Skip to content

x86: Implement masked vector operations - #6687

Merged
0xdaryl merged 11 commits into
eclipse-omr:masterfrom
BradleyWood:maskopi
Jan 13, 2023
Merged

0xdaryl merged 11 commits into
eclipse-omr:masterfrom
BradleyWood:maskopi

Conversation

@BradleyWood

@BradleyWood BradleyWood commented Sep 2, 2022 •

Copy link
Copy Markdown
Contributor

Goals

  • Generic implementation of binary vector operations with merge masking
  • Support different architectures with/without native masking (AVX-512)
  • Support 128, 256, 512-bit vector lengths

TRIL Example (AVX-2)

Add two 128-bit byte vectors according to a merge mask.

(method return= NoType args=[Address,Address,Address,Address]
  (block                                                        
     (vstoreiVector128Int8 offset=0                             
         (aload parm=0)                                         
            (vmaddVector128Int8                                 
                 (vloadiVector128Int8 (aload parm=1))           
                 (vloadiVector128Int8 (aload parm=2))           
                 (mloadiVector128Int8 (aload parm=3))))         
     (return)))                                                 

x86 assembly

mov     rax,  QWORD PTR [rsp+0x18]
mov     rcx,  QWORD PTR [rsp+0x10]
vmovdqu xmm3, XMMWORD PTR [rcx]        ; vloadiVector128Int8 (aload parm=1) lhs
mov     rcx,  QWORD PTR [rsp+0x8]
vmovdqu xmm1, XMMWORD PTR [rcx]        ; vloadiVector128Int8 (aload parm=2) rhs
mov     rcx,  QWORD PTR [rsp]
vmovdqu xmm2, XMMWORD PTR [rcx]        ; mloadiVector128Int8 (aload parm=3) mask
vmovdqu xmm0, xmm3                     ; Copy lhs operand into result
vpaddb  xmm1, xmm3, xmm1               ; 8-bit add lhs + rhs
vpand   xmm1, xmm1, xmm2               ; 'and' away bits excluded from this operation
vpor    xmm0, xmm0, xmm2               ; 'or' mask bits into result register
vpxor   xmm0, xmm0, xmm2               ; 'zero' out bits we want to overwrite
vpor    xmm0, xmm0, xmm1               ; merge the results
vmovdqu XMMWORD PTR [rax], xmm0        ; vstoreiVector128Int8 (aload parm=1) result
add     rsp, 0x20
ret

AVX-2 Example Trace Log

TRIL Example (AVX-512)

Add two 512-bit int vectors according to a merge mask.

(method return= NoType args=[Address,Address,Address,Address]
  (block                                                        
     (vstoreiVector512Int32 offset=0                             
         (aload parm=0)                                         
            (vmaddVector512Int32                                 
                 (vloadiVector512Int32 (aload parm=1))           
                 (vloadiVector512Int32 (aload parm=2))           
                 (mloadiVector512Int32 (aload parm=3))))         
     (return)))                                                 
mov       rax,  QWORD PTR [rsp+0x18]
mov       rcx,  QWORD PTR [rsp+0x10]
vmovdqu32 zmm2, ZMMWORD PTR [rcx]
mov       rcx,  QWORD PTR [rsp+0x8]
vmovdqu32 zmm1, ZMMWORD PTR [rcx]
mov       rcx,  QWORD PTR [rsp]
kmovq     k1,   QWORD PTR [rcx]         ; load mask register
vmovdqu32 zmm0, zmm2                    ; copy lhs into result for merging
vpaddd    zmm0{k1}, zmm2, zmm1          ; perform add with write mask (merge)
vmovdqu32 ZMMWORD PTR [rax], zmm0       ; store result
add       rsp, 0x20
ret

AVX-512 Example Trace Log

Todo

  • Bitwise operations
  • Support ternary operations
  • Optimize loads
  • TRIL tests

BradleyWood added 10 commits January 9, 2023 07:39
Signed-off-by: BradleyWood <bradley.wood@ibm.com>
Signed-off-by: BradleyWood <bradley.wood@ibm.com>
Signed-off-by: BradleyWood <bradley.wood@ibm.com>
Signed-off-by: BradleyWood <bradley.wood@ibm.com>
Signed-off-by: BradleyWood <bradley.wood@ibm.com>
Signed-off-by: BradleyWood <bradley.wood@ibm.com>
Signed-off-by: BradleyWood <bradley.wood@ibm.com>
Signed-off-by: BradleyWood <bradley.wood@ibm.com>
Signed-off-by: BradleyWood <bradley.wood@ibm.com>
Signed-off-by: BradleyWood <bradley.wood@ibm.com>
@0xdaryl

0xdaryl commented Jan 11, 2023

Copy link
Copy Markdown
Contributor

Jenkins build all

@0xdaryl 0xdaryl self-assigned this Jan 11, 2023
@0xdaryl

0xdaryl commented Jan 12, 2023

Copy link
Copy Markdown
Contributor

FYI this is in draft state.

@BradleyWood
BradleyWood marked this pull request as ready for review January 12, 2023 16:04
@0xdaryl

0xdaryl commented Jan 13, 2023

Copy link
Copy Markdown
Contributor

I will merge this now. Although I know you have tested this with OpenJ9, please keep an eye out for any failures appearing on the weekend in OpenJ9.

@pshipton FYI

@0xdaryl
0xdaryl merged commit ec81a72 into eclipse-omr:master Jan 13, 2023
@pshipton

Copy link
Copy Markdown
Contributor

This wouldn't normally be in an OpenJ9 acceptance build until Monday 1pm, but I'll start one now.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants