tensorflow/compiler/xla/stream_executor/dnn.cc

/* Copyright 2015 The TensorFlow 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 "tensorflow/compiler/xla/stream_executor/dnn.h"

#include "absl/hash/hash.h"
#include "absl/strings/str_cat.h"
#include "absl/strings/str_format.h"
#include "absl/strings/str_join.h"
#include "tensorflow/core/lib/strings/proto_serialization.h"

namespace stream_executor {
namespace dnn {

namespace {

bool ProtoMapIsSubset(const google::protobuf::Map<int64_t, int64_t>& x,
                      const google::protobuf::Map<int64_t, int64_t>& y) {
  for (const auto& ypair : y) {
    const auto it = x.find(ypair.first);
    if (it == x.end() || it->second != ypair.second) return false;
  }
  return true;
}

bool ProtoMapsEqual(const google::protobuf::Map<int64_t, int64_t>& x,
                    const google::protobuf::Map<int64_t, int64_t>& y) {
  return ProtoMapIsSubset(x, y) && ProtoMapIsSubset(y, x);
}

}  // namespace

constexpr DataType ToDataType<float>::value;
constexpr DataType ToDataType<double>::value;
constexpr DataType ToDataType<Eigen::half>::value;
constexpr DataType ToDataType<Eigen::bfloat16>::value;
constexpr DataType ToDataType<int8>::value;
constexpr DataType ToDataType<int32>::value;
constexpr DataType ToDataType<std::complex<float>>::value;
constexpr DataType ToDataType<std::complex<double>>::value;

AlgorithmDesc::AlgorithmDesc(
    int64_t engine_id,
    const std::vector<std::pair<int64_t, int64_t>>& tuning_knobs,
    std::optional<uint64_t> workspace_size) {
  proto_.set_is_cudnn_frontend(true);
  proto_.set_algo_id(engine_id);
  if (workspace_size) {
    proto_.mutable_workspace_size()->set_value(*workspace_size);
  }
  for (const auto& pair : tuning_knobs) {
    (*proto_.mutable_tuning_knobs())[pair.first] = pair.second;
  }
}

uint64_t AlgorithmDesc::hash() const {
  return tensorflow::DeterministicProtoHash64(proto_);
}

bool AlgorithmDesc::operator==(const AlgorithmDesc& other) const {
  if (is_cudnn_frontend()) {
    return other.is_cudnn_frontend() && algo_id() == other.algo_id() &&
           ProtoMapsEqual(proto_.tuning_knobs(), other.proto_.tuning_knobs());
  }
  return !other.is_cudnn_frontend() && algo_id() == other.algo_id() &&
         tensor_ops_enabled() == other.tensor_ops_enabled();
}

std::string AlgorithmDesc::ToString() const {
  if (is_cudnn_frontend()) {
    // Format similarly to cudnn_frontend::ExecutionPlan::getTag(), e.g.
    // "eng2{k1=2,k3=4}".
    return absl::StrFormat(
        "eng%d{%s}", proto_.algo_id(),
        absl::StrJoin(
            proto_.tuning_knobs(), ",",
            [](std::string* out,
               const google::protobuf::Map<int64_t, int64_t>::value_type& pair) {
              absl::StrAppendFormat(out, "k%d=%d", pair.first, pair.second);
            }));
  }
  if (tensor_ops_enabled()) {
    return absl::StrCat(algo_id(), "#TC");
  } else {
    return absl::StrCat(algo_id());
  }
}

std::vector<std::pair<int64_t, int64_t>> AlgorithmDesc::TuningKnobs() const {
  std::vector<std::pair<int64_t, int64_t>> result;
  result.reserve(proto_.tuning_knobs().size());
  for (const auto& pair : proto_.tuning_knobs()) {
    result.emplace_back(pair.first, pair.second);
  }
  return result;
}

bool DnnSupport::GetConvolveAlgorithms(
    CudaComputeCapability cuda_compute_capability,
    std::vector<AlgorithmDesc>* out_algorithms) {
  return false;
}

port::Status DnnSupport::GetConvolveRunners(
    bool /* use_cudnn_frontend */, dnn::ConvolutionKind /*kind*/,
    dnn::DataType /*input_type*/, dnn::DataType /*output_type*/,
    Stream* /*stream*/, const dnn::BatchDescriptor& /*input_descriptor*/,
    DeviceMemoryBase /*input_data*/,
    const dnn::FilterDescriptor& /*filter_descriptor*/,
    DeviceMemoryBase /*filter_data*/,
    const dnn::BatchDescriptor& /*output_descriptor*/,
    DeviceMemoryBase /*output_data*/,
    const dnn::ConvolutionDescriptor& /*convolution_descriptor*/,
    bool /*use_fallback*/, ScratchAllocator* /*scratch_allocator*/,
    std::vector<std::unique_ptr<const dnn::ConvRunner>>* /*exec_plans*/) {
  return port::UnimplementedError("GetConvolveRunners not implemented.");
}

port::StatusOr<std::unique_ptr<const dnn::ConvRunner>>
DnnSupport::ConvolveRunnerFromDesc(
    Stream* stream, const dnn::AlgorithmDesc& algorithm_desc,
    dnn::ConvolutionKind kind, dnn::DataType element_type,
    dnn::DataType output_type, const dnn::BatchDescriptor& input_descriptor,
    const dnn::FilterDescriptor& filter_descriptor,
    const dnn::BatchDescriptor& output_descriptor,
    const dnn::ConvolutionDescriptor& convolution_descriptor) {
  return port::UnimplementedError("ConvolveRunnerFromDesc not implemented.");
}

port::Status DnnSupport::GetFusedConvolveRunners(
    bool use_cudnn_frontend, dnn::ConvolutionKind kind,
    dnn::DataType element_type, dnn::DataType bias_type,
    dnn::DataType output_type, double conv_input_scale, double side_input_scale,
    double leakyrelu_alpha, Stream* stream,
    const dnn::BatchDescriptor& input_descriptor,
    const dnn::FilterDescriptor& filter_descriptor,
    const dnn::BatchDescriptor& bias_descriptor,
    const dnn::BatchDescriptor& output_descriptor,
    const dnn::ConvolutionDescriptor& convolution_descriptor, bool use_fallback,
    dnn::ActivationMode activation_mode,
    std::vector<std::unique_ptr<const dnn::FusedConvRunner>>* out_exec_plans) {
  return port::UnimplementedError("GetFusedConvolveRunners not implemented.");
}

port::StatusOr<std::unique_ptr<const dnn::FusedConvRunner>>
DnnSupport::FusedConvolveRunnerFromDesc(
    Stream* stream, const dnn::AlgorithmDesc& algorithm_desc,
    dnn::ConvolutionKind kind, dnn::DataType element_type,
    dnn::DataType bias_type, dnn::DataType output_type, double conv_scale,
    double side_input_scale, double leakyrelu_alpha,
    const dnn::BatchDescriptor& input_descriptor,
    const dnn::FilterDescriptor& filter_descriptor,
    const dnn::BatchDescriptor& bias_descriptor,
    const dnn::BatchDescriptor& output_descriptor,
    const dnn::ConvolutionDescriptor& convolution_descriptor,
    dnn::ActivationMode activation_mode) {
  return port::UnimplementedError(
      "FusedConvolveRunnerFromDesc not implemented.");
}

bool DnnSupport::GetMIOpenConvolveAlgorithms(
    dnn::ConvolutionKind /*kind*/, dnn::DataType /*element_type*/,
    Stream* /*stream*/, const dnn::BatchDescriptor& /*input_descriptor*/,
    DeviceMemoryBase input_data,
    const dnn::FilterDescriptor& /*filter_descriptor*/,
    DeviceMemoryBase filter_data,
    const dnn::BatchDescriptor& /*output_descriptor*/,
    DeviceMemoryBase output_data,
    const dnn::ConvolutionDescriptor& /*convolution_descriptor*/,
    ScratchAllocator* scratch_allocator,
    std::vector<ProfileResult>* /*out_algorithms*/) {
  return false;
}

bool DnnSupport::GetRnnAlgorithms(std::vector<AlgorithmDesc>* out_algorithms) {
  return false;
}

bool DnnSupport::GetConvolveBackwardDataAlgorithms(
    CudaComputeCapability cuda_compute_capability,
    std::vector<AlgorithmDesc>* out_algorithms) {
  return false;
}

bool DnnSupport::GetConvolveBackwardFilterAlgorithms(
    CudaComputeCapability cuda_compute_capability,
    std::vector<AlgorithmDesc>* out_algorithms) {
  return false;
}

std::string QuantizedActivationModeString(QuantizedActivationMode mode) {
  switch (mode) {
    case dnn::QuantizedActivationMode::k8Bit:
      return "uint8";
    case dnn::QuantizedActivationMode::k16Bit:
      return "uint16";
    case dnn::QuantizedActivationMode::k32Bit:
      return "int32";
    default:
      return absl::StrCat("unknown: ", static_cast<int32_t>(mode));
  }
}

std::string ActivationModeString(ActivationMode mode) {
  switch (mode) {
    case ActivationMode::kNone:
      return "none";
    case ActivationMode::kSigmoid:
      return "sigmoid";
    case ActivationMode::kRelu:
      return "relu";
    case ActivationMode::kRelu6:
      return "relu6";
    case ActivationMode::kReluX:
      return "reluX";
    case ActivationMode::kTanh:
      return "tanh";
    case ActivationMode::kBandPass:
      return "bandpass";
    default:
      return absl::StrCat("unknown: ", static_cast<int32_t>(mode));
  }
}

std::string ElementwiseOperationString(ElementwiseOperation op) {
  switch (op) {
    case ElementwiseOperation::kAdd:
      return "add";
    case ElementwiseOperation::kMultiply:
      return "multiply";
    default:
      return absl::StrCat("unknown: ", static_cast<int32_t>(op));
  }
}

std::string DataLayoutString(DataLayout layout) {
  switch (layout) {
    case DataLayout::kYXDepthBatch:
      return "YXDepthBatch";
    case DataLayout::kYXBatchDepth:
      return "YXBatchDepth";
    case DataLayout::kBatchYXDepth:
      return "BatchYXDepth";
    case DataLayout::kBatchDepthYX:
      return "BatchDepthYX";
    case DataLayout::kBatchDepthYX4:
      return "BatchDepthYX4";
    case DataLayout::kBatchDepthYX32:
      return "BatchDepthYX32";
    default:
      return absl::StrCat("unknown: ", static_cast<int32_t>(layout));
  }
}

std::string FilterLayoutString(FilterLayout layout) {
  switch (layout) {
    case FilterLayout::kOutputInputYX:
      return "OutputInputYX";
    case FilterLayout::kOutputYXInput:
      return "OutputYXInput";
    case FilterLayout::kOutputInputYX4:
      return "OutputInputYX4";
    case FilterLayout::kOutputInputYX32:
      return "OutputInputYX32";
    case FilterLayout::kInputYXOutput:
      return "InputYXOutput";
    case FilterLayout::kYXInputOutput:
      return "YXInputOutput";
    default:
      return absl::StrCat("unknown: ", static_cast<int32_t>(layout));
  }
}

std::string PadAlignmentString(PadAlignment alignment) {
  switch (alignment) {
    case PadAlignment::kDefault:
      return "default";
    case PadAlignment::kCudnnPadding:
      return "cuDNN padding";
    case PadAlignment::kTensorFlowPadding:
      return "TensorFlow padding";
    default:
      return absl::StrCat("unknown: ", static_cast<int32_t>(alignment));
  }
}

std::ostream& operator<<(std::ostream& str, dnn::PadAlignment alignment) {
  return str << PadAlignmentString(alignment);
}

std::string ShortPoolingModeString(PoolingMode mode) {
  switch (mode) {
    case PoolingMode::kMaximum:
      return "Max";
    case PoolingMode::kAverage:
      return "Avg";
    default:
      return absl::StrCat("unknown: ", static_cast<int32_t>(mode));
  }
}

struct ConvDimIndices {
  union {
    struct {
      int depth_idx;
      int batch_idx;
      int spatial_idx;
    } data;
    struct {
      int output_idx;
      int input_idx;
      int spatial_idx;
    } filter;
  };
};

ConvDimIndices GetDimIndices(const DataLayout& layout, const int data_dims) {
  ConvDimIndices dim_indices;
  switch (layout) {
    case DataLayout::kYXBatchDepth:
      dim_indices.data.depth_idx = data_dims - 1;
      dim_indices.data.batch_idx = data_dims - 2;
      dim_indices.data.spatial_idx = 0;
      break;

    case DataLayout::kYXDepthBatch:
      dim_indices.data.depth_idx = data_dims - 2;
      dim_indices.data.batch_idx = data_dims - 1;
      dim_indices.data.spatial_idx = 0;
      break;

    case DataLayout::kBatchYXDepth:
      dim_indices.data.depth_idx = data_dims - 1;
      dim_indices.data.batch_idx = 0;
      dim_indices.data.spatial_idx = 1;
      break;

    case DataLayout::kBatchDepthYX:
    case DataLayout::kBatchDepthYX4:
    case DataLayout::kBatchDepthYX32:
      dim_indices.data.depth_idx = 1;
      dim_indices.data.batch_idx = 0;
      dim_indices.data.spatial_idx = 2;
      break;

    default:
      LOG(FATAL) << "Unknown layout " << layout;
  }

  return dim_indices;
}

ConvDimIndices GetDimIndices(const FilterLayout& layout, const int data_dims) {
  ConvDimIndices dim_indices;
  switch (layout) {
    case FilterLayout::kOutputInputYX:
    case FilterLayout::kOutputInputYX4:
    case FilterLayout::kOutputInputYX32:
      dim_indices.filter.input_idx = 1;
      dim_indices.filter.output_idx = 0;
      dim_indices.filter.spatial_idx = 2;
      break;

    case FilterLayout::kOutputYXInput:
      dim_indices.filter.input_idx = data_dims - 1;
      dim_indices.filter.output_idx = 0;
      dim_indices.filter.spatial_idx = 1;
      break;

    case FilterLayout::kInputYXOutput:
      dim_indices.filter.input_idx = 0;
      dim_indices.filter.output_idx = data_dims - 1;
      dim_indices.filter.spatial_idx = 1;
      break;

    case FilterLayout::kYXInputOutput:
      dim_indices.filter.input_idx = data_dims - 2;
      dim_indices.filter.output_idx = data_dims - 1;
      dim_indices.filter.spatial_idx = 0;
      break;

    default:
      LOG(FATAL) << "Unknown layout " << layout;
  }

  return dim_indices;
}

std::vector<int64_t> ReorderDims(const std::vector<int64_t>& input,
                                 const DataLayout& from, const DataLayout& to) {
  if (from == to) return input;

  ConvDimIndices from_indices = GetDimIndices(from, input.size());
  ConvDimIndices to_indices = GetDimIndices(to, input.size());

  std::vector<int64_t> reordered(input.size());
  reordered[to_indices.data.batch_idx] = input[from_indices.data.batch_idx];
  reordered[to_indices.data.depth_idx] = input[from_indices.data.depth_idx];

  int spatial_idx_from = from_indices.data.spatial_idx;
  int spatial_idx_to = to_indices.data.spatial_idx;
  for (size_t i = 0; i < input.size() - 2;
       i++, spatial_idx_from++, spatial_idx_to++) {
    reordered[spatial_idx_to] = input[spatial_idx_from];
  }

  return reordered;
}

std::vector<int64_t> ReorderDims(const std::vector<int64_t>& input,
                                 const FilterLayout& from,
                                 const FilterLayout& to) {
  if (from == to) return input;

  ConvDimIndices from_indices = GetDimIndices(from, input.size());
  ConvDimIndices to_indices = GetDimIndices(to, input.size());

  std::vector<int64_t> reordered(input.size());
  reordered[to_indices.filter.output_idx] =
      input[from_indices.filter.output_idx];
  reordered[to_indices.filter.input_idx] = input[from_indices.filter.input_idx];

  int spatial_idx_from = from_indices.filter.spatial_idx;
  int spatial_idx_to = to_indices.filter.spatial_idx;
  for (size_t i = 0; i < input.size() - 2;
       i++, spatial_idx_from++, spatial_idx_to++) {
    reordered[spatial_idx_to] = input[spatial_idx_from];
  }

  return reordered;
}

// -- AlgorithmConfig

std::string AlgorithmConfig::ToString() const {
  std::string algo = "none";
  if (algorithm().has_value()) {
    algo = algorithm()->ToString();
  }
  std::string algo_no_scratch = "none";
  if (algorithm_no_scratch().has_value()) {
    algo_no_scratch = algorithm_no_scratch()->ToString();
  }
  return absl::StrCat(algo, ", ", algo_no_scratch);
}

// -- BatchDescriptor

BatchDescriptor::BatchDescriptor(int ndims)
    : value_max_(0.0),
      value_min_(0.0),
      quantized_activation_mode_(QuantizedActivationMode::k8Bit) {
  tensor_.mutable_dimensions()->Resize(ndims + 2, 0);
  set_layout(DataLayout::kYXDepthBatch);
}

BatchDescriptor::BatchDescriptor() : BatchDescriptor(/*ndims=*/2) {}

std::vector<int64_t> BatchDescriptor::full_dims(
    const DataLayout& layout) const {
  std::vector<int64_t> bdyx_dims(ndims() + 2);
  bdyx_dims[0] = count();
  bdyx_dims[1] = feature_map_count();
  std::copy(spatial_size().begin(), spatial_size().end(),
            bdyx_dims.begin() + 2);
  return ReorderDims(bdyx_dims, DataLayout::kBatchDepthYX, layout);
}

std::vector<int64_t> BatchDescriptor::full_strides(
    const DataLayout& layout) const {
  std::vector<int64_t> phys_dims = full_dims(this->layout());
  std::vector<int64_t> phys_strides(phys_dims.size());
  phys_strides[ndims() + 1] = 1;
  for (int i = ndims(); i >= 0; i--) {
    phys_strides[i] = phys_strides[i + 1] * phys_dims[i + 1];
  }
  return ReorderDims(phys_strides, this->layout(), layout);
}

std::vector<int64_t> BatchDescriptor::vectorized_dims(const DataLayout& layout,
                                                      int vector_size,
                                                      int vector_dim) const {
  std::vector<int64_t> bdyx_dims = full_dims(dnn::DataLayout::kBatchDepthYX);
  if (vector_dim != -1) {
    bdyx_dims[vector_dim] /= vector_size;
  }
  return dnn::ReorderDims(bdyx_dims, dnn::DataLayout::kBatchDepthYX, layout);
}

std::vector<int64_t> BatchDescriptor::vectorized_strides(
    const DataLayout& layout, int vector_size, int vector_dim) const {
  std::vector<int64_t> phys_dims =
      vectorized_dims(this->layout(), vector_size, vector_dim);
  std::vector<int64_t> phys_strides(phys_dims.size());
  phys_strides[phys_dims.size() - 1] = 1;
  for (int i = phys_dims.size() - 2; i >= 0; i--) {
    phys_strides[i] = phys_strides[i + 1] * phys_dims[i + 1];
  }
  return ReorderDims(phys_strides, this->layout(), layout);
}

void BatchDescriptor::CloneFrom(const BatchDescriptor& other) {
  tensor_ = other.tensor_;
  value_max_ = other.value_max_;
  value_min_ = other.value_min_;
  quantized_activation_mode_ = other.quantized_activation_mode_;
}

std::string BatchDescriptor::ToString() const {
  std::string spatial;
  for (int i = 0; i < ndims(); i++) {
    absl::StrAppendFormat(&spatial, "%d ", spatial_size()[i]);
  }
  return absl::StrFormat(
      "{count: %d feature_map_count: %d spatial: %s "
      "value_min: %f value_max: %f layout: %s}",
      count(), feature_map_count(), spatial, value_min_, value_max_,
      DataLayoutString(layout()));
}

std::string BatchDescriptor::ToShortString() const {
  // All the constituent strings are less than 15 characters, so the
  // small string optimization ensures that there will be at most one
  // heap memory allocation.
  std::string depth = absl::StrCat("d", feature_map_count());
  std::string batch = absl::StrCat("b", count());

  std::string spatial = "s";
  for (int i = 0; i < ndims(); i++) {
    absl::StrAppendFormat(&spatial, "%d ", spatial_size()[i]);
  }

  std::string suffix;
  if (value_min() != value_max()) {
    absl::StrAppend(&suffix, "[", value_min(), ";", value_max(), "]");
  }
  if (quantized_activation_mode() == QuantizedActivationMode::k16Bit) {
    suffix += "_16bit";
  }

  switch (layout()) {
    case DataLayout::kYXDepthBatch:
      return absl::StrCat(spatial, depth, batch, suffix);
    case DataLayout::kYXBatchDepth:
      return absl::StrCat(spatial, batch, depth, suffix);
    case DataLayout::kBatchYXDepth:
      return absl::StrCat(batch, spatial, depth, suffix);
    case DataLayout::kBatchDepthYX:
      return absl::StrCat(batch, depth, spatial, suffix);
    case DataLayout::kBatchDepthYX4:
    case DataLayout::kBatchDepthYX32:
      return absl::StrCat(batch, depth, spatial, suffix, "(VECT_C)");
    default:
      LOG(FATAL) << "Unknown layout " << static_cast<int32>(layout());
      return "";  // Avoid return warning (unreachable)
  }
}

int64_t BatchDescriptor::NodesPerFeatureMap() const {
  int64_t ret = 1;
  for (int i = 0; i < ndims(); i++) {
    ret *= spatial_size()[i];
  }
  return ret;
}

int64_t BatchDescriptor::NodesAcrossFeatureMaps() const {
  return NodesPerFeatureMap() * feature_map_count();
}

int64_t BatchDescriptor::ElementCount() const {
  return count() * feature_map_count() * NodesPerFeatureMap();
}

int64_t BatchDescriptor::FullyConnectedWeightCount(
    const BatchDescriptor& input, const BatchDescriptor& output) {
  return input.NodesAcrossFeatureMaps() * output.NodesAcrossFeatureMaps();
}

int64_t BatchDescriptor::FullyConnectedBiasCount(
    const BatchDescriptor& output) {
  return output.NodesAcrossFeatureMaps();
}

BatchDescriptor BatchDescriptor::DepthConcatenateOutputDescriptor(
    absl::Span<const dnn::BatchDescriptor> inputs) {
  if (inputs.empty()) {
    return BatchDescriptor();
  }
  int feature_map_count = 0;
  for (const auto& dimensions : inputs) {
    feature_map_count += dimensions.feature_map_count();
  }
  BatchDescriptor output = inputs[0];
  output.set_feature_map_count(feature_map_count);
  return output;
}

TensorDescriptorProto BatchDescriptor::ToProto(DataType data_type) const {
  CHECK_EQ(0.0, value_max_);
  CHECK_EQ(0.0, value_min_);
  CHECK(quantized_activation_mode_ == QuantizedActivationMode::k8Bit);

  TensorDescriptorProto ret = tensor_;
  ret.set_data_type(data_type);
  return ret;
}

// -- FilterDescriptor

FilterDescriptor::FilterDescriptor(int ndims) {
  tensor_.mutable_dimensions()->Resize(ndims + 2, 0);
  set_layout(FilterLayout::kOutputInputYX);
}

FilterDescriptor::FilterDescriptor() : FilterDescriptor(/*ndims=*/2) {}

FilterDescriptor::~FilterDescriptor() {}

void FilterDescriptor::CloneFrom(const FilterDescriptor& other) {
  tensor_ = other.tensor_;
}

std::string FilterDescriptor::ToString() const {
  std::string desc = absl::StrFormat(
      "{output_feature_map_count: %d input_feature_map_count: %d "
      "layout: %s shape: ",
      output_feature_map_count(), input_feature_map_count(),
      FilterLayoutString(layout()));
  for (int i = 0; i < ndims(); i++) {
    absl::StrAppendFormat(&desc, "%d ", input_filter_dims()[i]);
  }
  absl::StrAppend(&desc, "}");

  return desc;
}

std::string FilterDescriptor::ToShortString() const {
  // All the constituent strings are less than 15 characters, so the
  // small string optimization ensures that there will be at most one
  // heap memory allocation.
  std::string od = absl::StrCat("od", output_feature_map_count());
  std::string id = absl::StrCat("id", input_feature_map_count());

  std::string spatial = "s";
  for (int i = 0; i < ndims(); i++) {
    absl::StrAppendFormat(&spatial, "%d ", input_filter_dims()[i]);
  }

  switch (layout()) {
    case FilterLayout::kOutputInputYX:
      return absl::StrCat(od, id, spatial);
    case FilterLayout::kOutputYXInput:
      return absl::StrCat(od, spatial, id);
    case FilterLayout::kOutputInputYX4:
    case FilterLayout::kOutputInputYX32:
      return absl::StrCat(od, id, spatial, "(VECT_C)");
    case FilterLayout::kInputYXOutput:
      return absl::StrCat(id, spatial, od);
    case FilterLayout::kYXInputOutput:
      return absl::StrCat(spatial, id, od);
    default:
      LOG(FATAL) << "Unknown layout " << static_cast<int32>(layout());
      return "";  // Avoid return warning (unreachable)
  }
}

int64_t FilterDescriptor::ComputeWeightCount() const {
  int64_t ret = output_feature_map_count() * input_feature_map_count();
  for (int i = 0; i < ndims(); i++) {
    ret *= input_filter_dims()[i];
  }
  return ret;
}

std::vector<int64_t> FilterDescriptor::full_dims(
    const FilterLayout& layout) const {
  std::vector<int64_t> oiyx_dims(ndims() + 2);
  oiyx_dims[0] = output_feature_map_count();
  oiyx_dims[1] = input_feature_map_count();
  std::copy(input_filter_dims().begin(), input_filter_dims().end(),
            oiyx_dims.begin() + 2);
  return ReorderDims(oiyx_dims, FilterLayout::kOutputInputYX, layout);
}

std::vector<int64_t> FilterDescriptor::full_strides(
    const FilterLayout& layout) const {
  std::vector<int64_t> phys_dims = full_dims(this->layout());
  std::vector<int64_t> phys_strides(phys_dims.size());
  phys_strides[ndims() + 1] = 1;
  for (int i = ndims(); i >= 0; i--) {
    phys_strides[i] = phys_strides[i + 1] * phys_dims[i + 1];
  }
  return ReorderDims(phys_strides, this->layout(), layout);
}

std::vector<int64_t> FilterDescriptor::vectorized_dims(
    const FilterLayout& layout, int vector_size, int vector_dim) const {
  std::vector<int64_t> oiyx_dims = full_dims(dnn::FilterLayout::kOutputInputYX);
  if (vector_dim != -1) {
    oiyx_dims[vector_dim] /= vector_size;
  }
  return ReorderDims(oiyx_dims, FilterLayout::kOutputInputYX, layout);
}

std::vector<int64_t> FilterDescriptor::vectorized_strides(
    const FilterLayout& layout, int vector_size, int vector_dim) const {
  std::vector<int64_t> phys_dims =
      vectorized_dims(this->layout(), vector_size, vector_dim);
  std::vector<int64_t> phys_strides(phys_dims.size());
  phys_strides[phys_dims.size() - 1] = 1;
  for (int i = phys_dims.size() - 2; i >= 0; i--) {
    phys_strides[i] = phys_strides[i + 1] * phys_dims[i + 1];
  }
  return ReorderDims(phys_strides, this->layout(), layout);
}

TensorDescriptorProto FilterDescriptor::ToProto(DataType data_type) const {
  TensorDescriptorProto ret = tensor_;
  ret.set_data_type(data_type);
  return ret;
}

// -- ConvolutionDescriptor

ConvolutionDescriptor::ConvolutionDescriptor(int ndims) {
  proto_.mutable_paddings()->Resize(ndims, 0);
  proto_.mutable_strides()->Resize(ndims, 1);
  proto_.mutable_dilations()->Resize(ndims, 1);
  proto_.set_group_count(1);
  proto_.set_convolution_mode(ConvolutionMode::CROSS_CORRELATION);
}

ConvolutionDescriptor::ConvolutionDescriptor()
    : ConvolutionDescriptor(/*ndims=*/2) {}

ConvolutionDescriptor::~ConvolutionDescriptor() {}

std::string ConvolutionDescriptor::ToString() const {
  std::string padding;
  std::string strides;
  std::string dilations;
  for (int i = 0; i < ndims(); i++) {
    absl::StrAppendFormat(&padding, "%d ", this->padding()[i]);
    absl::StrAppendFormat(&strides, "%d ", this->strides()[i]);
    absl::StrAppendFormat(&dilations, "%d ", this->dilations()[i]);
  }

  return absl::StrFormat(
      "{zero_padding: %s pad_alignment: %s filter_strides: %s dilation_rates: "
      "%s}",
      padding, PadAlignmentString(pad_alignment()), strides, dilations);
}

std::string ConvolutionDescriptor::ToShortString() const {
  std::string desc;
  for (int i = 0; i < ndims(); i++) {
    if (i > 0) absl::StrAppend(&desc, "_");
    absl::StrAppendFormat(&desc, "p%d:%d", i, padding()[i]);
  }
  for (int i = 0; i < ndims(); i++) {
    absl::StrAppendFormat(&desc, "_s%d:%d", i, strides()[i]);
  }
  for (int i = 0; i < ndims(); i++) {
    absl::StrAppendFormat(&desc, "_d%d:%d", i, dilations()[i]);
  }
  return desc;
}

// -- PoolingDescriptor

PoolingDescriptor::PoolingDescriptor(int ndims)
    : mode_(dnn::PoolingMode::kMaximum),
      ndims_(ndims),
      propagate_nans_(false),
      window_(ndims, 0),
      padding_(ndims, 0),
      strides_(ndims, 1) {}

PoolingDescriptor::PoolingDescriptor() : PoolingDescriptor(/*ndims=*/2) {}

void PoolingDescriptor::CloneFrom(const PoolingDescriptor& other) {
  mode_ = other.mode_;
  ndims_ = other.ndims_;
  window_ = other.window_;
  padding_ = other.padding_;
  strides_ = other.strides_;
  propagate_nans_ = other.propagate_nans_;
}

std::string PoolingDescriptor::ToString() const {
  const char* mode_string =
      mode_ == dnn::PoolingMode::kMaximum ? "kMaximum" : "kAverage";

  std::string window, strides, padding;
  for (int i = 0; i < ndims_; i++) {
    absl::StrAppendFormat(&window, "%d ", window_[i]);
    absl::StrAppendFormat(&strides, "%d ", strides_[i]);
    absl::StrAppendFormat(&padding, "%d", padding_[i]);
  }

  const char* propagate_string = propagate_nans_ ? "Yes" : "No";

  return absl::StrFormat(
      "{mode: %s window: %s strides: %s padding: %s propagate NaNs: %s}",
      mode_string, window, strides, padding, propagate_string);
}

std::string PoolingDescriptor::ToShortString() const {
  std::string window, strides, padding;
  for (int i = 0; i < ndims_; i++) {
    absl::StrAppendFormat(&window, "_w%d:%d", i, window_[i]);
    absl::StrAppendFormat(&strides, "_s%d:%d", i, strides_[i]);
    absl::StrAppendFormat(&padding, "_p%d:%d", i, padding_[i]);
  }
  return absl::StrCat(mode_ == dnn::PoolingMode::kMaximum ? "max" : "avg",
                      window, strides, padding,
                      propagate_nans_ ? "propagate_nans" : "ignore_nans");
}

// -- NormalizeDescriptor

NormalizeDescriptor::NormalizeDescriptor()
    : bias_(0.0),
      range_(0),
      alpha_(0.0),
      beta_(0.0),
      wrap_around_(false),
      segment_size_(0) {}

void NormalizeDescriptor::CloneFrom(const NormalizeDescriptor& other) {
  bias_ = other.bias_;
  range_ = other.range_;
  alpha_ = other.alpha_;
  beta_ = other.beta_;
  wrap_around_ = other.wrap_around_;
  segment_size_ = other.segment_size_;
}

std::string NormalizeDescriptor::ToString() const {
  return absl::StrFormat(
      "{bias: %f range: %d alpha: %f beta: %f wrap_around: %d "
      "segment_size: %d}",
      bias_, range_, alpha_, beta_, wrap_around_, segment_size_);
}

std::string NormalizeDescriptor::ToShortString() const {
  return absl::StrCat("bias:", bias_, "_range:", range_, "_alpha:", alpha_,
                      "_beta:", beta_, "_wrap:", wrap_around_,
                      "_size:", segment_size_);
}

bool DnnSupport::IsStatusOk(const port::Status& status, bool report_error) {
  if (status.ok()) {
    return true;
  }
  if (report_error) {
    LOG(ERROR) << status.error_message();
  }
  return false;
}

port::Status DnnSupport::DoCtcLoss(
    Stream* stream, dnn::DataType element_type,
    const RnnStateTensorDescriptor& probs_desc,
    const DeviceMemoryBase probs_data, absl::Span<const int> labels_data,
    absl::Span<const int> labels_lengths_data,
    absl::Span<const int> input_lengths_data, DeviceMemoryBase costs_data,
    const RnnStateTensorDescriptor& grads_desc, DeviceMemoryBase grads_data,
    DeviceMemory<uint8> scratch_memory, int ctc_loss_algo_id) {
  return port::UnimplementedError("CtcLoss not implemented");
}

}  // namespace dnn
}  // namespace stream_executor