!40230 Add dynamic shape support for PReLU operator.

Merge pull request !40230 from hezhenhao1/add_prelu

mindspore/ccsrc/plugin/device/gpu/kernel/nn/prelu_gpu_kernel.cc

@@ -18,14 +18,70 @@
 
 namespace mindspore {
 namespace kernel {
-MS_REG_GPU_KERNEL_ONE(
+bool PReLUGpuKernelMod::Init(const BaseOperatorPtr &base_operator, const std::vector<KernelTensorPtr> &inputs,
-  PReLU,
+                             const std::vector<KernelTensorPtr> &outputs) {
-  KernelAttr().AddInputAttr(kNumberTypeFloat16).AddInputAttr(kNumberTypeFloat16).AddOutputAttr(kNumberTypeFloat16),
+  constexpr size_t input_num = 2;
-  PReLUGpuKernelMod, half)
+  constexpr size_t output_num = 1;
+  CHECK_KERNEL_INPUTS_NUM(inputs.size(), input_num, kernel_name_);
+  CHECK_KERNEL_OUTPUTS_NUM(outputs.size(), output_num, kernel_name_);
+  kernel_name_ = base_operator->GetPrim()->name();
+  auto kernel_attr = GetKernelAttrFromTensors(inputs, outputs);
+  auto [is_match, index] = MatchKernelAttr(kernel_attr, GetOpSupport());
+  if (!is_match) {
+    MS_LOG(ERROR) << "For '" << kernel_name_ << "', it does not support this kernel data type: " << kernel_attr;
+    return false;
+  }
+  kernel_func_ = func_list_[index].second;
+  return true;
+}
 
-MS_REG_GPU_KERNEL_ONE(
+int PReLUGpuKernelMod::Resize(const BaseOperatorPtr &base_operator, const std::vector<KernelTensorPtr> &inputs,
-  PReLU,
+                              const std::vector<KernelTensorPtr> &outputs,
-  KernelAttr().AddInputAttr(kNumberTypeFloat32).AddInputAttr(kNumberTypeFloat32).AddOutputAttr(kNumberTypeFloat32),
+                              const std::map<uint32_t, tensor::TensorPtr> &inputsOnHost) {
-  PReLUGpuKernelMod, float)
+  if (auto ret = KernelMod::Resize(base_operator, inputs, outputs, inputsOnHost); ret != KRET_OK) {
+    return ret;
+  }
+  auto input_shape = LongVecToSizeVec(inputs[kIndex0]->GetShapeVector());
+  auto weight_shape = LongVecToSizeVec(inputs[kIndex1]->GetShapeVector());
+  is_null_input_ =
+    CHECK_SHAPE_NULL(input_shape, kernel_name_, "x") || CHECK_SHAPE_NULL(weight_shape, kernel_name_, "weight");
+  input_length_ = std::accumulate(input_shape.begin(), input_shape.end(), size_t(1), std::multiplies<>());
+  per_channel_length_ =
+    input_shape.size() <= 1 ? input_length_ : input_length_ / (input_shape[kIndex0] * input_shape[kIndex1]);
+  weight_length_ = weight_shape[0];
+  return KRET_OK;
+}
+
+template <typename T>
+bool PReLUGpuKernelMod::LaunchKernel(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace,
+                                     const std::vector<AddressPtr> &outputs, void *stream_ptr) {
+  if (is_null_input_) {
+    return true;
+  }
+  auto input = GetDeviceAddress<T>(inputs, 0);
+  auto weight = GetDeviceAddress<T>(inputs, 1);
+  auto output = GetDeviceAddress<T>(outputs, 0);
+
+  CalPReLU(input_length_, weight_length_, per_channel_length_, input, weight, output,
+           reinterpret_cast<cudaStream_t>(stream_ptr));
+  return true;
+}
+
+std::vector<std::pair<KernelAttr, PReLUGpuKernelMod::PReLULaunchFunc>> PReLUGpuKernelMod::func_list_ = {
+  {KernelAttr().AddInputAttr(kNumberTypeFloat16).AddInputAttr(kNumberTypeFloat16).AddOutputAttr(kNumberTypeFloat16),
+   &PReLUGpuKernelMod::LaunchKernel<half>},
+  {KernelAttr().AddInputAttr(kNumberTypeFloat32).AddInputAttr(kNumberTypeFloat32).AddOutputAttr(kNumberTypeFloat32),
+   &PReLUGpuKernelMod::LaunchKernel<float>},
+};
+
+std::vector<KernelAttr> PReLUGpuKernelMod::GetOpSupport() {
+  std::vector<KernelAttr> support_list;
+  (void)std::transform(
+    func_list_.begin(), func_list_.end(), std::back_inserter(support_list),
+    [](const std::pair<KernelAttr, PReLUGpuKernelMod::PReLULaunchFunc> &pair) { return pair.first; });
+  return support_list;
+}
+
+MS_KERNEL_FACTORY_REG(NativeGpuKernelMod, PReLU, PReLUGpuKernelMod);
 }  // namespace kernel
 }  // namespace mindspore

mindspore/ccsrc/plugin/device/gpu/kernel/nn/prelu_gpu_kernel.h

@@ -19,6 +19,7 @@
 
 #include <vector>
 #include <map>
+#include <utility>
 #include <functional>
 
 #include "plugin/device/gpu/kernel/gpu_kernel.h"
@@ -27,91 +28,34 @@
 
 namespace mindspore {
 namespace kernel {
-template <typename T>
+class PReLUGpuKernelMod : public NativeGpuKernelMod {
-class PReLUGpuKernelMod : public DeprecatedNativeGpuKernelMod {
  public:
   PReLUGpuKernelMod() = default;
   ~PReLUGpuKernelMod() override = default;
 
-  bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &,
+  bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace,
               const std::vector<AddressPtr> &outputs, void *stream_ptr) override {
-    if (is_null_input_) {
+    return kernel_func_(this, inputs, workspace, outputs, stream_ptr);
-      return true;
-    }
-    auto *input = GetDeviceAddress<T>(inputs, 0);
-    auto *weight = GetDeviceAddress<T>(inputs, 1);
-    auto *output = GetDeviceAddress<T>(outputs, 0);
-
-    CalPReLU(input_length_, weight_length_, per_channel_length_, input, weight, output,
-             reinterpret_cast<cudaStream_t>(stream_ptr));
-    return true;
   }
 
-  bool Init(const CNodePtr &kernel_node) override {
+  bool Init(const BaseOperatorPtr &base_operator, const std::vector<KernelTensorPtr> &inputs,
-    auto kernel_name = common::AnfAlgo::GetCNodeName(kernel_node);
+            const std::vector<KernelTensorPtr> &outputs) override;
-    kernel_node_ = kernel_node;
-    ResetResource();
-    size_t input_num = common::AnfAlgo::GetInputTensorNum(kernel_node);
-    if (input_num != 2) {
-      MS_LOG(EXCEPTION) << "For '" << kernel_name << "', the number of inputs must be 2, but got " << input_num;
-    }
-    size_t output_num = common::AnfAlgo::GetOutputTensorNum(kernel_node);
-    if (output_num != 1) {
-      MS_LOG(EXCEPTION) << "For '" << kernel_name << "', the number of outputs must be 1, but got " << output_num;
-    }
 
-    auto input_shape = AnfAlgo::GetInputDeviceShape(kernel_node, 0);
+  int Resize(const BaseOperatorPtr &base_operator, const std::vector<KernelTensorPtr> &inputs,
-    auto weight_shape = AnfAlgo::GetInputDeviceShape(kernel_node, 1);
+             const std::vector<KernelTensorPtr> &outputs, const std::map<uint32_t, tensor::TensorPtr> &) override;
-    is_null_input_ =
-      CHECK_SHAPE_NULL(input_shape, kernel_name, "x") || CHECK_SHAPE_NULL(weight_shape, kernel_name, "weight");
-    if (is_null_input_) {
-      InitSizeLists();
-      return true;
-    }
-    input_length_ = std::accumulate(input_shape.begin(), input_shape.end(), size_t(1), std::multiplies<>());
-    size_t input_rank = input_shape.size();
-    int64_t channel_num;
-    if (input_rank == 0) {
-      channel_num = 1;
-      per_channel_length_ = 1;
-    } else if (input_rank == 1) {
-      channel_num = 1;
-      per_channel_length_ = input_shape[0];
-    } else {
-      channel_num = input_shape[1];
-      per_channel_length_ = std::accumulate(input_shape.begin() + 2, input_shape.end(), size_t(1), std::multiplies<>());
-    }
-
-    if (weight_shape.size() != 1 || (weight_shape[0] != 1 && weight_shape[0] != channel_num)) {
-      MS_LOG(EXCEPTION) << "For '" << kernel_name << "', the dimension of weight must be equal to 1 and "
-                        << "weight.shape[0] must be equal to 1 or the channel number, but got the dimension of "
-                        << "weight: " << weight_shape.size() << ", weight.shape[0]: " << weight_shape[0]
-                        << ", the channel num: " << channel_num;
-    }
-    weight_length_ = LongToSizeClipNeg(weight_shape[0]);
-    InitSizeLists();
-    return true;
-  }
-
-  void ResetResource() noexcept override {
-    input_length_ = 0;
-    weight_length_ = 0;
-    per_channel_length_ = 0;
-    is_null_input_ = false;
-    input_size_list_.clear();
-    output_size_list_.clear();
-    workspace_size_list_.clear();
-  }
 
  protected:
-  void InitSizeLists() override {
+  std::vector<KernelAttr> GetOpSupport() override;
-    size_t data_size = sizeof(T);
-    input_size_list_.push_back(input_length_ * data_size);
-    input_size_list_.push_back(weight_length_ * data_size);
-    output_size_list_.push_back(input_length_ * data_size);
-  }
 
  private:
+  template <typename T>
+  bool LaunchKernel(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace,
+                    const std::vector<AddressPtr> &outputs, void *stream_ptr);
+
+  using PReLULaunchFunc = std::function<bool(PReLUGpuKernelMod *, const std::vector<AddressPtr> &,
+                                             const std::vector<AddressPtr> &, const std::vector<AddressPtr> &, void *)>;
+  static std::vector<std::pair<KernelAttr, PReLULaunchFunc>> func_list_;
+  PReLULaunchFunc kernel_func_;
   bool is_null_input_{false};
   size_t input_length_{0};
   size_t weight_length_{0};

mindspore/core/ops/prelu.cc

@@ -1,5 +1,5 @@
 /**
- * Copyright 2020-2021 Huawei Technologies Co., Ltd
+ * Copyright 2020-2022 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.
@@ -20,10 +20,75 @@
 #include "ops/primitive_c.h"
 #include "utils/check_convert_utils.h"
 #include "mindapi/src/helper.h"
+#include "utils/ms_context.h"
+#include "ops/op_utils.h"
 
 namespace mindspore {
 namespace ops {
+bool IsAscend() {
+  auto context = MsContext::GetInstance();
+  MS_EXCEPTION_IF_NULL(context);
+  return context->get_param<std::string>(MS_CTX_DEVICE_TARGET) == kAscendDevice;
+}
+
+abstract::ShapePtr PReLUInferShape(const PrimitivePtr &primitive, const std::vector<AbstractBasePtr> &input_args) {
+  auto prim_name = primitive->name();
+  auto x_shape_ptr = input_args[kInputIndex0]->BuildShape();
+  auto weight_shape_ptr = input_args[kInputIndex1]->BuildShape();
+  // Dynamic rank.
+  if (x_shape_ptr->IsDimUnknown() || weight_shape_ptr->IsDimUnknown()) {
+    return x_shape_ptr->cast<abstract::ShapePtr>();
+  }
+
+  auto x_shape = CheckAndConvertUtils::ConvertShapePtrToShapeMap(x_shape_ptr)[kShape];
+  auto weight_shape = CheckAndConvertUtils::ConvertShapePtrToShapeMap(weight_shape_ptr)[kShape];
+  auto x_rank = x_shape.size();
+  auto weight_rank = weight_shape.size();
+  auto channel_num = x_rank <= 1 ? 1 : x_shape[1];
+  if (IsAscend() && x_rank <= 1) {
+    MS_EXCEPTION(ValueError)
+      << "For '" << prim_name
+      << "', the dimension of 'x' can not be 0-D or 1-D when the platform is \"Ascend\", but got dimension of 'x' is "
+      << x_rank << ".";
+  }
+  (void)CheckAndConvertUtils::CheckInteger("dimension of 'weight'", SizeToLong(weight_rank), kEqual, 1, prim_name);
+  if (weight_shape[0] != 1 && weight_shape[0] != channel_num) {
+    MS_EXCEPTION(ValueError)
+      << "For '" << prim_name
+      << "', the first dimension of 'weight' must be (1,) or it must be equal to number of channels: " << channel_num
+      << ", but got " << weight_shape << ".";
+  }
+  return x_shape_ptr->cast<abstract::ShapePtr>();
+}
+
+TypePtr PReLUInferType(const PrimitivePtr &primitive, const std::vector<AbstractBasePtr> &input_args) {
+  auto prim_name = primitive->name();
+  auto x_type = input_args[kInputIndex0]->BuildType();
+  auto weight_type = input_args[kInputIndex1]->BuildType();
+  auto valid_types = {kFloat16, kFloat32};
+
+  if (IsAscend()) {
+    (void)CheckAndConvertUtils::CheckTensorTypeValid("x", x_type, valid_types, prim_name);
+    (void)CheckAndConvertUtils::CheckTensorTypeValid("weight", weight_type, valid_types, prim_name);
+  } else {
+    std::map<std::string, TypePtr> args;
+    (void)args.emplace("x", x_type);
+    (void)args.emplace("weight", weight_type);
+    (void)CheckAndConvertUtils::CheckTensorTypeSame(args, valid_types, prim_name);
+  }
+  return x_type;
+}
+
 MIND_API_OPERATOR_IMPL(PReLU, BaseOperator);
-REGISTER_PRIMITIVE_C(kNamePReLU, PReLU);
+AbstractBasePtr PReLUInfer(const abstract::AnalysisEnginePtr &, const PrimitivePtr &primitive,
+                           const std::vector<AbstractBasePtr> &input_args) {
+  MS_EXCEPTION_IF_NULL(primitive);
+  constexpr size_t input_num = 2;
+  CheckAndConvertUtils::CheckInputArgs(input_args, kEqual, input_num, primitive->name());
+  auto type = PReLUInferType(primitive, input_args);
+  auto shape = PReLUInferShape(primitive, input_args);
+  return abstract::MakeAbstract(shape, type);
+}
+REGISTER_PRIMITIVE_EVAL_IMPL(PReLU, prim::kPrimPRelu, PReLUInfer, nullptr, true);
 }  // namespace ops
 }  // namespace mindspore

mindspore/core/ops/prelu.h

@@ -1,5 +1,5 @@
 /**
- * Copyright 2020 Huawei Technologies Co., Ltd
+ * Copyright 2020-2022 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.
@@ -32,8 +32,8 @@ class MIND_API PReLU : public BaseOperator {
  public:
   MIND_API_BASE_MEMBER(PReLU);
   /// \brief Constructor.
-  PReLU() : BaseOperator(kNamePReLU) { InitIOName({"x"}, {"y"}); }
+  PReLU() : BaseOperator(kNamePReLU) { InitIOName({"x", "weight"}, {"output"}); }
-  explicit PReLU(const std::string k_name) : BaseOperator(k_name) { InitIOName({"x"}, {"y"}); }
+  explicit PReLU(const std::string k_name) : BaseOperator(k_name) { InitIOName({"x", "weight"}, {"output"}); }
   /// \brief Init. Refer to the parameters of Python API @ref mindspore.ops.PReLU for the inputs.
   void Init() const {}
 };

mindspore/python/mindspore/ops/_op_impl/tbe/__init__.py

@@ -417,6 +417,7 @@ from .unpack import _unpack_tbe
 from .unpack_ds import _unpack_ds_tbe
 from .scatter_update import _scatter_update_tbe
 from .prelu import _prelu_tbe
+from .prelu_ds import _prelu_ds_tbe
 from .prelu_grad import _prelu_grad_tbe
 from .binary_cross_entropy_ds import _binary_cross_entropy_ds_tbe
 from .binary_cross_entropy import _binary_cross_entropy_tbe

mindspore/python/mindspore/ops/_op_impl/tbe/prelu_ds.py

@@ -0,0 +1,38 @@
+# Copyright 2022 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.
+# ============================================================================
+
+"""PReLU op"""
+from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
+
+prelu_ds_op_info = TBERegOp("PReLU") \
+    .fusion_type("ELEMWISE") \
+    .async_flag(False) \
+    .binfile_name("prelu.so") \
+    .compute_cost(10) \
+    .kernel_name("prelu") \
+    .partial_flag(True) \
+    .dynamic_shape(True) \
+    .input(0, "x", False, "required", "all") \
+    .input(1, "weight", False, "required", "all") \
+    .output(0, "y", False, "required", "all") \
+    .is_dynamic_format(True) \
+    .dtype_format(DataType.None_None, DataType.None_None, DataType.None_None) \
+    .get_op_info()
+
+
+@op_info_register(prelu_ds_op_info)
+def _prelu_ds_tbe():
+    """PReLU TBE register"""
+    return

mindspore/python/mindspore/ops/operations/nn_ops.py

@@ -4107,35 +4107,7 @@ class PReLU(PrimitiveWithInfer):
 
     @prim_attr_register
     def __init__(self):
-        pass
+        self.init_prim_io_names(inputs=['x', 'weight'], outputs=['output'])
-
-    def infer_shape(self, input_x_shape, weight_shape):
-        input_x_dim = len(input_x_shape)
-        if input_x_dim in (0, 1):
-            if context.get_context("device_target") == "Ascend":
-                raise ValueError(f"For '{self.name}', the dimension of 'x' can not be 0-D or 1-D when the platform is "
-                                 f"\"Ascend\", but got dimension of 'x' is {input_x_dim}.")
-            channel_num = 1
-        else:
-            channel_num = input_x_shape[1]
-
-        weight_dim = len(weight_shape)
-        if weight_dim != 1:
-            raise ValueError(f"For '{self.name}', the dimension of 'weight' must be 1, while got {weight_dim}.")
-        if weight_shape[0] != 1 and weight_shape[0] != channel_num:
-            raise ValueError(f"For '{self.name}', the first dimension of 'weight' must be (1,) or "
-                             f"it must be equal to number of channels: {channel_num}, but got {weight_shape}")
-        return input_x_shape
-
-    def infer_dtype(self, input_x_dtype, weight_dtype):
-        valid_dtypes = (mstype.float16, mstype.float32)
-        args = {"input_x": input_x_dtype, "weight": weight_dtype}
-        if context.get_context("device_target") == "GPU":
-            validator.check_tensors_dtypes_same_and_valid(args, valid_dtypes, self.name)
-        else:
-            validator.check_tensor_dtype_valid("input_x", input_x_dtype, valid_dtypes, self.name)
-            validator.check_tensor_dtype_valid("weight", weight_dtype, valid_dtypes, self.name)
-        return input_x_dtype
 
 
 class LSTM(PrimitiveWithInfer):

tests/st/ops/dynamic_shape/test_dynamic_shape_prelu.py

@@ -0,0 +1,76 @@
+# Copyright 2022 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.nn as nn
+from mindspore import Tensor, context
+from mindspore.ops import operations as P
+from mindspore.common import dtype as mstype
+
+
+class GetDynamicInputNet(nn.Cell):
+    def __init__(self, axis=0):
+        super(GetDynamicInputNet, self).__init__()
+        self.unique = P.Unique()
+        self.gather = P.Gather()
+        self.cast = P.Cast()
+        self.axis = axis
+
+    def construct(self, x, indices):
+        unique_indices, _ = self.unique(indices)
+        x_dtype = x.dtype
+        x = self.cast(x, mstype.float32)
+        real_x = self.gather(x, unique_indices, self.axis)
+        return self.cast(real_x, x_dtype)
+
+
+class PReLUDyNet(nn.Cell):
+    def __init__(self):
+        super(PReLUDyNet, self).__init__()
+        self.op = P.PReLU()
+        self.transformer = GetDynamicInputNet()
+
+    def construct(self, indices, x, weight):
+        real_x = self.transformer(x, indices)
+        out = self.op(real_x, weight)
+        return real_x, weight, out
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+@pytest.mark.parametrize("data_shape", [((8, 6, 7), (1,))])
+@pytest.mark.parametrize("data_type", [np.float16, np.float32])
+def test_dynamic_shape_prelu(data_shape, data_type):
+    """
+    Feature: PReLU DynamicShape.
+    Description: Test case of dynamic shape for PReLU operator.
+    Expectation: success.
+    """
+    context.set_context(mode=context.GRAPH_MODE)
+    x_shape, weight_shape = data_shape
+    x = Tensor(np.random.random(size=x_shape).astype(data_type))
+    weight = Tensor(np.random.random(size=weight_shape).astype(data_type))
+    indices = Tensor(np.random.randint(0, x_shape[0], size=(5,)).astype(np.int32))
+
+    dy_net = PReLUDyNet()
+    real_x, real_weight, output = dy_net(indices, x, weight)
+    x, weight = real_x.asnumpy(), real_weight.asnumpy()
+    expect = np.where(x >= 0, x, weight * x)
+
+    np.testing.assert_allclose(expect, output.asnumpy())