Add SquareDifference kernel for GPU

添加测试用例

添加测试用例

添加测试用例

添加测试用例
清除告警

添加测试用例

添加测试用例

清除告警

mindspore/ccsrc/backend/kernel_compiler/cpu/random_cpu_kernel.cc

@@ -105,6 +105,5 @@ bool RandomCPUKernel::Launch(const std::vector<kernel::AddressPtr> &inputs,
   }
   return true;
 }
-
 }  // namespace kernel
 }  // namespace mindspore

mindspore/ccsrc/backend/kernel_compiler/cpu/select_cpu_kernel.cc

@@ -48,6 +48,5 @@ bool SelectCPUKernel<T>::Launch(const std::vector<AddressPtr> &inputs, const std
   }
   return true;
 }
-
 }  // namespace kernel
 }  // namespace mindspore

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

@@ -192,6 +192,14 @@ struct AbsGradFunc<half2> {
   }
 };
 
+template <typename T>
+struct SquaredDifferenceFunc {
+  __device__ __forceinline__ T operator()(const T &lhs, const T &rhs) {
+    T diff = lhs - rhs;
+    return diff * diff;
+  }
+};
+
 // Element-wise Comparation
 template <typename T, typename Func>
 __global__ void ElewiseCmpKernel(const int nums, const T *x0, const T *x1, bool *y) {
@@ -260,6 +268,8 @@ void ElewiseArithKernel(const int &nums, enum BroadcastOpType op, const T *x0, c
       return ElewiseArithKernel<T, DivFunc<T>><<<(nums + 255) / 256, 256, 0, stream>>>(nums, x0, x1, y);
     case BROADCAST_TYPE_DIVNONAN:
       return ElewiseArithKernel<T, DivNoNanFunc<T>><<<(nums + 255) / 256, 256, 0, stream>>>(nums, x0, x1, y);
+    case BROADCAST_TYPE_SQUARED_DIFFERENCE:
+      return ElewiseArithKernel<T, SquaredDifferenceFunc<T>><<<(nums + 255) / 256, 256, 0, stream>>>(nums, x0, x1, y);
     default:
       break;
   }
@@ -481,6 +491,11 @@ void BroadcastArith(const std::vector<size_t> &x0_dims, const std::vector<size_t
         x0_dims[0], x0_dims[1], x0_dims[2], x0_dims[3], x0_dims[4], x0_dims[5], x0_dims[6], x1_dims[0], x1_dims[1],
         x1_dims[2], x1_dims[3], x1_dims[4], x1_dims[5], x1_dims[6], y_dims[0], y_dims[1], y_dims[2], y_dims[3],
         y_dims[4], y_dims[5], y_dims[6], x0, x1, y);
+    case BROADCAST_TYPE_SQUARED_DIFFERENCE:
+      return BroadcastArithKernel<T, SquaredDifferenceFunc<T>><<<(size + 255) / 256, 256, 0, stream>>>(
+        x0_dims[0], x0_dims[1], x0_dims[2], x0_dims[3], x0_dims[4], x0_dims[5], x0_dims[6], x1_dims[0], x1_dims[1],
+        x1_dims[2], x1_dims[3], x1_dims[4], x1_dims[5], x1_dims[6], y_dims[0], y_dims[1], y_dims[2], y_dims[3],
+        y_dims[4], y_dims[5], y_dims[6], x0, x1, y);
     default:
       break;
   }

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

@@ -37,6 +37,7 @@ enum BroadcastOpType {
   BROADCAST_TYPE_DIV = 11,
   BROADCAST_TYPE_DIVNONAN = 12,
   BROADCAST_TYPE_EQUAL = 13,
+  BROADCAST_TYPE_SQUARED_DIFFERENCE = 14,
   BROADCAST_TYPE_INVALID = 0xffffffff,
 };
 

mindspore/ccsrc/backend/kernel_compiler/gpu/math/squared_difference_kernel.cc

@@ -0,0 +1,40 @@
+/**
+ * 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/math/squared_difference_kernel.h"
+
+namespace mindspore {
+namespace kernel {
+// fp32
+MS_REG_GPU_KERNEL_ONE(
+  SquaredDifference,
+  KernelAttr().AddInputAttr(kNumberTypeFloat32).AddInputAttr(kNumberTypeFloat32).AddOutputAttr(kNumberTypeFloat32),
+  SquaredDifferenceOpGpuKernel, float)
+
+// fp16
+MS_REG_GPU_KERNEL_ONE(
+  SquaredDifference,
+  KernelAttr().AddInputAttr(kNumberTypeFloat16).AddInputAttr(kNumberTypeFloat16).AddOutputAttr(kNumberTypeFloat16),
+  SquaredDifferenceOpGpuKernel, half)
+
+// int32
+MS_REG_GPU_KERNEL_ONE(
+  SquaredDifference,
+  KernelAttr().AddInputAttr(kNumberTypeInt32).AddInputAttr(kNumberTypeInt32).AddOutputAttr(kNumberTypeInt32),
+  SquaredDifferenceOpGpuKernel, int)
+
+}  // namespace kernel
+}  // namespace mindspore

mindspore/ccsrc/backend/kernel_compiler/gpu/math/squared_difference_kernel.h

@@ -0,0 +1,144 @@
+/**
+ * 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_SQUARED_DIFFERENCE_GPU_KERNEL_H_
+#define MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_SQUARED_DIFFERENCE_GPU_KERNEL_H_
+
+#include <cuda_runtime_api.h>
+#include <vector>
+#include <string>
+#include <map>
+#include "backend/kernel_compiler/gpu/gpu_kernel.h"
+#include "backend/kernel_compiler/gpu/gpu_kernel_factory.h"
+#include "backend/kernel_compiler/gpu/cuda_impl/broadcast_impl.cuh"
+#include "backend/kernel_compiler/gpu/kernel_constants.h"
+namespace mindspore {
+namespace kernel {
+constexpr int MAX_DIMS = 7;
+template <typename T>
+class SquaredDifferenceOpGpuKernel : public GpuKernel {
+ public:
+  SquaredDifferenceOpGpuKernel() { ResetResource(); }
+  ~SquaredDifferenceOpGpuKernel() 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 {
+    T *lhs = GetDeviceAddress<T>(inputs, 0);
+    T *rhs = GetDeviceAddress<T>(inputs, 1);
+    T *output = GetDeviceAddress<T>(outputs, 0);
+    if (need_broadcast_) {
+      BroadcastArith(lhs_shape_, rhs_shape_, output_shape_, op_type_, lhs, rhs, output,
+                     reinterpret_cast<cudaStream_t>(stream_ptr));
+    } else {
+      ElewiseArith(output_num_, op_type_, lhs, rhs, output, reinterpret_cast<cudaStream_t>(stream_ptr));
+    }
+    return true;
+  }
+
+  bool Init(const CNodePtr &kernel_node) override {
+    auto input_shape1 = AnfAlgo::GetInputRealDeviceShapeIfExist(kernel_node, 0);
+    auto input_shape2 = AnfAlgo::GetInputRealDeviceShapeIfExist(kernel_node, 1);
+    auto output_shape = AnfAlgo::GetOutputRealDeviceShapeIfExist(kernel_node, 0);
+    need_broadcast_ = IsBroadcast(input_shape1, input_shape2);
+    if (need_broadcast_ && output_shape.size() > MAX_DIMS) {
+      MS_LOG(EXCEPTION) << "Broadcast operation not support dim greater than 7";
+    }
+
+    lhs_shape_.resize(MAX_DIMS, 1);
+    rhs_shape_.resize(MAX_DIMS, 1);
+    output_shape_.resize(MAX_DIMS, 1);
+    for (size_t i = 0; i < output_shape.size(); i++) {
+      if (need_broadcast_) {
+        output_shape_[i] = output_shape[i];
+      }
+      output_num_ *= output_shape[i];
+    }
+    int lhs_offset = output_shape.size() - input_shape1.size();
+    for (size_t j = 0; j < input_shape1.size(); j++) {
+      if (need_broadcast_) {
+        lhs_shape_[j + lhs_offset] = input_shape1[j];
+      }
+      input1_num_ *= input_shape1[j];
+    }
+    int rhs_offset = output_shape.size() - input_shape2.size();
+    for (size_t k = 0; k < input_shape2.size(); k++) {
+      if (need_broadcast_) {
+        rhs_shape_[k + rhs_offset] = input_shape2[k];
+      }
+      input2_num_ *= input_shape2[k];
+    }
+
+    InitSizeLists();
+    return true;
+  }
+
+  void ResetResource() noexcept override {
+    op_type_ = BROADCAST_TYPE_SQUARED_DIFFERENCE;
+    need_broadcast_ = false;
+    input1_num_ = 1;
+    input2_num_ = 1;
+    output_num_ = 1;
+    lhs_shape_.clear();
+    rhs_shape_.clear();
+    output_shape_.clear();
+    input_size_list_.clear();
+    output_size_list_.clear();
+    workspace_size_list_.clear();
+  }
+
+ protected:
+  void InitResource() override { return; }
+  void InitSizeLists() override {
+    input_size_list_.push_back(input1_num_ * sizeof(T));
+    input_size_list_.push_back(input2_num_ * sizeof(T));
+    output_size_list_.push_back(output_num_ * sizeof(T));
+  }
+
+ private:
+  bool IsBroadcast(const std::vector<size_t> &lhs, const std::vector<size_t> &rhs) {
+    if (lhs.size() != rhs.size()) {
+      return true;
+    }
+    for (size_t i = 0; i < lhs.size(); i++) {
+      if (lhs[i] != rhs[i]) {
+        return true;
+      }
+    }
+    return false;
+  }
+
+  BroadcastOpType op_type_;
+  bool need_broadcast_;
+  bool is_comp_op_;
+  size_t input1_num_;
+  size_t input2_num_;
+  size_t output_num_;
+  std::vector<size_t> lhs_shape_;
+  std::vector<size_t> rhs_shape_;
+  std::vector<size_t> output_shape_;
+
+  std::vector<size_t> input_size_list_;
+  std::vector<size_t> output_size_list_;
+  std::vector<size_t> workspace_size_list_;
+};  // namespace kernel
+}  // namespace kernel
+}  // namespace mindspore
+
+#endif  // MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_SQUARED_DIFFERENCE_GPU_KERNEL_H_