Merge pull request PaddlePaddle#687 from joey12300/add_tile_linespace…

…_functions

[Functions] Add fdtensor functions

fastdeploy/function/cumprod.cc

@@ -0,0 +1,78 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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 "fastdeploy/function/cumprod.h"
+
+namespace fastdeploy {
+namespace function {
+
+void GetCumprodDimInfo(const std::vector<int64_t>& dim, int cumprod_dim,
+                       size_t* outer_dim, size_t* mid_dim, size_t* inner_dim) {
+  int dim_size = dim.size();
+  FDASSERT(cumprod_dim >= -dim_size,
+           "The input dim of CumprodOp should be larger than the opposite "
+           "rank of input x which is %d. But received dim = %d",
+           -dim_size, cumprod_dim);
+  FDASSERT(cumprod_dim < dim_size,
+           "The input dim of CumprodOp should be smaller than the "
+           "rank of input x which is %d. But received dim = %d",
+           dim_size, cumprod_dim);
+  if (cumprod_dim < 0)
+    cumprod_dim += dim_size;
+
+  *outer_dim = 1;
+  for (int i = 0; i < cumprod_dim; ++i) {
+    *outer_dim *= dim[i];
+  }
+  *mid_dim = dim[cumprod_dim];
+  *inner_dim = 1;
+  for (int i = cumprod_dim + 1; i < dim_size; ++i) {
+    *inner_dim *= dim[i];
+  }
+}
+
+template <typename T>
+void CumprodKernel(const FDTensor& x, FDTensor* out, int axis) {
+  auto* x_data = reinterpret_cast<const T*>(x.Data());
+  auto shape = x.Shape();
+
+  size_t outer_dim = 1;
+  size_t mid_dim = 1;
+  size_t inner_dim = 1;
+  GetCumprodDimInfo(shape, axis, &outer_dim, &mid_dim, &inner_dim);
+
+  out->Allocate(x.Shape(), x.Dtype());
+  auto* out_data = reinterpret_cast<T*>(out->Data());
+
+  for (size_t i = 0; i < outer_dim; i++) {
+    for (size_t j = 0; j < mid_dim; j++) {
+      for (size_t k = 0; k < inner_dim; k++) {
+        size_t pos = i * mid_dim * inner_dim + j * inner_dim + k;
+        if (j == 0) {
+          out_data[pos] = x_data[pos];
+        } else {
+          out_data[pos] = out_data[pos - inner_dim] * x_data[pos];
+        }
+      }
+    }
+  }
+}
+
+void Cumprod(const FDTensor& x, FDTensor* out, int axis) {
+  FD_VISIT_INT_FLOAT_TYPES(x.dtype, "CumprodKernel",
+                           ([&] { CumprodKernel<data_t>(x, out, axis); }));
+}
+
+}  // namespace function
+}  // namespace fastdeploy

fastdeploy/function/cumprod.h

@@ -0,0 +1,31 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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.
+
+#pragma once
+
+#include "fastdeploy/core/fd_tensor.h"
+
+namespace fastdeploy {
+namespace function {
+
+/** Excute the concatenate operation for input FDTensor along given axis.
+    @param x The input tensor.
+    @param out The output tensor which stores the result.
+    @param axisi Axis which will be concatenated.
+*/
+
+FASTDEPLOY_DECL void Cumprod(const FDTensor& x, FDTensor* out, int axis = 0);
+
+}  // namespace function
+}  // namespace fastdeploy

fastdeploy/function/linspace.cc

@@ -0,0 +1,52 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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 "fastdeploy/function/linspace.h"
+
+namespace fastdeploy {
+namespace function {
+
+template <typename T>
+void LinspaceKernel(double start, double end, int num, FDTensor* out) {
+  FDASSERT(
+      num > 0,
+      "The num of linspace op should be larger than 0, but received num is %d",
+      num);
+  out->Allocate({num}, TypeToDataType<T>::dtype);
+  T* out_data = reinterpret_cast<T*>(out->Data());
+  if (num > 1) {
+    // step should be of double type for all types
+    double step = (static_cast<double>(end - start)) / (num - 1);
+    int half_num = num / 2;
+    for (int i = 0; i < num; ++i) {
+      if (i < half_num) {
+        out_data[i] = static_cast<T>(start + step * i);
+      } else {
+        out_data[i] = static_cast<T>(end - step * (num - i - 1));
+      }
+    }
+  } else {
+    out_data[0] = static_cast<T>(start);
+  }
+}
+
+void Linspace(double start, double end, int num, FDTensor* out,
+              FDDataType dtype) {
+  FD_VISIT_INT_FLOAT_TYPES(dtype, "LinspaceKernel", ([&] {
+                             LinspaceKernel<data_t>(start, end, num, out);
+                           }));
+}
+
+}  // namespace function
+}  // namespace fastdeploy

fastdeploy/function/linspace.h

@@ -0,0 +1,33 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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.
+
+#pragma once
+
+#include "fastdeploy/core/fd_tensor.h"
+
+namespace fastdeploy {
+namespace function {
+
+/** Return fixed number of evenly spaced values within a given interval.
+    @param start The input start is start variable of range.
+    @param end The input stop is start variable of range.
+    @param num The input num is given num of the sequence.
+    @param out The output tensor which stores the result.
+    @param dtype The data type of output tensor, default to float32.
+*/
+FASTDEPLOY_DECL void Linspace(double start, double end, int num, FDTensor* out,
+                              FDDataType dtype = FDDataType::FP32);
+
+}  // namespace function
+}  // namespace fastdeploy

fastdeploy/function/math.cc

@@ -39,6 +39,7 @@ DEFINE_ACTIVATION_KERNEL(Sqrt, SqrtFunctor)
 DEFINE_ACTIVATION_KERNEL(Log, LogFunctor)
 DEFINE_ACTIVATION_KERNEL(Round, RoundFunctor)
 DEFINE_ACTIVATION_KERNEL(Exp, ExpFunctor)
+DEFINE_ACTIVATION_KERNEL(Abs, AbsFunctor)
 
 void Sqrt(const FDTensor& x, FDTensor* out) {
   FD_VISIT_FLOAT_TYPES(x.dtype, "SqrtKernel",
@@ -60,5 +61,10 @@ void Exp(const FDTensor& x, FDTensor* out) {
                        ([&] { ExpKernel<data_t>(x, out); }));
 }
 
+void Abs(const FDTensor& x, FDTensor* out) {
+  FD_VISIT_FLOAT_TYPES(x.dtype, "AbsKernel",
+                       ([&] { AbsKernel<data_t>(x, out); }));
+}
+
 }  // namespace function
 }  // namespace fastdeploy

fastdeploy/function/math.h

@@ -43,5 +43,11 @@ FASTDEPLOY_DECL void Round(const FDTensor& x, FDTensor* out);
 */
 FASTDEPLOY_DECL void Exp(const FDTensor& x, FDTensor* out);
 
+/** This operator is used to perform elementwise abs for input X. Only for float type FDTensor
+    @param x The input tensor.
+    @param out The output tensor which stores the result.
+*/
+FASTDEPLOY_DECL void Abs(const FDTensor& x, FDTensor* out);
+
 }  // namespace function
 }  // namespace fastdeploy

fastdeploy/function/math_functor.h

@@ -52,5 +52,14 @@ template <typename T> struct SqrtFunctor {
   }
 };
 
+// abs(x) = x if x > 0 else -x
+template <typename T> struct AbsFunctor {
+  template <typename Device, typename X, typename Out>
+  void operator()(Device d, X x, Out out) const {
+    out.device(d) =
+        x.unaryExpr([](T v) { return v > static_cast<T>(0) ? v : -v; });
+  }
+};
+
 }  // namespace function
 }  // namespace fastdeploy