Code

Add support for AMX instructions#5818

Code

Merged

steven-johnson merged 68 commits intohalide:masterhalide/Halide:masterfrom mcleary:tile_matmulmcleary/Halide:tile_matmulCopy head branch name to clipboardOct 21, 2021

+1,186-8

Commits

Commits on Mar 16, 2021

• Add support for AMX tile instructions

  

  jwlawson

  authored and

  Thales Sabino

  committed
  036b037

  Copy full SHA for 036b037

• Make AMX transform opt-in with memory type

  

  jwlawson

  authored and

  Thales Sabino

  committed
  2bd3453

  Copy full SHA for 2bd3453

• Clean up tiled_matmul test

  

  jwlawson

  authored and

  Thales Sabino

  committed
  34e1a4c

  Copy full SHA for 34e1a4c

• Handle AMX intrinsic attributes better

  

  jwlawson

  authored and

  Thales Sabino

  committed
  dfeac55

  Copy full SHA for dfeac55

• Format

  

  jwlawson

  authored and

  Thales Sabino

  committed
  a9f84de

  Copy full SHA for a9f84de

• Fix test to behave like other tests

  

  jwlawson

  authored and

  Thales Sabino

  committed
  da04b0a

  Copy full SHA for da04b0a

• Add doc and missing load check

  

  jwlawson

  authored and

  Thales Sabino

  committed
  9923040

  Copy full SHA for 9923040

• Format

  

  jwlawson

  authored and

  Thales Sabino

  committed
  1a1d10a

  Copy full SHA for 1a1d10a

• Throw error if user requests AMX for invalid operation

  

  jwlawson

  authored and

  Thales Sabino

  committed
  d54fe24

  Copy full SHA for d54fe24

• Add Tile lowering pass to makefile

  

  jwlawson

  authored and

  Thales Sabino

  committed
  228beda

  Copy full SHA for 228beda

• Use spaces in Makefile

  

  jwlawson

  authored and

  Thales Sabino

  committed
  673480a

  Copy full SHA for 673480a

Commits on Mar 17, 2021

• Place AMX instrinsics into a separate module (x86_amx.ll)

  Show description for f91d79f
  

  Thales Sabino

  committed
  f91d79f

  Copy full SHA for f91d79f

• Merge branch 'master' into pr/5818

  steven-johnson

  committed
  9fc4366

  Copy full SHA for 9fc4366

• Fix CreateAlignedLoad() call in CodeGen_X86

  Show description for 16a1c7b
  steven-johnson

  committed
  16a1c7b

  Copy full SHA for 16a1c7b

Commits on Mar 22, 2021

• Merge branch 'master' into tile_matmul

  

  Thales Sabino

  committed
  b3b1dfa

  Copy full SHA for b3b1dfa

Commits on Mar 24, 2021

• Merge branch 'master' into tile_matmul

  

  Thales Sabino

  committed
  03f894a

  Copy full SHA for 03f894a

Commits on Mar 31, 2021

• Merge branch 'master' into tile_matmul

  

  Thales Sabino

  committed
  97890dd

  Copy full SHA for 97890dd

Commits on Apr 1, 2021

• Merge branch 'master' into tile_matmul

  

  Thales Sabino

  committed
  293ca90

  Copy full SHA for 293ca90

Commits on Apr 7, 2021

• Merge branch 'master' into tile_matmul

  

  Thales Sabino

  committed
  4b950e5

  Copy full SHA for 4b950e5

Commits on Apr 20, 2021

• fix exporting to module

  frengels

  committed
  d0e5123

  Copy full SHA for d0e5123

• add llvm funcs for su, us, uu amx variants

  frengels

  committed
  9189120

  Copy full SHA for 9189120

• add other amx intrinsics to intrinsic_defs

  frengels

  committed
  75c4262

  Copy full SHA for 75c4262

• match with unsigned 8 bit integers

  Show description for 09f7551
  frengels

  committed
  09f7551

  Copy full SHA for 09f7551

Commits on Apr 27, 2021

• match for 32 bit integer and guard unsigned amx on llvm 13

  frengels

  committed
  5dd1471

  Copy full SHA for 5dd1471

Commits on Apr 29, 2021

• adjust test to cover unsigned tile operations

  frengels

  committed
  7e45c29

  Copy full SHA for 7e45c29

• guard properly with llvm 12

  frengels

  committed
  4ab681b

  Copy full SHA for 4ab681b

• create explicit error if failed to use tile operations

  frengels

  committed
  6339ae7

  Copy full SHA for 6339ae7

• pass types as template params rather than boolean

  Show description for 525e11e
  frengels

  committed
  525e11e

  Copy full SHA for 525e11e

Commits on Apr 30, 2021

• clang-format patch

  

  Thales Sabino

  authored and

  frengels

  committed
  5a21484

  Copy full SHA for 5a21484

• add x86_amx to makefile's runtime components

  frengels

  committed
  7950614

  Copy full SHA for 7950614

• make tiled_matmul compatible with c++11

  frengels

  committed
  4a6c10c

  Copy full SHA for 4a6c10c

• Merge remote-tracking branch 'upstream/master' into tile_matmul

  frengels

  committed
  02c375a

  Copy full SHA for 02c375a

Commits on May 3, 2021

• add mattrs required for amx

  frengels

  committed
  f985644

  Copy full SHA for f985644

• Merge pull request #3 from frengels/tile_matmul

  Show description for 7df3d5e
  mcleary

  authored
  7df3d5e

  Copy full SHA for 7df3d5e

Commits on May 5, 2021

• fix formatting issues

  frengels

  committed
  e3f1ef6

  Copy full SHA for e3f1ef6

Commits on May 10, 2021

• remove outdated FIXME comments

  frengels

  committed
  57b6080

  Copy full SHA for 57b6080

• Merge pull request #4 from frengels/tile_matmul

  Show description for ef9d544
  mcleary

  authored
  ef9d544

  Copy full SHA for ef9d544

Commits on May 11, 2021

• Merge branch 'master' into tile_matmul

  

  Thales Sabino

  committed
  9c6bfbc

  Copy full SHA for 9c6bfbc

Commits on May 13, 2021

• Merge branch 'master' into tile_matmul

  

  Thales Sabino

  committed
  b5c46d2

  Copy full SHA for b5c46d2

Commits on May 24, 2021

• Merge remote-tracking branch 'upstream/master' into tile_matmul

  frengels

  committed
  5af3dae

  Copy full SHA for 5af3dae

• add bf16 tile operations to the runtime

  frengels

  committed
  55098f3

  Copy full SHA for 55098f3

• create a schedule that should map to amx

  frengels

  committed
  9f078f0

  Copy full SHA for 9f078f0

• create full amx-bf16 schedule

  frengels

  committed
  e73702d

  Copy full SHA for e73702d

• allow amx operations to yield f32s

  frengels

  committed
  16217b4

  Copy full SHA for 16217b4

• accept 32 bit float stores

  frengels

  committed
  c722610

  Copy full SHA for c722610

• add support for bf16

  frengels

  committed
  66885f1

  Copy full SHA for 66885f1

• add missing bf16 intrinsics

  frengels

  committed
  97fb022

  Copy full SHA for 97fb022

• fix striding error when loading matrix

  frengels

  committed
  f6ba739

  Copy full SHA for f6ba739

• add checks to verify bf16 result

  frengels

  committed
  c7278b6

  Copy full SHA for c7278b6

• fix scaling of col_bytes on matmul call

  frengels

  committed
  5e81a72

  Copy full SHA for 5e81a72

Commits on Jul 8, 2021

• move brace to previous line

  frengels

  committed
  ea74fe2

  Copy full SHA for ea74fe2

Commits on Jul 9, 2021

• derive result type using a function rather than lambda

  frengels

  committed
  a854dc9

  Copy full SHA for a854dc9

• Merge remote-tracking branch 'upstream/master' into tile_matmul_bf16

  frengels

  committed
  d820737

  Copy full SHA for d820737

• run clang tidy and format

  frengels

  committed
  26014d2

  Copy full SHA for 26014d2

Commits on Sep 2, 2021

• have tile_store return i32

  frengels

  committed
  34557cb

  Copy full SHA for 34557cb

Commits on Sep 8, 2021

• make is_3d_tile_index robust to indexing changes

  frengels

  committed
  5ad06e0

  Copy full SHA for 5ad06e0

• Merge remote-tracking branch 'upstream/master' into tile_matmul_bf16

  frengels

  committed
  f0f9f3e

  Copy full SHA for f0f9f3e

• apply formatting suggestions

  frengels

  committed
  7cab155

  Copy full SHA for 7cab155

Commits on Sep 9, 2021

• both first and second can be const qualified

  frengels

  committed
  8f83544

  Copy full SHA for 8f83544

• remove trailing whitespace in unformatted section

  frengels

  committed
  b1e1452

  Copy full SHA for b1e1452

Commits on Sep 17, 2021

• Merge branch 'master' into pr/5818

  steven-johnson

  committed
  95d38f0

  Copy full SHA for 95d38f0

Commits on Sep 21, 2021

• make requested style changes

  frengels

  committed
  14df0bc

  Copy full SHA for 14df0bc

Commits on Oct 1, 2021

• rename NewMatmul -> Matmul

  frengels

  committed
  8b63d77

  Copy full SHA for 8b63d77

• fix warning about missing return value

  frengels

  committed
  6a5eeaa

  Copy full SHA for 6a5eeaa

• use get_1d_tile_index to handle special case

  Show description for cc7c97d
  frengels

  committed
  cc7c97d

  Copy full SHA for cc7c97d

Commits on Oct 5, 2021

• add correctness test for AMX instructions

  frengels

  committed
  014f0c6

  Copy full SHA for 014f0c6

• correctness part has been separated out

  frengels

  committed
  655dbdf

  Copy full SHA for 655dbdf

Commits on Oct 6, 2021

• remove unused variables

  frengels

  committed
  abc660b

  Copy full SHA for abc660b