Merge branch 'depth_to_space' into 'master'

add DepthToSpace / SpaceToDepth Ops

See merge request !319

mace/core/operator.cc

@@ -63,7 +63,6 @@ std::unique_ptr<OperatorBase> OperatorRegistry::CreateOperator(
 }
 
 namespace ops {
-
 // Keep in lexicographical order
 extern void Register_Activation(OperatorRegistry *op_registry);
 extern void Register_AddN(OperatorRegistry *op_registry);
@@ -74,6 +73,7 @@ extern void Register_BufferToImage(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_DepthToSpace(OperatorRegistry *op_registry);
 extern void Register_DepthwiseConv2d(OperatorRegistry *op_registry);
 extern void Register_Eltwise(OperatorRegistry *op_registry);
 extern void Register_FoldedBatchNorm(OperatorRegistry *op_registry);
@@ -90,6 +90,7 @@ extern void Register_ResizeBilinear(OperatorRegistry *op_registry);
 extern void Register_Slice(OperatorRegistry *op_registry);
 extern void Register_Softmax(OperatorRegistry *op_registry);
 extern void Register_SpaceToBatchND(OperatorRegistry *op_registry);
+extern void Register_SpaceToDepth(OperatorRegistry *op_registry);
 extern void Register_WinogradInverseTransform(OperatorRegistry *op_registry);
 extern void Register_WinogradTransform(OperatorRegistry *op_registry);
 
@@ -107,6 +108,7 @@ OperatorRegistry::OperatorRegistry() {
   ops::Register_ChannelShuffle(this);
   ops::Register_Concat(this);
   ops::Register_Conv2D(this);
+  ops::Register_DepthToSpace(this);
   ops::Register_DepthwiseConv2d(this);
   ops::Register_Eltwise(this);
   ops::Register_FoldedBatchNorm(this);
@@ -123,6 +125,7 @@ OperatorRegistry::OperatorRegistry() {
   ops::Register_Slice(this);
   ops::Register_Softmax(this);
   ops::Register_SpaceToBatchND(this);
+  ops::Register_SpaceToDepth(this);
   ops::Register_WinogradInverseTransform(this);
   ops::Register_WinogradTransform(this);
 }

mace/kernels/depth_to_space.h

+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#ifndef MACE_KERNELS_DEPTH_TO_SPACE_H_
+#define MACE_KERNELS_DEPTH_TO_SPACE_H_
+#include <vector>
+
+#include "mace/core/future.h"
+#include "mace/core/runtime/opencl/cl2_header.h"
+#include "mace/core/tensor.h"
+#include "mace/public/mace.h"
+
+namespace mace {
+namespace kernels {
+
+template <DeviceType D, typename T>
+struct DepthToSpaceOpFunctor {
+  explicit DepthToSpaceOpFunctor(const int block_size, bool d2s)
+      : block_size_(block_size), d2s_(d2s) {}
+  void operator()(const Tensor *input, Tensor *output, StatsFuture *future) {
+    const int batch_size = input->dim(0);
+    const int input_height = input->dim(1);
+    const int input_width = input->dim(2);
+    const int input_depth = input->dim(3);
+
+    index_t output_depth, output_width, output_height;
+
+    if (d2s_) {
+      output_depth = input_depth / (block_size_ * block_size_);
+      output_width = input_width * block_size_;
+      output_height = input_height * block_size_;
+    } else {
+      output_depth = input_depth * block_size_ * block_size_;
+      output_width = input_width / block_size_;
+      output_height = input_height / block_size_;
+    }
+    std::vector<index_t> output_shape = {batch_size, output_height,
+                                         output_width, output_depth};
+
+    output->Resize(output_shape);
+
+    Tensor::MappingGuard logits_guard(input);
+    Tensor::MappingGuard output_guard(output);
+    const T *input_ptr = input->data<T>();
+    T *output_ptr = output->mutable_data<T>();
+
+    if (d2s_) {
+#pragma omp parallel for
+      for (int b = 0; b < batch_size; ++b) {
+        for (int h = 0; h < output_height; ++h) {
+          const int in_h = h / block_size_;
+          const int offset_h = (h % block_size_);
+          for (int w = 0; w < output_width; ++w) {
+            const int in_w = w / block_size_;
+            const int offset_w = w % block_size_;
+            const int offset_d =
+                (offset_h * block_size_ + offset_w) * output_depth;
+            for (int d = 0; d < output_depth; ++d) {
+              const int in_d = d + offset_d;
+              const int o_index =
+                  ((b * output_height + h) * output_width + w) * output_depth +
+                  d;
+              const int i_index =
+                  ((b * input_height + in_h) * input_width + in_w) *
+                      input_depth +
+                  in_d;
+              output_ptr[o_index] = input_ptr[i_index];
+            }
+          }
+        }
+      }
+    } else {
+#pragma omp parallel for
+      for (int b = 0; b < batch_size; ++b) {
+        for (int h = 0; h < input_height; ++h) {
+          const int out_h = h / block_size_;
+          const int offset_h = (h % block_size_);
+          for (int w = 0; w < input_width; ++w) {
+            const int out_w = w / block_size_;
+            const int offset_w = (w % block_size_);
+            const int offset_d =
+                (offset_h * block_size_ + offset_w) * input_depth;
+            for (int d = 0; d < input_depth; ++d) {
+              const int out_d = d + offset_d;
+              const int o_index =
+                  ((b * output_height + out_h) * output_width + out_w) *
+                      output_depth +
+                  out_d;
+              const int i_index =
+                  ((b * input_height + h) * input_width + w) * input_depth + d;
+              output_ptr[o_index] = input_ptr[i_index];
+            }
+          }
+        }
+      }
+    }
+  }
+
+  const int block_size_;
+  bool d2s_;
+};
+
+template <typename T>
+struct DepthToSpaceOpFunctor<DeviceType::OPENCL, T> {
+  DepthToSpaceOpFunctor(const int block_size, bool d2s)
+      : block_size_(block_size), d2s_(d2s) {}
+  void operator()(const Tensor *input, Tensor *output, StatsFuture *future);
+
+  cl::Kernel kernel_;
+  const int block_size_;
+  bool d2s_;
+  std::vector<index_t> input_shape_;
+};
+
+}  // namespace kernels
+}  // namespace mace
+
+#endif  // MACE_KERNELS_DEPTH_TO_SPACE_H_

mace/kernels/opencl/cl/depth_to_space.cl

+#include <common.h>
+
+__kernel void depth_to_space(__read_only image2d_t input,
+                             __private const int block_size,
+                             __private const int output_depth,
+                             __write_only image2d_t output) {
+  const int out_d = get_global_id(0);
+  const int out_w = get_global_id(1);
+  const int out_h = get_global_id(2);
+  const int output_width = get_global_size(1);
+
+  const int out_pos = mad24(out_d, output_width, out_w);
+
+  const int input_width = output_width / block_size;
+
+  const int in_h = out_h / block_size;
+  const int offset_h = out_h % block_size;
+  const int in_w = out_w / block_size;
+  const int offset_w = out_w % block_size;
+
+  const int offset_d = (offset_h * block_size + offset_w) * output_depth;
+  const int in_d = out_d + offset_d;
+
+  const int in_pos = mad24(in_d, input_width, in_w);
+
+  DATA_TYPE4 in_data = READ_IMAGET(input, SAMPLER, (int2)(in_pos, in_h));
+  WRITE_IMAGET(output, (int2)(out_pos, out_h), in_data);
+}
+
+__kernel void space_to_depth(__read_only image2d_t input,
+                             __private const int block_size,
+                             __private const int input_depth,
+                             __write_only image2d_t output) {
+  const int d = get_global_id(0);
+  const int w = get_global_id(1);
+  const int h = get_global_id(2);
+  const int input_width = get_global_size(1);
+  const int in_pos = mad24(d, input_width, w);
+  const int output_width = input_width / block_size;
+
+  const int out_h = h / block_size;
+  const int offset_h = h % block_size;
+  const int out_w = w / block_size;
+  const int offset_w = w % block_size;
+  const int offset_d = (offset_h * block_size + offset_w) * input_depth;
+  const int out_d = d + offset_d;
+  const int out_pos = mad24(out_d, output_width, out_w);
+
+  DATA_TYPE4 in_data = READ_IMAGET(input, SAMPLER, (int2)(in_pos, h));
+
+  WRITE_IMAGET(output, (int2)(out_pos, out_h), in_data);
+}

mace/kernels/opencl/depth_to_space_opencl.cc

+//
+// Copyright (c) 2018 XiaoMi All rights reserved.
+//
+
+#include "mace/kernels/depth_to_space.h"
+#include "mace/core/runtime/opencl/cl2_header.h"
+#include "mace/core/runtime/opencl/opencl_runtime.h"
+#include "mace/kernels/opencl/helper.h"
+#include "mace/utils/tuner.h"
+#include "mace/utils/utils.h"
+
+namespace mace {
+namespace kernels {
+
+template <typename T>
+void DepthToSpaceOpFunctor<DeviceType::OPENCL, T>::operator()(
+    const Tensor *input, Tensor *output, StatsFuture *future) {
+  const index_t batch = input->dim(0);
+  const index_t input_height = input->dim(1);
+  const index_t input_width = input->dim(2);
+  const index_t input_depth = input->dim(3);
+
+  int depth_blocks = 1;
+  const char *kernel_name = nullptr;
+
+  index_t output_height, output_width, output_depth;
+  if (d2s_) {
+    output_height = input_height * block_size_;
+    output_width = input_width * block_size_;
+    output_depth = input_depth / (block_size_ * block_size_);
+    depth_blocks = RoundUpDiv4(output_depth);
+    kernel_name = "depth_to_space";
+  } else {
+    output_height = input_height / block_size_;
+    output_width = input_width / block_size_;
+    output_depth = input_depth * block_size_ * block_size_;
+    depth_blocks = RoundUpDiv4(input_depth);
+    kernel_name = "space_to_depth";
+  }
+
+  std::vector<index_t> output_shape = {batch, output_height, output_width,
+                                       output_depth};
+
+  std::vector<size_t> image_shape;
+  CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL, &image_shape);
+  output->ResizeImage(output_shape, image_shape);
+
+  if (kernel_.get() == nullptr) {
+    auto runtime = OpenCLRuntime::Global();
+    std::set<std::string> built_options;
+    std::string obfuscated_kernel_name = MACE_OBFUSCATE_SYMBOL(kernel_name);
+    std::stringstream kernel_name_ss;
+    kernel_name_ss << "-D" << kernel_name << "=" << obfuscated_kernel_name;
+    built_options.emplace(kernel_name_ss.str());
+    auto dt = DataTypeToEnum<T>::value;
+    built_options.emplace("-DDATA_TYPE=" + DtToUpstreamCLDt(dt));
+    built_options.emplace("-DCMD_DATA_TYPE=" + DtToUpstreamCLCMDDt(dt));
+    kernel_ =
+        runtime->BuildKernel("depth_to_space", kernel_name, built_options);
+  }
+  if (!IsVecEqual(input_shape_, input->shape())) {
+    uint32_t idx = 0;
+    kernel_.setArg(idx++, *(input->opencl_image()));
+    kernel_.setArg(idx++, block_size_);
+    kernel_.setArg(idx++, depth_blocks);
+    kernel_.setArg(idx++, *(output->opencl_image()));
+    input_shape_ = input->shape();
+  }
+
+  if (d2s_) {
+    const uint32_t gws[3] = {static_cast<uint32_t>(depth_blocks),
+                             static_cast<uint32_t>(output_width),
+                             static_cast<uint32_t>(output_height * batch)};
+    const std::vector<uint32_t> lws = {8, 16, 8, 1};
+    std::stringstream ss;
+    ss << "depth_to_space_opencl_kernel_" << output->dim(0) << "_"
+       << output->dim(1) << "_" << output->dim(2) << "_" << output->dim(3);
+
+    TuningOrRun3DKernel(kernel_, ss.str(), gws, lws, future);
+  } else {
+    const uint32_t gws[3] = {static_cast<uint32_t>(depth_blocks),
+                             static_cast<uint32_t>(input_width),
+                             static_cast<uint32_t>(input_height * batch)};
+    const std::vector<uint32_t> lws = {8, 16, 8, 1};
+    std::stringstream ss;
+    ss << "space_to_depth_opencl_kernel_" << input->dim(0) << "_"
+       << input->dim(1) << "_" << input->dim(2) << "_" << input->dim(3);
+    TuningOrRun3DKernel(kernel_, ss.str(), gws, lws, future);
+  }
+}
+
+template struct DepthToSpaceOpFunctor<DeviceType::OPENCL, float>;
+template struct DepthToSpaceOpFunctor<DeviceType::OPENCL, half>;
+
+}  // namespace kernels
+}  // namespace mace

mace/ops/depth_to_space.cc

+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#include "mace/ops/depth_to_space.h"
+
+namespace mace {
+namespace ops {
+
+void Register_DepthToSpace(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("DepthToSpace")
+                                     .Device(DeviceType::CPU)
+                                     .TypeConstraint<float>("T")
+                                     .Build(),
+                    DepthToSpaceOp<DeviceType::CPU, float>);
+
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("DepthToSpace")
+                                     .Device(DeviceType::OPENCL)
+                                     .TypeConstraint<float>("T")
+                                     .Build(),
+                    DepthToSpaceOp<DeviceType::OPENCL, float>);
+
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("DepthToSpace")
+                                     .Device(DeviceType::OPENCL)
+                                     .TypeConstraint<half>("T")
+                                     .Build(),
+                    DepthToSpaceOp<DeviceType::OPENCL, half>);
+}
+
+}  // namespace ops
+}  // namespace mace

mace/ops/depth_to_space.h

+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#ifndef MACE_OPS_DEPTH_TO_SPACE_H_
+#define MACE_OPS_DEPTH_TO_SPACE_H_
+
+#include <memory>
+#include <vector>
+
+#include "mace/core/operator.h"
+#include "mace/kernels/depth_to_space.h"
+
+namespace mace {
+namespace ops {
+
+template <DeviceType D, typename T>
+class DepthToSpaceOp : public Operator<D, T> {
+ public:
+  DepthToSpaceOp(const OperatorDef &op_def, Workspace *ws)
+      : Operator<D, T>(op_def, ws),
+        functor_(OperatorBase::GetSingleArgument<int>("block_size", 1), true) {}
+
+  bool Run(StatsFuture *future) override {
+    const Tensor *input = this->Input(INPUT);
+    Tensor *output = this->Output(OUTPUT);
+    MACE_CHECK(input->dim_size() == 4, "input dim should be 4");
+
+    const int block_size =
+        OperatorBase::GetSingleArgument<int>("block_size", 1);
+
+    int input_depth = input->dim(3);
+    MACE_CHECK(input_depth % (block_size * block_size) == 0,
+               "input depth should be dividable by block_size * block_size",
+               input->dim(3));
+    MACE_CHECK((input_depth % 4) == 0,
+               "input channel should be dividable by 4");
+    functor_(input, output, future);
+    return true;
+  }
+
+ protected:
+  OP_INPUT_TAGS(INPUT);
+  OP_OUTPUT_TAGS(OUTPUT);
+
+ private:
+  kernels::DepthToSpaceOpFunctor<D, T> functor_;
+};
+
+}  // namespace ops
+}  // namespace mace
+
+#endif  // MACE_OPS_DEPTH_TO_SPACE_H_

mace/ops/depth_to_space_benchmark.cc

+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#include "mace/core/operator.h"
+#include "mace/core/testing/test_benchmark.h"
+#include "mace/ops/ops_test_util.h"
+
+namespace mace {
+namespace ops {
+namespace test {
+
+template <DeviceType D, typename T>
+static void DepthToSpace(
+    int iters, int batch, int channels, int height, int width, int block_size) {
+  mace::testing::StopTiming();
+
+  OpsTestNet net;
+
+  // Add input data
+  net.AddRandomInput<D, float>("Input", {batch, height, width, channels});
+
+  if (D == DeviceType::OPENCL) {
+    BufferToImage<D, float>(&net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT_CHANNEL);
+
+    OpDefBuilder("DepthToSpace", "DepthToSpaceBM")
+        .Input("InputImage")
+        .Output("Output")
+        .AddIntArg("block_size", block_size)
+        .Finalize(net.NewOperatorDef());
+  } else {
+    OpDefBuilder("DepthToSpace", "DepthToSpaceBM")
+        .Input("Input")
+        .Output("Output")
+        .Finalize(net.NewOperatorDef());
+  }
+
+  // Warm-up
+  for (int i = 0; i < 5; ++i) {
+    net.RunOp(D);
+  }
+  net.Sync();
+
+  mace::testing::StartTiming();
+  while (iters--) {
+    net.RunOp(D);
+  }
+  net.Sync();
+}
+
+#define BM_DEPTH_TO_SPACE_MACRO(N, C, H, W, G, TYPE, DEVICE)             \
+  static void                                                            \
+      BM_DEPTH_TO_SPACE_##N##_##C##_##H##_##W##_##G##_##TYPE##_##DEVICE( \
+          int iters) {                                                   \
+    const int64_t tot = static_cast<int64_t>(iters) * N * C * H * W;     \
+    mace::testing::MaccProcessed(tot);                                   \
+    mace::testing::BytesProcessed(tot *(sizeof(TYPE)));                  \
+    DepthToSpace<DEVICE, TYPE>(iters, N, C, H, W, G);                    \
+  }                                                                      \
+  BENCHMARK(BM_DEPTH_TO_SPACE_##N##_##C##_##H##_##W##_##G##_##TYPE##_##DEVICE)
+
+#define BM_DEPTH_TO_SPACE(N, C, H, W, G)                 \
+  BM_DEPTH_TO_SPACE_MACRO(N, C, H, W, G, float, CPU);    \
+  BM_DEPTH_TO_SPACE_MACRO(N, C, H, W, G, float, OPENCL); \
+  BM_DEPTH_TO_SPACE_MACRO(N, C, H, W, G, half, OPENCL);
+
+BM_DEPTH_TO_SPACE(1, 64, 64, 64, 4);
+BM_DEPTH_TO_SPACE(1, 64, 128, 128, 4);
+BM_DEPTH_TO_SPACE(1, 64, 256, 256, 4);
+
+}  // namespace test
+}  // namespace ops
+}  // namespace mace

mace/ops/depth_to_space_test.cc

+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#include "mace/core/operator.h"
+#include "mace/ops/ops_test_util.h"
+
+namespace mace {
+namespace ops {
+namespace test {
+
+template <DeviceType D>
+void RunDepthToSpace(const bool d2s,
+                     const std::vector<index_t> &input_shape,
+                     const std::vector<float> &input_data,
+                     const int block_size,
+                     const std::vector<index_t> &expected_shape,
+                     const std::vector<float> &expected_data) {
+  OpsTestNet net;
+  net.AddInputFromArray<D, float>("Input", input_shape, input_data);
+  const char *ops_name = (d2s) ? "DepthToSpace" : "SpaceToDepth";
+  const char *ops_test_name = (d2s) ? "DepthToSpaceTest" : "SpaceToDepthTest";
+  // Construct graph
+  if (D == DeviceType::CPU) {
+    OpDefBuilder(ops_name, ops_test_name)
+        .Input("Input")
+        .Output("Output")
+        .AddIntArg("block_size", block_size)
+        .Finalize(net.NewOperatorDef());
+
+  } else {
+    BufferToImage<D, float>(&net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT_CHANNEL);
+    OpDefBuilder(ops_name, ops_test_name)
+        .Input("InputImage")
+        .Output("OutputImage")
+        .AddIntArg("block_size", block_size)
+        .Finalize(net.NewOperatorDef());
+  }
+  // Run
+  net.RunOp(D);
+
+  if (D == DeviceType::OPENCL) {
+    ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "Output",
+        kernels::BufferType::IN_OUT_CHANNEL);
+  }
+  auto expected = CreateTensor<float>(expected_shape, expected_data);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+}
+
+class SpaceToDepthOpTest : public OpsTestBase {};
+
+TEST_F(SpaceToDepthOpTest, Input2x4x4_B2_CPU) {
+  RunDepthToSpace<DeviceType::CPU>(false, {1, 2, 4, 4},
+      {0, 1, 2,  3,  4,  5,  6,  7,  16, 17, 18, 19, 20, 21, 22, 23,
+       8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31},
+      2,
+      {1, 1, 2, 16},
+      {0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+      16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31});
+}
+
+TEST_F(SpaceToDepthOpTest, Input2x4x4_B2_OPENCL) {
+  RunDepthToSpace<DeviceType::OPENCL>(false, {1, 2, 4, 4},
+      {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23,
+       8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31},
+      2,
+      {1, 1, 2, 16},
+      {0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+      16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31});
+}
+
+TEST_F(SpaceToDepthOpTest, Input2x2x4_B2_CPU) {
+  RunDepthToSpace<DeviceType::CPU>(false, {1, 2, 2, 4},
+      {1, 2, 3, 4, 5, 6, 7, 8,
+       9, 10, 11, 12, 13, 14, 15, 16},
+      2,
+      {1, 1, 1, 16},
+      {1, 2, 3, 4, 5, 6, 7, 8,
+       9, 10, 11, 12, 13, 14, 15, 16});
+}
+
+TEST_F(SpaceToDepthOpTest, Input4x4x1_B2_OPENCL) {
+  RunDepthToSpace<DeviceType::OPENCL>(false, {1, 2, 2, 4},
+      {1, 2, 3, 4, 5, 6, 7, 8,
+       9, 10, 11, 12, 13, 14, 15, 16},
+      2,
+      {1, 1, 1, 16},
+      {1, 2, 3, 4, 5, 6, 7, 8,
+       9, 10, 11, 12, 13, 14, 15, 16});
+}
+
+class DepthToSpaceOpTest : public OpsTestBase {};
+
+TEST_F(DepthToSpaceOpTest, Input1x2x16_B2_CPU) {
+  RunDepthToSpace<DeviceType::CPU>(true, {1, 1, 2, 16},
+      {0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+      16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31},
+      2,
+      {1, 2, 4, 4},
+      {0, 1, 2,  3,  4,  5,  6,  7,  16, 17, 18, 19, 20, 21, 22, 23,
+      8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31});
+}
+
+TEST_F(DepthToSpaceOpTest, Input1x2x16_B2_OPENCL) {
+  RunDepthToSpace<DeviceType::OPENCL>(true, {1, 1, 2, 16},
+      {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
+      16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31},
+      2,
+      {1, 2, 4, 4},
+      {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23,
+      8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31});
+}
+
+TEST_F(DepthToSpaceOpTest, Input1x1x16_B2_CPU) {
+  RunDepthToSpace<DeviceType::CPU>(true, {1, 1, 1, 16},
+      {1,  2,  3,  4,  5,  6,  7,  8,
+       9,  10, 11, 12, 13, 14, 15, 16},
+      2,
+      {1, 2, 2, 4},
+      {1, 2, 3, 4, 5, 6, 7, 8,
+       9, 10, 11, 12, 13, 14, 15, 16});
+}
+
+TEST_F(DepthToSpaceOpTest, Input1x1x16_B2_OPENCL) {
+  RunDepthToSpace<DeviceType::OPENCL>(true, {1, 1, 1, 16},
+      {1,  2,  3,  4,  5,  6,  7,  8,
+       9,  10, 11, 12, 13, 14, 15, 16},
+      2,
+      {1, 2, 2, 4},
+      {1, 2, 3, 4, 5, 6, 7, 8,
+       9, 10, 11, 12, 13, 14, 15, 16});
+}
+
+/*
+TEST_F(DepthToSpaceOpTest, Input2x2x3_B2_CPU) {
+
+  RunDepthToSpace<DeviceType::CPU>({1, 2, 2, 3},
+                                   {1, 2, 3, 4, 5, 6,
+									7, 8, 9, 10, 11, 12},
+                                   2,
+                                   {1, 1, 1, 12},
+                                   {1,  2,  3,  4,  5,  6,  7,  8,  
+									9,  10, 11, 12});
+}
+
+TEST_F(DepthToSpaceOpTest, Input2x2x3_B2_OPENCL) {
+  RunDepthToSpace<DeviceType::OPENCL>({1, 2, 2, 6},
+                                   {1, 2, 3, 4, 5, 6,
+									7, 8, 9, 10, 11, 12
+									},
+                                   2,
+                                   {1, 1, 1, 12},
+                                   {1,  2,  3,  4,  5,  6,  7,  8,  
+									9,  10, 11, 12});
+}
+
+TEST_F(DepthToSpaceOpTest, Input2x2x2_B2_CPU) {
+
+  RunDepthToSpace<DeviceType::CPU>({1, 2, 2, 2},
+                                   {1, 10,  2,  20,  3,  30,  4, 40},
+                                   2,
+                                   {1, 1, 1, 8},
+                                   {1, 10,  2,  20,  3,  30,  4, 40});
+}
+
+TEST_F(DepthToSpaceOpTest, Input2x2x2_B2_OPENCL) {
+
+  RunDepthToSpace<DeviceType::OPENCL>({1, 2, 2, 2},
+                                      {1, 10,  2,  20,  3,  30,  4, 40},
+                                      2,
+                                      {1, 1, 1, 8},
+                                      {1, 10,  2,  20,  3,  30,  4, 40});
+}*/
+}  // namespace test
+}  // namespace ops
+}  // namespace mace

mace/ops/space_to_depth.cc

+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#include "mace/ops/space_to_depth.h"
+
+namespace mace {
+namespace ops {
+
+void Register_SpaceToDepth(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("SpaceToDepth")
+                                     .Device(DeviceType::CPU)
+                                     .TypeConstraint<float>("T")
+                                     .Build(),
+                    SpaceToDepthOp<DeviceType::CPU, float>);
+
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("SpaceToDepth")
+                                     .Device(DeviceType::OPENCL)
+                                     .TypeConstraint<float>("T")
+                                     .Build(),
+                    SpaceToDepthOp<DeviceType::OPENCL, float>);
+
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("SpaceToDepth")
+                                     .Device(DeviceType::OPENCL)
+                                     .TypeConstraint<half>("T")
+                                     .Build(),
+                    SpaceToDepthOp<DeviceType::OPENCL, half>);
+}
+
+}  // namespace ops
+}  // namespace mace

mace/ops/space_to_depth.h

+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#ifndef MACE_OPS_SPACE_TO_DEPTH_H_
+#define MACE_OPS_SPACE_TO_DEPTH_H_
+
+#include <memory>
+#include <vector>
+
+#include "mace/core/operator.h"
+#include "mace/kernels/depth_to_space.h"
+
+namespace mace {
+namespace ops {
+
+template <DeviceType D, typename T>
+class SpaceToDepthOp : public Operator<D, T> {
+ public:
+  SpaceToDepthOp(const OperatorDef &op_def, Workspace *ws)
+      : Operator<D, T>(op_def, ws),
+        functor_(OperatorBase::GetSingleArgument<int>("block_size", 1), false) {
+  }
+
+  bool Run(StatsFuture *future) override {
+    const Tensor *input = this->Input(INPUT);
+    Tensor *output = this->Output(OUTPUT);
+    MACE_CHECK(input->dim_size() == 4, "input dim should be 4");
+    const int block_size =
+        OperatorBase::GetSingleArgument<int>("block_size", 1);
+    const int input_height = input->dim(1);
+    const int input_width = input->dim(2);
+    const int input_depth = input->dim(3);
+    MACE_CHECK((input_depth % 4) == 0,
+               "input channel should be dividable by 4");
+    MACE_CHECK(
+        (input_width%block_size == 0)&&(input_height%block_size == 0),
+        "input width and height should be dividable by block_size",
+        input->dim(3));
+    functor_(input, output, future);
+    return true;
+  }
+
+ protected:
+  OP_INPUT_TAGS(INPUT);
+  OP_OUTPUT_TAGS(OUTPUT);
+
+ private:
+  kernels::DepthToSpaceOpFunctor<D, T> functor_;
+};
+
+}  // namespace ops
+}  // namespace mace
+
+#endif  // MACE_OPS_SPACE_TO_DEPTH_H_

mace/ops/space_to_depth_benchmark.cc

+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#include "mace/core/operator.h"
+#include "mace/core/testing/test_benchmark.h"
+#include "mace/ops/ops_test_util.h"
+
+namespace mace {
+namespace ops {
+namespace test {
+
+template <DeviceType D, typename T>
+static void SpaceToDepth(
+    int iters, int batch, int channels, int height, int width, int block_size) {
+  mace::testing::StopTiming();
+
+  OpsTestNet net;
+
+  // Add input data
+  net.AddRandomInput<D, float>("Input", {batch, height, width, channels});
+
+  if (D == DeviceType::OPENCL) {
+    BufferToImage<D, float>(&net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT_CHANNEL);
+
+    OpDefBuilder("SpaceToDepth", "SpaceToDepthBM")
+        .Input("InputImage")
+        .Output("Output")
+        .AddIntArg("block_size", block_size)
+        .Finalize(net.NewOperatorDef());
+  } else {
+    OpDefBuilder("SpaceToDepth", "SpaceToDepthBM")
+        .Input("Input")
+        .Output("Output")
+        .Finalize(net.NewOperatorDef());
+  }
+
+  // Warm-up
+  for (int i = 0; i < 5; ++i) {
+    net.RunOp(D);
+  }
+  net.Sync();
+
+  mace::testing::StartTiming();
+  while (iters--) {
+    net.RunOp(D);
+  }
+  net.Sync();
+}
+
+#define BM_SPACE_TO_DEPTH_MACRO(N, C, H, W, G, TYPE, DEVICE)             \
+  static void                                                            \
+      BM_SPACE_TO_DEPTH_##N##_##C##_##H##_##W##_##G##_##TYPE##_##DEVICE( \
+          int iters) {                                                   \
+    const int64_t tot = static_cast<int64_t>(iters) * N * C * H * W;     \
+    mace::testing::MaccProcessed(tot);                                   \
+    mace::testing::BytesProcessed(tot *(sizeof(TYPE)));                  \
+    SpaceToDepth<DEVICE, TYPE>(iters, N, C, H, W, G);                    \
+  }                                                                      \
+  BENCHMARK(BM_SPACE_TO_DEPTH_##N##_##C##_##H##_##W##_##G##_##TYPE##_##DEVICE)
+
+#define BM_SPACE_TO_DEPTH(N, C, H, W, G)                 \
+  BM_SPACE_TO_DEPTH_MACRO(N, C, H, W, G, float, CPU);    \
+  BM_SPACE_TO_DEPTH_MACRO(N, C, H, W, G, float, OPENCL); \
+  BM_SPACE_TO_DEPTH_MACRO(N, C, H, W, G, half, OPENCL);
+
+BM_SPACE_TO_DEPTH(1, 64, 64, 64, 4);
+BM_SPACE_TO_DEPTH(1, 64, 128, 128, 4);
+BM_SPACE_TO_DEPTH(1, 64, 256, 256, 4);
+
+}  // namespace test
+}  // namespace ops
+}  // namespace mace