** Additional Installation Instructions **


__________________ Ubuntu 22.04 and more _________________

To compile nvBowtie, it is essential to install the CUDA development suite. You can find many pages explaining how to 
install cuda on your linux distribution. If you're using Ubuntu 24.04 or higher, you can also use the installation_cuda.sh 
script located in the 'inst/scripts' directory.
This script will erase all traces of drivers and older versions of cuda before you start. This can be useful if you are 
unable to install cuda correctly after several attempts.
In addition, this script also offers to install the default version of your distribution, or to install the latest version (12.8) 
if you're having problems with the older version. Use the default version first.


An additional script, called check_requirements.sh, verifies your installation and determines if everything went smoothly.

Architecture handling (auto-detection + PTX fallback):
By default nvBowtie auto-detects the GPU(s) on the build host via nvidia-smi and compiles native SASS
for each detected compute capability, PLUS a PTX JIT fallback for the oldest architecture supported by
the installed CUDA toolkit. The resulting fatbin therefore runs natively on the build GPU and via driver
JIT on any other GPU at or above that floor, so a package built on one card works across varied GPUs.

Floor (oldest target) by toolkit: CUDA 13.x -> sm_75 (Turing); CUDA <= 12.x -> sm_70 (Volta).
Note: CUDA 13 removed Pascal/Maxwell/Volta, so nvcc 13 rejects sm_60/sm_70 targets outright.

To override auto-detection, set CMAKE_CUDA_ARCHITECTURES when configuring src/nvbio, e.g.:
    cmake -DCMAKE_CUDA_ARCHITECTURES="120;75-virtual" ..
(native sm_120 SASS + compute_75 PTX JIT fallback). Keep at least one "-virtual"/PTX entry for portability.
Refer to the NVIDIA compatibility guide for matching cards:
(https://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards).

It is also necessary to install the thrust and cub libraries by typing in a terminal :
sudo apt install libthrust-dev libcub-dev

_____________________ Windows __________________

** Cuda installation on Windows **

You must have a developer account on the NVidia website (https://developer.nvidia.com).
If you don’t already have one, you will need to create an account.

Next, download the CUDA Toolkit 11.8, which is available here:

https://developer.nvidia.com/cuda-11-8-0-download-archive?target_os=Windows&target_arch=x86_64

For detailed instructions on installing CUDA on Windows, refer to this guide:
 
https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html

Note that RbowtieCuda is compatible with Cuda versions >= 10.

** Installing RTools on Windows **

RTools must also be installed on Windows. You can download it here:

https://cran.r-project.org/bin/windows/Rtools/

** Installing Build Tools for Visual Studio 2022 **

The Visual Studio Build Tools are required and can be downloaded here:

https://visualstudio.microsoft.com/en/downloads/

** Note: Compilation on Windows **

The package compilation process on Windows can be lengthy (up to 15 minutes).

________________ Recommended runtime options _________________

For paired-end FASTQ workloads, we recommend running nvBowtie with:

    --cache-writes --batch-size 2048

This overlaps output writing with GPU compute through a dedicated write thread.
Measured on a mouse dataset (5M read pairs, GRCm39 index, RTX 5070, SATA SSD):
total wall time drops from ~36 s to ~28 s (compute ~20 s + index load ~4 s + I/O).

Notes:
- Index load is ~4 s per call (bwt/rbwt/occ/ssa); the occurrence tables are built on the fly
  (~1.3 s each). Reusing a warm OS page cache between calls keeps this low.
- SAM output for 5M pairs is ~3.4 GB; plan disk space accordingly. Writing to tmpfs (e.g. /tmp)
  avoids the write tax when the SAM file is consumed immediately downstream.
- BAM output (-S out.bam) is supported and bit-exact with the SAM records, but currently slower
  overall (~41 s vs ~28 s on the same run): BGZF compression happens inline instead of on the
  shared write thread. Prefer SAM unless BAM size is the bottleneck.