Code

Add FP8 rocblas gemm support#2473

Code

Merged

causten merged 97 commits intodevelopROCm/AMDMIGraphX:developfrom rocblas_fp8ROCm/AMDMIGraphX:rocblas_fp8Copy head branch name to clipboardDec 6, 2023

+490-312

Commits

Commits on Nov 9, 2023

• changes for the FP8 ref implementation

  umangyadav

  committed
  df7f8a3

  Copy full SHA for df7f8a3

• cppcheck fixes

  umangyadav

  committed
  9bc1828

  Copy full SHA for 9bc1828

Commits on Nov 10, 2023

• move FNUZ as template parameter

  umangyadav

  committed
  155a2b1

  Copy full SHA for 155a2b1

• Fix numeric limits

  umangyadav

  committed
  d9f11e3

  Copy full SHA for d9f11e3

• Working FNUZ and FN

  umangyadav

  committed
  4e9d51f

  Copy full SHA for 4e9d51f

• use float equal

  umangyadav

  committed
  7639c28

  Copy full SHA for 7639c28

• add test for fp8e5m2

  umangyadav

  committed
  a6372c5

  Copy full SHA for a6372c5

• add test for fp8e5m2fnuz

  umangyadav

  committed
  439ea40

  Copy full SHA for 439ea40

• refactor add some comments

  umangyadav

  committed
  183db78

  Copy full SHA for 183db78

Commits on Nov 13, 2023

• Review updates

  umangyadav

  committed
  ab653af

  Copy full SHA for ab653af

Commits on Nov 14, 2023

• Fix tidy

  umangyadav

  committed
  8319e01

  Copy full SHA for 8319e01

• Fix test failure

  umangyadav

  committed
  9ee0418

  Copy full SHA for 9ee0418

• fix isfinite

  umangyadav

  committed
  355e4f6

  Copy full SHA for 355e4f6

• Merge remote-tracking branch 'origin/develop' into ref_fp8

  umangyadav

  committed
  ba471f4

  Copy full SHA for ba471f4

• fix test for neg inf

  umangyadav

  committed
  6aec703

  Copy full SHA for 6aec703

• fix warning

  umangyadav

  committed
  12aac37

  Copy full SHA for 12aac37

• add tests

  umangyadav

  committed
  6009232

  Copy full SHA for 6009232

• Fix tests

  umangyadav

  committed
  03f7139

  Copy full SHA for 03f7139

• add stringstream tests

  umangyadav

  committed
  1e220c0

  Copy full SHA for 1e220c0

Commits on Nov 15, 2023

• Remove clang diagnostics

  umangyadav

  committed
  a83e9dc

  Copy full SHA for a83e9dc

• Merge remote-tracking branch 'origin/develop' into ref_fp8

  umangyadav

  committed
  dfb35a6

  Copy full SHA for dfb35a6

• Remove NOLINTS

  umangyadav

  committed
  26956f1

  Copy full SHA for 26956f1

Commits on Nov 16, 2023

• Bugfixes and additional tests

  umangyadav

  committed
  269ce6d

  Copy full SHA for 269ce6d

• Fix undoing

  umangyadav

  committed
  6414ee3

  Copy full SHA for 6414ee3

• Handle underflow case separately to avoid sanitization errors

  umangyadav

  committed
  cd26ada

  Copy full SHA for cd26ada

• use std::min to avoid sanitization errors

  umangyadav

  committed
  1cf87ef

  Copy full SHA for 1cf87ef

• Merge branch 'develop' into ref_fp8

  umangyadav

  authored
  e7e5ba2

  Copy full SHA for e7e5ba2

• formatting

  umangyadav

  committed
  98a838f

  Copy full SHA for 98a838f

• use 31 for min value

  umangyadav

  committed
  61e4e1d

  Copy full SHA for 61e4e1d

• add note

  umangyadav

  committed
  a5c38eb

  Copy full SHA for a5c38eb

• Merge branch 'ref_fp8' of github.com:ROCmSoftwarePlatform/AMDMIGraphX into ref_fp8

  umangyadav

  committed
  61775ea

  Copy full SHA for 61775ea

• Merge branch 'develop' into ref_fp8

  umangyadav

  authored
  3806427

  Copy full SHA for 3806427

Commits on Nov 17, 2023

• add some more comments

  umangyadav

  committed
  017d67e

  Copy full SHA for 017d67e

• Merge branch 'ref_fp8' of github.com:ROCmSoftwarePlatform/AMDMIGraphX into ref_fp8

  umangyadav

  committed
  9e6d866

  Copy full SHA for 9e6d866

• port gpu changes

  umangyadav

  committed
  a9dd42f

  Copy full SHA for a9dd42f

• use bit cast

  umangyadav

  committed
  d7339e8

  Copy full SHA for d7339e8

• Make FNUZ template param and add numeric limits

  umangyadav

  committed
  6094234

  Copy full SHA for 6094234

• only compile for device

  umangyadav

  committed
  78ec77e

  Copy full SHA for 78ec77e

• remove non-JIT related code

  umangyadav

  committed
  3411649

  Copy full SHA for 3411649

• Remove FP8_Lowest/Max

  umangyadav

  committed
  d2c25a0

  Copy full SHA for d2c25a0

• remove using for dtypes

  umangyadav

  committed
  5da68df

  Copy full SHA for 5da68df

• Update float8_impl

  umangyadav

  committed
  b36f72d

  Copy full SHA for b36f72d

• constructor from float works with constexpr

  umangyadav

  committed
  85ba819

  Copy full SHA for 85ba819

• Remove unnecessary pragmas

  umangyadav

  committed
  aed1922

  Copy full SHA for aed1922

• Remove clang diagnostics

  umangyadav

  committed
  f975c63

  Copy full SHA for f975c63

• Add back floatequal

  umangyadav

  committed
  32033d8

  Copy full SHA for 32033d8

• disable DPP For FP8

  umangyadav

  committed
  e88d46a

  Copy full SHA for e88d46a

• Merge remote-tracking branch 'origin/develop' into gpu_fp8

  umangyadav

  committed
  3ae93ca

  Copy full SHA for 3ae93ca

• formatting

  umangyadav

  committed
  60dd1f4

  Copy full SHA for 60dd1f4

• revert unwanted changes

  umangyadav

  committed
  ef425d0

  Copy full SHA for ef425d0

• Merge branch 'gpu_fp8' of https://github.com/ROCmSoftwarePlatform/AMDMIGraphX into gpu_fp8

  umangyadav

  committed
  76f0318

  Copy full SHA for 76f0318

• add some more tests

  umangyadav

  committed
  bd0ae5f

  Copy full SHA for bd0ae5f

Commits on Nov 18, 2023

• Add math and reduce tests

  umangyadav

  committed
  91cc9c7

  Copy full SHA for 91cc9c7

• Fix tidy and other errors

  umangyadav

  committed
  e2b0c40

  Copy full SHA for e2b0c40

• fixes

  umangyadav

  committed
  9f50051

  Copy full SHA for 9f50051

• add nolint

  umangyadav

  committed
  249464c

  Copy full SHA for 249464c

• tidy fix

  umangyadav

  committed
  1be9587

  Copy full SHA for 1be9587

Commits on Nov 20, 2023

• roialign, softmax, pow, acosh, atanh,pad tests are enabled now

  umangyadav

  committed
  13403ab

  Copy full SHA for 13403ab

• add layernorm, remove constexpr for 1/r

  umangyadav

  committed
  f550f81

  Copy full SHA for f550f81

• tidy fixes

  umangyadav

  committed
  7e3444c

  Copy full SHA for 7e3444c

• use __builtin_is_constant_evaluated

  umangyadav

  committed
  6155c78

  Copy full SHA for 6155c78

• add test for rsqrt and remove old-styple-cast

  umangyadav

  committed
  13ef414

  Copy full SHA for 13ef414

• add comment about c++20 extensions

  umangyadav

  committed
  8660572

  Copy full SHA for 8660572

• Remove old cast

  umangyadav

  committed
  6fbd997

  Copy full SHA for 6fbd997

• Remove DPP

  umangyadav

  committed
  2acd265

  Copy full SHA for 2acd265

• Remove MIN max overloads

  umangyadav

  committed
  836e201

  Copy full SHA for 836e201

• Put numeric_max and numeeric lowest into float8

  umangyadav

  committed
  f9542d5

  Copy full SHA for f9542d5

Commits on Nov 21, 2023

• use void for highest to match template candidates

  umangyadav

  committed
  480288f

  Copy full SHA for 480288f

• add float8 for tensorview

  umangyadav

  committed
  a6c5772

  Copy full SHA for a6c5772

Commits on Nov 26, 2023

• compiles all right

  umangyadav

  committed
  3aa465f

  Copy full SHA for 3aa465f

• Works now

  umangyadav

  committed
  037205c

  Copy full SHA for 037205c

• add ifdef to compile

  umangyadav

  committed
  87548b5

  Copy full SHA for 87548b5

• add tests and fix cmake

  umangyadav

  committed
  d473b80

  Copy full SHA for d473b80

• add tests

  umangyadav

  committed
  4604f2e

  Copy full SHA for 4604f2e

• add eliminate_fp8 pass

  umangyadav

  committed
  ad9c25e

  Copy full SHA for ad9c25e

• remove convert from lowering

  umangyadav

  committed
  8734ffa

  Copy full SHA for 8734ffa

• Fix eliminate_fp8 pass

  umangyadav

  committed
  f014fb9

  Copy full SHA for f014fb9

• Move pass before optimize module

  umangyadav

  committed
  83ce487

  Copy full SHA for 83ce487

• formatting

  umangyadav

  committed
  9a9e964

  Copy full SHA for 9a9e964

• fix cppcheck

  umangyadav

  committed
  c40a39c

  Copy full SHA for c40a39c

Commits on Dec 1, 2023

• Merge branch 'develop' into rocblas_fp8

  umangyadav

  authored
  c4cee34

  Copy full SHA for c4cee34

• merge changes

  umangyadav

  committed
  f155b0e

  Copy full SHA for f155b0e

• few changes

  umangyadav

  committed
  38218ed

  Copy full SHA for 38218ed

• few more cosmetic changes

  umangyadav

  committed
  379692f

  Copy full SHA for 379692f

Commits on Dec 2, 2023

• add half tests

  umangyadav

  committed
  381b2d9

  Copy full SHA for 381b2d9

Commits on Dec 4, 2023

• use updated eliminate_fp8 pass

  umangyadav

  committed
  5423577

  Copy full SHA for 5423577

Commits on Dec 5, 2023

• use eliminate_data_type pass instead of eliminate_fp8 pass

  umangyadav

  committed
  402c66a

  Copy full SHA for 402c66a

• Merge branch 'develop' into rocblas_fp8

  umangyadav

  authored
  8738f3b

  Copy full SHA for 8738f3b

• remove older files

  umangyadav

  committed
  4ca90ec

  Copy full SHA for 4ca90ec

• remove header

  umangyadav

  committed
  b099a7d

  Copy full SHA for b099a7d

• fix typo

  umangyadav

  committed
  7d6e6ad

  Copy full SHA for 7d6e6ad

• add changes for the eliminate_data_type pass

  umangyadav

  committed
  cf91c2b

  Copy full SHA for cf91c2b

• add comments

  umangyadav

  committed
  82f9847

  Copy full SHA for 82f9847

• fix typo

  umangyadav

  committed
  a9db2bf

  Copy full SHA for a9db2bf

• remove else

  umangyadav

  committed
  aeaac20

  Copy full SHA for aeaac20

• disable tests that uses CK

  umangyadav

  committed
  a196e90

  Copy full SHA for a196e90

• formatting

  umangyadav

  committed
  7e80f62

  Copy full SHA for 7e80f62