Code

Half precision support#1257

Code

Closed

yhmtsai wants to merge 62 commits intodevelopginkgo-project/ginkgo:developfrom halfginkgo-project/ginkgo:halfCopy head branch name to clipboard

+3,862-1,466

Commits

Commits on Oct 22, 2024

• only can compile cuda/omp

  yhmtsai

  committed
  4ed36a8

  Copy full SHA for 4ed36a8

• next_precision to itself when complex only float, double add empty conditional

  yhmtsai

  committed
  ddc7c16

  Copy full SHA for ddc7c16

• can compile with cuda/omp/ref (without test)

  yhmtsai

  committed
  501e6c7

  Copy full SHA for 501e6c7

• compile for cuda/sycl/test/mpi (hip needs trick)

  yhmtsai

  committed
  957d29c

  Copy full SHA for 957d29c

• hip finally

  yhmtsai

  committed
  a0c389c

  Copy full SHA for a0c389c

• fix the narrow issue and atomic support

  yhmtsai

  committed
  fe5e491

  Copy full SHA for fe5e491

• fixed more error

  yhmtsai

  committed
  6c3c12b

  Copy full SHA for 6c3c12b

• fix the op order and gdb

  Show description for a0ee872
  

  yhmtsai

  and

  MarcelKoch

  committed
  a0ee872

  Copy full SHA for a0ee872

• add the rand template not_implemented

  yhmtsai

  committed
  5acbf27

  Copy full SHA for 5acbf27

• this version can compile/run complex<half> on cuda114

  Show description for b171312
  yhmtsai

  committed
  b171312

  Copy full SHA for b171312

• does not work for the other executor

  yhmtsai

  committed
  6c17701

  Copy full SHA for 6c17701

• fix complex issue and sqrt issue

  Show description for cdbf0a0
  yhmtsai

  committed
  cdbf0a0

  Copy full SHA for cdbf0a0

• try fix the compilation issue from MSVC and MacOS

  Show description for 209c799
  yhmtsai

  committed
  209c799

  Copy full SHA for 209c799

• move the half to public and use sycl::half for dpcpp

  yhmtsai

  committed
  75b54fa

  Copy full SHA for 75b54fa

• limit the next precision in test and benchmark

  Show description for 48ea338
  yhmtsai

  committed
  48ea338

  Copy full SHA for 48ea338

• allow disable half operation

  yhmtsai

  committed
  fdcc066

  Copy full SHA for fdcc066

• fix macro

  yhmtsai

  committed
  f0a8a07

  Copy full SHA for f0a8a07

• clean and refine the code

  yhmtsai

  committed
  f041b4a

  Copy full SHA for f041b4a

• move half.hpp out of type.hpp

  yhmtsai

  committed
  3154a04

  Copy full SHA for 3154a04

• enable half for testing

  yhmtsai

  committed
  58784ab

  Copy full SHA for 58784ab

• __habs is added in cuda10.2

  Show description for 9b2465b
  yhmtsai

  committed
  9b2465b

  Copy full SHA for 9b2465b

• fix nullptr and missing instantiation.

  Show description for e2a6c9a
  yhmtsai

  committed
  e2a6c9a

  Copy full SHA for e2a6c9a

• fix missing device_type and ptr_param

  yhmtsai

  committed
  51cf597

  Copy full SHA for 51cf597

• update rounding

  yhmtsai

  committed
  0a42796

  Copy full SHA for 0a42796

Commits on Oct 23, 2024

• do not use distribution with half

  yhmtsai

  committed
  e3b81df

  Copy full SHA for e3b81df

• WIP fix half of failed test

  yhmtsai

  committed
  c9fd747

  Copy full SHA for c9fd747

• fix/skip half test and fix numeric_limit on device

  yhmtsai

  committed
  684cadb

  Copy full SHA for 684cadb

• mkl csr does not support half

  yhmtsai

  committed
  60767ed

  Copy full SHA for 60767ed

• add half to batch_vector

  yhmtsai

  committed
  d65255a

  Copy full SHA for d65255a

• fix hip thrust complex op, avoid const in nvhpc, reduce job in windows

  yhmtsai

  committed
  5c0454f

  Copy full SHA for 5c0454f

• fix nvc++ atomic, dpcpp half

  Show description for da15916
  yhmtsai

  committed
  da15916

  Copy full SHA for da15916

• make half test optional

  yhmtsai

  committed
  5f9e3ff

  Copy full SHA for 5f9e3ff

• nvhpc optimization/computation error workaround

  yhmtsai

  committed
  fe45560

  Copy full SHA for fe45560

• some math func is not defined if nvhpc is for host

  yhmtsai

  committed
  c7f0d2a

  Copy full SHA for c7f0d2a

• add half spmv benchmark (with cusparse for cuda)

  yhmtsai

  committed
  710e037

  Copy full SHA for 710e037

• fixes batched support for half

  

  MarcelKoch

  authored and

  yhmtsai

  committed
  34845f3

  Copy full SHA for 34845f3

• generate PTX load/stores for half

  

  MarcelKoch

  authored and

  yhmtsai

  committed
  48afbb5

  Copy full SHA for 48afbb5

• fix mc64 for half

  Show description for a51f136
  

  MarcelKoch

  authored and

  yhmtsai

  committed
  a51f136

  Copy full SHA for a51f136

• fix hip memory.hip.hpp for half

  

  MarcelKoch

  authored and

  yhmtsai

  committed
  60123dc

  Copy full SHA for 60123dc

• WIP: can compile but three tests are still failed

  yhmtsai

  committed
  8f1e28f

  Copy full SHA for 8f1e28f

• fix config, ambiguous namespace, and batch

  yhmtsai

  committed
  6dbd616

  Copy full SHA for 6dbd616

• update format

  yhmtsai

  committed
  cd270e1

  Copy full SHA for cd270e1

• check the failed tests

  Show description for 69d5b59
  

  yhmtsai

  and

  MarcelKoch

  committed
  69d5b59

  Copy full SHA for 69d5b59

• fix windows and icpx

  yhmtsai

  committed
  57fc170

  Copy full SHA for 57fc170

• hip does not support atomic on 16 bits

  yhmtsai

  committed
  18e825f

  Copy full SHA for 18e825f

• fix batch

  yhmtsai

  committed
  825f76f

  Copy full SHA for 825f76f

• add miss instantiation

  yhmtsai

  committed
  81d63ac

  Copy full SHA for 81d63ac

• update documentation, remove half.hpp

  Show description for 2a6d382
  

  3 peoplecommitted
  2a6d382

  Copy full SHA for 2a6d382

• put function in gko not std

  Show description for 8731fc3
  

  yhmtsai

  and

  Thomas Grützmacher

  committed
  8731fc3

  Copy full SHA for 8731fc3

• fix after rebase

  yhmtsai

  committed
  64406f3

  Copy full SHA for 64406f3

• hip does not support 16bit shuffle

  yhmtsai

  committed
  baa95f7

  Copy full SHA for baa95f7

• merge two #if block

  yhmtsai

  committed
  4bb8093

  Copy full SHA for 4bb8093

• do not use attributes in sqrt and abs

  yhmtsai

  committed
  c539398

  Copy full SHA for c539398

• make half constexpr

  yhmtsai

  committed
  d0e2446

  Copy full SHA for d0e2446

• isolate half out of device completely

  yhmtsai

  committed
  0d777df

  Copy full SHA for 0d777df

• bits constexpr construct half and make numeric_limit in half

  yhmtsai

  committed
  56e2af8

  Copy full SHA for 56e2af8

• refine the code and fix error without half

  yhmtsai

  committed
  6cc26d7

  Copy full SHA for 6cc26d7

• reduce abs/sqrt location

  yhmtsai

  committed
  3d15350

  Copy full SHA for 3d15350

• move the math function to math

  yhmtsai

  committed
  c4697a5

  Copy full SHA for c4697a5

• nohalf

  yhmtsai

  committed
  377432a

  Copy full SHA for 377432a

• cbgmres without half

  yhmtsai

  committed
  e806a0a

  Copy full SHA for e806a0a

• direct without half

  yhmtsai

  committed
  3e49252

  Copy full SHA for 3e49252