melonedo

"""Decode JPEG data directly into CUDA tensors with nvImageCodec.

The decoder enables fancy chroma upsampling and asks nvImageCodec to produce the
requested color order and layout directly. Compressed JPEG data remains on the
host, while decoded pixels are allocated in GPU memory and exposed to PyTorch
without copying through DLPack/CUDA array interoperability.

```
import cv2
import numpy as np

melonedo

# The algorithm proposed by https://github.com/triton-lang/triton/pull/7558

from dataclasses import dataclass
from enum import Enum


def inverse_permutation(P):
    """
    Given a permutation P, return its inverse permutation.
    """

melonedo

### A Pluto.jl notebook ###
# v0.17.4

using Markdown
using InteractiveUtils

# ╔═╡ 72851118-6cb1-11ec-3338-8d9690c97989
begin
	using Plots,LaTeXStrings
	import PlotlyJS