conv_op.cc

// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include "lite/operators/conv_op.h"
#include <iostream>
#include <vector>
#include "lite/kernels/apu/bridges/graph.h"
#include "lite/kernels/apu/bridges/utility.h"
#include "lite/kernels/npu/bridges/registry.h"

namespace paddle {
namespace lite {
namespace subgraph {
namespace apu {

int ConvConverter(void* ctx, OpLite* op, KernelBase* kernel) {
  CHECK(ctx != nullptr);
  CHECK(op != nullptr);
  auto graph = static_cast<Graph*>(ctx);
  auto model = graph->model();
  auto op_info = op->op_info();
  auto op_type = op_info->Type();
  auto scope = op->scope();
  int neuron_errCode;
  VLOG(3) << "[APU] Converting [" << op_type << "]";

  // Get input and output vars and op attributes
  auto input_name = op_info->Input("Input").front();
  auto input = scope->FindMutableTensor(input_name);
  auto input_dims = input->dims();

  auto filter_name = op_info->Input("Filter").front();
  auto filter = scope->FindMutableTensor(filter_name);
  auto filter_dims = filter->dims();

  auto output_name = op_info->Output("Output").front();
  auto output = scope->FindMutableTensor(output_name);
  auto output_dims = output->dims();

  auto bs = input_dims[0];
  auto ic = input_dims[1];
  auto oc = filter_dims[0];
  CHECK_EQ(input_dims.size(), 4L);
  CHECK_EQ(output_dims.size(), 4L);
  CHECK_EQ(filter_dims.size(), 4L);
  CHECK_EQ(output_dims[0], bs);
  CHECK_EQ(output_dims[1], oc);
  auto strides = op_info->GetAttr<std::vector<int>>("strides");
  auto paddings = op_info->GetAttr<std::vector<int>>("paddings");
  auto groups = op_info->GetAttr<int>("groups");
  auto dilations = op_info->GetAttr<std::vector<int>>("dilations");
  bool with_act =
      op_info->HasAttr("with_act") && op_info->GetAttr<bool>("with_act");
  std::string act_type =
      with_act ? op_info->GetAttr<std::string>("act_type") : "";
  float leaky_relu_alpha = act_type == "leaky_relu"
                               ? op_info->GetAttr<float>("leaky_relu_alpha")
                               : 0.f;
  CHECK_EQ(strides.size(), 2L);
  CHECK_EQ(dilations.size(), 2L);
  bool is_depthwise_mode = ic == groups && oc == groups;
  VLOG(3) << "is_depthwise_mode" << is_depthwise_mode;

  if (paddings.size() == 2L) {
    for (size_t i = 0; i < strides.size(); ++i) {
      int copy_pad = *(paddings.begin() + 2 * i);
      paddings.insert(paddings.begin() + 2 * i + 1, copy_pad);
    }
  }

  CHECK_EQ(paddings.size(), 4L)
      << "[APU] Paddings size should be the same or twice as the input size."
      << paddings.size();

  std::string padding_algorithm("");
  if (op_info->HasAttr("padding_algorithm")) {
    padding_algorithm = op_info->GetAttr<std::string>("padding_algorithm");
  }
  operators::UpdatePaddingAndDilation(&paddings,
                                      &dilations,
                                      strides,
                                      padding_algorithm,
                                      input_dims,
                                      filter_dims);

  float input_scale;
  float output_scale;
  std::vector<float> weight_scale;
  if (op_info->HasAttr("enable_int8")) {
    if (op_info->GetAttr<bool>("enable_int8")) {
      if (op_info->HasAttr("input_scale"))
        input_scale = op_info->GetAttr<float>("input_scale");
      if (op_info->HasAttr("weight_scale"))
        weight_scale = op_info->GetAttr<std::vector<float>>("weight_scale");
      if (op_info->HasAttr("output_scale"))
        output_scale = op_info->GetAttr<float>("output_scale");
      VLOG(3) << "has output scale:" << output_scale;
    } else {
      return FAILED;
    }
  } else {
    return FAILED;
  }

  VLOG(3) << "strides.size(): " << strides.size() << " ,groups: " << groups
          << " ,dilations: " << dilations[0] << ":" << dilations[1];
  VLOG(3) << "with_act: " << with_act << " ,act_type:" << act_type;
  VLOG(3) << "input_dims: " << input_dims << " ,output_dims: " << output_dims
          << " ,weight_scale size: " << weight_scale.size();
  VLOG(3) << "filter_dims: " << filter_dims
          << " ,memory_size: " << filter->memory_size()
          << " ,data_size: " << filter->data_size();

  // Add input tensor type
  NeuronOperandType inType;
  inType.type = NEURON_TENSOR_QUANT8_ASYMM;
  inType.scale = input_scale;
  inType.zeroPoint = 128;
  inType.dimensionCount = input_dims.size();
  std::vector<uint32_t> dims_in = {(uint32_t)input_dims[0],
                                   (uint32_t)input_dims[2],
                                   (uint32_t)input_dims[3],
                                   (uint32_t)input_dims[1]};
  inType.dimensions = &dims_in[0];

  std::shared_ptr<Node> input_node = nullptr;
  if (graph->Has(input_name)) {
    VLOG(3) << "Graph has " << input_name;
    // input operand already exist
    input_node = graph->Get(input_name);
  } else {
    // add input operand
    if (graph->IsInput(input_name)) {
      // Insert transpose for NCHW -> NHWC
      insert_transpose_node(
          ctx,
          input_name,
          "transpose_" + input_name,
          {input_dims[0], input_dims[1], input_dims[2], input_dims[3]},
          dims_in,
          {0, 2, 3, 1},
          inType.scale,
          inType.zeroPoint);

      // change input_name
      input_name = "transpose_" + input_name;
      input_node = graph->Get(input_name);
      if (input_node == nullptr) return subgraph::FAILED;
    } else {
      NeuronModel_addOperand(model, &inType);  // input
      input_node = graph->Add(input_name, dims_in);
    }
  }
  VLOG(3) << "input node idx" << input_node->index()
          << ": input_scale: " << input_scale
          << ", inType: " << inType.dimensions[0] << ":" << inType.dimensions[1]
          << ":" << inType.dimensions[2] << ":" << inType.dimensions[3];

  // Add bias type
  NeuronOperandType biasType;

  // Add filter type
  // filter NCHW -> NHWC
  Tensor transpose_filter;
  std::vector<uint32_t> dims_filter;

  if (is_depthwise_mode) {
    transpose_filter.Resize({1,
                             (uint32_t)filter_dims[2],
                             (uint32_t)filter_dims[3],
                             (uint32_t)filter_dims[0]});
    dims_filter = {1,
                   (uint32_t)filter_dims[0],
                   (uint32_t)filter_dims[2],
                   (uint32_t)filter_dims[3]};
    transpose(filter->data<int8_t>(),
              transpose_filter.mutable_data<uint8_t>(),
              dims_filter,
              {0, 2, 3, 1});

    dims_filter = {(uint32_t)filter_dims[1],
                   (uint32_t)filter_dims[2],
                   (uint32_t)filter_dims[3],
                   (uint32_t)filter_dims[0]};
  } else {
    transpose_filter.Resize({(uint32_t)filter_dims[0],
                             (uint32_t)filter_dims[2],
                             (uint32_t)filter_dims[3],
                             (uint32_t)filter_dims[1]});
    dims_filter = {(uint32_t)filter_dims[0],
                   (uint32_t)filter_dims[1],
                   (uint32_t)filter_dims[2],
                   (uint32_t)filter_dims[3]};
    transpose(filter->data<int8_t>(),
              transpose_filter.mutable_data<uint8_t>(),
              dims_filter,
              {0, 2, 3, 1});

    dims_filter = {(uint32_t)filter_dims[0],
                   (uint32_t)filter_dims[2],
                   (uint32_t)filter_dims[3],
                   (uint32_t)filter_dims[1]};
  }

  NeuronOperandType filterType;
  NeuronOperandType channelFilterType;
  NeuronSymmPerChannelQuantParams symmPerChannelQuantParams;
  if (1 == weight_scale.size()) {
    // Per layer type
    filterType.type = NEURON_TENSOR_QUANT8_ASYMM;
    filterType.scale = weight_scale[0];
    filterType.zeroPoint = 128;
    filterType.dimensionCount = filter_dims.size();
    filterType.dimensions = &dims_filter[0];
    biasType.scale = inType.scale * filterType.scale;
  } else {
    // Per channel type
    channelFilterType.type = NEURON_TENSOR_QUANT8_SYMM_PER_CHANNEL;
    channelFilterType.scale = 0.0f;
    channelFilterType.zeroPoint = 0;
    channelFilterType.dimensionCount = filter_dims.size();
    channelFilterType.dimensions = &dims_filter[0];

    // Per channel setting
    if (is_depthwise_mode)
      symmPerChannelQuantParams.channelDim = 3;
    else
      symmPerChannelQuantParams.channelDim = 0;
    symmPerChannelQuantParams.scaleCount = weight_scale.size();
    symmPerChannelQuantParams.scales = weight_scale.data();
    biasType.scale = 0;
  }

  std::shared_ptr<Node> filter_node = nullptr;
  if (1 == weight_scale.size()) {
    NeuronModel_addOperand(model, &filterType);  // 1: filter
    filter_node = graph->Add(filter_name, dims_filter);
    VLOG(3) << "filter node idx: " << filter_node->index() << "w_scale[0]"
            << weight_scale[0] << ": filterType: " << filterType.dimensions[0]
            << ":" << filterType.dimensions[1] << ":"
            << filterType.dimensions[2] << ":" << filterType.dimensions[3];
    memcpy(filter->mutable_data<int8_t>(),
           transpose_filter.mutable_data<uint8_t>(),
           filter->memory_size());
    neuron_errCode = NeuronModel_setOperandValue(
        model, filter_node->index(), filter->raw_data(), filter->memory_size());
    if (NEURON_NO_ERROR != neuron_errCode) {
      LOG(WARNING) << "Set filter operand value fail:" << neuron_errCode;
      return subgraph::FAILED;
    }
  } else {
    NeuronModel_addOperand(model, &channelFilterType);  // 1: filter
    filter_node = graph->Add(filter_name, dims_filter);
    VLOG(3) << "chennel filter node idx: " << filter_node->index()
            << " ,scale_count:" << weight_scale.size()
            << " weight_scale[0]:" << weight_scale.data()[0]
            << " ,channelFilterType: " << channelFilterType.dimensions[0] << ":"
            << channelFilterType.dimensions[1] << ":"
            << channelFilterType.dimensions[2] << ":"
            << channelFilterType.dimensions[3];
    memcpy(filter->mutable_data<int8_t>(),
           transpose_filter.mutable_data<uint8_t>(),
           filter->memory_size());
    neuron_errCode = NeuronModel_setOperandValue(
        model, filter_node->index(), filter->raw_data(), filter->memory_size());
    if (NEURON_NO_ERROR != neuron_errCode) {
      LOG(WARNING) << "Set filter operand value fail:" << neuron_errCode;
      return subgraph::FAILED;
    }
    neuron_errCode = NeuronModel_setOperandSymmPerChannelQuantParams(
        model, filter_node->index(), &symmPerChannelQuantParams);
    if (NEURON_NO_ERROR != neuron_errCode) {
      LOG(WARNING) << "Set per channel filter params fail:" << neuron_errCode;
      return subgraph::FAILED;
    }
  }

  // Add biasType node value
  // A 1-D tensor, of shape [depth_out], specifying the bias.
  // For filter tensor of NEURON_TENSOR_QUANT8_SYMM_PER_CHANNEL, the bias
  // should be of ANEURALNETWORKS_TENSOR_INT32, with zeroPoint of 0
  // and bias_scale of 0. The actual scale of each value 'i' is equal
  // to bias_scale[i] = input_scale * filter_scale[i].
  biasType.type = NEURON_TENSOR_INT32;
  biasType.zeroPoint = 0;
  std::vector<uint32_t> dims_bias;
  std::shared_ptr<Node> bias_node = nullptr;
  if (HasInputArg(op_info, scope, "Bias")) {
    auto bias_name = op_info->Input("Bias").front();
    auto bias_type = kernel->GetInputDeclType("Bias");
    auto bias = scope->FindMutableTensor(bias_name);
    auto bias_dims = bias->dims();

    biasType.dimensionCount = bias_dims.size();
    for (int i = 0; i < bias_dims.size(); i++)
      dims_bias.push_back(bias_dims[i]);
    biasType.dimensions = &dims_bias[0];
    NeuronModel_addOperand(model, &biasType);  // 2: bias
    bias_node = graph->Add(bias_name, dims_bias);
    VLOG(3) << "node idx" << bias_node->index() << ": Bias name: " << bias_name
            << " ,bias scale: " << biasType.scale
            << " ,dimensions: " << bias_dims;
  } else {
    biasType.dimensionCount = 1;
    dims_bias = {(uint32_t)output_dims[1]};
    biasType.dimensions = &dims_bias[0];
    NeuronModel_addOperand(model, &biasType);  // 2: bias
    bias_node = graph->Add(filter_name + "_default_bias", dims_bias);
    VLOG(3) << "node idx" << bias_node->index() << ": Bias name: default_bias "
            << " ,bias scale: " << biasType.scale
            << " ,dimensions: " << dims_bias.size();
  }

  NeuronOperandType int32Type;
  int32Type.type = NEURON_INT32;
  int32Type.dimensionCount = 0;
  std::vector<uint32_t> dims_int32 = {1};

  std::shared_ptr<Node> paddingL_node = nullptr;
  NeuronModel_addOperand(model, &int32Type);  // 3: padding left
  paddingL_node = graph->Add(filter_name + "_padding_left", dims_int32);

  std::shared_ptr<Node> paddingR_node = nullptr;
  NeuronModel_addOperand(model, &int32Type);  // 4: padding right
  paddingR_node = graph->Add(filter_name + "_padding_right", dims_int32);

  std::shared_ptr<Node> paddingT_node = nullptr;
  NeuronModel_addOperand(model, &int32Type);  // 5: padding top
  paddingT_node = graph->Add(filter_name + "_padding_top", dims_int32);

  std::shared_ptr<Node> paddingB_node = nullptr;
  NeuronModel_addOperand(model, &int32Type);  // 6: padding bottom
  paddingB_node = graph->Add(filter_name + "_padding_bottom", dims_int32);

  std::shared_ptr<Node> strideW_node = nullptr;
  NeuronModel_addOperand(model, &int32Type);  // 7: stride width
  strideW_node = graph->Add(filter_name + "_stride_width", dims_int32);

  std::shared_ptr<Node> strideH_node = nullptr;
  NeuronModel_addOperand(model, &int32Type);  // 8: stride height
  strideH_node = graph->Add(filter_name + "_stride_height", dims_int32);

  std::shared_ptr<Node> dm_node = nullptr;
  if (is_depthwise_mode) {
    NeuronModel_addOperand(model, &int32Type);  // 9: depthwise multiplier
    dm_node = graph->Add(filter_name + "_dm", dims_int32);
  }

  std::shared_ptr<Node> fuse_node = nullptr;
  NeuronModel_addOperand(model, &int32Type);  // 9/10: fuse
  fuse_node = graph->Add(filter_name + "_fuse", dims_int32);

  // Add output tensor type
  NeuronOperandType outType;
  outType.type = NEURON_TENSOR_QUANT8_ASYMM;
  if (graph->IsOutput(output_name))
    outType.scale = output_scale / 127;
  else
    outType.scale = output_scale;
  outType.zeroPoint = 128;
  outType.dimensionCount = output_dims.size();
  std::vector<uint32_t> dims_out = {(uint32_t)output_dims[0],
                                    (uint32_t)output_dims[2],
                                    (uint32_t)output_dims[3],
                                    (uint32_t)output_dims[1]};
  outType.dimensions = &dims_out[0];
  std::shared_ptr<Node> output_node = nullptr;
  if (graph->Has(output_name)) {
    output_node = graph->Get(output_name);
  } else {
    // add output operand
    if (graph->IsOutput(output_name)) {
      NeuronModel_addOperand(model, &outType);  // output
      output_node = graph->Add("transpose_" + output_name, dims_out);
    } else {
      NeuronModel_addOperand(model, &outType);  // output
      output_node = graph->Add(output_name, dims_out);
    }
  }
  VLOG(3) << "output node idx: " << output_node->index()
          << ": output_scale: " << outType.scale
          << ", outType: " << outType.dimensions[0] << ":"
          << outType.dimensions[1] << ":" << outType.dimensions[2] << ":"
          << outType.dimensions[3];

  // Add bias value
  if (HasInputArg(op_info, scope, "Bias")) {
    auto bias_name = op_info->Input("Bias").front();
    auto bias = scope->FindMutableTensor(bias_name);
    int32_t* int32_bias_data =
        reinterpret_cast<int32_t*>(bias->mutable_data<float>());
    float2int32(
        bias->data<float>(), input_scale, weight_scale, int32_bias_data);

    VLOG(3) << "int32_bias_data: " << int32_bias_data[0] << " : "
            << int32_bias_data[1] << " : " << int32_bias_data[2] << " : "
            << int32_bias_data[3];
    neuron_errCode = NeuronModel_setOperandValue(
        model, bias_node->index(), bias->raw_data(), bias->memory_size());
  } else {
    auto int32_bias = std::make_shared<Tensor>();
    int32_bias->Resize({1, output_dims[1]});
    int32_bias->mutable_data<int32_t>();
    VLOG(3) << "bais_default: " << int32_bias->memory_size();
    memset(int32_bias->mutable_data<int32_t>(), 0, int32_bias->memory_size());
    neuron_errCode = NeuronModel_setOperandValue(model,
                                                 bias_node->index(),
                                                 int32_bias->raw_data(),
                                                 int32_bias->memory_size());
    bias_node->set_data(int32_bias);
  }
  if (NEURON_NO_ERROR != neuron_errCode) {
    LOG(WARNING) << "Set bias operand value fail:" << neuron_errCode;
    return subgraph::FAILED;
  }

  VLOG(3) << "paddings: " << paddings[0] << ":" << paddings[1] << ":"
          << paddings[2] << ":" << paddings[3];
  // Add padding value
  int32_t padding_val[1];
  padding_val[0] = paddings[2];
  NeuronModel_setOperandValue(
      model, paddingL_node->index(), padding_val, sizeof(int32_t) * 1);
  padding_val[0] = paddings[3];
  NeuronModel_setOperandValue(
      model, paddingR_node->index(), padding_val, sizeof(int32_t) * 1);
  padding_val[0] = paddings[0];
  NeuronModel_setOperandValue(
      model, paddingT_node->index(), padding_val, sizeof(int32_t) * 1);
  padding_val[0] = paddings[1];
  NeuronModel_setOperandValue(
      model, paddingB_node->index(), padding_val, sizeof(int32_t) * 1);

  VLOG(3) << " stride width:" << strides[1] << " height:" << strides[0];

  // Add Stride
  int32_t stride_val[1];
  stride_val[0] = strides[1];  // width
  NeuronModel_setOperandValue(
      model, strideW_node->index(), stride_val, sizeof(int32_t) * 1);
  stride_val[0] = strides[0];  // height
  NeuronModel_setOperandValue(
      model, strideH_node->index(), stride_val, sizeof(int32_t) * 1);

  // Add fuse
  int32_t fuse_val[1] = {0};
  if (act_type == "relu") {
    fuse_val[0] = 1;
  } else if (act_type == "relu1") {
    fuse_val[0] = 2;
  } else if (act_type == "relu6") {
    fuse_val[0] = 3;
  } else if (!act_type.empty()) {
    fuse_val[0] = 0;
    LOG(WARNING) << "Support act_type: " << act_type;
    return FAILED;
  }

  if (is_depthwise_mode) {
    int32_t dm = oc / ic;
    NeuronModel_setOperandValue(
        model, dm_node->index(), &dm, sizeof(int32_t) * 1);
    VLOG(3) << "depthwise multiplier:" << dm;

    // Depthwise conv
    NeuronModel_setOperandValue(
        model, fuse_node->index(), fuse_val, sizeof(int32_t) * 1);
    std::vector<uint32_t> addInIndex = {
        input_node->index(),     // 0: input
        filter_node->index(),    // 1: filter
        bias_node->index(),      // 2: bias
        paddingL_node->index(),  // 3: padding left
        paddingR_node->index(),  // 4: padding right
        paddingT_node->index(),  // 5: padding top
        paddingB_node->index(),  // 6: padding bottom
        strideW_node->index(),   // 7: stride width
        strideH_node->index(),   // 8: stride height
        dm_node->index(),        // 9: depthwise multiplier
        fuse_node->index()};     // 10 : fuse

    std::vector<uint32_t> addOutIndex = {output_node->index()};
    neuron_errCode = NeuronModel_addOperation(model,
                                              NEURON_DEPTHWISE_CONV_2D,
                                              addInIndex.size(),
                                              &addInIndex[0],
                                              addOutIndex.size(),
                                              &addOutIndex[0]);
  } else {
    NeuronModel_setOperandValue(
        model, fuse_node->index(), fuse_val, sizeof(int32_t) * 1);
    std::vector<uint32_t> addInIndex = {
        input_node->index(),     // 0: input
        filter_node->index(),    // 1: filter
        bias_node->index(),      // 2: bias
        paddingL_node->index(),  // 3: padding left
        paddingR_node->index(),  // 4: padding right
        paddingT_node->index(),  // 5: padding top
        paddingB_node->index(),  // 6: padding bottom
        strideW_node->index(),   // 7: stride width
        strideH_node->index(),   // 8: stride height
        fuse_node->index()};     // 9: fuse

    std::vector<uint32_t> addOutIndex = {output_node->index()};
    neuron_errCode = NeuronModel_addOperation(model,
                                              NEURON_CONV_2D,
                                              addInIndex.size(),
                                              &addInIndex[0],
                                              addOutIndex.size(),
                                              &addOutIndex[0]);
  }

  if (NEURON_NO_ERROR != neuron_errCode) {
    LOG(WARNING) << "Add op fail:" << op_type;
    return FAILED;
  }

  if (graph->IsOutput(output_name)) {
    // Insert transpose for NHWC -> NCHW
    insert_transpose_node(
        ctx,
        "transpose_" + output_name,
        output_name,
        dims_out,
        {output_dims[0], output_dims[1], output_dims[2], output_dims[3]},
        {0, 3, 1, 2},
        outType.scale,
        outType.zeroPoint);
    output_node = graph->Get(output_name);
    if (output_node == nullptr) return subgraph::FAILED;
  }

  return REBUILD_WHEN_SHAPE_CHANGED;
}

}  // namespace apu
}  // namespace subgraph
}  // namespace lite
}  // namespace paddle

REGISTER_SUBGRAPH_BRIDGE(conv2d,
                         kAPU,
                         paddle::lite::subgraph::apu::ConvConverter);
REGISTER_SUBGRAPH_BRIDGE(depthwise_conv2d,
                         kAPU,
                         paddle::lite::subgraph::apu::ConvConverter);