FM / core

/usr/local/lib64/python3.6/site-packages/torch/include/caffe2/core/

/usr/local/lib64/python3.6/site-packages/torch/include/caffe2/core

Name	Size	Mode	Actions
allocator.h	136	0644	edit dl rm
blob.h	4168	0644	edit dl rm
blob_serialization.h	10791	0644	edit dl rm
blob_serializer_base.h	3905	0644	edit dl rm
blob_stats.h	1127	0644	edit dl rm
common.h	4329	0644	edit dl rm
common_cudnn.h	9893	0644	edit dl rm
common_gpu.h	21414	0644	edit dl rm
common_omp.h	156	0644	edit dl rm
context.h	6174	0644	edit dl rm
context_base.h	4382	0644	edit dl rm
context_gpu.h	11014	0644	edit dl rm
cudnn_wrappers.h	6956	0644	edit dl rm
db.h	9352	0644	edit dl rm
distributions_stubs.h	2161	0644	edit dl rm
event.h	12420	0644	edit dl rm
event_cpu.h	1192	0644	edit dl rm
export_c10_op_to_caffe2.h	9487	0644	edit dl rm
export_caffe2_op_to_c10.h	11101	0644	edit dl rm
flags.h	74	0644	edit dl rm
graph.h	5258	0644	edit dl rm
init.h	6496	0644	edit dl rm
logging.h	75	0644	edit dl rm
macros.h	3426	0644	edit dl rm
memonger.h	817	0644	edit dl rm
module.h	2473	0644	edit dl rm
net.h	4634	0644	edit dl rm
net_async_base.h	7397	0644	edit dl rm
net_async_scheduling.h	993	0644	edit dl rm
net_async_task.h	833	0644	edit dl rm
net_async_task_future.h	1925	0644	edit dl rm
net_async_task_graph.h	2253	0644	edit dl rm
net_async_tracing.h	5093	0644	edit dl rm
net_dag_utils.h	2146	0644	edit dl rm
net_parallel.h	2144	0644	edit dl rm
net_simple.h	2606	0644	edit dl rm
net_simple_refcount.h	2097	0644	edit dl rm
numa.h	72	0644	edit dl rm
observer.h	3809	0644	edit dl rm
operator.h	58872	0644	edit dl rm
operator_gradient.h	10222	0644	edit dl rm
operator_schema.h	18477	0644	edit dl rm
plan_executor.h	219	0644	edit dl rm
prof_dag_counters.h	2751	0644	edit dl rm
qtensor.h	6615	0644	edit dl rm
qtensor_serialization.h	2624	0644	edit dl rm
scope_guard.h	4675	0644	edit dl rm
static_tracepoint.h	398	0644	edit dl rm
static_tracepoint_elfx86.h	5555	0644	edit dl rm
stats.h	10365	0644	edit dl rm
storage.h	733	0644	edit dl rm
tensor.h	18668	0644	edit dl rm
tensor_impl.h	351	0644	edit dl rm
tensor_int8.h	450	0644	edit dl rm
test_utils.h	6285	0644	edit dl rm
timer.h	1218	0644	edit dl rm
transform.h	5741	0644	edit dl rm
types.h	2248	0644	edit dl rm
workspace.h	11305	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/include/caffe2/core/qtensor.h (6615B)

#ifndef CAFFE2_CORE_QTENSOR_H_
#define CAFFE2_CORE_QTENSOR_H_

#include "caffe2/core/common.h"
#include "caffe2/core/context.h"
#include "caffe2/core/tensor.h"
#include <c10/util/accumulate.h>
#include <c10/util/typeid.h>

#include <algorithm>
#include <climits>
#include <cstddef>
#include <vector>

namespace caffe2 {

template <class Context>
class C10_EXPORT QTensor {
 public:
  QTensor() {}
  virtual ~QTensor() {}
  /**
   * @brief Creates a quantized tensor of the given dimension.
   *
   * Note that the actual data allocation is not going to be carried out until
   * the first time mutable_data() is called.
   *
   * The underlying storage of the quantized tensor interleaves elements
   * by bit depth.
   *
   * Labeled memory for tensor of size 6, precision 3
   *   [ E1[0] E2[0] E3[0] E4[0] E5[0] E6[0] ] // Least significant Bits
   *   [ E1[1] E2[1] E3[1] E4[1] E5[1] E6[1] ]
   *   [ E1[2] E2[2] E3[2] E4[2] E5[2] E6[2] ]
   *
   * In the case of sign bits (see enable_sign argument), an extra bit
   * per element is added:
   *
   * Labeled memory for tensor of size 6, precision 3, sign bit enabled
   *   [ E1[0] E2[0] E3[0] E4[0] E5[0] E6[0] ]
   *   [ E1[1] E2[1] E3[1] E4[1] E5[1] E6[1] ]
   *   [ E1[2] E2[2] E3[2] E4[2] E5[2] E6[2] ]
   *   [ E1[s] E2[s] E3[s] E4[s] E5[s] E6[s] ]
   *   Where 's' is 1 if E is negative
   *
   * The reason for this layout is the ability to efficiently multiply
   * many low precision integers as a sum of popcnt(A & B) * 1 << bit.
   * Explained here: https://arxiv.org/abs/1606.06160
   */
  // TODO: changing at::ArrayRef<int> to at::ArrayRef<int64_t>?
  explicit QTensor(
      at::ArrayRef<int> dims,
      const unsigned char precision,
      const bool signbit = false)
      : precision_(precision), signed_(signbit) {
    Resize(dims);
  }

void Resize(at::ArrayRef<int> dim_source) {
    if (dims_ != dim_source) {
      const auto source_size = c10::multiply_integers(dim_source);
      // NOLINTNEXTLINE(clang-diagnostic-sign-compare)
      if ((source_size * (precision_ + signed_)) > capacity_) {
        data_ptr_.clear();
        capacity_ = 0;
      }
      dims_ = dim_source.vec();
      size_ = source_size;
    }
  }

void
  SetBitAtIndex(const unsigned char bit, const size_t index, const bool value) {
    // Get the mutable data at bit depth `bit`.
    unsigned char* d = mutable_data();

CAFFE_ENFORCE(
        bit < precision_ + signed_,
        "Attempted to a set a bit that is not allocated.");
    CAFFE_ENFORCE(bit * aligned_size() < capacity_);

auto idx = (aligned_size() * bit) / CHAR_BIT;
    d = &d[idx];

idx = index / CHAR_BIT;
    auto shift = CHAR_BIT - (index % CHAR_BIT) - 1;

if (value) {
      d[idx] |= 1 << shift;
    } else {
      d[idx] &= ~(1 << shift);
    }
  }

bool GetBitAtIndex(const unsigned char bit, const size_t index) const {
    // Get the data at bit depth `bit`
    const unsigned char* d = data();
    auto idx = (aligned_size() * bit) / CHAR_BIT;
    d = &d[idx];

idx = index / CHAR_BIT;
    auto shift = CHAR_BIT - (index % CHAR_BIT) - 1;

return d[idx] & (1 << shift);
  }

void SetPrecision(const unsigned char precision) {
    precision_ = precision;
    data_ptr_.clear();
  }

void SetSigned(const bool make_signed = true) {
    signed_ = make_signed;
    data_ptr_.clear();
  }

void SetScale(const double scale) {
    scale_ = scale;
  }

void SetBias(const double bias) {
    bias_ = bias;
  }

unsigned char* mutable_data() {
    if (!data_ptr_) {
      data_ptr_ = Context::New(nbytes());
      capacity_ = nbytes() * CHAR_BIT;
    }
    CAFFE_ENFORCE(capacity_ == nbytes() * CHAR_BIT);
    return static_cast<unsigned char*>(data_ptr_.get());
  }

inline const unsigned char* data() const {
    return static_cast<unsigned char*>(data_ptr_.get());
  }

inline size_t size() const {
    return size_;
  }

inline unsigned char alignment() const {
    return alignment_;
  }

inline unsigned char precision() const {
    return precision_;
  }

inline at::ArrayRef<int> sizes() const {
    return dims_;
  }

// TODO: deprecate?
  inline at::ArrayRef<int> dims() const {
    return dims_;
  }

inline bool is_signed() const {
    return signed_;
  }

/**
   * Returns the number of dimensions of the data.
   */
  inline int ndim() const {
    return dims_.size();
  }

inline size_t aligned_size() const {
    return alignment_ * ((size_ + alignment_ - 1) / alignment_);
  }

inline size_t nbytes() const {
    return (aligned_size() * (precision_ + signed_)) / CHAR_BIT;
  }

inline double scale() const {
    return scale_;
  }

inline double bias() const {
    return bias_;
  }

/**
   * Returns the i-th dimension of the qtensor in int.
   */
  inline int dim32(const int i) const {
    DCHECK_LT(i, dims_.size()) << "Exceeding ndim limit " << dims_.size();
    DCHECK_GE(i, 0) << "Cannot have negative index";
    CAFFE_ENFORCE_LT(dims_[i], std::numeric_limits<int>::max());
    return static_cast<int>(dims_[i]);
  }

/**
   * Returns the 'canonical' version of a (usually)  user-specified axis,
   * allowing for negative indexing (e.g., -1 for the last axis).
   *
   * @param axis_index the axis index.
   *        If 0 <= index < ndim(), return index.
   *        If -ndim <= index <= -1, return (ndim() - (-index)),
   *        e.g., the last axis index (ndim() - 1) if index == -1,
   *        the second to last if index == -2, etc.
   *        Dies on out of range index.
   */
  inline int canonical_axis_index(int axis_index) const {
    CAFFE_ENFORCE_GE(axis_index, -ndim());
    CAFFE_ENFORCE_LT(axis_index, ndim());
    if (axis_index < 0) {
      return axis_index + ndim();
    }
    return axis_index;
  }

/**
   * Return product of all dimensions starting from K.
   */
  inline int64_t size_from_dim(int k) const {
    int64_t r = 1;
    for (int i = k; i < dims_.size(); ++i) {
      r *= dims_[i];
    }
    return r;
  }

/**
   * Product of all dims up to.
   */
  inline int64_t size_to_dim(int k) const {
    CAFFE_ENFORCE(k < dims_.size());
    int64_t r = 1;
    for (int i = 0; i < k; ++i) {
      r *= dims_[i];
    }
    return r;
  }

protected:
  std::vector<int> dims_;
  size_t size_ = 0;

// Precision in bits.
  unsigned char precision_ = CHAR_BIT;
  // Bit alignment.
  unsigned char alignment_ = CHAR_BIT;

// Allocated data.
  at::DataPtr data_ptr_;

// value = scale_ * (x + bias_)
  double scale_;
  double bias_;
  bool signed_ = false;

// Capacity in bits.
  size_t capacity_ = 0;
};

} // namespace caffe2
#endif // CAFFE2_CORE_QTENSOR_H_

cmd: