Code

Bfloat16 #1709

Code

Draft

yhmtsai wants to merge 48 commits intodevelopginkgo-project/ginkgo:developfrom bfloat16ginkgo-project/ginkgo:bfloat16Copy head branch name to clipboard

+5,356-1,239

Commits

Commits on Sep 29, 2023

• only can compile cuda/omp

  yhmtsai

  committed
  c0f2d9a

  Copy full SHA for c0f2d9a

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

  yhmtsai

  committed
  ac1dbb2

  Copy full SHA for ac1dbb2

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

  yhmtsai

  committed
  7bffc78

  Copy full SHA for 7bffc78

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

  yhmtsai

  committed
  c517644

  Copy full SHA for c517644

• hip finally

  yhmtsai

  committed
  6b5a4ec

  Copy full SHA for 6b5a4ec

• fix the narrow issue and atomic support

  yhmtsai

  committed
  add37ab

  Copy full SHA for add37ab

• fixed more error

  yhmtsai

  committed
  2f53fce

  Copy full SHA for 2f53fce

• fix the op order and gdb

  Show description for cdc4d6b
  

  yhmtsai

  and

  MarcelKoch

  committed
  cdc4d6b

  Copy full SHA for cdc4d6b

• add the rand template not_implemented

  yhmtsai

  committed
  c1c1551

  Copy full SHA for c1c1551

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

  Show description for 3f11657
  yhmtsai

  committed
  3f11657

  Copy full SHA for 3f11657

• does not work for the other executor

  yhmtsai

  committed
  6b9546b

  Copy full SHA for 6b9546b

• fix complex issue and sqrt issue

  Show description for 384371b
  yhmtsai

  committed
  384371b

  Copy full SHA for 384371b

• try fix the compilation issue from MSVC and MacOS

  Show description for c04d7d0
  yhmtsai

  committed
  c04d7d0

  Copy full SHA for c04d7d0

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

  yhmtsai

  committed
  efb9aea

  Copy full SHA for efb9aea

• limit the next precision in test and benchmark

  Show description for 9480b50
  yhmtsai

  committed
  9480b50

  Copy full SHA for 9480b50

• allow disable half operation

  yhmtsai

  committed
  9625e11

  Copy full SHA for 9625e11

• fix macro

  yhmtsai

  committed
  620aef5

  Copy full SHA for 620aef5

• clean and refine the code

  yhmtsai

  committed
  06a6f46

  Copy full SHA for 06a6f46

• move half.hpp out of type.hpp

  yhmtsai

  committed
  814dca4

  Copy full SHA for 814dca4

• enable half for testing

  yhmtsai

  committed
  54755b4

  Copy full SHA for 54755b4

• __habs is added in cuda10.2

  Show description for cc3a7d5
  yhmtsai

  committed
  cc3a7d5

  Copy full SHA for cc3a7d5

• fix nullptr and missing instantiation.

  Show description for 7af58c9
  yhmtsai

  committed
  7af58c9

  Copy full SHA for 7af58c9

• fix missing device_type and ptr_param

  yhmtsai

  committed
  35a47fa

  Copy full SHA for 35a47fa

• update rounding

  yhmtsai

  committed
  faea48d

  Copy full SHA for faea48d

• do not use distribution with half

  yhmtsai

  committed
  f1c1181

  Copy full SHA for f1c1181

• WIP fix half of failed test

  yhmtsai

  committed
  d450dc5

  Copy full SHA for d450dc5

• fix/skip half test and fix numeric_limit on device

  yhmtsai

  committed
  e83d53d

  Copy full SHA for e83d53d

• mkl csr does not support half

  yhmtsai

  committed
  2d7f4cb

  Copy full SHA for 2d7f4cb

• add half to batch_vector

  yhmtsai

  committed
  b04c993

  Copy full SHA for b04c993

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

  yhmtsai

  committed
  10bb4ae

  Copy full SHA for 10bb4ae

• fix nvc++ atomic, dpcpp half

  Show description for bf25352
  yhmtsai

  committed
  bf25352

  Copy full SHA for bf25352

• make half test optional

  yhmtsai

  committed
  cf9c218

  Copy full SHA for cf9c218

• nvhpc optimization/computation error workaround

  yhmtsai

  committed
  1e4b68b

  Copy full SHA for 1e4b68b

• disable mpi half test

  yhmtsai

  committed
  c6bbf20

  Copy full SHA for c6bbf20

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

  yhmtsai

  committed
  18d323e

  Copy full SHA for 18d323e

Commits on Sep 30, 2023

• add half spmv benchmark (with cusparse for cuda)

  yhmtsai

  committed
  8b27e3c

  Copy full SHA for 8b27e3c

Commits on Oct 1, 2023

• add bfloat16

  yhmtsai

  committed
  4eb3b53

  Copy full SHA for 4eb3b53

• this can be compiled after cuda arch 80

  yhmtsai

  committed
  f3913a2

  Copy full SHA for f3913a2

• some trick for cuda arch < 8.0

  yhmtsai

  committed
  d87dc17

  Copy full SHA for d87dc17

• fix the missing type conversion

  yhmtsai

  committed
  5f48681

  Copy full SHA for 5f48681

• enable bfloat16 test and fix/skip

  yhmtsai

  committed
  316775f

  Copy full SHA for 316775f

Commits on Oct 2, 2023

• fix missing type, use bfloat16 op from cudaarch80

  yhmtsai

  committed
  4f0d12a

  Copy full SHA for 4f0d12a

• fix hip

  Show description for 8364204
  yhmtsai

  committed
  8364204

  Copy full SHA for 8364204

• fix dpcpp

  yhmtsai

  committed
  95ac5ea

  Copy full SHA for 95ac5ea

• fix nvhpc

  yhmtsai

  committed
  653cd36

  Copy full SHA for 653cd36

Commits on Oct 3, 2023

• add the casting

  yhmtsai

  committed
  50cf5b2

  Copy full SHA for 50cf5b2

Commits on Oct 4, 2023

• use float as the bridge between bfloat16 and half

  yhmtsai

  committed
  51ab0b0

  Copy full SHA for 51ab0b0

• fix ell accessor type

  yhmtsai

  committed
  b531ba3

  Copy full SHA for b531ba3