num_segments = num_segments
+
+    def construct(self, data, ids):
+        dyn_data = self.gpu_convert_to_dynamic_shape(data)
+        dyn_ids = self.gpu_convert_to_dynamic_shape(ids)
+        return self.unsorted_segment_max(dyn_data, dyn_ids, self.num_segments)
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_3d_float32_dyn():
+    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 = UnsortedSegmentMaxDynNet(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_dyn():
+    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)
+    num_segments = 4
+    net = UnsortedSegmentMaxDynNet(num_segments)
+    output = net(input_x, segment_ids).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
+    net = UnsortedSegmentMaxDynNet(num_segments)
+    output = net(input_x, segment_ids).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)