chromium/third_party/tflite/src/tensorflow/lite/kernels/internal/optimized/sparse_ops/fully_connected.h

/* Copyright 2020 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.
==============================================================================*/
#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_OPTIMIZED_SPARSE_OPS_FULLY_CONNECTED_H_
#define TENSORFLOW_LITE_KERNELS_INTERNAL_OPTIMIZED_SPARSE_OPS_FULLY_CONNECTED_H_

#include <algorithm>
#include <cstdint>

#include "ruy/profiler/instrumentation.h"  // from @ruy
#include "tensorflow/lite/core/c/common.h"
#include "tensorflow/lite/kernels/cpu_backend_threadpool.h"
#include "tensorflow/lite/kernels/internal/common.h"
#include "tensorflow/lite/kernels/internal/cppmath.h"
#include "tensorflow/lite/kernels/internal/optimized/optimized_ops.h"
#include "tensorflow/lite/kernels/internal/quantization_util.h"
#include "tensorflow/lite/kernels/internal/tensor_utils.h"
#include "tensorflow/lite/kernels/internal/types.h"

namespace tflite {
namespace optimized_ops {

inline void FullyConnectedSparseWeight(
    const TfLiteSparsity& sparsity, const FullyConnectedParams& params,
    const RuntimeShape& input_shape, const float* input_data,
    const RuntimeShape& weights_shape, const float* weights_data,
    const RuntimeShape& bias_shape, const float* bias_data,
    const RuntimeShape& output_shape, float* output_data) { … }

inline void FullyConnectedSparseWeight1x16Impl(
    const TfLiteSparsity& sparsity, const FullyConnectedParams& params,
    const RuntimeShape& input_shape, const int8_t* input_data,
    const RuntimeShape& weights_shape, const int8_t* weights_data,
    const int32_t* per_channel_scale, const int32_t* per_channel_shift,
    const RuntimeShape& bias_shape, const int32_t* bias_data,
    const RuntimeShape& output_shape, int8_t* output_data, int thread_start,
    int thread_end, const CpuBackendContext& cpu_backend_context) { … }

inline void FullyConnectedSparseWeight1x4Impl(
    const TfLiteSparsity& sparsity, const FullyConnectedParams& params,
    const RuntimeShape& input_shape, const float* input_data,
    const RuntimeShape& weights_shape, const float* weights_data,
    const RuntimeShape& bias_shape, const float* bias_data,
    const RuntimeShape& output_shape, float* output_data, int thread_start,
    int thread_end, const CpuBackendContext& cpu_backend_context) { … }

struct FullyConnectedSparseWeight1x4Task : cpu_backend_threadpool::Task { … };

inline void FullyConnectedSparseWeight1x16(
    const TfLiteSparsity& sparsity, const FullyConnectedParams& params,
    const RuntimeShape& input_shape, const int8_t* input_data,
    const RuntimeShape& weights_shape, const int8_t* weights_data,
    const int32_t* per_channel_scale, const int32_t* per_channel_shift,
    const RuntimeShape& bias_shape, const int32_t* bias_data,
    const RuntimeShape& output_shape, int8_t* output_data,
    CpuBackendContext* cpu_backend_context) { … }

// The multi-threaded kernel slices the workload along the batch dimension. If
// there's not enough batches of data, the number of threads used is equal to
// the batch size. We can improve this later with slicing along the row
// dimension of the weight.
inline void FullyConnectedSparseWeight1x4(
    const TfLiteSparsity& sparsity, const FullyConnectedParams& params,
    const RuntimeShape& input_shape, const float* input_data,
    const RuntimeShape& weights_shape, const float* weights_data,
    const RuntimeShape& bias_shape, const float* bias_data,
    const RuntimeShape& output_shape, float* output_data,
    CpuBackendContext* cpu_backend_context) { … }

}  // namespace optimized_ops
}  // namespace tflite
#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_OPTIMIZED_SPARSE_OPS_FULLY_CONNECTED_H_