FM / ATen

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

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

Name	Size	Mode	Actions
core/	-	0755	rm
cpu/	-	0755	rm
cuda/	-	0755	rm
cudnn/	-	0755	rm
detail/	-	0755	rm
hip/	-	0755	rm
native/	-	0755	rm
quantized/	-	0755	rm
AccumulateType.h	4438	0644	edit dl rm
ArrayRef.h	44	0644	edit dl rm
ATen.h	998	0644	edit dl rm
autocast_mode.h	6716	0644	edit dl rm
Backend.h	43	0644	edit dl rm
Backtrace.h	46	0644	edit dl rm
BatchedFallback.h	965	0644	edit dl rm
BatchedTensorImpl.h	5383	0644	edit dl rm
CompositeExplicitAutogradFunctions.h	1622	0644	edit dl rm
CompositeExplicitAutogradFunctions_inl.h	54075	0644	edit dl rm
CompositeImplicitAutogradFunctions.h	1622	0644	edit dl rm
CompositeImplicitAutogradFunctions_inl.h	142082	0644	edit dl rm
Config.h	734	0644	edit dl rm
Context.h	12767	0644	edit dl rm
cpp_custom_type_hack.h	5326	0644	edit dl rm
CPUApplyUtils.h	12582	0644	edit dl rm
CPUFixedAllocator.h	830	0644	edit dl rm
CPUFunctions.h	1600	0644	edit dl rm
CPUFunctions_inl.h	171924	0644	edit dl rm
CPUGeneratorImpl.h	1431	0644	edit dl rm
CUDAFunctions.h	1601	0644	edit dl rm
CUDAFunctions_inl.h	185696	0644	edit dl rm
CUDAGeneratorImpl.h	4695	0644	edit dl rm
Device.h	42	0644	edit dl rm
DeviceGuard.h	1134	0644	edit dl rm
Dimname.h	31	0644	edit dl rm
DimVector.h	46	0644	edit dl rm
Dispatch.h	52137	0644	edit dl rm
div_rtn.h	204	0644	edit dl rm
DLConvertor.h	576	0644	edit dl rm
dlpack.h	5244	0644	edit dl rm
DynamicLibrary.h	369	0644	edit dl rm
ExpandUtils.h	14506	0644	edit dl rm
Formatting.h	34	0644	edit dl rm
Functions.h	846326	0644	edit dl rm
Generator.h	46	0644	edit dl rm
InferSize.h	2143	0644	edit dl rm
InitialTensorOptions.h	445	0644	edit dl rm
Layout.h	42	0644	edit dl rm
MapAllocator.h	2999	0644	edit dl rm
MatrixRef.h	3016	0644	edit dl rm
MemoryOverlap.h	1117	0644	edit dl rm
MetaFunctions.h	1601	0644	edit dl rm
MetaFunctions_inl.h	84006	0644	edit dl rm
NamedTensor.h	35	0644	edit dl rm
NamedTensorUtils.h	5747	0644	edit dl rm
NativeFunctions.h	354651	0644	edit dl rm
NativeMetaFunctions.h	35445	0644	edit dl rm
NumericUtils.h	2787	0644	edit dl rm
OpaqueTensorImpl.h	6080	0644	edit dl rm
Operators.h	1707199	0644	edit dl rm
OpMathType.h	460	0644	edit dl rm
Parallel.h	4875	0644	edit dl rm
ParallelNative.h	2443	0644	edit dl rm
ParallelNativeTBB.h	2934	0644	edit dl rm
ParallelOpenMP.h	3049	0644	edit dl rm
PTThreadPool.h	394	0644	edit dl rm
record_function.h	24044	0644	edit dl rm
RedispatchFunctions.h	1112886	0644	edit dl rm
RegistrationDeclarations.h	545777	0644	edit dl rm
SavedTensorHooks.h	328	0644	edit dl rm
Scalar.h	44	0644	edit dl rm
ScalarOps.h	2272	0644	edit dl rm
ScalarType.h	129	0644	edit dl rm
SequenceNumber.h	373	0644	edit dl rm
SmallVector.h	47	0644	edit dl rm
SparseCsrTensorImpl.h	2045	0644	edit dl rm
SparseCsrTensorUtils.h	523	0644	edit dl rm
SparseTensorImpl.h	12417	0644	edit dl rm
SparseTensorUtils.h	4219	0644	edit dl rm
Storage.h	43	0644	edit dl rm
Tensor.h	48	0644	edit dl rm
TensorAccessor.h	51	0644	edit dl rm
TensorGeometry.h	1855	0644	edit dl rm
TensorIndexing.h	21923	0644	edit dl rm
TensorIterator.h	29962	0644	edit dl rm
TensorIteratorInternal.h	1862	0644	edit dl rm
TensorMeta.h	2917	0644	edit dl rm
TensorNames.h	2519	0644	edit dl rm
TensorOperators.h	3275	0644	edit dl rm
TensorOptions.h	49	0644	edit dl rm
TensorUtils.h	5687	0644	edit dl rm
ThreadLocalState.h	3289	0644	edit dl rm
TracerMode.h	5576	0644	edit dl rm
TypeDefault.h	680	0644	edit dl rm
Utils.h	5993	0644	edit dl rm
Version.h	340	0644	edit dl rm
VmapMode.h	952	0644	edit dl rm
VmapTransforms.h	7654	0644	edit dl rm
WrapDimUtils.h	3438	0644	edit dl rm
WrapDimUtilsMulti.h	768	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/include/ATen/autocast_mode.h (6716B)

#pragma once

namespace at {
namespace autocast {

TORCH_API bool is_enabled();
TORCH_API void set_enabled(bool enabled);
TORCH_API void clear_cache();
TORCH_API int increment_nesting();
TORCH_API int decrement_nesting();
TORCH_API bool is_cpu_enabled();
TORCH_API void set_cpu_enabled(bool enabled);
TORCH_API at::ScalarType get_autocast_gpu_dtype();
TORCH_API at::ScalarType get_autocast_cpu_dtype();
TORCH_API void set_autocast_gpu_dtype(at::ScalarType dtype);
TORCH_API void set_autocast_cpu_dtype(at::ScalarType dtype);
TORCH_API bool is_autocast_cache_enabled();
TORCH_API void set_autocast_cache_enabled(bool enabled);

namespace {
  bool is_autocast_eligible(const Tensor& tensor, DeviceType device_type) {
    return device_type == DeviceType::CUDA
        ? (tensor.is_cuda() || tensor.is_xla()) && tensor.is_floating_point()
        : (tensor.is_cpu() || tensor.is_mkldnn()) && tensor.is_floating_point();
  }
} // namespace

inline DispatchKey get_autocast_dispatch_key_from_device_type(
    DeviceType device_type) {
  switch (device_type) {
    case DeviceType::CUDA:
      return DispatchKey::Autocast;
    case DeviceType::CPU:
      return DispatchKey::AutocastCPU;
    default:
      throw std::runtime_error(
          "unknown device type for autocast in get_autocast_dispatch_key_from_device_type");
  }
}

inline at::ScalarType get_lower_precision_fp_from_device_type(
    DeviceType device_type) {
  switch (device_type) {
    case DeviceType::CUDA:
      return get_autocast_gpu_dtype();
    case DeviceType::CPU:
      return get_autocast_cpu_dtype();
    default:
      throw std::runtime_error(
          "unknown device type for autocast in get_lower_precision_fp_from_device_type");
  }
}

/********************************************************************
Logic to extract the promote type from any Tensor or TensorList args.
********************************************************************/

// Overload to catch Tensor args.
// If nextArg is floating-point, compare its scalar_type with our
// current best guess for the promote type, and update if necessary.
inline at::ScalarType prioritize(
    at::ScalarType current,
    const Tensor& nextArg,
    DeviceType device_type=DeviceType::CUDA) {
  if (current == at::kDouble) {
    AT_ERROR("promote type is double in at::autocast::prioritize");
    return current;
  }
  at::ScalarType lower_precision_fp =
      get_lower_precision_fp_from_device_type(device_type);
  if (is_autocast_eligible(nextArg, device_type)) {
    auto next = nextArg.scalar_type();
    if (next == at::kDouble) {
      return current; // ignores double tensors
    } else if (current == at::kFloat || next == at::kFloat) {
      return at::kFloat; // prioritizes float over lower_precision_fp
    } else if (current == lower_precision_fp && next == lower_precision_fp) {
      return lower_precision_fp;
    } else {
      AT_ERROR("Unexpected floating ScalarType in at::autocast::prioritize");
      return current;
    }
  } else {
    return current;
  }
}

// Overload to catch TensorList args (for e.g. cat, stack).
// Reuses the overload above to process each Tensor in the list.
inline at::ScalarType prioritize(
    at::ScalarType current,
    const TensorList& list,
    DeviceType device_type=DeviceType::CUDA) {
  for (const auto& tensor : list) {
    current = prioritize(current, tensor, device_type);
  }
  return current;
}

// Template to catch non-Tensor args (no-op that returns current best guess)
template<typename T>
inline at::ScalarType prioritize(
    at::ScalarType current,
    T nextArg,
    DeviceType device_type=DeviceType::CUDA) {
  return current;
}

// Overload for the tail case.
inline at::ScalarType promote_type(
    at::ScalarType current,
    DeviceType device_type) {
  return current;
}

// Unpack args and determine if incoming lower_precision_fp tensors need to be promoted to float32.
// Non-Tensor arguments are ignored.
template<typename Arg0, typename... Args>
inline at::ScalarType promote_type(
    at::ScalarType current,
    DeviceType device_type,
    Arg0 arg0,
    Args... args) {
  auto new_current = prioritize(current, arg0, device_type);
  return promote_type(new_current, device_type, args...);
}

/****************************************************
Logic to apply cached casting to any Tensor argument.
****************************************************/
inline bool is_eligible(
    const Tensor& arg,
    DeviceType device_type=DeviceType::CUDA) {
  return (arg.defined() &&
          is_autocast_eligible(arg, device_type) &&
          (arg.scalar_type() != at::kDouble));
}

// Overload to catch Tensor args
TORCH_API Tensor cached_cast(
    at::ScalarType to_type,
    const Tensor& arg,
    DeviceType device_type=DeviceType::CUDA);

// Overload to process optional<Tensor>
inline c10::optional<Tensor> cached_cast(
    at::ScalarType to_type,
    const c10::optional<Tensor>& arg,
    DeviceType device_type=DeviceType::CUDA) {
  if (arg.has_value()) {
    return cached_cast(to_type, *arg, device_type);
  } else {
    return c10::nullopt;
  }
}

// Overload to process TensorLists
inline std::vector<Tensor> cached_cast(
    at::ScalarType to_type,
    const TensorList& arg,
    DeviceType device_type=DeviceType::CUDA) {
  std::vector<Tensor> vec;
  vec.reserve(arg.size());
  for (const auto& t : arg) {
    vec.push_back(cached_cast(to_type, t, device_type));
  }
  return vec;
}

// Template to catch non-Tensor args.
template<typename T>
inline T cached_cast(
    at::ScalarType to_type,
    T arg,
    DeviceType device_type=DeviceType::CUDA) {
  return arg;
}

/*******************************************************
Logic to flip an output dtype flag.
Keep it simple for now by assuming only one such flag is
present in the argument list.  If I ever need a function
with more than flag I'll figure out something else.
The policy is:
If the user has explicity specified a dtype, respect it.
Otherwise, set it to the autocast type.
********************************************************/

// Overload to catch dtype flags
c10::optional<ScalarType> inline set_opt_dtype(at::ScalarType to_type, const c10::optional<ScalarType>& dtype) {
  return dtype.has_value() ? dtype : to_type;
}

// Template to catch other args
template<typename T>
inline T set_opt_dtype(at::ScalarType to_type, T arg) {
  return arg;
}

template<typename... Args>
inline bool firstarg_is_eligible(const Tensor& arg, Args... args) {
  return is_eligible(arg);
}

template<typename... Args>
inline at::ScalarType type_from_firstarg(at::ScalarType to_type, const Tensor& arg, Args... args) {
  return (is_eligible(arg) ? to_type : arg.scalar_type());
}

} // namespace autocast
} // namespace at

cmd: