FM / cuda

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

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

Name	Size	Mode	Actions
BatchLinearAlgebraLib.h	3114	0644	edit dl rm
block_reduce.cuh	2549	0644	edit dl rm
CompositeRandomAccessor.h	929	0644	edit dl rm
CUDALoops.cuh	7598	0644	edit dl rm
CuFFTPlanCache.h	19282	0644	edit dl rm
CuFFTUtils.h	1892	0644	edit dl rm
DeviceSqrt.cuh	585	0644	edit dl rm
DistributionTemplates.h	27435	0644	edit dl rm
EmbeddingBackwardKernel.cuh	715	0644	edit dl rm
ForeachFunctors.cuh	16851	0644	edit dl rm
GridSampler.cuh	11316	0644	edit dl rm
im2col.cuh	6577	0644	edit dl rm
KernelUtils.cuh	2553	0644	edit dl rm
LaunchUtils.h	306	0644	edit dl rm
Loops.cuh	9997	0644	edit dl rm
Math.cuh	13840	0644	edit dl rm
MemoryAccess.cuh	12463	0644	edit dl rm
MiscUtils.h	3341	0644	edit dl rm
MultiTensorApply.cuh	7552	0644	edit dl rm
Normalization.cuh	74441	0644	edit dl rm
PersistentSoftmax.cuh	14635	0644	edit dl rm
Randperm.cuh	2114	0644	edit dl rm
Reduce.cuh	38784	0644	edit dl rm
Resize.cuh	1919	0644	edit dl rm
ROCmLoops.cuh	13526	0644	edit dl rm
SortingCommon.cuh	5688	0644	edit dl rm
SortingRadixSelect.cuh	11918	0644	edit dl rm
SortUtils.cuh	5549	0644	edit dl rm
TensorModeKernel.cuh	14391	0644	edit dl rm
UniqueCub.cuh	345	0644	edit dl rm
UpSample.cuh	7552	0644	edit dl rm
vol2col.cuh	8297	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/include/ATen/native/cuda/Resize.cuh (1919B)

#pragma once

#include <ATen/ATen.h>
#include <THC/THCTensor.hpp>
#include <ATen/native/ResizeCommon.h>

#include <c10/cuda/CUDAGuard.h>

namespace at { namespace native {

// These functions are called by native::resize_ as well as (legacy) THC resize.
// They are not in THC/THCTensor.cpp because the at namespace is easier
// to benchmark than THC; I can't get gbenchmark to call fns from THTensor.cpp

static inline void maybe_resize_storage_cuda(TensorImpl* self, uint64_t new_size) {
  // It does not make sense to try to resize a storage
  // to hold 0 elements, and this can break
  // if storage_offset is positive but
  // new_size is 0, so just bail in that case
  // (same comment is in Resize.h)
  if (new_size == 0) {
    return;
  }
  if (!THTensor_getStoragePtr(self)) {
    TORCH_CHECK(false, "Tensor: invalid null storage");
  }
  uint64_t new_size_bytes = (new_size + self->storage_offset()) * self->dtype().itemsize();
  if (new_size_bytes > self->storage().nbytes()) {
    THCStorage_resizeBytes(
        globalContext().getTHCState(),
        THTensor_getStoragePtr(self),
        new_size_bytes
    );
  }
}

inline TensorImpl* resize_impl_cuda_(
    TensorImpl* self,
    IntArrayRef size,
    c10::optional<IntArrayRef> stride,
    bool device_guard = true) {
  if (self->sizes() == size && (!stride || self->strides() == stride)) {
    return self;
  }

// NB: We don't need to hold the device guard when calling from TH
  cuda::OptionalCUDAGuard guard;
  if (device_guard) {
    guard.set_index(self->storage().device().index());
  }

int64_t storage_size = 1;
  if (stride) {
    self->set_sizes_and_strides(size, *stride);
    // NB: storage size can be different from numel.
    storage_size = storage_size_for(size, *stride);
  } else {
    self->set_sizes_contiguous(size);
    storage_size = self->numel();
  }
  maybe_resize_storage_cuda(self, storage_size);

return self;
}

}}

cmd: