spacetodepth and spacetobatch fp32

dequantize and quantize

mindspore/lite/src/ops/dequantize.cc

@@ -0,0 +1,34 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/ops/ops.h"
+#include "include/errorcode.h"
+#include "utils/log_adapter.h"
+#include "src/ir/tensor.h"
+
+namespace mindspore::lite {
+int Dequantize::InferShape(std::vector<tensor::Tensor *> inputs_, std::vector<tensor::Tensor *> outputs_) {
+  MS_ASSERT(this->primitive != nullptr);
+  auto input = inputs_.front();
+  MS_ASSERT(input != nullptr);
+  auto output = outputs_.front();
+  MS_ASSERT(output != nullptr);
+  output->set_shape(input->shape());
+  output->set_data_type(kNumberTypeFloat32);
+  return RET_OK;
+}
+}  // namespace mindspore::lite
+

mindspore/lite/src/ops/ops.cc

@@ -133,6 +133,14 @@ Primitive *Primitive::CreatePrimitive(schema::Primitive *primitive) {
       return new lite::OneHot(const_cast<schema::Primitive *>(primitive));
     case schema::PrimitiveType_PriorBox:
       return new lite::PriorBox(const_cast<schema::Primitive *>(primitive));
+    case schema::PrimitiveType_SpaceToDepth:
+      return new lite::SpaceToDepth(const_cast<schema::Primitive *>(primitive));
+    case schema::PrimitiveType_SpaceToBatch:
+      return new lite::SpaceToBatch(const_cast<schema::Primitive *>(primitive));
+    case schema::PrimitiveType_OnnxInt8Dequantize:
+      return new lite::Dequantize(const_cast<schema::Primitive *>(primitive));
+    case schema::PrimitiveType_OnnxInt8Quantize:
+      return new lite::Quantize(const_cast<schema::Primitive *>(primitive));
     default:
       break;
   }

mindspore/lite/src/ops/ops.h

@@ -596,6 +596,22 @@ class BatchToSpace : public Primitive {
   int InferShape(std::vector<tensor::Tensor *> inputs, std::vector<tensor::Tensor *> outputs) override;
 };
 
+class SpaceToBatch : public Primitive {
+ public:
+  explicit SpaceToBatch(schema::Primitive *primitive) : Primitive(primitive) {}
+  const schema::SpaceToBatch *GetAttribute() const { return this->primitive->value_as_SpaceToBatch(); }
+  int InferShape(std::vector<tensor::Tensor *> inputs, std::vector<tensor::Tensor *> outputs) override;
+  std::vector<int> BlockSizes() {return block_sizes_;}
+  std::vector<int> Paddings() {return block_sizes_;}
+  std::vector<int> InShape() {return block_sizes_;}
+  std::vector<int> PaddedInShape() {return block_sizes_;}
+ private:
+  std::vector<int> block_sizes_;
+  std::vector<int> paddings_;
+  std::vector<int> in_shape_;
+  std::vector<int> padded_in_shape_;
+};
+
 class Crop : public Primitive {
  public:
   explicit Crop(schema::Primitive *primitive) : Primitive(primitive) {}
@@ -668,6 +684,26 @@ class PriorBox : public Primitive {
   int InferShape(std::vector<tensor::Tensor *> inputs, std::vector<tensor::Tensor *> outputs) override;
 };
 
+class SpaceToDepth : public Primitive {
+ public:
+  explicit SpaceToDepth(schema::Primitive *primitive) : Primitive(primitive) {}
+  const schema::SpaceToDepth *GetAttribute() const { return this->primitive->value_as_SpaceToDepth(); }
+  int InferShape(std::vector<tensor::Tensor *> inputs, std::vector<tensor::Tensor *> outputs) override;
+};
+
+class Dequantize : public Primitive {
+ public:
+  explicit Dequantize(schema::Primitive *primitive) : Primitive(primitive) {}
+  const schema::OnnxInt8Dequantize *GetAttribute() const { return this->primitive->value_as_OnnxInt8Dequantize(); }
+  int InferShape(std::vector<tensor::Tensor *> inputs, std::vector<tensor::Tensor *> outputs) override;
+};
+
+class Quantize : public Primitive {
+ public:
+  explicit Quantize(schema::Primitive *primitive) : Primitive(primitive) {}
+  const schema::OnnxInt8Quantize *GetAttribute() const { return this->primitive->value_as_OnnxInt8Quantize(); }
+  int InferShape(std::vector<tensor::Tensor *> inputs, std::vector<tensor::Tensor *> outputs) override;
+};
 }  // namespace lite
 }  // namespace mindspore
 #endif  // MINDSPORE_LITE_SRC_OPS_OPS_H_

mindspore/lite/src/ops/quantize.cc

@@ -0,0 +1,34 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/ops/ops.h"
+#include "include/errorcode.h"
+#include "utils/log_adapter.h"
+#include "src/ir/tensor.h"
+
+namespace mindspore::lite {
+int Quantize::InferShape(std::vector<tensor::Tensor *> inputs_, std::vector<tensor::Tensor *> outputs_) {
+  MS_ASSERT(this->primitive != nullptr);
+  auto input = inputs_.front();
+  MS_ASSERT(input != nullptr);
+  auto output = outputs_.front();
+  MS_ASSERT(output != nullptr);
+  output->set_shape(input->shape());
+  output->set_data_type(kNumberTypeInt8);
+  return RET_OK;
+}
+}  // namespace mindspore::lite
+

mindspore/lite/src/ops/space_to_batch.cc

@@ -0,0 +1,91 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <vector>
+#include "src/ops/ops.h"
+#include "include/errorcode.h"
+#include "utils/log_adapter.h"
+#include "src/ir/tensor.h"
+
+namespace mindspore::lite {
+namespace {
+constexpr int kSpaceToBatchNDOutputNum = 1;
+constexpr int kSpaceToBatchNDInputNum = 1;
+constexpr int kBlockSizesSize = 2;
+constexpr int kPaddingsSize = 4;
+}  // namespace
+
+int SpaceToBatch::InferShape(std::vector<lite::tensor::Tensor *> inputs, std::vector<lite::tensor::Tensor *> outputs) {
+  MS_ASSERT(this->primitive != nullptr);
+  if (outputs.size() != kSpaceToBatchNDOutputNum || inputs.size() != kSpaceToBatchNDInputNum) {
+    MS_LOG(ERROR) << "Invalid output/input size! output size: " << outputs.size() << ",input size: " << inputs.size();
+    return RET_PARAM_INVALID;
+  }
+
+  auto input = inputs.at(0);
+  if (input->GetFormat() != schema::Format_NHWC) {
+    MS_LOG(ERROR) << "space_to_batch only support NHWC now!";
+    return RET_FORMAT_ERR;
+  }
+  auto input_shape = input->shape();
+  if (input_shape.size() != kDimension_4d) {
+    MS_LOG(ERROR) << "input shape dimension size should == " << kDimension_4d;
+    return RET_PARAM_INVALID;
+  }
+
+  auto prim = this->primitive->value_as_SpaceToBatch();
+  if (prim->blockShape()->size() != kBlockSizesSize) {
+    MS_LOG(ERROR) << "Block shape size should be " << kBlockSizesSize;
+    return RET_PARAM_INVALID;
+  }
+  if (prim->paddings()->size() != kPaddingsSize) {
+    MS_LOG(ERROR) << "Crops size should be " << kPaddingsSize;
+    return RET_PARAM_INVALID;
+  }
+
+  for (auto iter = prim->blockShape()->begin(); iter != prim->blockShape()->end(); ++iter) {
+    block_sizes_.emplace_back(*iter);
+  }
+
+  in_shape_.clear();
+  padded_in_shape_.clear();
+  paddings_.clear();
+  in_shape_.emplace_back(input_shape.at(kNHWC_n_index));
+  padded_in_shape_.emplace_back(input_shape.at(kNHWC_n_index));
+  for (int i = 0; i < kBlockSizesSize; i++) {
+    in_shape_.emplace_back(input_shape.at(i + 1));
+    padded_in_shape_.emplace_back(input_shape.at(i + 1) + (paddings_.at(2 * i) + paddings_.at(2 * i + 1)));
+    paddings_.emplace_back(paddings_.at(2 * i));
+    paddings_.emplace_back(paddings_.at(2 * i + 1));
+    if (paddings_.back() % block_sizes_.at(i)) {
+      MS_LOG(ERROR) << "Padded shape does not divide block size " << block_sizes_.at(i);
+      return RET_PARAM_INVALID;
+    }
+  }
+  in_shape_.emplace_back(input_shape.at(kNHWC_c_index));
+  padded_in_shape_.emplace_back(input_shape.at(kNHWC_c_index));
+
+  std::vector<int32_t> output_shape(input_shape.size());
+  output_shape[kNHWC_n_index] =
+    input_shape[kNHWC_n_index] * (block_sizes_[kNHWC_n_index] * block_sizes_[kNHWC_h_index]);
+  output_shape[kNHWC_h_index] = input_shape[kNHWC_h_index] / block_sizes_[kNHWC_n_index];
+  output_shape[kNHWC_w_index] = input_shape[kNHWC_w_index] / block_sizes_[kNHWC_h_index];
+  output_shape[kNHWC_c_index] = input_shape[kNHWC_c_index];
+  outputs[0]->set_shape(output_shape);
+  outputs[0]->set_data_type(input->data_type());
+  return RET_OK;
+}
+}  // namespace mindspore::lite

mindspore/lite/src/ops/space_to_depth.cc

@@ -0,0 +1,62 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <vector>
+#include "src/ops/ops.h"
+#include "include/errorcode.h"
+#include "utils/log_adapter.h"
+#include "src/ir/tensor.h"
+
+namespace mindspore::lite {
+namespace {
+constexpr int kSpaceToDepthOutputNum = 1;
+constexpr int kSpaceToDepthInputNum = 1;
+}
+
+int SpaceToDepth::InferShape(std::vector<lite::tensor::Tensor *> inputs, std::vector<lite::tensor::Tensor *> outputs) {
+  MS_ASSERT(this->primitive != nullptr);
+  if (outputs.size() != kSpaceToDepthOutputNum || inputs.size() != kSpaceToDepthInputNum) {
+    MS_LOG(ERROR) << "Invalid output/input size! output size: " << outputs.size() << ",input size: " << inputs.size();
+    return RET_PARAM_INVALID;
+  }
+
+  auto input = inputs.at(0);
+  if (input->GetFormat() != schema::Format_NHWC) {
+    MS_LOG(ERROR) << "space_to_depth only support NHWC now!";
+    return RET_FORMAT_ERR;
+  }
+  auto input_shape = input->shape();
+  if (input_shape.size() != kDimension_4d) {
+    MS_LOG(ERROR) << "input shape dimension size should == " << kDimension_4d;
+    return RET_PARAM_INVALID;
+  }
+  auto prim = this->primitive->value_as_SpaceToDepth();
+  int32_t block_size = prim->blockSize();
+  if (input_shape[kNHWC_c_index] % (block_size * block_size) != 0 || input_shape[kNHWC_c_index] == 0) {
+    MS_LOG(ERROR) << "input dimension c size " << input_shape[kNHWC_c_index] << " should be mulitple of block_size("
+                  << block_size << ") * block_size)!";
+    return RET_PARAM_INVALID;
+  }
+  std::vector<int32_t> output_shape(input_shape.size());
+  output_shape[kNHWC_n_index] = input_shape[kNHWC_n_index];
+  output_shape[kNHWC_h_index] = input_shape[kNHWC_h_index] / block_size;
+  output_shape[kNHWC_w_index] = input_shape[kNHWC_w_index] / block_size;
+  output_shape[kNHWC_c_index] = input_shape[kNHWC_c_index] * (block_size * block_size);
+  outputs[0]->set_shape(output_shape);
+  outputs[0]->set_data_type(input->data_type());
+  return RET_OK;
+}
+}  // namespace mindspore::lite

mindspore/lite/src/populate_parameter.cc

@@ -63,6 +63,10 @@
 #include "src/runtime/kernel/arm/opclib/fp32/one_hot.h"
 #include "src/runtime/kernel/arm/opclib/fp32/strided_slice.h"
 #include "src/runtime/kernel/arm/base/prior_box.h"
+#include "src/runtime/kernel/arm/opclib/fp32/space_to_depth.h"
+#include "src/runtime/kernel/arm/opclib/fp32/space_to_batch.h"
+#include "src/runtime/kernel/arm/opclib/int8/dequantize.h"
+#include "src/runtime/kernel/arm/opclib/fp32/quantize.h"
 
 namespace mindspore::kernel {
 FillParameter *PopulateFillParam(const lite::Primitive *primitive) {
@@ -873,6 +877,22 @@ DepthToSpaceParameter *PopulateDepthToSpaceParam(const lite::Primitive *primitiv
   return parameter;
 }
 
+SpaceToDepthParameter *PopulateSpaceToDepthParam(const lite::Primitive *primitive) {
+  SpaceToDepthParameter *parameter = new (std::nothrow) SpaceToDepthParameter();
+  if (parameter == nullptr) {
+    MS_LOG(ERROR) << "new SpaceToDepthParameter failed.";
+    return nullptr;
+  }
+  auto param = primitive->Value()->value_as_DepthToSpace();
+  parameter->op_parameter_.type_ = primitive->Type();
+  parameter->block_size_ = param->blockSize();
+  if (param->format() != schema::Format_NHWC) {
+    MS_LOG(ERROR) << "Currently only NHWC format is supported.";
+    return nullptr;
+  }
+  return parameter;
+}
+
 ResizeParameter *PopulateResizeParameter(const lite::Primitive *primitive) {
   ResizeParameter *parameter = new (std::nothrow) ResizeParameter();
   if (parameter == nullptr) {
@@ -961,6 +981,24 @@ FlattenParameter *PopulateFlattenParameter(const lite::Primitive *primitive) {
   return parameter;
 }
 
+DequantizeParameter *PopulateDequantizeParameter(const lite::Primitive *primitive) {
+  DequantizeParameter *parameter = new (std::nothrow) DequantizeParameter();
+  if (parameter == nullptr) {
+    MS_LOG(ERROR) << "new DequantizeParameter fail!";
+    return nullptr;
+  }
+    return parameter;
+}
+
+QuantizeParameter *PopulateQuantizeParameter(const lite::Primitive *primitive) {
+  QuantizeParameter *parameter = new (std::nothrow) QuantizeParameter();
+  if (parameter == nullptr) {
+    MS_LOG(ERROR) << "new QuantizeParameter fail!";
+    return nullptr;
+  }
+    return parameter;
+}
+
 StridedSliceParameter *PopulateStridedSliceParam(const lite::Primitive *primitive) {
   StridedSliceParameter *parameter = new (std::nothrow) StridedSliceParameter();
   if (parameter == nullptr) {
@@ -1045,6 +1083,24 @@ PriorBoxParameter *PopulatePriorBoxParameter(const lite::Primitive *primitive) {
   return param;
 }
 
+SpaceToBatchParameter *PopulateSpaceToBatchParam(const lite::Primitive *primitive) {
+  SpaceToBatchParameter *parameter = new (std::nothrow) SpaceToBatchParameter();
+  if (parameter == nullptr) {
+    MS_LOG(ERROR) << "new SpaceToBatchParameter failed.";
+    return nullptr;
+  }
+  parameter->op_parameter_.type_ = primitive->Type();
+  auto block_sizes = ((lite::SpaceToBatch *)primitive)->BlockSizes();
+  (void)memcpy(parameter->block_sizes_, (block_sizes.data()), block_sizes.size() * sizeof(int));
+  auto paddings = ((lite::SpaceToBatch *)primitive)->Paddings();
+  (void)memcpy(parameter->paddings_, (paddings.data()), paddings.size() * sizeof(int));
+  auto in_shape = ((lite::SpaceToBatch *)primitive)->InShape();
+  (void)memcpy(parameter->in_shape_, (in_shape.data()), in_shape.size() * sizeof(int));
+  auto padded_in_shape = ((lite::SpaceToBatch *)primitive)->PaddedInShape();
+  (void)memcpy(parameter->padded_in_shape_, (padded_in_shape.data()), padded_in_shape.size() * sizeof(int));
+  return parameter;
+}
+
 OpParameter *PopulateParameter(const lite::Primitive *primitive) {
   MS_EXCEPTION_IF_NULL(primitive);
   auto op_type = primitive->Type();
@@ -1166,6 +1222,10 @@ OpParameter *PopulateParameter(const lite::Primitive *primitive) {
       return reinterpret_cast<OpParameter *>(PopulateAddNParam(primitive));
     case schema::PrimitiveType_PriorBox:
       return reinterpret_cast<OpParameter *>(PopulatePriorBoxParameter(primitive));
+    case schema::PrimitiveType_OnnxInt8Dequantize:
+      return reinterpret_cast<OpParameter *>(PopulateDequantizeParameter(primitive));
+    case schema::PrimitiveType_OnnxInt8Quantize:
+      return reinterpret_cast<OpParameter *>(PopulateQuantizeParameter(primitive));
     default:
       break;
   }

mindspore/lite/src/runtime/kernel/arm/fp32/quantize.cc

@@ -0,0 +1,118 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "src/runtime/kernel/arm/fp32/quantize.h"
+#include <vector>
+#include "src/runtime/kernel/arm/opclib/fp32/quantize.h"
+#include "src/kernel_registry.h"
+#include "src/runtime/runtime_api.h"
+#include "schema/model_generated.h"
+#include "include/errorcode.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_OnnxInt8Quantize;
+
+namespace mindspore::kernel {
+namespace {
+constexpr int kQuantizeInputNum = 1;
+constexpr int kQuantizeOutputNum = 1;
+}  // namespace
+
+int QuantizeCPUKernel::Init() {
+  if (inputs_.size() != 1) {
+    MS_LOG(ERROR) << "inputs number should be 1, but " << inputs_.size() << " is given.";
+    return RET_ERROR;
+  }
+  if (outputs_.size() != 1) {
+    MS_LOG(ERROR) << "outputs number should be 1, but " << inputs_.size() << " is given.";
+    return RET_ERROR;
+  }
+  auto in_tensor = inputs_.front();
+  num_unit_ = static_cast<int>(in_tensor->DataSize());
+  thread_n_num_ = MSMIN(thread_num_, num_unit_);
+  thread_n_stride_ = UP_DIV(num_unit_, thread_n_num_);
+
+  return RET_OK;
+}
+
+int QuantizeCPUKernel::ReSize() { return RET_OK; }
+
+int QuantizeCPUKernel::Quantize(int task_id) {
+  int num_unit_thread = MSMIN(thread_n_stride_, num_unit_ - task_id * thread_n_stride_);
+  if (num_unit_thread <= 0) {
+    return RET_OK;
+  }
+  int thread_offset = task_id * thread_n_stride_;
+  auto quant_arg = inputs_.front()->GetQuantParams().front();
+  int ret = QuantizeToInt8(input_ptr_ + thread_offset, output_ptr_ + thread_offset, quant_arg.scale,
+                           quant_arg.zeroPoint, num_unit_thread);
+
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Quantize error task_id[" << task_id << "] error_code[" << ret << "]";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int QuantizeRun(int task_id, LiteParallelGroupEnv *penv, void *cdata) {
+  auto g_kernel = reinterpret_cast<QuantizeCPUKernel *>(cdata);
+  auto ret = g_kernel->Quantize(task_id);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "QuantizeRun error task_id[" << task_id << "] error_code[" << ret << "]";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int QuantizeCPUKernel::Run() {
+  input_ptr_ = reinterpret_cast<float *>(inputs_[0]->Data());
+  output_ptr_ = reinterpret_cast<int8_t *>(outputs_[0]->Data());
+  int ret = LiteBackendParallelLaunch(QuantizeRun, this, thread_n_num_);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Scale error error_code[" << ret << "]";
+    return RET_ERROR;
+  }
+
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuQuantizeFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                 const std::vector<lite::tensor::Tensor *> &outputs,
+                                                 OpParameter *opParameter, const lite::Context *ctx,
+                                                 const kernel::KernelKey &desc) {
+  if (opParameter == nullptr) {
+    MS_LOG(ERROR) << "Input opParameter is nullptr!";
+    return nullptr;
+  }
+  auto *kernel = new (std::nothrow) QuantizeCPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new QuantizeCPUKernel fail!";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Init kernel failed! name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_OnnxInt8Quantize, CpuQuantizeFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/quantize.h

@@ -0,0 +1,46 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_QUANTIZE_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_QUANTIZE_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+
+namespace mindspore::kernel {
+class QuantizeCPUKernel : public LiteKernel {
+ public:
+  QuantizeCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                    const std::vector<lite::tensor::Tensor *> &outputs, const lite::Context *ctx)
+      : LiteKernel(parameter, inputs, outputs), thread_num_(ctx->threadNum) {}
+  ~QuantizeCPUKernel() = default;
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+  int Quantize(int task_id);
+
+ private:
+  int thread_num_;
+  int thread_n_num_;
+  int thread_n_stride_;
+  int num_unit_;
+  float *input_ptr_;
+  int8_t *output_ptr_;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_QUANTIZE_H_

mindspore/lite/src/runtime/kernel/arm/fp32/space_to_batch.cc

@@ -0,0 +1,106 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "src/runtime/kernel/arm/fp32/space_to_batch.h"
+#include <vector>
+#include "schema/ops_generated.h"
+#include "schema/model_generated.h"
+#include "src/kernel_registry.h"
+#include "src/runtime/kernel/arm/opclib/fp32/space_to_batch.h"
+#include "src/runtime/kernel/arm/opclib/errorcode.h"
+#include "include/errorcode.h"
+
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_FORMAT_ERR;
+using mindspore::lite::RET_OK;
+using mindspore::lite::RET_OP_EXECUTE_FAILURE;
+using mindspore::schema::PrimitiveType_SpaceToBatch;
+
+namespace mindspore::kernel {
+
+int SpaceToBatchCPUKernel::Init() {
+  if (inputs_[0]->GetFormat() != schema::Format_NHWC) {
+    MS_LOG(ERROR) << "space_to_batch only support NHWC now!";
+    return RET_FORMAT_ERR;
+  }
+  SpaceToBatchParameter *param = reinterpret_cast<SpaceToBatchParameter *>(this->opParameter);
+  for (int i = 0; i < SPACE_TO_BATCH_PADDINGS_SIZE; ++i) {
+    if (param->paddings_[i] != 0) {
+      param->need_paddings_ = true;
+      break;
+    }
+  }
+  param->n_dims_ = DIMENSION_4D;
+  param->n_space_dims_ = SPACE_TO_BATCH_BLOCK_SIZES_SIZE;
+  param->num_elements_ = EnumElement(param->in_shape_, param->n_dims_);
+  param->num_elements_padded_ = EnumElement(param->padded_in_shape_, param->n_dims_);
+  return RET_OK;
+}
+
+int SpaceToBatchCPUKernel::Run() {
+  auto input = inputs_[0];
+  auto output = outputs_[0];
+  input_ptr_ = reinterpret_cast<const float *>(input->Data());
+  output_ptr_ = reinterpret_cast<float *>(output->Data());
+  SpaceToBatchParameter *param = reinterpret_cast<SpaceToBatchParameter *>(this->opParameter);
+
+  int ret;
+  float *tmp_space[3] = {nullptr, nullptr, nullptr};
+  if (param->need_paddings_) {
+    tmp_space[0] = reinterpret_cast<float *>(malloc(param->num_elements_padded_ * sizeof(float)));
+    (void)memset(tmp_space[0], 0, param->num_elements_padded_);
+    tmp_space[1] = reinterpret_cast<float *>(malloc(param->num_elements_padded_ * sizeof(float)));
+    (void)memset(tmp_space[1], 0, param->num_elements_padded_);
+    tmp_space[2] = reinterpret_cast<float *>(malloc(param->num_elements_padded_ * sizeof(float)));
+    (void)memset(tmp_space[2], 0, param->num_elements_padded_);
+
+    ret = SpaceToBatch(input_ptr_, output_ptr_, *param, tmp_space);
+  } else {
+    ret = SpaceToBatch(input_ptr_, output_ptr_, *param, tmp_space);
+  }
+  if (ret != OPCLIB_OK) {
+    MS_LOG(ERROR) << "Do space to batch fails!";
+    return RET_OP_EXECUTE_FAILURE;
+  }
+
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuSpaceToBatchFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                     const std::vector<lite::tensor::Tensor *> &outputs,
+                                                     OpParameter *opParameter, const lite::Context *ctx,
+                                                     const kernel::KernelKey &desc) {
+  if (opParameter == nullptr) {
+    MS_LOG(ERROR) << "Input opParameter is nullptr!";
+    return nullptr;
+  }
+  auto *kernel = new (std::nothrow) SpaceToBatchCPUKernel(opParameter, inputs, outputs);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new SpaceToBatchCPUKernel fail!";
+    return nullptr;
+  }
+
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_SpaceToBatch, CpuSpaceToBatchFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/space_to_batch.h

@@ -0,0 +1,41 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_SPACE_TO_BATCH_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_SPACE_TO_BATCH_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+
+namespace mindspore::kernel {
+class SpaceToBatchCPUKernel : public LiteKernel {
+ public:
+  SpaceToBatchCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                        const std::vector<lite::tensor::Tensor *> &outputs)
+      : LiteKernel(parameter, inputs, outputs) {}
+
+  ~SpaceToBatchCPUKernel() = default;
+
+  int Init() override;
+  int ReSize() override { return 0; }
+  int Run() override;
+
+ private:
+  const float *input_ptr_;
+  float *output_ptr_;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_SPACE_TO_BATCH_H_

mindspore/lite/src/runtime/kernel/arm/fp32/space_to_depth.cc

@@ -0,0 +1,89 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/runtime/kernel/arm/fp32/space_to_depth.h"
+#include <vector>
+#include "schema/ops_generated.h"
+#include "schema/model_generated.h"
+#include "src/kernel_registry.h"
+#include "src/runtime/kernel/arm/opclib/fp32/space_to_depth.h"
+#include "include/errorcode.h"
+
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_FORMAT_ERR;
+using mindspore::lite::RET_OK;
+using mindspore::lite::RET_PARAM_INVALID;
+using mindspore::schema::PrimitiveType_SpaceToDepth;
+
+namespace mindspore::kernel {
+
+int SpaceToDepthCPUKernel::Init() {
+  if (inputs_[0]->GetFormat() != schema::Format_NHWC) {
+    MS_LOG(ERROR) << "space_to_depth only support NHWC now!";
+    return RET_FORMAT_ERR;
+  }
+  SpaceToDepthParameter *param = reinterpret_cast<SpaceToDepthParameter *>(opParameter);
+  if (param->block_size_ <= 0) {
+    MS_LOG(ERROR) << "Input block_size should > 0!";
+    return RET_PARAM_INVALID;
+  }
+  return RET_OK;
+}
+
+int SpaceToDepthCPUKernel::Run() {
+  auto input = inputs_[0];
+  auto output = outputs_[0];
+  const float *input_data = static_cast<const float *>(input->Data());
+  float *output_data = static_cast<float *>(output->Data());
+  auto in_shape = input->shape();
+  auto out_shape = output->shape();
+  SpaceToDepthParameter *param = reinterpret_cast<SpaceToDepthParameter *>(opParameter);
+  if (input->GetFormat() == schema::Format_NHWC) {
+    auto ret = SpaceToDepthForNHWC(input_data, output_data, in_shape.data(), out_shape.data(), in_shape.size(),
+                                   param->block_size_);
+    return ret;
+  } else {
+    MS_LOG(ERROR) << "Only support NHWC now!";
+    return RET_ERROR;
+  }
+}
+kernel::LiteKernel *CpuSpaceToDepthFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                     const std::vector<lite::tensor::Tensor *> &outputs,
+                                                     OpParameter *opParameter, const lite::Context *ctx,
+                                                     const kernel::KernelKey &desc) {
+  if (opParameter == nullptr) {
+    MS_LOG(ERROR) << "Input opParameter is nullptr!";
+    return nullptr;
+  }
+  auto *kernel = new (std::nothrow) SpaceToDepthCPUKernel(opParameter, inputs, outputs);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new SpaceToDepthCPUKernel fail!";
+    return nullptr;
+  }
+
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_SpaceToDepth, CpuSpaceToDepthFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/space_to_depth.h

@@ -0,0 +1,37 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MINDSPORE_LITE_SRC_BACKEND_ARM_FP32_SPACE_TO_DEPTH_H_
+#define MINDSPORE_LITE_SRC_BACKEND_ARM_FP32_SPACE_TO_DEPTH_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+
+namespace mindspore::kernel {
+class SpaceToDepthCPUKernel : public LiteKernel {
+ public:
+  SpaceToDepthCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                        const std::vector<lite::tensor::Tensor *> &outputs)
+      : LiteKernel(parameter, inputs, outputs) {}
+  ~SpaceToDepthCPUKernel() = default;
+
+  int Init() override;
+  int ReSize() override { return 0; };
+  int Run() override;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_BACKEND_ARM_FP32_SPACE_TO_DEPTH_H_

mindspore/lite/src/runtime/kernel/arm/int8/dequantize.cc

@@ -0,0 +1,118 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "src/runtime/kernel/arm/int8/dequantize.h"
+#include <vector>
+#include "src/runtime/kernel/arm/opclib/int8/dequantize.h"
+#include "src/runtime/runtime_api.h"
+#include "src/kernel_registry.h"
+#include "schema/model_generated.h"
+#include "include/errorcode.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_OnnxInt8Dequantize;
+
+namespace mindspore::kernel {
+namespace {
+constexpr int kDequantizeInputNum = 1;
+constexpr int kDequantizeOutputNum = 1;
+}  // namespace
+
+int DequantizeCPUKernel::Init() {
+  if (inputs_.size() != 1) {
+    MS_LOG(ERROR) << "inputs number should be 1, but " << inputs_.size() << " is given.";
+    return RET_ERROR;
+  }
+  if (outputs_.size() != 1) {
+    MS_LOG(ERROR) << "outputs number should be 1, but " << inputs_.size() << " is given.";
+    return RET_ERROR;
+  }
+  auto in_tensor = inputs_.front();
+  num_unit_ = static_cast<int>(in_tensor->DataSize());
+  thread_n_num_ = MSMIN(thread_num_, num_unit_);
+  thread_n_stride_ = UP_DIV(num_unit_, thread_n_num_);
+
+  return RET_OK;
+}
+
+int DequantizeCPUKernel::ReSize() { return RET_OK; }
+
+int DequantizeCPUKernel::Dequantize(int task_id) {
+  int num_unit_thread = MSMIN(thread_n_stride_, num_unit_ - task_id * thread_n_stride_);
+  if (num_unit_thread <= 0) {
+    return RET_OK;
+  }
+  int thread_offset = task_id * thread_n_stride_;
+  auto quant_arg = inputs_.front()->GetQuantParams().front();
+  int ret = DequantizeInt8(input_ptr_ + thread_offset, output_ptr_ + thread_offset, quant_arg.scale,
+                           quant_arg.zeroPoint, num_unit_thread);
+
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Dequantize error task_id[" << task_id << "] error_code[" << ret << "]";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int DequantizeRun(int task_id, LiteParallelGroupEnv *penv, void *cdata) {
+  auto g_kernel = reinterpret_cast<DequantizeCPUKernel *>(cdata);
+  auto ret = g_kernel->Dequantize(task_id);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "DequantizeRun error task_id[" << task_id << "] error_code[" << ret << "]";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int DequantizeCPUKernel::Run() {
+  input_ptr_ = reinterpret_cast<int8_t *>(inputs_[0]->Data());
+  output_ptr_ = reinterpret_cast<float *>(outputs_[0]->Data());
+  int ret = LiteBackendParallelLaunch(DequantizeRun, this, thread_n_num_);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Scale error error_code[" << ret << "]";
+    return RET_ERROR;
+  }
+
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuDequantizeFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                   const std::vector<lite::tensor::Tensor *> &outputs,
+                                                   OpParameter *opParameter, const lite::Context *ctx,
+                                                   const kernel::KernelKey &desc) {
+  if (opParameter == nullptr) {
+    MS_LOG(ERROR) << "Input opParameter is nullptr!";
+    return nullptr;
+  }
+  auto *kernel = new (std::nothrow) DequantizeCPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new DequantizeCPUKernel fail!";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Init kernel failed! name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_OnnxInt8Dequantize, CpuDequantizeFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/int8/dequantize.h

@@ -0,0 +1,46 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_INT8_DEQUANTIZE_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_INT8_DEQUANTIZE_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+
+namespace mindspore::kernel {
+class DequantizeCPUKernel : public LiteKernel {
+ public:
+  DequantizeCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                      const std::vector<lite::tensor::Tensor *> &outputs, const lite::Context *ctx)
+      : LiteKernel(parameter, inputs, outputs), thread_num_(ctx->threadNum) {}
+  ~DequantizeCPUKernel() = default;
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+  int Dequantize(int task_id);
+
+ private:
+  int thread_num_;
+  int thread_n_num_;
+  int thread_n_stride_;
+  int num_unit_;
+  int8_t *input_ptr_;
+  float *output_ptr_;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_INT8_DEQUANTIZE_H_

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/quantize.cc

@@ -0,0 +1,29 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/runtime/kernel/arm/opclib/fp32/quantize.h"
+#include "src/runtime/kernel/arm/opclib/errorcode.h"
+
+int QuantizeToInt8(float *real_values, int8_t *quant_values, float scale, int32_t zp, int size) {
+  if (quant_values == nullptr || real_values == nullptr) {
+    return OPCLIB_PARAM_INVALID;
+  }
+
+  for (int i = 0; i < size; ++i) {
+    quant_values[i] = (int8_t)round(real_values[i] / scale + zp);
+  }
+  return OPCLIB_OK;
+}

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/quantize.h

@@ -0,0 +1,28 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_QUANTIZE_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_QUANTIZE_H_
+
+#include "src/runtime/kernel/arm/opclib/op_base.h"
+
+struct QuantizeParameter {
+  OpParameter op_parameter_;
+};
+
+int QuantizeToInt8(float *real_values, int8_t *quant_values, float scale, int32_t zp, int size);
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_QUANTIZE_H_

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/space_to_batch.cc

@@ -0,0 +1,157 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "src/runtime/kernel/arm/opclib/fp32/space_to_batch.h"
+#include "src/runtime/kernel/arm/opclib/arithmetic_common.h"
+#include "src/runtime/kernel/arm/opclib/errorcode.h"
+#include "src/runtime/kernel/arm/opclib/fp32/concat.h"
+#include "src/runtime/kernel/arm/opclib/op_base.h"
+
+int EnumElement(int *shape, int n_dims) {
+  int total = 1;
+  for (int i = 0; i < n_dims; i++) {
+    total *= shape[i];
+  }
+  return total;
+}
+
+void TransposeForNHWC(const float *in_data, float *out_data, int *strides, int *out_strides, int *perm,
+                      int *output_shape) {
+  const int stride0 = strides[perm[0]];
+  const int stride1 = strides[perm[1]];
+  const int stride2 = strides[perm[2]];
+  const int stride3 = strides[perm[3]];
+  const int stride4 = strides[perm[4]];
+  const int out_stride0 = out_strides[0];
+  const int out_stride1 = out_strides[1];
+  const int out_stride2 = out_strides[2];
+  const int out_stride3 = out_strides[3];
+  const int out_stride4 = out_strides[4];
+  const int output0 = output_shape[0];
+  const int output1 = output_shape[1];
+  const int output2 = output_shape[2];
+  const int output3 = output_shape[3];
+  const int output4 = output_shape[4];
+
+  for (int i = 0; i < output0; ++i) {
+    int out_stride0_i = i * out_stride0;
+    int stride0_i = i * stride0;
+    for (int j = 0; j < output1; ++j) {
+      int out_stride1_j = j * out_stride1;
+      int stride1_j = j * stride1;
+      for (int k = 0; k < output2; ++k) {
+        int out_stride2_k = k * out_stride2;
+        int stride2_k = k * stride2;
+        for (int m = 0; m < output3; ++m) {
+          int out_stride3_m = m * out_stride3;
+          int stride3_m = m * stride3;
+          for (int n = 0; n < output4; ++n) {
+            int out_stride4_n = n * out_stride4;
+            int stride4_n = n * stride4;
+            memcpy(out_data + out_stride0_i + out_stride1_j + out_stride2_k + out_stride3_m + out_stride4_n,
+                   in_data + stride0_i + stride1_j + stride2_k + stride3_m + stride4_n, stride4 * sizeof(float));
+          }
+        }
+      }
+    }
+  }
+}
+
+int SpaceToBatchForNHWC(const float *input, float *output, int *in_shape, int shape_size, int *block_sizes) {
+  int trans_in_shape[6] = {in_shape[0],    in_shape[1] / block_sizes[0],
+                           block_sizes[0], in_shape[2] / block_sizes[1],
+                           block_sizes[1], in_shape[3]};
+  int trans_out_shape[6] = {
+    in_shape[0], block_sizes[0], block_sizes[1], in_shape[1] / block_sizes[0], in_shape[2] / block_sizes[1],
+    in_shape[3]};
+  int in_strides[C4NUM + 2];
+  ComputeStrides(trans_in_shape, in_strides, shape_size + 2);
+  int out_strides[C4NUM + 2];
+  ComputeStrides(trans_out_shape, out_strides, shape_size + 2);
+
+  int perm[6] = {0, 2, 4, 1, 3, 5};
+  TransposeForNHWC(input, output, in_strides, out_strides, perm, trans_out_shape);
+  return OPCLIB_OK;
+}
+
+void DoPadding(const float *input, float *padded_input, SpaceToBatchParameter param, float *tmp_space[]) {
+  float *tmp = padded_input;
+  (void)memcpy(tmp, input, param.num_elements_ * sizeof(float));
+  float *target = tmp_space[0];
+  float *tmp_zeros = tmp_space[1];
+  float *tmp2 = nullptr;
+  int cur_shape[param.n_dims_], cur_start_shape[param.n_dims_], cur_end_shape[param.n_dims_],
+    cur_target_shape[param.n_dims_];
+  float *concat_inputs[3];
+  int *concat_shapes[4];
+
+  for (int i = 0; i < param.n_dims_; i++) {
+    cur_shape[i] = param.in_shape_[i];
+    cur_start_shape[i] = param.in_shape_[i];
+    cur_end_shape[i] = param.in_shape_[i];
+    cur_target_shape[i] = param.in_shape_[i];
+  }
+  for (int i = 0; i < param.n_space_dims_; ++i) {
+    if (param.padded_in_shape_[i + 1] > param.in_shape_[i + 1]) {
+      int concat_idx = 0;
+      cur_target_shape[i + 1] = 0;
+      if (param.paddings_[2 * i] != 0) {
+        cur_start_shape[i + 1] = param.paddings_[2 * i];
+        concat_inputs[concat_idx] = tmp_zeros;
+        concat_shapes[concat_idx++] = cur_start_shape;
+        cur_target_shape[i + 1] += cur_start_shape[i + 1];
+      }
+
+      concat_inputs[concat_idx] = tmp;
+      concat_shapes[concat_idx++] = cur_shape;
+      cur_target_shape[i + 1] += cur_shape[i + 1];
+      if (param.paddings_[2 * i + 1] != 0) {
+        cur_end_shape[i + 1] = param.paddings_[2 * i + 1];
+        concat_inputs[concat_idx] = tmp_zeros;
+        concat_shapes[concat_idx++] = cur_end_shape;
+        cur_target_shape[i + 1] += cur_end_shape[i + 1];
+      }
+      concat_shapes[concat_idx] = cur_target_shape;
+      Concat((void **)concat_inputs, concat_idx, i + 1, concat_shapes, param.n_dims_, target);
+
+      tmp2 = tmp;
+      tmp = target;
+      target = tmp2;
+      cur_start_shape[i + 1] = cur_end_shape[i + 1] = cur_shape[i + 1] = concat_shapes[concat_idx][i + 1];
+    }
+  }
+  if (padded_input != tmp) {
+    memcpy(padded_input, tmp, param.num_elements_padded_ * sizeof(float));
+  }
+}
+
+int SpaceToBatch(const float *input, float *output, SpaceToBatchParameter param, float *tmp_space[3]) {
+  float *padded_input;
+  int ret;
+  if (param.need_paddings_) {
+    if (tmp_space[0] == nullptr || tmp_space[1] == nullptr || tmp_space[2] == nullptr) {
+      return OPCLIB_NULL_PTR;
+    }
+    padded_input = tmp_space[0];
+    DoPadding(input, padded_input, param, tmp_space + 1);
+  }
+
+  if (param.need_paddings_) {
+    ret = SpaceToBatchForNHWC(padded_input, output, param.padded_in_shape_, param.n_dims_, param.block_sizes_);
+  } else {
+    ret = SpaceToBatchForNHWC(input, output, param.padded_in_shape_, param.n_dims_, param.block_sizes_);
+  }
+  return ret;
+}

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/space_to_batch.h

@@ -0,0 +1,42 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MINDSPORE_LITE_SRC_BACKEND_ARM_OPCLIB_FP32_SPACE_TO_BATCH_H_
+#define MINDSPORE_LITE_SRC_BACKEND_ARM_OPCLIB_FP32_SPACE_TO_BATCH_H_
+#include "src/runtime/kernel/arm/opclib/op_base.h"
+
+#define SPACE_TO_BATCH_BLOCK_SIZES_SIZE 2
+#define SPACE_TO_BATCH_PADDINGS_SIZE 4
+
+struct SpaceToBatchParameter {
+  OpParameter op_parameter_;
+  int block_sizes_[8];
+  int paddings_[8];
+  int n_dims_;
+  int num_elements_;
+  int num_elements_padded_;
+  int n_space_dims_;
+  int in_shape_[8];
+  int padded_in_shape_[8];
+  bool need_paddings_ = false;
+};
+
+int SpaceToBatch(const float *input, float *output, SpaceToBatchParameter param, float *tmp_space[3]);
+int SpaceToBatchForNHWC(const float *input, float *output, int *in_shape, int shape_size, int *block_size);
+void TransposeForNHWC(const float *in_data, float *out_data, int *strides, int *out_strides, int *perm,
+                      int *output_shape);
+void DoPadding(const float *input, float *padded_input, SpaceToBatchParameter param);
+int EnumElement(int *shape, int n_dims);
+#endif  // MINDSPORE_LITE_SRC_BACKEND_ARM_OPCLIB_FP32_SPACE_TO_BATCH_H_

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/space_to_depth.cc

@@ -0,0 +1,50 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "src/runtime/kernel/arm/opclib/fp32/space_to_depth.h"
+#include "src/runtime/kernel/arm/opclib/arithmetic_common.h"
+#include "src/runtime/kernel/arm/opclib/errorcode.h"
+#include "src/runtime/kernel/arm/opclib/op_base.h"
+
+int SpaceToDepthForNHWC(const float *input, float *output, int *in_shape, int *out_shape, int shape_size,
+                        int block_size) {
+  if (input == nullptr || output == nullptr) {
+    return OPCLIB_NULL_PTR;
+  }
+  if (shape_size != C4NUM) {
+    return OPCLIB_PARAM_INVALID;
+  }
+  int in_strides[C4NUM];
+  ComputeStrides(in_shape, in_strides, shape_size);
+  int out_strides[C4NUM];
+  ComputeStrides(out_shape, out_strides, shape_size);
+  for (int i = 0; i < out_shape[0]; ++i) {
+    size_t in_offset_n = i * in_strides[0];
+    size_t out_offset_n = i * out_strides[0];
+    for (int j = 0; j < out_shape[1]; ++j) {
+      size_t in_offset_h = in_offset_n + j * block_size * in_strides[1];
+      size_t out_offset_h = out_offset_n + j * out_strides[1];
+      for (int k = 0; k < out_shape[2]; ++k) {
+        size_t in_offset_w = in_offset_h + k * block_size * in_strides[2];
+        size_t out_offset_w = out_offset_h + k * out_strides[2];
+        for (int l = 0; l < block_size; ++l) {
+          memcpy(output + out_offset_w + l * block_size * in_strides[2], input + in_offset_w + l * in_strides[1],
+                 block_size * in_strides[2] * sizeof(float));
+        }
+      }
+    }
+  }
+  return OPCLIB_OK;
+}

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/space_to_depth.h

@@ -0,0 +1,27 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MINDSPORE_LITE_SRC_BACKEND_ARM_OPCLIB_FP32_SPACE_TO_DEPTH_H_
+#define MINDSPORE_LITE_SRC_BACKEND_ARM_OPCLIB_FP32_SPACE_TO_DEPTH_H_
+#include "src/runtime/kernel/arm/opclib/op_base.h"
+
+struct SpaceToDepthParameter {
+  OpParameter op_parameter_;
+  int32_t block_size_;
+};
+
+int SpaceToDepthForNHWC(const float *input, float *output, int *in_shape, int *out_shape, int shape_size,
+                        int block_size);
+#endif  // MINDSPORE_LITE_SRC_BACKEND_ARM_OPCLIB_FP32_SPACE_TO_DEPTH_H_

mindspore/lite/src/runtime/kernel/arm/opclib/int8/dequantize.cc

@@ -0,0 +1,29 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/runtime/kernel/arm/opclib/int8/dequantize.h"
+#include "src/runtime/kernel/arm/opclib/errorcode.h"
+
+int DequantizeInt8(int8_t *quant_values, float *real_values, float scale, int32_t zp, int size) {
+  if (quant_values == nullptr || real_values == nullptr) {
+    return OPCLIB_PARAM_INVALID;
+  }
+
+  for (int i = 0; i < size; ++i) {
+    real_values[i] = (quant_values[i] + zp) * scale;
+  }
+  return OPCLIB_OK;
+}

mindspore/lite/src/runtime/kernel/arm/opclib/int8/dequantize.h

@@ -0,0 +1,28 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_INT8_DEQUANTIZE_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_INT8_DEQUANTIZE_H_
+
+#include "src/runtime/kernel/arm/opclib/op_base.h"
+
+struct DequantizeParameter {
+  OpParameter op_parameter_;
+};
+
+int DequantizeInt8(int8_t *quant_values, float *real_values, float scale, int32_t zp, int size);
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_INT8_DEQUANTIZE_H_

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32/quantize_fp32_tests.cc

@@ -0,0 +1,77 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include <iostream>
+#include <memory>
+#include "utils/log_adapter.h"
+#include "common/common_test.h"
+#include "mindspore/lite/src/runtime/kernel/arm/fp32/quantize.h"
+#include "mindspore/lite/src/runtime/kernel/arm/opclib/fp32/quantize.h"
+#include "mindspore/lite/src/kernel_registry.h"
+#include "mindspore/lite/src/lite_kernel.h"
+
+namespace mindspore {
+
+class QuantizeTestFp32 : public mindspore::Common {
+ public:
+  QuantizeTestFp32() {}
+};
+
+TEST_F(QuantizeTestFp32, QuantizeTest1) {
+  const lite::tensor::QuantArg quant_arg = {0.3515625, -57};
+  QuantizeParameter param;
+  param.op_parameter_.type_ = schema::PrimitiveType_OnnxInt8Quantize;
+
+  std::vector<float> input = {1, 2, 5, 6, 10, -20, 3, 8, 18, 10, 3, 4, 11, 16, 15, 25};
+  std::vector<int> in_shape = {1, 4, 4, 1};
+  lite::tensor::Tensor input_tensor;
+  input_tensor.SetData(input.data());
+  input_tensor.set_shape(in_shape);
+  input_tensor.SetFormat(schema::Format_NHWC);
+  input_tensor.set_data_type(kNumberTypeFloat32);
+  input_tensor.AddQuantParam(quant_arg);
+  std::vector<lite::tensor::Tensor *> inputs_tensor;
+  inputs_tensor.emplace_back(&input_tensor);
+
+  const int out_size = 16;
+  int8_t expect_out[16] = {-54, -51, -43, -40, -29, -114, -48, -34, -6, -29, -48, -46, -26, -11, -14, 14};
+  std::vector<int8_t> output(16);
+  std::vector<int> out_shape = {1, 4, 4, 1};
+  lite::tensor::Tensor output_tensor;
+  output_tensor.SetData(output.data());
+  output_tensor.set_shape(out_shape);
+  output_tensor.SetFormat(schema::Format_NHWC);
+  output_tensor.set_data_type(kNumberTypeInt8);
+  std::vector<lite::tensor::Tensor *> outputs_tensor;
+  outputs_tensor.emplace_back(&output_tensor);
+
+  lite::Context ctx;
+  ctx.threadNum = 3;
+  kernel::KernelKey desc = {kernel::KERNEL_ARCH::kCPU, kNumberTypeFloat32, schema::PrimitiveType_OnnxInt8Quantize};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
+  kernel::LiteKernel *kernel =
+    creator(inputs_tensor, outputs_tensor, reinterpret_cast<OpParameter *>(&param), &ctx, desc);
+  ASSERT_NE(kernel, nullptr);
+  kernel->Run();
+
+  for (int i = 0; i < out_size; ++i) {
+    std::cout << output[i] << " ";
+  }
+  std::cout << "\n";
+  CompareOutputData(output.data(), expect_out, out_size, 0.000001);
+}
+
+}  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32/space_to_batch_fp32_tests.cc

@@ -0,0 +1,163 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include <iostream>
+#include <memory>
+#include "utils/log_adapter.h"
+#include "common/common_test.h"
+#include "mindspore/lite/src/runtime/kernel/arm/opclib/fp32/space_to_batch.h"
+#include "mindspore/lite/src/kernel_registry.h"
+#include "mindspore/lite/src/lite_kernel.h"
+
+namespace mindspore {
+
+class SpaceToBatchTestFp32 : public mindspore::Common {
+ public:
+  SpaceToBatchTestFp32() {}
+};
+
+int InitSpaceToBatchParameter(SpaceToBatchParameter *param) {
+  param->n_dims_ = 4;
+  param->n_space_dims_ = 2;
+
+  param->block_sizes_[0] = 2;
+  param->block_sizes_[1] = 2;
+
+  param->paddings_[0] = 2;
+  param->paddings_[1] = 0;
+  param->paddings_[2] = 2;
+  param->paddings_[3] = 2;
+
+  param->in_shape_[0] = 1;
+  param->in_shape_[1] = 4;
+  param->in_shape_[2] = 4;
+  param->in_shape_[3] = 1;
+
+  param->padded_in_shape_[0] = 1;
+  param->padded_in_shape_[1] = 6;
+  param->padded_in_shape_[2] = 8;
+  param->padded_in_shape_[3] = 1;
+
+  param->num_elements_ = 16;
+  param->num_elements_padded_ = 48;
+
+  param->need_paddings_ = true;
+}
+
+int InitSpaceToBatchParameter2(SpaceToBatchParameter *param) {
+  param->block_sizes_[0] = 2;
+  param->block_sizes_[1] = 2;
+
+  param->paddings_[0] = 2;
+  param->paddings_[1] = 0;
+  param->paddings_[2] = 2;
+  param->paddings_[3] = 2;
+
+  param->in_shape_[0] = 1;
+  param->in_shape_[1] = 4;
+  param->in_shape_[2] = 4;
+  param->in_shape_[3] = 1;
+
+  param->padded_in_shape_[0] = 1;
+  param->padded_in_shape_[1] = 6;
+  param->padded_in_shape_[2] = 8;
+  param->padded_in_shape_[3] = 1;
+}
+
+TEST_F(SpaceToBatchTestFp32, SpaceToBatchTest1) {
+  float input[16] = {1, 2, 5, 6, 10, 20, 3, 8, 18, 10, 3, 4, 11, 55, 15, 25};
+  const int out_size = 16;
+  float expect_out[16] = {1, 5, 18, 3, 2, 6, 10, 4, 10, 3, 11, 15, 20, 8, 55, 25};
+
+  float output[16];
+  int in_shape[4] = {1, 4, 4, 1};
+  int out_shape[4] = {4, 2, 2, 1};
+  int block_sizes[2] = {2, 2};
+  SpaceToBatchForNHWC((const float *)input, output, in_shape, 4, block_sizes);
+  for (int i = 0; i < out_size; ++i) {
+    std::cout << output[i] << " ";
+  }
+  std::cout << "\n";
+  CompareOutputData(output, expect_out, out_size, 0.000001);
+}
+
+TEST_F(SpaceToBatchTestFp32, SpaceToBatchTest2) {
+  SpaceToBatchParameter param;
+  InitSpaceToBatchParameter(&param);
+  float input[16] = {1, 2, 5, 6, 10, 20, 3, 8, 18, 10, 3, 4, 11, 55, 15, 25};
+  const int out_size = 48;
+  float expect_out[48] = {0, 0, 0, 0, 0, 1,  5, 0, 0, 18, 3,  0, 0, 0, 0, 0, 0, 2,  6, 0, 0, 10, 4,  0,
+                          0, 0, 0, 0, 0, 10, 3, 0, 0, 11, 15, 0, 0, 0, 0, 0, 0, 20, 8, 0, 0, 55, 25, 0};
+  float output[48];
+  int in_shape[4] = {1, 4, 4, 1};
+  int out_shape[4] = {4, 3, 4, 1};
+  int block_sizes[2] = {2, 2};
+
+  float padded_input[48]{}, tmp[48]{}, tmp_zero[48]{};
+  float *tmp_space[3] = {padded_input, tmp, tmp_zero};
+  auto ret = SpaceToBatch((const float *)input, output, param, tmp_space);
+  std::cout << "return " << ret << std::endl;
+  for (int i = 0; i < out_size; ++i) {
+    std::cout << output[i] << " ";
+  }
+  std::cout << "\n";
+  CompareOutputData(output, expect_out, out_size, 0.000001);
+}
+
+TEST_F(SpaceToBatchTestFp32, SpaceToBatchTest3) {
+  SpaceToBatchParameter param;
+  InitSpaceToBatchParameter2(&param);
+  param.op_parameter_.type_ = schema::PrimitiveType_SpaceToBatch;
+
+  std::vector<float> input = {1, 2, 5, 6, 10, 20, 3, 8, 18, 10, 3, 4, 11, 55, 15, 25};
+  std::vector<int> in_shape = {1, 4, 4, 1};
+  lite::tensor::Tensor input_tensor;
+  input_tensor.SetData(input.data());
+  input_tensor.set_shape(in_shape);
+  input_tensor.SetFormat(schema::Format_NHWC);
+  input_tensor.set_data_type(kNumberTypeFloat32);
+  std::vector<lite::tensor::Tensor *> inputs_tensor;
+  inputs_tensor.emplace_back(&input_tensor);
+
+  const int out_size = 48;
+  float expect_out[48] = {0, 0, 0, 0, 0, 1,  5, 0, 0, 18, 3,  0, 0, 0, 0, 0, 0, 2,  6, 0, 0, 10, 4,  0,
+                          0, 0, 0, 0, 0, 10, 3, 0, 0, 11, 15, 0, 0, 0, 0, 0, 0, 20, 8, 0, 0, 55, 25, 0};
+  std::vector<float> output(48);
+  std::vector<int> out_shape = {4, 3, 4, 1};
+  lite::tensor::Tensor output_tensor;
+  output_tensor.SetData(output.data());
+  output_tensor.set_shape(out_shape);
+  output_tensor.SetFormat(schema::Format_NHWC);
+  output_tensor.set_data_type(kNumberTypeFloat32);
+  std::vector<lite::tensor::Tensor *> outputs_tensor;
+  outputs_tensor.emplace_back(&output_tensor);
+
+  lite::Context ctx;
+  kernel::KernelKey desc = {kernel::KERNEL_ARCH::kCPU, kNumberTypeFloat32, schema::PrimitiveType_SpaceToBatch};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
+  kernel::LiteKernel *kernel =
+    creator(inputs_tensor, outputs_tensor, reinterpret_cast<OpParameter *>(&param), &ctx, desc);
+  ASSERT_NE(kernel, nullptr);
+  kernel->Run();
+
+  for (int i = 0; i < out_size; ++i) {
+    std::cout << output[i] << " ";
+  }
+  std::cout << "\n";
+  CompareOutputData(output.data(), expect_out, out_size, 0.000001);
+}
+
+}  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32/space_to_depth_fp32_tests.cc

@@ -0,0 +1,91 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include <vector>
+#include <iostream>
+#include <memory>
+#include "utils/log_adapter.h"
+#include "common/common_test.h"
+#include "mindspore/lite/src/runtime/kernel/arm/opclib/fp32/space_to_depth.h"
+#include "mindspore/lite/src/kernel_registry.h"
+#include "mindspore/lite/src/lite_kernel.h"
+
+namespace mindspore {
+
+class SpaceToDepthTestFp32 : public mindspore::Common {
+ public:
+  SpaceToDepthTestFp32() {}
+};
+
+TEST_F(SpaceToDepthTestFp32, SpaceToDepthTest1) {
+  float input[16] = {1, 2, 5, 6, 10, 20, 3, 8, 18, 10, 3, 4, 11, 55, 15, 25};
+  const int out_size = 16;
+  float expect_out[16] = {1, 2, 10, 20, 5, 6, 3, 8, 18, 10, 11, 55, 3, 4, 15, 25};
+
+  float output[16];
+  int in_shape[4] = {1, 4, 4, 1};
+  int out_shape[4] = {1, 2, 2, 4};
+  SpaceToDepthForNHWC((const float *)input, output, in_shape, out_shape, 4, 2);
+  for (int i = 0; i < out_size; ++i) {
+    std::cout << output[i] << " ";
+  }
+  std::cout << "\n";
+  CompareOutputData(output, expect_out, out_size, 0.000001);
+}
+
+TEST_F(SpaceToDepthTestFp32, SpaceToDepthTest2) {
+  std::vector<float> input = {1, 2, 5, 6, 10, 20, 3, 8, 18, 10, 3, 4, 11, 55, 15, 25};
+  std::vector<int> in_shape = {1, 4, 4, 1};
+  lite::tensor::Tensor input_tensor;
+  input_tensor.SetData(input.data());
+  input_tensor.set_shape(in_shape);
+  input_tensor.SetFormat(schema::Format_NHWC);
+  input_tensor.set_data_type(kNumberTypeFloat32);
+  std::vector<lite::tensor::Tensor *> inputs_tensor;
+  inputs_tensor.emplace_back(&input_tensor);
+
+  const int out_size = 16;
+  float expect_out[16] = {1, 2, 10, 20, 5, 6, 3, 8, 18, 10, 11, 55, 3, 4, 15, 25};
+  std::vector<float> output(16);
+  std::vector<int> out_shape = {1, 2, 2, 4};
+  lite::tensor::Tensor output_tensor;
+  output_tensor.SetData(output.data());
+  output_tensor.set_shape(out_shape);
+  output_tensor.SetFormat(schema::Format_NHWC);
+  output_tensor.set_data_type(kNumberTypeFloat32);
+  std::vector<lite::tensor::Tensor *> outputs_tensor;
+  outputs_tensor.emplace_back(&output_tensor);
+
+  SpaceToDepthParameter op_param;
+  op_param.op_parameter_.type_ = schema::PrimitiveType_SpaceToBatch;
+  op_param.block_size_ = 2;
+
+  lite::Context ctx;
+  kernel::KernelKey desc = {kernel::KERNEL_ARCH::kCPU, kNumberTypeFloat32, schema::PrimitiveType_SpaceToDepth};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
+  kernel::LiteKernel *kernel =
+    creator(inputs_tensor, outputs_tensor, reinterpret_cast<OpParameter *>(&op_param), &ctx, desc);
+  ASSERT_NE(kernel, nullptr);
+  kernel->Run();
+
+  for (int i = 0; i < out_size; ++i) {
+    std::cout << output[i] << " ";
+  }
+  std::cout << "\n";
+  CompareOutputData(output.data(), expect_out, out_size, 0.000001);
+}
+
+}  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/arm/int8/dequantize_int8_tests.cc

@@ -0,0 +1,83 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include <iostream>
+#include <memory>
+#include "utils/log_adapter.h"
+#include "common/common_test.h"
+#include "mindspore/lite/src/runtime/kernel/arm/int8/dequantize.h"
+#include "mindspore/lite/src/runtime/kernel/arm/opclib/int8/dequantize.h"
+#include "mindspore/lite/src/kernel_registry.h"
+#include "mindspore/lite/src/lite_kernel.h"
+
+namespace mindspore {
+
+class DequantizeTestFp32 : public mindspore::Common {
+ public:
+  DequantizeTestFp32() {}
+};
+
+TEST_F(DequantizeTestFp32, DequantizeTest1) {
+  const lite::tensor::QuantArg quant_arg{0.21176, 5};
+  // quant_arg.scale = 100.0;
+  // quant_arg.zeroPoint = 20;
+  DequantizeParameter param;
+  param.op_parameter_.type_ = schema::PrimitiveType_OnnxInt8Dequantize;
+
+  std::vector<int8_t> input = {10, 14, 29, 33, 52, 99, 19, 43, 90, 52, 19, 24, 57, 127, 76, 123};
+  // int8_t input0[] = {1, 2, 10};
+  // int32_t a = input0[0] + 2;
+  std::vector<int> in_shape = {1, 4, 4, 1};
+  lite::tensor::Tensor input_tensor;
+  input_tensor.SetData(input.data());
+  input_tensor.set_shape(in_shape);
+  input_tensor.set_data_type(kNumberTypeInt8);
+  input_tensor.SetFormat(schema::Format_NHWC);
+
+  input_tensor.AddQuantParam(quant_arg);
+  std::vector<lite::tensor::Tensor *> inputs_tensor;
+  inputs_tensor.emplace_back(&input_tensor);
+
+  const int out_size = 16;
+  float expect_out[16] = {3.1764,  4.02344,  7.19984,  8.04688, 12.07032, 22.02304, 5.08224,  10.16448,
+                          20.1172, 12.07032, 5.082240, 6.14104, 13.12912, 27.95232, 17.15256, 27.10528};
+  std::vector<float> output(16);
+  std::vector<int> out_shape = {1, 4, 4, 1};
+  lite::tensor::Tensor output_tensor;
+  output_tensor.SetData(output.data());
+  output_tensor.set_shape(out_shape);
+  output_tensor.set_data_type(kNumberTypeFloat32);
+  output_tensor.SetFormat(schema::Format_NHWC);
+  std::vector<lite::tensor::Tensor *> outputs_tensor;
+  outputs_tensor.emplace_back(&output_tensor);
+
+  lite::Context ctx;
+  ctx.threadNum = 3;
+  kernel::KernelKey desc = {kernel::KERNEL_ARCH::kCPU, kNumberTypeInt8, schema::PrimitiveType_OnnxInt8Dequantize};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
+  kernel::LiteKernel *kernel =
+    creator(inputs_tensor, outputs_tensor, reinterpret_cast<OpParameter *>(&param), &ctx, desc);
+  ASSERT_NE(kernel, nullptr);
+  kernel->Run();
+
+  for (int i = 0; i < out_size; ++i) {
+    std::cout << output[i] << " ";
+  }
+  std::cout << "\n";
+  CompareOutputData(output.data(), expect_out, out_size, 0.000001);
+}
+
+}  // namespace mindspore