Code

CUDA#368

Code

Draft

ajwheeler wants to merge 40 commits intomainajwheeler/Korg.jl:mainfrom cudaajwheeler/Korg.jl:cudaCopy head branch name to clipboard

+352-81

Commits

Commits on Nov 18, 2024

• set up

  ajwheeler

  committed
  3deb82d

  Copy full SHA for 3deb82d

• add CUDA version of line absorption code

  ajwheeler

  committed
  ec5c401

  Copy full SHA for ec5c401

• it is technically on GPU!

  ajwheeler

  committed
  f7814c4

  Copy full SHA for f7814c4

• add sync (still not that slow???)

  ajwheeler

  committed
  9d5a822

  Copy full SHA for 9d5a822

• do a bit more on the GPU

  ajwheeler

  committed
  068d1c4

  Copy full SHA for 068d1c4

• add kwarg for turning cuda on and off; make it slow, finally

  ajwheeler

  committed
  722c702

  Copy full SHA for 722c702

• even more on the GPU

  ajwheeler

  committed
  ceadb07

  Copy full SHA for ceadb07

• still working

  ajwheeler

  committed
  4c4c387

  Copy full SHA for 4c4c387

• compute n/Z on cpu first

  ajwheeler

  committed
  95f8574

  Copy full SHA for 95f8574

Commits on Nov 19, 2024

• simplify

  ajwheeler

  committed
  51aa87c

  Copy full SHA for 51aa87c

• use cntm alpha at the nearest coarse wl value; don't interpolate

  ajwheeler

  committed
  5e39b2f

  Copy full SHA for 5e39b2f

• break loop body into separte function

  ajwheeler

  committed
  5b19a0d

  Copy full SHA for 5b19a0d

• only pass the needed line params

  ajwheeler

  committed
  68a530c

  Copy full SHA for 68a530c

• Writing a kernel

  ajwheeler

  committed
  a1e1bd0

  Copy full SHA for a1e1bd0

• more stuff in the kernel

  ajwheeler

  committed
  e9b030b

  Copy full SHA for e9b030b

• more in kernel

  ajwheeler

  committed
  25aa06c

  Copy full SHA for 25aa06c

• more in the kernel

  ajwheeler

  committed
  3ed7ba7

  Copy full SHA for 3ed7ba7

• use grid (warp?)-stride loop over layers

  ajwheeler

  committed
  22c69bc

  Copy full SHA for 22c69bc

• everything up to the reduction

  ajwheeler

  committed
  61be6da

  Copy full SHA for 61be6da

• for mysterious reaons, searchsortedlast is behaving differently on device

  ajwheeler

  committed
  49f4291

  Copy full SHA for 49f4291

• remove some stuff

  ajwheeler

  committed
  d6207ae

  Copy full SHA for d6207ae

Commits on Nov 20, 2024

• converted Γ from an array to a local scalar

  ajwheeler

  committed
  341c332

  Copy full SHA for 341c332

• fuse the loops within the kernel

  ajwheeler

  committed
  72bade7

  Copy full SHA for 72bade7

• convert levels_factor from an array into a scalar

  ajwheeler

  committed
  a56d010

  Copy full SHA for a56d010

• get mass inside kernel

  ajwheeler

  committed
  7aafac7

  Copy full SHA for 7aafac7

• don't allocate rho_crit as a CuVector, use a scalar

  ajwheeler

  committed
  1803634

  Copy full SHA for 1803634

• introduce n_gpu_block; remove errant index

  ajwheeler

  committed
  862fbff

  Copy full SHA for 862fbff

• do it all in a single kernel

  ajwheeler

  committed
  6ab3cce

  Copy full SHA for 6ab3cce

• CUDA sync

  ajwheeler

  committed
  9b7156d

  Copy full SHA for 9b7156d

• atomic add to alpha

  ajwheeler

  committed
  b5ab8b3

  Copy full SHA for b5ab8b3

• inverse densities as scalars

  ajwheeler

  committed
  a3c299a

  Copy full SHA for a3c299a

• remove @inbounds; fix bounds errors (oops)

  ajwheeler

  committed
  f1f0ff5

  Copy full SHA for f1f0ff5

• make Wavelengths more parametric

  ajwheeler

  committed
  fe0327c

  Copy full SHA for fe0327c

• make Wavelengths more parametric

  ajwheeler

  committed
  4fdc2bf

  Copy full SHA for 4fdc2bf

• get the Wavelengths type on the GPU. Why is it broken?

  ajwheeler

  committed
  8e7cdf0

  Copy full SHA for 8e7cdf0

• fix bug (now workign with fast searchsortedfirst); remove print statements

  ajwheeler

  committed
  656da8d

  Copy full SHA for 656da8d

• use an iterator; delete old code

  ajwheeler

  committed
  181f4f7

  Copy full SHA for 181f4f7

• make Line a bitstype by making vdW always a tuple

  ajwheeler

  committed
  e8cdc16

  Copy full SHA for e8cdc16

• Merge branch 'line_bitstype' into cuda.

  ajwheeler

  committed
  e5dd873

  Copy full SHA for e5dd873

• delete unnecessarily redundant code; make it work with the merged new Line implementation

  ajwheeler

  committed
  6bd810e

  Copy full SHA for 6bd810e