Skip to content

use bitshifts for int to int in convert_dtype #6978

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Nov 28, 2022
Merged

use bitshifts for int to int in convert_dtype #6978

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d5da6e4
use bitshifts for int to int in convert_dtype
pmeier Nov 25, 2022
5cf3fa4
Merge branch 'main' into convert-dtype
pmeier Nov 28, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 1 addition & 9 deletions torchvision/prototype/transforms/functional/_meta.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -379,15 +379,7 @@ def convert_dtype_image_tensor(image: torch.Tensor, dtype: torch.dtype = torch.f
if num_value_bits_input > num_value_bits_output:
return image.bitwise_right_shift(num_value_bits_input - num_value_bits_output).to(dtype)
else:
# The bitshift kernel is not vectorized
# https://github.com/pytorch/pytorch/blob/703c19008df4700b6a522b0ae5c4b6d5ffc0906f/aten/src/ATen/native/cpu/BinaryOpsKernel.cpp#L315-L322
# This results in the multiplication actually being faster.
# TODO: If the bitshift kernel is optimized in core, replace the computation below with
# `image.to(dtype).bitwise_left_shift_(num_value_bits_output - num_value_bits_input)`
max_value_input = float(_FT._max_value(dtype))
max_value_output = float(_FT._max_value(image.dtype))
factor = int((max_value_input + 1) // (max_value_output + 1))
return image.to(dtype).mul_(factor)
return image.to(dtype).bitwise_left_shift_(num_value_bits_output - num_value_bits_input)


# We changed the name to align it with the new naming scheme. Still, `convert_image_dtype` is
Expand Down