Code

GPU Support#67

Code

Closed

TeachRaccooon wants to merge 117 commits intomainBallisticLA/RandLAPACK:mainfrom GPU_SUPPORT_NEWBallisticLA/RandLAPACK:GPU_SUPPORT_NEWCopy head branch name to clipboard

+4,323-280

Commits

Commits on Sep 9, 2024

• Begin adding GPU support into RandLAPACK

  TeachRaccooon

  committed
  ed82b76

  Copy full SHA for ed82b76

• Porting files from Parth's project

  TeachRaccooon

  committed
  bc1476e

  Copy full SHA for bc1476e

• Making sure GPU functions work

  TeachRaccooon

  committed
  f86a091

  Copy full SHA for f86a091

• CQRRPT GPU test

  TeachRaccooon

  committed
  92502f7

  Copy full SHA for 92502f7

• Update

  TeachRaccooon

  committed
  b658c8e

  Copy full SHA for b658c8e

• Update

  TeachRaccooon

  committed
  1c07f29

  Copy full SHA for 1c07f29

• Quick compilartion error comment-out

  TeachRaccooon

  committed
  5b606fd

  Copy full SHA for 5b606fd

• GPU-based CQRRPT finished

  TeachRaccooon

  committed
  7d87d79

  Copy full SHA for 7d87d79

• CQRRPT GPU benchmark

  TeachRaccooon

  committed
  94f908e

  Copy full SHA for 94f908e

• CQRRPT GPU benchmark

  TeachRaccooon

  committed
  046d71a

  Copy full SHA for 046d71a

• Fixing the RNG template issue that is visible on ISAAC

  TeachRaccooon

  committed
  7b01495

  Copy full SHA for 7b01495

• Fixing the RNG template issue that is visible on ISAAC

  TeachRaccooon

  committed
  f836b06

  Copy full SHA for f836b06

• Save before update

  TeachRaccooon

  committed
  7a14cfb

  Copy full SHA for 7a14cfb

• Attempt at making GPU kernels work

  TeachRaccooon

  committed
  8beb5e9

  Copy full SHA for 8beb5e9

• Update

  TeachRaccooon

  committed
  c47b0ec

  Copy full SHA for c47b0ec

• Update core-linux.yaml

  Show description for e92afdf
  TeachRaccooon

  committed
  e92afdf

  Copy full SHA for e92afdf

• Update core-linux.yaml

  TeachRaccooon

  committed
  9e4cebc

  Copy full SHA for 9e4cebc

• Resolved the issue of cuda kernels being undefined/attempted to be compiled with gcc.

  TeachRaccooon

  committed
  ef2ae3b

  Copy full SHA for ef2ae3b

• Issue not fully fixed.

  TeachRaccooon

  committed
  88d0684

  Copy full SHA for 88d0684

• Issue not fully fixed.

  TeachRaccooon

  committed
  51e0d6b

  Copy full SHA for 51e0d6b

• Added a probably temporary fix

  TeachRaccooon

  committed
  88e3f6a

  Copy full SHA for 88e3f6a

• Issue with col_swap persists.

  TeachRaccooon

  committed
  c11ca90

  Copy full SHA for c11ca90

• Update

  TeachRaccooon

  committed
  6b60f39

  Copy full SHA for 6b60f39

• Rebased from main

  TeachRaccooon

  committed
  778b53d

  Copy full SHA for 778b53d

• Update

  TeachRaccooon

  committed
  8c17cf8

  Copy full SHA for 8c17cf8

• Linking OMP properly

  TeachRaccooon

  committed
  e570ce7

  Copy full SHA for e570ce7

• Linking OMP properly

  TeachRaccooon

  committed
  a76e91b

  Copy full SHA for a76e91b

• Implemented a basic version of column swapping. Still experiencing issues,

  TeachRaccooon

  committed
  ef05d4e

  Copy full SHA for ef05d4e

• The issue is due to our choice of the number of threads per block.

  TeachRaccooon

  committed
  d764504

  Copy full SHA for d764504

• Push before update

  TeachRaccooon

  committed
  3047527

  Copy full SHA for 3047527

• Fixed a couple of issues.

  TeachRaccooon

  committed
  576b846

  Copy full SHA for 576b846

• Column swapping works properly, transposition fails for matrices larger than 640x640

  TeachRaccooon

  committed
  ce778cb

  Copy full SHA for ce778cb

• GPU kernels are finally set up properly

  TeachRaccooon

  committed
  a689662

  Copy full SHA for a689662

• Unknown bug in prev commit. In this one, all GPU tests pass

  TeachRaccooon

  committed
  e34de76

  Copy full SHA for e34de76

• Bug resolved

  TeachRaccooon

  committed
  8898106

  Copy full SHA for 8898106

• Trying to figure out if the issue is in the synch or orhr_col

  TeachRaccooon

  committed
  c3dec77

  Copy full SHA for c3dec77

• Need to fix gpu vector pivoting

  TeachRaccooon

  committed
  c3ee672

  Copy full SHA for c3ee672

• ICQRRP_GPU segfaults at size 36

  TeachRaccooon

  committed
  918d166

  Copy full SHA for 918d166

• It seems that my bug in ICQRRP_GPU is yet another synch issue

  TeachRaccooon

  committed
  6345e99

  Copy full SHA for 6345e99

• ICQRRP_GPU workscd ../../randlibs/RandLAPACK_GPU/!

  TeachRaccooon

  committed
  43cd360

  Copy full SHA for 43cd360

• Cleanup

  TeachRaccooon

  committed
  1ac72a7

  Copy full SHA for 1ac72a7

• Finsihed ICQRRP_GPU

  TeachRaccooon

  committed
  af46bae

  Copy full SHA for af46bae

• Updating col_swap test

  TeachRaccooon

  committed
  1f8ea4b

  Copy full SHA for 1f8ea4b

• Added a new ICQRRP test

  TeachRaccooon

  committed
  6666065

  Copy full SHA for 6666065

• reduce CMake required version and have tests print out more useful info

  

  Riley John Murray

  authored and

  TeachRaccooon

  committed
  2ad8a1f

  Copy full SHA for 2ad8a1f

• Update

  TeachRaccooon

  committed
  39b17e0

  Copy full SHA for 39b17e0

• Added a subvector test

  TeachRaccooon

  committed
  89d0f99

  Copy full SHA for 89d0f99

• better test

  

  Riley John Murray

  authored and

  TeachRaccooon

  committed
  5ffb7b7

  Copy full SHA for 5ffb7b7

• Update

  TeachRaccooon

  committed
  d176c85

  Copy full SHA for d176c85

• Update

  TeachRaccooon

  committed
  3bcae44

  Copy full SHA for 3bcae44

• benchmark added

  TeachRaccooon

  committed
  314847d

  Copy full SHA for 314847d

• Added a test for col_swap on wide matrices

  TeachRaccooon

  committed
  3b86264

  Copy full SHA for 3b86264

• Added a test for col_swap on wide matrices

  TeachRaccooon

  committed
  bacb8b1

  Copy full SHA for bacb8b1

• Fixed col swap yet again

  TeachRaccooon

  committed
  f63d1d7

  Copy full SHA for f63d1d7

• Debugging finished

  TeachRaccooon

  committed
  8d08762

  Copy full SHA for 8d08762

• Reverting the change

  TeachRaccooon

  committed
  b99611e

  Copy full SHA for b99611e

• Update

  TeachRaccooon

  committed
  bbdbec7

  Copy full SHA for bbdbec7

• Update

  TeachRaccooon

  committed
  92495c6

  Copy full SHA for 92495c6

• Update

  TeachRaccooon

  committed
  da697b0

  Copy full SHA for da697b0

• Update

  TeachRaccooon

  committed
  be91ad7

  Copy full SHA for be91ad7

• Benchmark fix; synch statements on

  TeachRaccooon

  committed
  4191f96

  Copy full SHA for 4191f96

• Got rid of synch

  TeachRaccooon

  committed
  ce7dc74

  Copy full SHA for ce7dc74

• Fixed

  TeachRaccooon

  committed
  562521d

  Copy full SHA for 562521d

• Added parametrized tests

  

  Max Melnichenko

  authored and

  TeachRaccooon

  committed
  1b5dda2

  Copy full SHA for 1b5dda2

• Optimizations

  TeachRaccooon

  committed
  9dc457e

  Copy full SHA for 9dc457e

• nvtx

  TeachRaccooon

  committed
  acdaff8

  Copy full SHA for acdaff8

• Runtime breakdown added to ICQRRP_GPU

  TeachRaccooon

  committed
  b1671af

  Copy full SHA for b1671af

• Update

  TeachRaccooon

  committed
  bb0a974

  Copy full SHA for bb0a974

• Update

  TeachRaccooon

  committed
  868f9e9

  Copy full SHA for 868f9e9

• Update

  TeachRaccooon

  committed
  455a048

  Copy full SHA for 455a048

• Update

  TeachRaccooon

  committed
  e7c5bdc

  Copy full SHA for e7c5bdc

• Update

  TeachRaccooon

  committed
  d89f55e

  Copy full SHA for d89f55e

• Update

  TeachRaccooon

  committed
  72bfc4e

  Copy full SHA for 72bfc4e

• Merged pivoting of work(A) and trail(R)

  TeachRaccooon

  committed
  bc79e8f

  Copy full SHA for bc79e8f

• Adding nvtx back

  TeachRaccooon

  committed
  2d121ac

  Copy full SHA for 2d121ac

• add qrf

  TeachRaccooon

  committed
  565e767

  Copy full SHA for 565e767

• Offload col swap on a CPU

  TeachRaccooon

  committed
  b5cf5f8

  Copy full SHA for b5cf5f8

• Update

  TeachRaccooon

  committed
  d6e8889

  Copy full SHA for d6e8889

• 1st attempt at resolving the col_swap bottleneck

  TeachRaccooon

  committed
  9b117ad

  Copy full SHA for 9b117ad

• Better parallelism in the new col_swap

  TeachRaccooon

  committed
  47d64e8

  Copy full SHA for 47d64e8

• Applied the new col_swap to ICQRRP_GPU

  TeachRaccooon

  committed
  6ba198a

  Copy full SHA for 6ba198a

• Updating the runtime breakdown

  TeachRaccooon

  committed
  8b96321

  Copy full SHA for 8b96321

• Updating the runtime breakdown

  TeachRaccooon

  committed
  105ba71

  Copy full SHA for 105ba71

• Update

  TeachRaccooon

  committed
  792ccd0

  Copy full SHA for 792ccd0

• Added an option to use qrf instead of cholqr

  TeachRaccooon

  committed
  8081600

  Copy full SHA for 8081600

• Added an option to use qrf instead of cholqr

  TeachRaccooon

  committed
  6f5a4f2

  Copy full SHA for 6f5a4f2

• Got rid of the Copy(A_sk)

  TeachRaccooon

  committed
  3e5821c

  Copy full SHA for 3e5821c

• Reverting the change (temporary)

  TeachRaccooon

  committed
  be785d9

  Copy full SHA for be785d9

• Update

  TeachRaccooon

  committed
  f4ba5c2

  Copy full SHA for f4ba5c2

• Update

  TeachRaccooon

  committed
  59a9a76

  Copy full SHA for 59a9a76

• Update

  TeachRaccooon

  committed
  f3aea26

  Copy full SHA for f3aea26

• Update

  TeachRaccooon

  committed
  f46f94d

  Copy full SHA for f46f94d

• try

  TeachRaccooon

  committed
  60a22a1

  Copy full SHA for 60a22a1

• Onto something

  TeachRaccooon

  committed
  dad4f02

  Copy full SHA for dad4f02

• Got rid of pivot idx copy

  TeachRaccooon

  committed
  22b3808

  Copy full SHA for 22b3808

• Col_swap_bupdate

  TeachRaccooon

  committed
  1bbb4a7

  Copy full SHA for 1bbb4a7

• Cusolver stream update

  TeachRaccooon

  committed
  f05957c

  Copy full SHA for f05957c

• Switched copy_mat_gpu to a standard one

  TeachRaccooon

  committed
  d6a5c4b

  Copy full SHA for d6a5c4b

• Reworked copy_mat_gpu

  TeachRaccooon

  committed
  f6f2696

  Copy full SHA for f6f2696

• Placing A copy on a separate stream

  TeachRaccooon

  committed
  d3bf1a8

  Copy full SHA for d3bf1a8

• Fixed J copy bug & reverted the col_swap change

  TeachRaccooon

  committed
  f3ae8a0

  Copy full SHA for f3ae8a0

• Synchronization bug fixed

  TeachRaccooon

  committed
  70c0aec

  Copy full SHA for 70c0aec

• Small fix in regards to the col_swap_gpu

  TeachRaccooon

  committed
  7927ef3

  Copy full SHA for 7927ef3

• fixed the issue with copying extra data

  TeachRaccooon

  committed
  fb3fe8a

  Copy full SHA for fb3fe8a

• This commit simply demonstrates that the numerical issues are due to POTRF+SYRK on a GPU not staying strictly in fp64 land. Thee code here is not polished and will be reverted

  TeachRaccooon

  committed
  c93ca35

  Copy full SHA for c93ca35

• Reverting the demonstrative change

  TeachRaccooon

  committed
  222596f

  Copy full SHA for 222596f

• Made sure the low-rank cases are handled properly

  TeachRaccooon

  committed
  ca9290a

  Copy full SHA for ca9290a

• Update

  TeachRaccooon

  committed
  f14ed67

  Copy full SHA for f14ed67

• Some fixes per Riley's comments

  TeachRaccooon

  committed
  4bd5d10

  Copy full SHA for 4bd5d10

• File add

  TeachRaccooon

  committed
  182e498

  Copy full SHA for 182e498

• reduce CMake version dependency. Move GPU benchmark tests into its own executable, so ./bin/RandLAPACK_tests_gpu doesnt dispatch benchmarks by default

  

  rileyjmurray

  authored and

  TeachRaccooon

  committed
  ea52ff1

  Copy full SHA for ea52ff1

• add checks for unnoticed cuda errors before returning from a test

  

  rileyjmurray

  authored and

  TeachRaccooon

  committed
  a2f96e2

  Copy full SHA for a2f96e2

• move error check, and error message

  

  rileyjmurray

  authored and

  TeachRaccooon

  committed
  25e8898

  Copy full SHA for 25e8898

• Fixed the bug with freeing A_sk_cpy space

  TeachRaccooon

  committed
  ab00b1e

  Copy full SHA for ab00b1e

• Nearly all fixes that Riley requested were applied.

  TeachRaccooon

  committed
  185380d

  Copy full SHA for 185380d

• Temporary solution tensure things work without CUDA support

  TeachRaccooon

  committed
  f286f94

  Copy full SHA for f286f94

• Rebased from main. Reverted the CPU fix.

  TeachRaccooon

  committed
  16cdbb8

  Copy full SHA for 16cdbb8