Merge branch 'add-deconv' into 'master'

Add deconv

See merge request !401

mace/core/operator.cc

@@ -82,6 +82,7 @@ extern void Register_BiasAdd(OperatorRegistry *op_registry);
 extern void Register_ChannelShuffle(OperatorRegistry *op_registry);
 extern void Register_Concat(OperatorRegistry *op_registry);
 extern void Register_Conv2D(OperatorRegistry *op_registry);
+extern void Register_Deconv2D(OperatorRegistry *op_registry);
 extern void Register_DepthToSpace(OperatorRegistry *op_registry);
 extern void Register_DepthwiseConv2d(OperatorRegistry *op_registry);
 extern void Register_Dequantize(OperatorRegistry *op_registry);
@@ -122,6 +123,7 @@ OperatorRegistry::OperatorRegistry() {
   ops::Register_ChannelShuffle(this);
   ops::Register_Concat(this);
   ops::Register_Conv2D(this);
+  ops::Register_Deconv2D(this);
   ops::Register_DepthToSpace(this);
   ops::Register_DepthwiseConv2d(this);
   ops::Register_Dequantize(this);

mace/kernels/conv_pool_2d_util.cc

@@ -14,6 +14,7 @@
 
 #include "mace/kernels/conv_pool_2d_util.h"
 
+#include <algorithm>
 #include <vector>
 
 namespace mace {
@@ -147,6 +148,8 @@ void CalcNHWCPaddingAndOutputSize(const index_t *input_shape,   // NHWC
   output_shape[3] = output_channels;
 }
 
+
+
 void CalcOutputSize(const index_t *input_shape,   // NHWC
                     const index_t *filter_shape,  // HWOI
                     const int *padding_size,
@@ -161,14 +164,7 @@ void CalcOutputSize(const index_t *input_shape,   // NHWC
              "If dilations > 1, strides should be 1");
   MACE_CHECK_NOTNULL(output_shape);
   MACE_CHECK_NOTNULL(padding_size);
-  /*
-  * Convlution arithmetic:
-  * o = floor((i + 2 * p - k - (k - 1) * (d - 1)) / s) + 1
-  * Pooling arithmetic:
-  * o = ceil((i + 2 * p - k - (k - 1) * (d - 1)) / s) + 1
-  * For details, see https://arxiv.org/pdf/1603.07285.pdf or
-  * http://deeplearning.net/software/theano/tutorial/conv_arithmetic.html
-  */
+
   output_shape[0] = input_shape[0];
   if (round_type == FLOOR) {
     output_shape[1] = static_cast<index_t>(
@@ -454,5 +450,6 @@ void ConstructNHWCInputWithPadding(const Tensor *input_tensor,
     }
   }
 }
+
 }  // namespace kernels
 }  // namespace mace

mace/kernels/deconv_2d.h

+// Copyright 2018 Xiaomi, Inc.  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.
+
+#ifndef MACE_KERNELS_DECONV_2D_H_
+#define MACE_KERNELS_DECONV_2D_H_
+
+#if defined(MACE_ENABLE_NEON) && defined(__aarch64__)
+#include <arm_neon.h>
+#endif
+#include <algorithm>
+#include <memory>
+#include <vector>
+
+#include "mace/core/future.h"
+#include "mace/core/runtime/opencl/cl2_header.h"
+#include "mace/core/tensor.h"
+#include "mace/kernels/activation.h"
+#include "mace/kernels/conv_pool_2d_util.h"
+#include "mace/utils/utils.h"
+
+namespace mace {
+namespace kernels {
+
+namespace deconv {
+
+template<typename T>
+void Deconv2dNCHW(const T *input,
+                  const T *filter,
+                  const T *bias,
+                  const index_t batch,
+                  const index_t in_height,
+                  const index_t in_width,
+                  const index_t in_channels,
+                  const index_t out_height,
+                  const index_t out_width,
+                  const index_t out_channels,
+                  const index_t filter_height,
+                  const index_t filter_width,
+                  const index_t stride_h,
+                  const index_t stride_w,
+                  const int padding_top,
+                  const int padding_left,
+                  float *output) {
+#pragma omp parallel for collapse(4)
+  for (index_t b = 0; b < batch; ++b) {
+    for (index_t oc = 0; oc < out_channels; ++oc) {
+      for (index_t oh = 0; oh < out_height; ++oh) {
+        for (index_t ow = 0; ow < out_width; ++ow) {
+          index_t filter_start_y, filter_start_x;
+          index_t start_x = std::max<int>(0, ow + stride_w -1 - padding_left);
+          index_t start_y = std::max<int>(0, oh + stride_h -1 - padding_top);
+          start_x /= stride_w;
+          start_y /= stride_h;
+          filter_start_x = padding_left + stride_w * start_x - ow;
+          filter_start_y = padding_top + stride_h * start_y - oh;
+          filter_start_x = filter_width - 1 - filter_start_x;
+          filter_start_y = filter_height - 1 - filter_start_y;
+          T out_value = 0;
+          index_t out_pos =
+              ((b * out_channels + oc) * out_height + oh) * out_width + ow;
+          for (index_t ic = 0; ic < in_channels; ++ic) {
+            for (index_t f_y = filter_start_y, ih = start_y;
+                 f_y >= 0 && ih < in_height; f_y -= stride_h, ++ih) {
+              for (index_t f_x = filter_start_x, iw = start_x;
+                  f_x >= 0 && iw < in_width; f_x -= stride_w, ++iw) {
+                  index_t weight_pos =
+                      ((oc * in_channels + ic) * filter_height + f_y)
+                          * filter_width + f_x;
+                  index_t in_pos =
+                      ((b * in_channels + ic) * in_height + ih)
+                          * in_width + iw;
+                  out_value += input[in_pos] * filter[weight_pos];
+              }
+            }
+          }
+          if (bias != nullptr)
+            out_value += bias[oc];
+          output[out_pos] = out_value;
+        }
+      }
+    }
+  }
+}
+}  // namespace deconv
+
+struct Deconv2dFunctorBase {
+  Deconv2dFunctorBase(const int *strides,
+                      const Padding &padding_type,
+                      const std::vector<int> &paddings,
+                      const std::vector<index_t> &output_shape,
+                      const ActivationType activation,
+                      const float relux_max_limit)
+      : strides_(strides),
+        padding_type_(padding_type),
+        paddings_(paddings),
+        output_shape_(output_shape),
+        activation_(activation),
+        relux_max_limit_(relux_max_limit) {}
+
+  static void CalcDeconvOutputSize(
+      const index_t *input_shape,   // NHWC
+      const index_t *filter_shape,  // OIHW
+      const int *strides,
+      index_t *output_shape,
+      const int *padding_size,
+      const bool isNCHW = false,
+      const bool isOIHW = false) {
+    MACE_CHECK_NOTNULL(output_shape);
+    MACE_CHECK_NOTNULL(padding_size);
+    MACE_CHECK_NOTNULL(input_shape);
+    MACE_CHECK_NOTNULL(filter_shape);
+    MACE_CHECK_NOTNULL(strides);
+
+    const index_t output_channel = isOIHW ? filter_shape[0] : filter_shape[2];
+
+    const index_t in_height = isNCHW ? input_shape[2] : input_shape[1];
+    const index_t in_width = isNCHW ? input_shape[3] : input_shape[2];
+    const index_t in_channels = isNCHW ? input_shape[1] : input_shape[3];
+
+    const index_t extended_input_height =
+        (in_height - 1) * strides[0] + 1 + padding_size[0];
+    const index_t extended_input_width =
+        (in_width - 1) * strides[1] + 1 + padding_size[1];
+
+    const index_t filter_h = isOIHW ? filter_shape[2] : filter_shape[0];
+    const index_t filter_w = isOIHW ? filter_shape[3] : filter_shape[1];
+
+    index_t out_height = extended_input_height - filter_h + 1;
+    index_t out_width = extended_input_width - filter_w + 1;
+
+    output_shape[0] = input_shape[0];
+    if (isNCHW) {
+      output_shape[1] = output_channel;
+      output_shape[2] = out_height;
+      output_shape[3] = out_width;
+    } else {
+      output_shape[1] = out_height;
+      output_shape[2] = out_width;
+      output_shape[3] = output_channel;
+    }
+  }
+
+  static void CalcDeconvPaddingAndInputSize(
+      const index_t *input_shape,   // NHWC
+      const index_t *filter_shape,  // OIHW
+      const int *strides,
+      Padding padding,
+      const index_t *output_shape,
+      int *padding_size,
+      const bool isNCHW = false,
+      const bool isOIHW = false) {
+    MACE_CHECK_NOTNULL(output_shape);
+    MACE_CHECK_NOTNULL(padding_size);
+    MACE_CHECK_NOTNULL(input_shape);
+    MACE_CHECK_NOTNULL(filter_shape);
+    MACE_CHECK_NOTNULL(strides);
+
+    const index_t in_height = isNCHW ? input_shape[2] : input_shape[1];
+    const index_t in_width = isNCHW ? input_shape[3] : input_shape[2];
+    const index_t in_channels = isNCHW ? input_shape[1] : input_shape[3];
+
+    const index_t out_height = isNCHW ? output_shape[2] : output_shape[1];
+    const index_t out_width = isNCHW ? output_shape[3] : output_shape[2];
+    const index_t out_channels = isNCHW ? output_shape[1] : output_shape[3];
+
+    const index_t extended_input_height = (in_height - 1) * strides[0] + 1;
+    const index_t extended_input_width = (in_width - 1) * strides[1] + 1;
+
+    const index_t filter_h = isOIHW ? filter_shape[2] : filter_shape[0];
+    const index_t filter_w = isOIHW ? filter_shape[3] : filter_shape[1];
+
+    index_t expected_input_height = 0, expected_input_width = 0;
+
+    switch (padding) {
+      case VALID:
+        expected_input_height =
+            (out_height - filter_h) / strides[0] + 1;
+        expected_input_width =
+            (out_width - filter_w) / strides[1] + 1;
+        break;
+      case SAME:
+        expected_input_height =
+            (out_height - 1) / strides[0] + 1;
+        expected_input_width =
+            (out_width - 1) / strides[1] + 1;
+        break;
+      default:
+        MACE_CHECK(false, "Unsupported padding type: ", padding);
+    }
+
+    MACE_CHECK(expected_input_height == in_height,
+               expected_input_height, "!=", in_height);
+    MACE_CHECK(expected_input_width == in_width,
+               expected_input_width, "!=", in_width);
+
+    const int p_h = static_cast<int>(out_height +
+        filter_h - 1 - extended_input_height);
+    const int p_w = static_cast<int>(out_width +
+        filter_w - 1 - extended_input_width);
+
+    padding_size[0] = std::max<int>(0, p_h);
+    padding_size[1] = std::max<int>(0, p_w);
+  }
+
+  const int *strides_;  // [stride_h, stride_w]
+  const Padding padding_type_;
+  std::vector<int> paddings_;
+  const ActivationType activation_;
+  const float relux_max_limit_;
+  std::vector<index_t> output_shape_;
+};
+
+template <DeviceType D, typename T>
+struct Deconv2dFunctor : Deconv2dFunctorBase {
+  Deconv2dFunctor(const int *strides,
+                  const Padding &padding_type,
+                  const std::vector<int> &paddings,
+                  const std::vector<index_t> &output_shape,
+                  const ActivationType activation,
+                  const float relux_max_limit,
+                  const bool is_filter_transformed,
+                  ScratchBuffer *scratch)
+      : Deconv2dFunctorBase(strides,
+                            padding_type,
+                            paddings,
+                            output_shape,
+                            activation,
+                            relux_max_limit) {}
+
+  void operator()(const Tensor *input,   // NCHW
+                  const Tensor *filter,  // OIHW
+                  const Tensor *bias,
+                  Tensor *output,
+                  StatsFuture *future) {
+    MACE_CHECK_NOTNULL(input);
+    MACE_CHECK_NOTNULL(filter);
+    MACE_CHECK_NOTNULL(output);
+
+    std::vector<index_t> output_shape(4);
+    if (output_shape_.size() == 4) {
+      output_shape[0] = output_shape_[0];
+      output_shape[1] = output_shape_[3];
+      output_shape[2] = output_shape_[1];
+      output_shape[3] = output_shape_[2];
+      paddings_.clear();
+      paddings_ = std::vector<int>(2, 0);
+      CalcDeconvPaddingAndInputSize(
+          input->shape().data(),
+          filter->shape().data(),
+          strides_, padding_type_,
+          output_shape.data(),
+          paddings_.data(), true, true);
+      output->Resize(output_shape);
+    } else {
+      output_shape_.clear();
+      output_shape_ = std::vector<index_t>(4, 0);
+      CalcDeconvOutputSize(input->shape().data(),
+                           filter->shape().data(),
+                           strides_,
+                           output_shape_.data(),
+                           paddings_.data(), true, true);
+      output->Resize(output_shape_);
+    }
+    index_t batch = output->dim(0);
+    index_t channels = output->dim(1);
+    index_t height = output->dim(2);
+    index_t width = output->dim(3);
+
+    index_t input_batch = input->dim(0);
+    index_t input_channels = input->dim(1);
+    index_t input_height = input->dim(2);
+    index_t input_width = input->dim(3);
+
+    index_t kernel_h = filter->dim(2);
+    index_t kernel_w = filter->dim(3);
+    MACE_CHECK(filter->dim(0) == channels, filter->dim(0), " != ", channels);
+    MACE_CHECK(filter->dim(1) == input_channels, filter->dim(1), " != ",
+               input_channels);
+
+    index_t stride_h = strides_[0];
+    index_t stride_w = strides_[1];
+
+    MACE_CHECK(batch == input_batch, "Input/Output batch size mismatch");
+    Tensor::MappingGuard input_mapper(input);
+    Tensor::MappingGuard filter_mapper(filter);
+    Tensor::MappingGuard bias_mapper(bias);
+    Tensor::MappingGuard output_mapper(output);
+    auto input_data = input->data<T>();
+    auto filter_data = filter->data<T>();
+    auto bias_data = bias == nullptr ? nullptr : bias->data<T>();
+    auto output_data = output->mutable_data<T>();
+    int padding_top = (paddings_[0] + 1) >> 1;
+    int padding_left = (paddings_[1] + 1) >> 1;
+
+    deconv::Deconv2dNCHW(input_data, filter_data, bias_data,
+                         batch, input_height, input_width, input_channels,
+                         height, width, channels,
+                         kernel_h, kernel_w,
+                         stride_h, stride_w, padding_top, padding_left,
+                         output_data);
+
+    DoActivation(output_data, output_data, output->size(), activation_,
+                 relux_max_limit_);
+  }
+};
+
+template <typename T>
+struct Deconv2dFunctor<DeviceType::GPU, T> : Deconv2dFunctorBase {
+  Deconv2dFunctor(const int *strides,
+                  const Padding &padding_type,
+                  const std::vector<int> &paddings,
+                  const std::vector<index_t> &output_shape,
+                  const ActivationType activation,
+                  const float relux_max_limit,
+                  const bool is_filter_transformed,
+                  ScratchBuffer *scratch)
+      : Deconv2dFunctorBase(strides,
+                            padding_type,
+                            paddings,
+                            output_shape,
+                            activation,
+                            relux_max_limit) {}
+
+  void operator()(const Tensor *input,
+                  const Tensor *filter,
+                  const Tensor *bias,
+                  Tensor *output,
+                  StatsFuture *future);
+
+  cl::Kernel kernel_;
+  uint32_t kwg_size_;
+  std::unique_ptr<BufferBase> kernel_error_;
+  std::vector<index_t> input_shape_;
+};
+
+}  // namespace kernels
+}  // namespace mace
+
+#endif  // MACE_KERNELS_DECONV_2D_H_

mace/kernels/eltwise.h

@@ -40,7 +40,8 @@ enum EltwiseType {
   NEG = 6,
   ABS = 7,
   SQR_DIFF = 8,
-  NONE = 9,
+  POW = 9,
+  NONE = 10,
 };
 
 inline void TensorScalar(const EltwiseType type,
@@ -103,12 +104,18 @@ inline void TensorScalar(const EltwiseType type,
         output[i] = std::pow(input0[i] - value, 2.f);
       }
       break;
+    case POW:
+#pragma omp parallel for
+      for (index_t i = 0; i < size; ++i) {
+        output[i] = std::pow(input0[i], value);
+      }
+      break;
     default:
       LOG(FATAL) << "Eltwise op not support type " << type;
   }
 }
 
-inline void TensorVector(const EltwiseType type,
+inline void TensorBatchVector(const EltwiseType type,
                               const float *input0,
                               const float *input1,
                               const index_t batch,
@@ -227,6 +234,153 @@ inline void TensorVector(const EltwiseType type,
         }
       }
       break;
+    case POW:
+#pragma omp parallel for collapse(3)
+      for (index_t b = 0; b < batch; ++b) {
+        for (index_t c = 0; c < channel; ++c) {
+          for (index_t i = 0; i < hw; ++i) {
+            const index_t idx0 = (b * channel + c) * hw + i;
+            const index_t idx1 = b * channel + c;
+            output[idx0] = std::pow(input0[idx0], input1[idx1]);
+          }
+        }
+      }
+      break;
+    default:
+      LOG(FATAL) << "Eltwise op not support type " << type;
+  }
+}
+inline void TensorVector(const EltwiseType type,
+                         const float *input0,
+                         const float *input1,
+                         const index_t batch,
+                         const index_t channel,
+                         const index_t hw,
+                         const bool swapped,
+                         float *output) {
+  switch (type) {
+    case SUM:
+#pragma omp parallel for collapse(3)
+      for (index_t b = 0; b < batch; ++b) {
+        for (index_t c = 0; c < channel; ++c) {
+          for (index_t i = 0; i < hw; ++i) {
+            const index_t idx0 = (b * channel + c) * hw + i;
+            const index_t idx1 = c;
+            output[idx0] = input0[idx0] + input1[idx1];
+          }
+        }
+      }
+      break;
+    case SUB:
+      if (swapped) {
+#pragma omp parallel for collapse(3)
+        for (index_t b = 0; b < batch; ++b) {
+          for (index_t c = 0; c < channel; ++c) {
+            for (index_t i = 0; i < hw; ++i) {
+              const index_t idx0 = (b * channel + c) * hw + i;
+              const index_t idx1 = c;
+              output[idx0] = input1[idx1] - input0[idx0];
+            }
+          }
+        }
+      } else {
+#pragma omp parallel for collapse(3)
+        for (index_t b = 0; b < batch; ++b) {
+          for (index_t c = 0; c < channel; ++c) {
+            for (index_t i = 0; i < hw; ++i) {
+              const index_t idx0 = (b * channel + c) * hw + i;
+              const index_t idx1 = c;
+              output[idx0] = input0[idx0] - input1[idx1];
+            }
+          }
+        }
+      }
+      break;
+    case PROD:
+#pragma omp parallel for collapse(3)
+      for (index_t b = 0; b < batch; ++b) {
+        for (index_t c = 0; c < channel; ++c) {
+          for (index_t i = 0; i < hw; ++i) {
+            const index_t idx0 = (b * channel + c) * hw + i;
+            const index_t idx1 = c;
+            output[idx0] = input0[idx0] * input1[idx1];
+          }
+        }
+      }
+      break;
+    case DIV:
+      if (swapped) {
+#pragma omp parallel for collapse(3)
+        for (index_t b = 0; b < batch; ++b) {
+          for (index_t c = 0; c < channel; ++c) {
+            for (index_t i = 0; i < hw; ++i) {
+              const index_t idx0 = (b * channel + c) * hw + i;
+              const index_t idx1 = c;
+              output[idx0] = input1[idx1] / input0[idx0];
+            }
+          }
+        }
+      } else {
+#pragma omp parallel for collapse(3)
+        for (index_t b = 0; b < batch; ++b) {
+          for (index_t c = 0; c < channel; ++c) {
+            for (index_t i = 0; i < hw; ++i) {
+              const index_t idx0 = (b * channel + c) * hw + i;
+              const index_t idx1 = c;
+              output[idx0] = input0[idx0] / input1[idx1];
+            }
+          }
+        }
+      }
+      break;
+    case MIN:
+#pragma omp parallel for collapse(3)
+      for (index_t b = 0; b < batch; ++b) {
+        for (index_t c = 0; c < channel; ++c) {
+          for (index_t i = 0; i < hw; ++i) {
+            const index_t idx0 = (b * channel + c) * hw + i;
+            const index_t idx1 = c;
+            output[idx0] = std::min<float>(input0[idx0], input1[idx1]);
+          }
+        }
+      }
+      break;
+    case MAX:
+#pragma omp parallel for collapse(3)
+      for (index_t b = 0; b < batch; ++b) {
+        for (index_t c = 0; c < channel; ++c) {
+          for (index_t i = 0; i < hw; ++i) {
+            const index_t idx0 = (b * channel + c) * hw + i;
+            const index_t idx1 = c;
+            output[idx0] = std::max<float>(input0[idx0], input1[idx1]);
+          }
+        }
+      }
+      break;
+    case SQR_DIFF:
+#pragma omp parallel for collapse(3)
+      for (index_t b = 0; b < batch; ++b) {
+        for (index_t c = 0; c < channel; ++c) {
+          for (index_t i = 0; i < hw; ++i) {
+            const index_t idx0 = (b * channel + c) * hw + i;
+            const index_t idx1 = c;
+            output[idx0] = std::pow(input0[idx0] - input1[idx1], 2.f);
+          }
+        }
+      }
+      break;
+    case POW:
+#pragma omp parallel for collapse(3)
+      for (index_t b = 0; b < batch; ++b) {
+        for (index_t c = 0; c < channel; ++c) {
+          for (index_t i = 0; i < hw; ++i) {
+            const index_t idx0 = (b * channel + c) * hw + i;
+            const index_t idx1 = c;
+            output[idx0] = std::pow(input0[idx0], input1[idx1]);
+          }
+        }
+      }
+      break;
     default:
       LOG(FATAL) << "Eltwise op not support type " << type;
   }
@@ -279,6 +433,12 @@ inline void TensorEltwise(const EltwiseType type,
         output[i] = std::pow(input0[i] - input1[i], 2.f);
       }
       break;
+    case POW:
+#pragma omp parallel for
+      for (index_t i = 0; i < size; ++i) {
+        output[i] = std::pow(input0[i], input1[i]);
+      }
+      break;
     default:
       LOG(FATAL) << "Eltwise op not support type " << type;
   }
@@ -312,20 +472,27 @@ struct EltwiseFunctor<DeviceType::CPU, float>: EltwiseFunctorBase {
                   StatsFuture *future) {
     bool swapped = false;
     if (input1 != nullptr) {
-      MACE_CHECK(input0->dim_size() == input1->dim_size())
+      MACE_CHECK(input0->dim_size() == input1->dim_size()
+                     || input0->dim_size() == 1
+                     || input1->dim_size() == 1)
         << "Inputs of Eltwise op must be same shape";
       if (input0->size() != input1->size()) {
         if (input0->size() < input1->size()) {
           std::swap(input0, input1);
           swapped = true;
         }
-        MACE_CHECK(input0->dim(0) == input1->dim(0) &&
-            input0->dim(1) == input1->dim(1) &&
-            input1->dim(2) == 1 &&
-            input1->dim(3) == 1)
+        if (input1->dim_size() == 1) {
+          MACE_CHECK(input0->dim(1) == input1->dim(0))
+            << "Element-Wise op only support channel dimension broadcast";
+        } else {
+          MACE_CHECK((input0->dim(0) == input1->dim(0) || input1->dim(0) == 1)
+                         && input0->dim(1) == input1->dim(1)
+                         && input1->dim(2) == 1
+                         && input1->dim(3) == 1)
             << "Element-Wise op only support channel dimension broadcast";
         }
       }
+    }
     output->ResizeLike(input0);
 
     Tensor::MappingGuard input0_guard(input0);
@@ -344,8 +511,12 @@ struct EltwiseFunctor<DeviceType::CPU, float>: EltwiseFunctorBase {
         const index_t batch = input0->dim(0);
         const index_t channel = input0->dim(1);
         const index_t hw = input0->dim(2) * input0->dim(3);
+        if (input1->dim(0) == 1 || input1->dim_size() == 1)
           TensorVector(type_, input0_ptr, input1_ptr,
                        batch, channel, hw, swapped, output_ptr);
+        else
+          TensorBatchVector(type_, input0_ptr, input1_ptr,
+                            batch, channel, hw, swapped, output_ptr);
       } else {
         if (!coeff_.empty() && type_ == SUM) {
 #pragma omp parallel for

mace/kernels/opencl/cl/deconv_2d.cl

+#include <common.h>
+
+__kernel void deconv_2d(KERNEL_ERROR_PARAMS
+                        GLOBAL_WORK_GROUP_SIZE_DIM3
+                        __read_only image2d_t input,
+                        __read_only image2d_t weights,
+#ifdef BIAS
+                        __read_only image2d_t bias,
+#endif
+                        __write_only image2d_t output,
+                        __private const float relux_max_limit,
+                        __private const int in_height,
+                        __private const int in_width,
+                        __private const int in_channels,
+                        __private const int out_height,
+                        __private const int out_width,
+                        __private const int out_channel,
+                        __private const int stride,
+                        __private const float stride_r,
+                        __private const int align_h,
+                        __private const int align_w,
+                        __private const int padding_h,
+                        __private const int padding_w,
+                        __private const int kernel_h,
+                        __private const int kernel_w,
+                        __private const int kernel_size,
+                        __private const int in_channel_blocks,
+                        __private const int out_channel_blocks)
+{
+  const int c = get_global_id(0);
+  const int w_id = get_global_id(1);
+  const int hb = get_global_id(2);
+
+#ifndef NON_UNIFORM_WORK_GROUP
+  if (c >= global_size_dim0 || w_id >= global_size_dim1
+      || hb >= global_size_dim2) {
+    return;
+  }
+#endif
+
+#ifdef BIAS
+  DATA_TYPE4 out0 =
+     READ_IMAGET(bias, SAMPLER, (int2)(c, 0));
+  DATA_TYPE4 out1 = out0;
+  DATA_TYPE4 out2 = out0;
+  DATA_TYPE4 out3 = out0;
+  DATA_TYPE4 out4 = out0;
+#else
+  DATA_TYPE4 out0 = 0;
+  DATA_TYPE4 out1 = 0;
+  DATA_TYPE4 out2 = 0;
+  DATA_TYPE4 out3 = 0;
+  DATA_TYPE4 out4 = 0;
+#endif
+
+  const int n_stride = mad(w_id, stride_r, 0);
+  const int mod_stride = w_id - mul24(n_stride, stride);
+  const int w = mad24(mul24(n_stride, 5), stride, mod_stride);
+  const int b = hb / out_height;
+  const int h = hb - mul24(b, out_height);
+  if (w < out_width) {
+    int start_x = floor((float) (w + align_w) * stride_r);
+    int start_y = (h + align_h) * stride_r;
+    start_y = max(0, start_y);
+
+    int f_start_x = mad24(start_x, stride, padding_w) - w;
+    int f_start_y = mad24(start_y, stride, padding_h) - h;
+    f_start_x = kernel_w - 1 - f_start_x;
+    f_start_y = kernel_h - 1 - f_start_y;
+
+    int2 in_pos;
+    int f_pos_x0, f_pos_x1, f_pos_x2, f_pos_x3, f_pos_y;
+    DATA_TYPE4 in0, in1, in2, in3, in4;
+    DATA_TYPE4 weight0, weight1, weight2, weight3;
+    int idx_w0, idx_w1, idx_w2, idx_w3, idx_w4;
+    int index_x, index_y;
+    for (int ic = 0; ic < in_channel_blocks; ++ic) {
+      f_pos_x0 = mul24(ic, 4);
+      f_pos_x1 = f_pos_x0 + 1;
+      f_pos_x2 = f_pos_x0 + 2;
+      f_pos_x3 = f_pos_x0 + 3;
+      for (int f_y = f_start_y, idx_h = start_y ; f_y >= 0; f_y -= stride, ++idx_h) {
+        index_y = mad24(b, in_height, idx_h);
+        in_pos.y = select(index_y, -1, idx_h < 0 || idx_h >= in_height);
+        for (int f_x = f_start_x, idx_w = start_x; f_x >= 0; f_x -= stride, ++idx_w) {
+          f_pos_y = mad24(f_y, kernel_w, f_x);
+          f_pos_y = mad24(c, kernel_size, f_pos_y);
+          weight0 = READ_IMAGET(weights, SAMPLER, (int2)(f_pos_x0, f_pos_y));
+          weight1 = READ_IMAGET(weights, SAMPLER, (int2)(f_pos_x1, f_pos_y));
+          weight2 = READ_IMAGET(weights, SAMPLER, (int2)(f_pos_x2, f_pos_y));
+          weight3 = READ_IMAGET(weights, SAMPLER, (int2)(f_pos_x3, f_pos_y));
+
+          idx_w0 = idx_w;
+          idx_w1 = idx_w + 1;
+          idx_w2 = idx_w + 2;
+          idx_w3 = idx_w + 3;
+          idx_w4 = idx_w + 4;
+
+#define READ_INPUT(i)                                                         \
+          index_x = mad24(ic, in_width, idx_w##i);                            \
+          in_pos.x =                                                          \
+            select(index_x, -1, idx_w##i < 0 || idx_w##i >= in_width);        \
+          in##i = READ_IMAGET(input, SAMPLER, in_pos);
+
+          READ_INPUT(0);
+          READ_INPUT(1);
+          READ_INPUT(2);
+          READ_INPUT(3);
+          READ_INPUT(4);
+#undef READ_INPUT
+
+#define CALC_OUTPUT(i)                                                        \
+          out##i = mad(in##i.x, weight0, out##i);                             \
+          out##i = mad(in##i.y, weight1, out##i);                             \
+          out##i = mad(in##i.z, weight2, out##i);                             \
+          out##i = mad(in##i.w, weight3, out##i);
+
+          CALC_OUTPUT(0);
+          CALC_OUTPUT(1);
+          CALC_OUTPUT(2);
+          CALC_OUTPUT(3);
+          CALC_OUTPUT(4);
+#undef CALC_OUTPUT
+        }
+      }
+    }
+
+#if defined(USE_RELU) || defined(USE_RELUX) || defined(USE_TANH) || defined(USE_SIGMOID)
+    out0 = do_activation(out0, relux_max_limit);
+    out1 = do_activation(out1, relux_max_limit);
+    out2 = do_activation(out2, relux_max_limit);
+    out3 = do_activation(out3, relux_max_limit);
+    out4 = do_activation(out4, relux_max_limit);
+#endif
+
+    int2 out_pos;
+    out_pos.y = hb;
+
+    int ow = w;
+    if (ow >= out_width) return;
+    out_pos.x = mad24(c, out_width, ow);
+    WRITE_IMAGET(output, out_pos, out0);
+
+    ow += stride;
+    if (ow >= out_width) return;
+    out_pos.x += stride;
+    WRITE_IMAGET(output, out_pos, out1);
+
+    ow += stride;
+    if (ow >= out_width) return;
+    out_pos.x += stride;
+    WRITE_IMAGET(output, out_pos, out2);
+
+    ow += stride;
+    if (ow >= out_width) return;
+    out_pos.x += stride;
+    WRITE_IMAGET(output, out_pos, out3);
+
+    ow += stride;
+    if (ow >= out_width) return;
+    out_pos.x += stride;
+    WRITE_IMAGET(output, out_pos, out4);
+  }
+}
\ No newline at end of file

mace/kernels/opencl/cl/eltwise.cl

@@ -33,6 +33,8 @@ __kernel void eltwise(KERNEL_ERROR_PARAMS
 #elif INPUT_TYPE == 2
   const int batch_idx = hb / height;
   DATA_TYPE4 in1 = READ_IMAGET(input1, SAMPLER, (int2)(chan_idx, batch_idx));
+#elif INPUT_TYPE == 3
+  DATA_TYPE4 in1 = READ_IMAGET(input1, SAMPLER, (int2)(chan_idx, 0));
 #else
   DATA_TYPE4 in1 = READ_IMAGET(input1, SAMPLER, (int2)(pos, hb));
 #endif
@@ -70,10 +72,17 @@ __kernel void eltwise(KERNEL_ERROR_PARAMS
 #elif ELTWISE_TYPE == 8
   DATA_TYPE4 diff = in0 - in1;
   out = diff * diff;
+#elif ELTWISE_TYPE == 9
+  #ifdef SWAPPED
+    out = pow(in0, in1);
+  #else
+    out = pow(in1, in0);
+  #endif
 #endif
 
 #if INPUT_TYPE == 1
-  #if ELTWISE_TYPE == 0 || ELTWISE_TYPE == 1 || ELTWISE_TYPE == 4 || ELTWISE_TYPE == 5 || ELTWISE_TYPE == 8
+  #if ELTWISE_TYPE == 0 || ELTWISE_TYPE == 1 || ELTWISE_TYPE == 4 ||          \
+      ELTWISE_TYPE == 5 || ELTWISE_TYPE == 8 || ELTWISE_TYPE == 9
     const int remain_channel = channel - 4 * chan_idx;
     if (remain_channel < 4) {
       switch (remain_channel) {

mace/kernels/opencl/deconv_2d_opencl.cc

+// Copyright 2018 Xiaomi, Inc.  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 "mace/kernels/deconv_2d.h"
+#include "mace/kernels/opencl/helper.h"
+
+namespace mace {
+namespace kernels {
+
+namespace {
+
+void Deconv2dOpencl(cl::Kernel *kernel,
+                    const Tensor *input,
+                    const Tensor *filter,
+                    const Tensor *bias,
+                    const int stride,
+                    const int *paddings,
+                    const ActivationType activation,
+                    const float relux_max_limit,
+                    const DataType dt,
+                    std::vector<index_t> *prev_input_shape,
+                    Tensor *output,
+                    StatsFuture *future,
+                    uint32_t *kwg_size,
+                    std::unique_ptr<BufferBase> *kernel_error) {
+  const index_t batch = output->dim(0);
+  const index_t height = output->dim(1);
+  const index_t width = output->dim(2);
+  const index_t channels = output->dim(3);
+  const index_t input_channels = input->dim(3);
+
+  const index_t channel_blocks = RoundUpDiv4(channels);
+  const index_t input_channel_blocks = RoundUpDiv4(input_channels);
+  MACE_CHECK(stride > 0, "stride should > 0.");
+#define WIDTH_BLK 5
+  const index_t n_strides = (width + stride - 1) / stride;
+  const index_t width_blocks = ((n_strides + WIDTH_BLK -1)/ WIDTH_BLK) * stride;
+  const float stride_r = 1.f / static_cast<float>(stride);
+  const int padding_h = (paddings[0]+1) >> 1;
+  const int padding_w = (paddings[0]+1) >> 1;
+
+  const int align_h = stride - 1 - padding_h;
+  const int align_w = stride - 1 - padding_w;
+  const int kernel_size = filter->dim(0) * filter->dim(1);
+
+  auto runtime = OpenCLRuntime::Global();
+
+  if (kernel->get() == nullptr) {
+    std::set<std::string> built_options;
+    std::string kernel_name = MACE_OBFUSCATE_SYMBOL("deconv_2d");
+    built_options.emplace("-Ddeconv_2d=" + kernel_name);
+    built_options.emplace("-DDATA_TYPE=" + DtToUpstreamCLDt(dt));
+    built_options.emplace("-DCMD_DATA_TYPE=" + DtToUpstreamCLCMDDt(dt));
+    if (runtime->IsOutOfRangeCheckEnabled()) {
+      built_options.emplace("-DOUT_OF_RANGE_CHECK");
+      *kernel_error = std::move(std::unique_ptr<Buffer>(
+          new Buffer(GetDeviceAllocator(DeviceType::GPU), 1)));
+      (*kernel_error)->Map(nullptr);
+      *((*kernel_error)->mutable_data<char>()) = 0;
+      (*kernel_error)->UnMap();
+    }
+    if (runtime->IsNonUniformWorkgroupsSupported()) {
+      built_options.emplace("-DNON_UNIFORM_WORK_GROUP");
+    }
+    built_options.emplace(bias != nullptr ? "-DBIAS" : "");
+    switch (activation) {
+      case NOOP:break;
+      case RELU:built_options.emplace("-DUSE_RELU");
+        break;
+      case RELUX:built_options.emplace("-DUSE_RELUX");
+        break;
+      case TANH:built_options.emplace("-DUSE_TANH");
+        break;
+      case SIGMOID:built_options.emplace("-DUSE_SIGMOID");
+        break;
+      default:LOG(FATAL) << "Unknown activation type: " << activation;
+    }
+
+    *kernel = runtime->BuildKernel("deconv_2d", kernel_name, built_options);
+
+    *kwg_size =
+        static_cast<uint32_t>(runtime->GetKernelMaxWorkGroupSize(*kernel));
+  }
+
+  const uint32_t gws[3] = {static_cast<uint32_t>(channel_blocks),
+                           static_cast<uint32_t>(width_blocks),
+                           static_cast<uint32_t>(height * batch)};
+
+  if (!IsVecEqual(*prev_input_shape, input->shape())) {
+    uint32_t idx = 0;
+    if (runtime->IsOutOfRangeCheckEnabled()) {
+      kernel->setArg(idx++,
+                     *(static_cast<cl::Buffer *>((*kernel_error)->buffer())));
+    }
+    if (!runtime->IsNonUniformWorkgroupsSupported()) {
+      kernel->setArg(idx++, gws[0]);
+      kernel->setArg(idx++, gws[1]);
+      kernel->setArg(idx++, gws[2]);
+    }
+    kernel->setArg(idx++, *(input->opencl_image()));
+    kernel->setArg(idx++, *(filter->opencl_image()));
+    if (bias != nullptr) {
+      kernel->setArg(idx++, *(bias->opencl_image()));
+    }
+    kernel->setArg(idx++, *(output->opencl_image()));
+    kernel->setArg(idx++, relux_max_limit);
+    kernel->setArg(idx++, static_cast<int32_t>(input->dim(1)));
+    kernel->setArg(idx++, static_cast<int32_t>(input->dim(2)));
+    kernel->setArg(idx++, static_cast<int32_t>(input->dim(3)));
+    kernel->setArg(idx++, static_cast<int32_t>(height));
+    kernel->setArg(idx++, static_cast<int32_t>(width));
+    kernel->setArg(idx++, static_cast<int32_t>(channels));
+    kernel->setArg(idx++, static_cast<int32_t>(stride));
+    kernel->setArg(idx++, stride_r);
+    kernel->setArg(idx++, static_cast<int32_t>(align_h));
+    kernel->setArg(idx++, static_cast<int32_t>(align_w));
+    kernel->setArg(idx++, static_cast<int32_t>(padding_h));
+    kernel->setArg(idx++, static_cast<int32_t>(padding_w));
+    kernel->setArg(idx++, static_cast<int32_t>(filter->dim(0)));
+    kernel->setArg(idx++, static_cast<int32_t>(filter->dim(1)));
+    kernel->setArg(idx++, static_cast<int32_t>(kernel_size));
+    kernel->setArg(idx++, static_cast<int32_t>(input_channel_blocks));
+    kernel->setArg(idx++, static_cast<int32_t>(channel_blocks));
+
+    *prev_input_shape = input->shape();
+  }
+
+  const std::vector<uint32_t> lws = {8, *kwg_size / 64, 8, 0};
+  std::string tuning_key =
+      Concat("deconv2d_opencl_kernel_", activation, output->dim(0),
+             output->dim(1), output->dim(2), output->dim(3));
+  TuningOrRun3DKernel(*kernel, tuning_key, gws, lws, future);
+
+  if (runtime->IsOutOfRangeCheckEnabled()) {
+    (*kernel_error)->Map(nullptr);
+    char *kerror_code = (*kernel_error)->mutable_data<char>();
+    MACE_CHECK(*kerror_code == 0) << "Kernel error code: " << *kerror_code;
+    (*kernel_error)->UnMap();
+  }
+}
+
+}  // namespace
+
+template <typename T>
+void Deconv2dFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
+                                                      const Tensor *filter,
+                                                      const Tensor *bias,
+                                                      Tensor *output,
+                                                      StatsFuture *future) {
+  MACE_CHECK_NOTNULL(input);
+  MACE_CHECK_NOTNULL(filter);
+  MACE_CHECK_NOTNULL(output);
+
+  if (output_shape_.size() == 4) {
+    paddings_.clear();
+    paddings_ = std::vector<int>(2, 0);
+    CalcDeconvPaddingAndInputSize(
+        input->shape().data(),
+        filter->shape().data(),
+        strides_, padding_type_,
+        output_shape_.data(),
+        paddings_.data());
+  } else {
+    output_shape_.clear();
+    output_shape_ = std::vector<index_t>(4, 0);
+    CalcDeconvOutputSize(input->shape().data(),
+                         filter->shape().data(),
+                         strides_,
+                         output_shape_.data(),
+                         paddings_.data());
+  }
+
+  std::vector<size_t> output_image_shape;
+  CalImage2DShape(output_shape_, BufferType::IN_OUT_CHANNEL,
+                  &output_image_shape);
+  output->ResizeImage(output_shape_, output_image_shape);
+
+  Deconv2dOpencl(&kernel_, input, filter, bias,
+                 strides_[0], paddings_.data(),
+                 activation_, relux_max_limit_,
+                 DataTypeToEnum<T>::value, &input_shape_,
+                 output, future, &kwg_size_, &kernel_error_);
+}
+
+template struct Deconv2dFunctor<DeviceType::GPU, float>;
+template struct Deconv2dFunctor<DeviceType::GPU, half>;
+
+}  // namespace kernels
+}  // namespace mace

mace/kernels/opencl/eltwise.cc

@@ -27,21 +27,40 @@ void EltwiseFunctor<DeviceType::GPU, T>::operator()(const Tensor *input0,
                                                        StatsFuture *future) {
   bool swapped = false;
   if (input1 != nullptr) {
-    MACE_CHECK(input0->dim_size() == input1->dim_size())
+    MACE_CHECK(input0->dim_size() == input1->dim_size()
+                   || input0->dim_size() == 1
+                   || input1->dim_size() == 1)
       << "Inputs of Eltwise op must be same shape";
     if (input0->size() != input1->size()) {
       if (input0->size() < input1->size()) {
         std::swap(input0, input1);
         swapped = true;
       }
-      MACE_CHECK(input0->dim(0) == input1->dim(0) &&
+      if (input1->dim_size() == 1) {
+        MACE_CHECK(input0->dim(3) == input1->dim(0))
+          << "Element-Wise op only support channel dimension broadcast";
+      } else {
+        MACE_CHECK((input0->dim(0) == input1->dim(0) || input1->dim(0) == 1) &&
+            input0->dim(3) == input1->dim(3) &&
             input1->dim(1) == 1 &&
-          input1->dim(2) == 1 &&
-          input0->dim(3) == input1->dim(3))
+            input1->dim(2) == 1)
           << "Element-Wise op only support channel dimension broadcast";
       }
     }
-  output->ResizeLike(input0);
+  }
+
+  std::vector<index_t > output_shape(4);
+  output_shape[0] = input0->dim(0);
+  output_shape[1] = input0->dim(1);
+  output_shape[2] = input0->dim(2);
+  output_shape[3] = input0->dim(3);
+
+  std::vector<size_t> output_image_shape;
+  CalImage2DShape(output_shape,
+                  BufferType::IN_OUT_CHANNEL,
+                  &output_image_shape);
+  output->ResizeImage(output_shape, output_image_shape);
+
   const index_t batch = output->dim(0);
   const index_t height = output->dim(1);
   const index_t width = output->dim(2);
@@ -66,6 +85,9 @@ void EltwiseFunctor<DeviceType::GPU, T>::operator()(const Tensor *input0,
     if (input1 == nullptr) {
       built_options.emplace("-DINPUT_TYPE=1");
     } else if (input0->size() != input1->size()) {
+      if (input1->dim(0) == 1 || input1->dim_size() == 1)
+        built_options.emplace("-DINPUT_TYPE=3");
+      else
         built_options.emplace("-DINPUT_TYPE=2");
       if (swapped) built_options.emplace("-DSWAPPED");
     }

mace/kernels/opencl/out_of_range_check_test.cc

@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <memory>
 #include <vector>
 
 #include "gtest/gtest.h"

mace/ops/conv_2d_test.cc

@@ -105,7 +105,8 @@ void TestNHWCSimple3x3SAME() {
     {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1});
   net.AddInputFromArray<D, T>(
     "Filter", {3, 3, 1, 2},
-    {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
+    {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
+     1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
      1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f});
   net.AddInputFromArray<D, T>("Bias", {1}, {0.1f});
 
@@ -191,7 +192,8 @@ void TestNHWCSimple3x3WithoutBias() {
     {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1});
   net.AddInputFromArray<D, T>(
     "Filter", {3, 3, 1, 2},
-    {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
+    {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
+     1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
      1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f});
 
   if (D == DeviceType::CPU) {
@@ -351,10 +353,12 @@ void TestFusedNHWCSimple3x3VALID() {
   // Add input data
   net.AddInputFromArray<D, float>(
       "Input", {1, 3, 3, 2},
-      {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1});
+      {-1, -1, -1, -1, -1, -1, -1, -1, -1,
+       -1, -1, -1, -1, -1, -1, -1, -1, -1});
   net.AddInputFromArray<D, float>(
       "Filter", {3, 3, 1, 2},
-      {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
+      {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
+       1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
        1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f});
   net.AddInputFromArray<D, float>("Bias", {1}, {-0.1f});
 
@@ -423,10 +427,13 @@ void TestFusedNHWCSimple3x3WithoutBias() {
   // Add input data
   net.AddInputFromArray<D, float>(
       "Input", {1, 3, 3, 2},
-      {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1});
+      {-1, -1, -1, -1, -1, -1,
+       -1, -1, -1, -1, -1, -1,
+       -1, -1, -1, -1, -1, -1});
   net.AddInputFromArray<D, float>(
       "Filter", {3, 3, 1, 2},
-      {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
+      {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
+       1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
        1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f});
 
   if (D == DeviceType::CPU) {

mace/ops/deconv_2d.cc

+// Copyright 2018 Xiaomi, Inc.  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 "mace/ops/deconv_2d.h"
+
+namespace mace {
+namespace ops {
+
+void Register_Deconv2D(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Deconv2D")
+                                     .Device(DeviceType::CPU)
+                                     .TypeConstraint<float>("T")
+                                     .Build(),
+                    Deconv2dOp<DeviceType::CPU, float>);
+
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Deconv2D")
+                                     .Device(DeviceType::GPU)
+                                     .TypeConstraint<float>("T")
+                                     .Build(),
+                    Deconv2dOp<DeviceType::GPU, float>);
+
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Deconv2D")
+                                     .Device(DeviceType::GPU)
+                                     .TypeConstraint<half>("T")
+                                     .Build(),
+                    Deconv2dOp<DeviceType::GPU, half>);
+}
+
+}  // namespace ops
+}  // namespace mace

mace/ops/deconv_2d.h

+// Copyright 2018 Xiaomi, Inc.  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.
+
+#ifndef MACE_OPS_DECONV_2D_H_
+#define MACE_OPS_DECONV_2D_H_
+
+#include <memory>
+
+#include "mace/core/operator.h"
+#include "mace/kernels/deconv_2d.h"
+#include "mace/ops/conv_pool_2d_base.h"
+
+namespace mace {
+namespace ops {
+
+template <DeviceType D, typename T>
+class Deconv2dOp : public ConvPool2dOpBase<D, T> {
+ public:
+  Deconv2dOp(const OperatorDef &op_def, Workspace *ws)
+      : ConvPool2dOpBase<D, T>(op_def, ws),
+        functor_(this->strides_.data(),
+                 this->padding_type_,
+                 this->paddings_,
+                 OperatorBase::GetRepeatedArgument<index_t>("output_shape"),
+                 kernels::ActivationType::NOOP,
+                 0.0f,
+                 static_cast<bool>(OperatorBase::GetSingleArgument<int>(
+                     "is_filter_transformed", false)),
+                 ws->GetScratchBuffer(D)) {}
+
+  bool Run(StatsFuture *future) override {
+    const Tensor *input = this->Input(INPUT);
+    const Tensor *filter = this->Input(FILTER);
+    const Tensor *bias = this->InputSize() >= 3 ? this->Input(BIAS) : nullptr;
+    Tensor *output = this->Output(OUTPUT);
+
+    functor_(input, filter, bias, output, future);
+
+    return true;
+  }
+
+ private:
+  kernels::Deconv2dFunctor<D, T> functor_;
+
+ protected:
+  OP_INPUT_TAGS(INPUT, FILTER, BIAS);
+  OP_OUTPUT_TAGS(OUTPUT);
+};
+
+}  // namespace ops
+}  // namespace mace
+
+#endif  // MACE_OPS_DECONV_2D_H_

mace/ops/deconv_2d_benchmark.cc

+// Copyright 2018 Xiaomi, Inc.  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 <algorithm>
+
+#include "mace/core/operator.h"
+#include "mace/core/testing/test_benchmark.h"
+#include "mace/ops/deconv_2d.h"
+#include "mace/ops/ops_test_util.h"
+
+namespace mace {
+namespace ops {
+namespace test {
+
+template <DeviceType D, typename T>
+static void Deconv2d(int iters,
+                   int batch,
+                   int channels,
+                   int height,
+                   int width,
+                   int kernel_h,
+                   int kernel_w,
+                   int stride,
+                   int out_h,
+                   int out_w,
+                   Padding padding,
+                   int output_channels) {
+  mace::testing::StopTiming();
+
+  OpsTestNet net;
+
+  // Add input data
+  if (D == DeviceType::CPU) {
+    net.AddRandomInput<D, float>("Input", {batch, channels, height, width});
+    net.AddRandomInput<D, float>("Filter",
+                                 {output_channels, channels, kernel_h,
+                                  kernel_w});
+  } else {
+    net.AddRandomInput<D, float>("Input", {batch, height, width, channels});
+    net.AddRandomInput<D, float>("Filter",
+                                 {kernel_h, kernel_w, output_channels,
+                                  channels});
+  }
+  if (D == DeviceType::GPU) {
+    BufferToImage<D, T>(&net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT_CHANNEL);
+    BufferToImage<D, T>(&net, "Filter", "FilterImage",
+                        kernels::BufferType::CONV2D_FILTER);
+    OpDefBuilder("Deconv2D", "Deconv2dTest")
+        .Input("InputImage")
+        .Input("FilterImage")
+        .Output("Output")
+        .AddIntsArg("strides", {stride, stride})
+        .AddIntArg("padding", padding)
+        .AddIntsArg("output_shape", {batch, out_h, out_w, output_channels})
+        .AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value))
+        .Finalize(net.NewOperatorDef());
+  } else {
+    OpDefBuilder("Deconv2D", "Deconv2dTest")
+        .Input("Input")
+        .Input("Filter")
+        .Output("Output")
+        .AddIntsArg("strides", {stride, stride})
+        .AddIntArg("padding", padding)
+        .AddIntsArg("output_shape", {batch, out_h, out_w, output_channels})
+        .AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value))
+        .Finalize(net.NewOperatorDef());
+  }
+
+  net.Setup(D);
+
+  // Warm-up
+  for (int i = 0; i < 2; ++i) {
+    net.Run();
+    net.Sync();
+  }
+
+  mace::testing::StartTiming();
+  while (iters--) {
+    net.Run();
+    net.Sync();
+  }
+}
+
+// In common network, there are usually more than 1 layers, this is used to
+// approximate the amortized latency. The OpenCL runtime for Mali/Adreno is
+// in-order.
+
+#define BM_DECONV_2D_MACRO(N, C, H, W, KH, KW, STRIDE, OH, OW, P, OC, TYPE,   \
+                         DEVICE)                                              \
+  static void                                                                 \
+    BM_DECONV_2D_##N##_##C##_##H##_##W##_##KH##_##KW##_##STRIDE##_##OH##_##OW\
+        ##_##P##_##OC##_##TYPE##_##DEVICE(                                    \
+          int iters) {                                                        \
+    const int64_t tot = static_cast<int64_t>(iters) * N * C * H * W;          \
+    int64_t oh = OH;                                                          \
+    int64_t ow = OW;                                                          \
+    const int64_t macc =                                                      \
+        static_cast<int64_t>(iters) * N * OC * oh * ow * (KH * KW * C + 1);   \
+    mace::testing::MaccProcessed(macc);                                       \
+    mace::testing::BytesProcessed(tot *(sizeof(TYPE)));                       \
+    Deconv2d<DEVICE, TYPE>(iters, N, C, H, W, KH, KW, STRIDE, OH, OW,         \
+                         mace::Padding::P, OC);                               \
+  }                                                                           \
+  BENCHMARK(                                                                  \
+    BM_DECONV_2D_##N##_##C##_##H##_##W##_##KH##_##KW##_##STRIDE##_##OH##_##OW##\
+    _##P##_##OC##_##TYPE##_##DEVICE)
+
+#define BM_DECONV_2D(N, C, H, W, KH, KW, S, OH, OW, P, OC)                 \
+  BM_DECONV_2D_MACRO(N, C, H, W, KH, KW, S, OH, OW, P, OC, float, CPU);    \
+  BM_DECONV_2D_MACRO(N, C, H, W, KH, KW, S, OH, OW, P, OC, float, GPU); \
+  BM_DECONV_2D_MACRO(N, C, H, W, KH, KW, S, OH, OW, P, OC, half, GPU);
+
+BM_DECONV_2D(1, 512, 15, 15, 1, 1, 1, 15, 15, VALID, 1024);
+BM_DECONV_2D(1, 32, 60, 60, 1, 1, 1, 60, 60, VALID, 128);
+
+BM_DECONV_2D(1, 128, 60, 60, 3, 3, 1, 62, 62, VALID, 128);
+BM_DECONV_2D(1, 32, 60, 60, 3, 3, 1, 60, 60, SAME, 32);
+BM_DECONV_2D(1, 3, 512, 512, 7, 7, 2, 1023, 1023, SAME, 64);
+BM_DECONV_2D(1, 128, 16, 16, 5, 5, 1, 20, 20, VALID, 32);
+BM_DECONV_2D(1, 128, 64, 64, 5, 5, 1, 68, 68, VALID, 32);
+
+BM_DECONV_2D(1, 3, 480, 480, 1, 1, 1, 480, 480, VALID, 3);
+
+BM_DECONV_2D(1, 64, 32, 32, 1, 1, 1, 32, 32, VALID, 128);
+BM_DECONV_2D(1, 64, 33, 32, 3, 3, 2, 65, 63, SAME, 128);
+BM_DECONV_2D(1, 3, 224, 224, 3, 3, 2, 447, 447, SAME, 32);
+BM_DECONV_2D(1, 3, 224, 224, 3, 3, 2, 449, 449, VALID, 32);
+
+}  // namespace test
+}  // namespace ops
+}  // namespace mace

mace/ops/eltwise_test.cc

@@ -553,6 +553,8 @@ TEST_F(EltwiseOpTest, RandomTensorVecFloat) {
                             {1, 32, 32, 16}, {1, 1, 1, 16});
   RandomTensorEltwise<float>(kernels::EltwiseType::SUB,
                             {5, 32, 32, 16}, {5, 1, 1, 16});
+  RandomTensorEltwise<float>(kernels::EltwiseType::SUB,
+                             {5, 32, 32, 16}, {1, 1, 1, 16});
   RandomTensorEltwise<float>(kernels::EltwiseType::SUB,
                              {5, 1, 1, 16}, {5, 32, 32, 16});
   RandomTensorEltwise<float>(kernels::EltwiseType::PROD,
@@ -574,12 +576,16 @@ TEST_F(EltwiseOpTest, RandomTensorVecHalf) {
                            {1, 32, 32, 16}, {1, 1, 1, 16});
   RandomTensorEltwise<half>(kernels::EltwiseType::SUB,
                            {3, 32, 32, 16}, {3, 1, 1, 16});
+  RandomTensorEltwise<half>(kernels::EltwiseType::SUB,
+                            {3, 32, 32, 16}, {1, 1, 1, 16});
   RandomTensorEltwise<half>(kernels::EltwiseType::SUB,
                             {3, 1, 1, 16}, {3, 32, 32, 16});
   RandomTensorEltwise<half>(kernels::EltwiseType::PROD,
                            {1, 1, 1, 17}, {1, 31, 37, 17});
   RandomTensorEltwise<half>(kernels::EltwiseType::DIV,
                             {5, 31, 37, 17}, {5, 1, 1, 17});
+  RandomTensorEltwise<half>(kernels::EltwiseType::DIV,
+                            {5, 31, 37, 17}, {1, 1, 1, 17});
   RandomTensorEltwise<half>(kernels::EltwiseType::DIV,
                            {5, 1, 1, 17}, {5, 31, 37, 17});
   RandomTensorEltwise<half>(kernels::EltwiseType::MIN,

mace/python/tools/tf_converter_lib.py

@@ -38,6 +38,8 @@ math_type_mode = {
     'MAX': 5,
     'NEG': 6,
     'ABS': 7,
+    'SQR_DIFF': 8,
+    'POW': 9,
 }
 
 buffer_type_map = {
@@ -528,6 +530,103 @@ class TFConverter(object):
         self.add_output_shape(final_op.outputs, op_def)
         self.net_def.op.extend([op_def])
 
+    def convert_deconv2d(self, op):
+        op_def = mace_pb2.OperatorDef()
+        arg = op_def.arg.add()
+        arg.name = 'T'
+        arg.i = self.dt
+        op_def.name = op.name
+        op_def.type = 'Deconv2D'
+
+        out_shape_value = None
+        if len(op.inputs) == 2:
+            out_shape_value = op.get_attr('output_shape')
+            if self.device == 'cpu':
+                self.transpose_filter_tensor[get_input_tensor(
+                    op, 1).name] = (3, 2, 0, 1)
+            else:
+                self.transpose_filter_tensor[get_input_tensor(
+                    op, 1).name] = (0, 1, 3, 2)
+            if self.device == 'gpu':
+                op_def.input.extend([op.inputs[0].name])
+                buffer_type = "CONV2D_FILTER"
+                output_name = self.add_buffer_to_image(
+                    get_input_tensor(op, 1).name, buffer_type)
+                op_def.input.extend([output_name])
+            else:
+                op_def.input.extend(
+                    [get_input_tensor(op, i).name
+                     for i in range(len(op.inputs))])
+        elif len(op.inputs) == 3:
+            out_shape_value = \
+                get_input_tensor(op, 0).eval().astype(np.int32).flat
+            self.unused_tensor.add(op.inputs[0].name)
+            if self.device == 'cpu':
+                self.transpose_filter_tensor[get_input_tensor(
+                    op, 1).name] = (2, 3, 0, 1)
+            else:
+                self.transpose_filter_tensor[get_input_tensor(
+                    op, 1).name] = (0, 1, 2, 3)
+            if self.device == 'gpu':
+                op_def.input.extend([op.inputs[2].name])
+                buffer_type = "CONV2D_FILTER"
+                output_name = self.add_buffer_to_image(
+                    get_input_tensor(op, 1).name, buffer_type)
+                op_def.input.extend([output_name])
+            else:
+                op_def.input.extend([op.inputs[2].name])
+                op_def.input.extend([op.inputs[1].name])
+        else:
+            raise Exception('Too many inputs. Op: %s, type: %s' % (op.name,
+                                                                   op.type))
+        if out_shape_value is not None:
+            out_shape_arg = op_def.arg.add()
+            out_shape_arg.name = 'output_shape'
+            out_shape_arg.ints.extend(out_shape_value)
+        padding_arg = op_def.arg.add()
+        padding_arg.name = 'padding'
+        padding_arg.i = padding_mode[op.get_attr('padding')]
+        strides_arg = op_def.arg.add()
+        strides_arg.name = 'strides'
+        strides_arg.ints.extend(op.get_attr('strides')[1:3])
+        data_format_arg = op_def.arg.add()
+        data_format_arg.name = 'data_format'
+        if self.device == 'cpu':
+            data_format_arg.s = 'NCHW'
+        else:
+            data_format_arg.s = 'NHWC'
+        final_op = op
+        self.resolved_ops[op.name] = 1
+
+        if len(self.tf_graph.get(op.name, [])) == 1 and \
+                self.tf_graph[op.name][0].type == 'BiasAdd':
+            bias_add_op = self.tf_graph[op.name][0]
+            if self.device == 'gpu':
+                output_name = self.add_buffer_to_image(
+                    get_input_tensor(bias_add_op, 1).name, "ARGUMENT")
+                op_def.input.extend([output_name])
+            else:
+                op_def.input.extend([get_input_tensor(bias_add_op, 1).name])
+            final_op = bias_add_op
+            self.resolved_ops[bias_add_op.name] = 1
+
+        if len(self.tf_graph.get(final_op.name, [])) == 1 and \
+                self.tf_graph[final_op.name][0].type in activation_name_map:
+            activation_op = self.tf_graph[final_op.name][0]
+            fused_act_arg = op_def.arg.add()
+            fused_act_arg.name = 'activation'
+            fused_act_arg.s = activation_name_map[activation_op.type]
+            if activation_op.type == 'Relu6':
+                max_limit_arg = op_def.arg.add()
+                max_limit_arg.name = 'max_limit'
+                max_limit_arg.f = 6
+            final_op = activation_op
+            self.resolved_ops[activation_op.name] = 1
+
+        op_def.output.extend([output.name for output in final_op.outputs])
+        self.add_output_shape(final_op.outputs, op_def)
+        self.net_def.op.extend([op_def])
+
     def check_conv_to_fc(self, op):
         if self.device != 'cpu' or op.type != "Conv2D":
             return False
@@ -857,6 +956,7 @@ class TFConverter(object):
         if len(op.inputs) == 2:
             input_tensor0 = get_input_tensor(op, 0)
             input_tensor1 = get_input_tensor(op, 1)
+
             x_value = None
             if np.asarray(input_tensor1.shape).size == 0:
                 x_value = input_tensor1.eval()
@@ -867,7 +967,22 @@ class TFConverter(object):
                 op_def.input.extend([op.inputs[1].name])
                 self.unused_tensor.add(input_tensor0.name)
             else:
-                op_def.input.extend([input.name for input in op.inputs])
+                if np.asarray(input_tensor0.shape).size == 1 \
+                        and input_tensor0.op.type == 'Const':
+                    if self.device == 'gpu':
+                        output_name = self.add_buffer_to_image(
+                            input_tensor0.name, "ARGUMENT")
+                        op_def.input.extend([output_name])
+                else:
+                    op_def.input.extend([input_tensor0.name])
+                if np.asarray(input_tensor1.shape).size == 1 \
+                        and input_tensor1.op.type == 'Const':
+                    if self.device == 'gpu':
+                        output_name = self.add_buffer_to_image(
+                            input_tensor1.name, "ARGUMENT")
+                        op_def.input.extend([output_name])
+                else:
+                    op_def.input.extend([input_tensor1.name])
             if x_value is not None:
                 x_arg = op_def.arg.add()
                 x_arg.name = 'x'
@@ -1150,6 +1265,8 @@ class TFConverter(object):
                     self.convert_winograd_conv_gpu(op)
                 else:
                     self.convert_conv2d(op)
+            elif op.type == 'Conv2DBackpropInput':
+                self.convert_deconv2d(op)
             elif op.type == 'FusedBatchNorm':
                 self.convert_fused_batchnorm(op)
             elif op.type == 'Mul' and op.name.find('batchnorm/mul') != -1:
@@ -1159,7 +1276,10 @@ class TFConverter(object):
             elif op.type == 'Relu6':
                 self.convert_relu6(op)
             elif op.type == 'Add':
+                if len(op.inputs) > 2:
                     self.convert_add(op)
+                else:
+                    self.convert_eltwise(op, 'ADD')
             elif op.type == 'ConcatV2':
                 self.convert_concat(op)
             elif op.type == 'ResizeBilinear':
@@ -1176,6 +1296,12 @@ class TFConverter(object):
                 self.convert_depth_to_space(op, False)
             elif op.type in ['Neg', 'neg', 'Negative', 'negative']:
                 self.convert_eltwise(op, 'NEG')
+            elif op.type in ['RealDiv', 'Div']:
+                self.convert_eltwise(op, 'DIV')
+            elif op.type in ['SquaredDifference']:
+                self.convert_eltwise(op, 'SQR_DIFF')
+            elif op.type in ['Pow']:
+                self.convert_eltwise(op, 'POW')
             elif op.type == 'Mul':
                 self.convert_eltwise(op, 'MUL')
             elif op.type == 'Sub':