Stowage/go
2
0
Fork 0
mirror of https://github.com/golang/go.git synced 2025-12-08 06:10:04 +00:00
Code Issues Projects Releases Packages Wiki Activity

[dev.simd] cmd/compile: add more dot products

Browse source 
This CL is generated by CL 680215.

Change-Id: Ie085e65e0473a8e96170702d7265d379ec8812ba
Reviewed-on: https://go-review.googlesource.com/c/go/+/681298
Reviewed-by: David Chase <drchase@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
This commit is contained in:
Junyang Shao 2025-06-12 16:43:10 +00:00
parent 3df41c856e
commit ded6e0ac71
8 changed files with 1439 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

24
src/cmd/compile/internal/ssa/_gen/simdAMD64ops.go
View file

@ -283,15 +283,23 @@ func simdAMD64Ops(fp11, fp21, fp2k1, fp1k1fp1, fp2k1fp1, fp2k1k1, fp31, fp3k1fp1
{name: "VPMINSDMasked512", argLength: 3, reg: fp2k1fp1, asm: "VPMINSD", commutative: true, typ: "Vec512", resultInArg0: false},
{name: "VPMULLDMasked512", argLength: 3, reg: fp2k1fp1, asm: "VPMULLD", commutative: true, typ: "Vec512", resultInArg0: false},
{name: "VPORDMasked512", argLength: 3, reg: fp2k1fp1, asm: "VPORD", commutative: true, typ: "Vec512", resultInArg0: false},
{name: "VPDPWSSDMasked512", argLength: 4, reg: fp3k1fp1, asm: "VPDPWSSD", commutative: false, typ: "Vec512", resultInArg0: true},
{name: "VPOPCNTDMasked512", argLength: 2, reg: fp1k1fp1, asm: "VPOPCNTD", commutative: false, typ: "Vec512", resultInArg0: false},
{name: "VPDPWSSDSMasked512", argLength: 4, reg: fp3k1fp1, asm: "VPDPWSSDS", commutative: false, typ: "Vec512", resultInArg0: true},
{name: "VPDPBUSDSMasked512", argLength: 4, reg: fp3k1fp1, asm: "VPDPBUSDS", commutative: false, typ: "Vec512", resultInArg0: true},
{name: "VPSUBDMasked512", argLength: 3, reg: fp2k1fp1, asm: "VPSUBD", commutative: false, typ: "Vec512", resultInArg0: false},
{name: "VPDPBUSDMasked512", argLength: 4, reg: fp3k1fp1, asm: "VPDPBUSD", commutative: false, typ: "Vec512", resultInArg0: true},
{name: "VPXORDMasked512", argLength: 3, reg: fp2k1fp1, asm: "VPXORD", commutative: true, typ: "Vec512", resultInArg0: false},
{name: "VPMAXSD512", argLength: 2, reg: fp21, asm: "VPMAXSD", commutative: true, typ: "Vec512", resultInArg0: false},
{name: "VPMINSD512", argLength: 2, reg: fp21, asm: "VPMINSD", commutative: true, typ: "Vec512", resultInArg0: false},
{name: "VPMULLD512", argLength: 2, reg: fp21, asm: "VPMULLD", commutative: true, typ: "Vec512", resultInArg0: false},
{name: "VPORD512", argLength: 2, reg: fp21, asm: "VPORD", commutative: true, typ: "Vec512", resultInArg0: false},
{name: "VPDPWSSD512", argLength: 3, reg: fp31, asm: "VPDPWSSD", commutative: false, typ: "Vec512", resultInArg0: true},
{name: "VPOPCNTD512", argLength: 1, reg: fp11, asm: "VPOPCNTD", commutative: false, typ: "Vec512", resultInArg0: false},
{name: "VPDPWSSDS512", argLength: 3, reg: fp31, asm: "VPDPWSSDS", commutative: false, typ: "Vec512", resultInArg0: true},
{name: "VPDPBUSDS512", argLength: 3, reg: fp31, asm: "VPDPBUSDS", commutative: false, typ: "Vec512", resultInArg0: true},
{name: "VPSUBD512", argLength: 2, reg: fp21, asm: "VPSUBD", commutative: false, typ: "Vec512", resultInArg0: false},
{name: "VPDPBUSD512", argLength: 3, reg: fp31, asm: "VPDPBUSD", commutative: false, typ: "Vec512", resultInArg0: true},
{name: "VPXORD512", argLength: 2, reg: fp21, asm: "VPXORD", commutative: true, typ: "Vec512", resultInArg0: false},
{name: "VPABSD128", argLength: 1, reg: fp11, asm: "VPABSD", commutative: false, typ: "Vec128", resultInArg0: false},
{name: "VPADDD128", argLength: 2, reg: fp21, asm: "VPADDD", commutative: true, typ: "Vec128", resultInArg0: false},
@ -307,18 +315,26 @@ func simdAMD64Ops(fp11, fp21, fp2k1, fp1k1fp1, fp2k1fp1, fp2k1k1, fp31, fp3k1fp1
{name: "VPMINSDMasked128", argLength: 3, reg: fp2k1fp1, asm: "VPMINSD", commutative: true, typ: "Vec128", resultInArg0: false},
{name: "VPMULLDMasked128", argLength: 3, reg: fp2k1fp1, asm: "VPMULLD", commutative: true, typ: "Vec128", resultInArg0: false},
{name: "VPORDMasked128", argLength: 3, reg: fp2k1fp1, asm: "VPORD", commutative: true, typ: "Vec128", resultInArg0: false},
{name: "VPDPWSSDMasked128", argLength: 4, reg: fp3k1fp1, asm: "VPDPWSSD", commutative: false, typ: "Vec128", resultInArg0: true},
{name: "VPOPCNTDMasked128", argLength: 2, reg: fp1k1fp1, asm: "VPOPCNTD", commutative: false, typ: "Vec128", resultInArg0: false},
{name: "VPDPWSSDSMasked128", argLength: 4, reg: fp3k1fp1, asm: "VPDPWSSDS", commutative: false, typ: "Vec128", resultInArg0: true},
{name: "VPDPBUSDSMasked128", argLength: 4, reg: fp3k1fp1, asm: "VPDPBUSDS", commutative: false, typ: "Vec128", resultInArg0: true},
{name: "VPSUBDMasked128", argLength: 3, reg: fp2k1fp1, asm: "VPSUBD", commutative: false, typ: "Vec128", resultInArg0: false},
{name: "VPDPBUSDMasked128", argLength: 4, reg: fp3k1fp1, asm: "VPDPBUSD", commutative: false, typ: "Vec128", resultInArg0: true},
{name: "VPXORDMasked128", argLength: 3, reg: fp2k1fp1, asm: "VPXORD", commutative: true, typ: "Vec128", resultInArg0: false},
{name: "VPMAXSD128", argLength: 2, reg: fp21, asm: "VPMAXSD", commutative: true, typ: "Vec128", resultInArg0: false},
{name: "VPMINSD128", argLength: 2, reg: fp21, asm: "VPMINSD", commutative: true, typ: "Vec128", resultInArg0: false},
{name: "VPMULDQ128", argLength: 2, reg: fp21, asm: "VPMULDQ", commutative: true, typ: "Vec128", resultInArg0: false},
{name: "VPMULLD128", argLength: 2, reg: fp21, asm: "VPMULLD", commutative: true, typ: "Vec128", resultInArg0: false},
{name: "VPDPWSSD128", argLength: 3, reg: fp31, asm: "VPDPWSSD", commutative: false, typ: "Vec128", resultInArg0: true},
{name: "VPHADDD128", argLength: 2, reg: fp21, asm: "VPHADDD", commutative: false, typ: "Vec128", resultInArg0: false},
{name: "VPHSUBD128", argLength: 2, reg: fp21, asm: "VPHSUBD", commutative: false, typ: "Vec128", resultInArg0: false},
{name: "VPOPCNTD128", argLength: 1, reg: fp11, asm: "VPOPCNTD", commutative: false, typ: "Vec128", resultInArg0: false},
{name: "VPDPWSSDS128", argLength: 3, reg: fp31, asm: "VPDPWSSDS", commutative: false, typ: "Vec128", resultInArg0: true},
{name: "VPDPBUSDS128", argLength: 3, reg: fp31, asm: "VPDPBUSDS", commutative: false, typ: "Vec128", resultInArg0: true},
{name: "VPSIGND128", argLength: 2, reg: fp21, asm: "VPSIGND", commutative: false, typ: "Vec128", resultInArg0: false},
{name: "VPSUBD128", argLength: 2, reg: fp21, asm: "VPSUBD", commutative: false, typ: "Vec128", resultInArg0: false},
{name: "VPDPBUSD128", argLength: 3, reg: fp31, asm: "VPDPBUSD", commutative: false, typ: "Vec128", resultInArg0: true},
{name: "VPABSD256", argLength: 1, reg: fp11, asm: "VPABSD", commutative: false, typ: "Vec256", resultInArg0: false},
{name: "VPADDD256", argLength: 2, reg: fp21, asm: "VPADDD", commutative: true, typ: "Vec256", resultInArg0: false},
{name: "VPCMPEQD256", argLength: 2, reg: fp21, asm: "VPCMPEQD", commutative: true, typ: "Vec256", resultInArg0: false},
@ -333,18 +349,26 @@ func simdAMD64Ops(fp11, fp21, fp2k1, fp1k1fp1, fp2k1fp1, fp2k1k1, fp31, fp3k1fp1
{name: "VPMINSDMasked256", argLength: 3, reg: fp2k1fp1, asm: "VPMINSD", commutative: true, typ: "Vec256", resultInArg0: false},
{name: "VPMULLDMasked256", argLength: 3, reg: fp2k1fp1, asm: "VPMULLD", commutative: true, typ: "Vec256", resultInArg0: false},
{name: "VPORDMasked256", argLength: 3, reg: fp2k1fp1, asm: "VPORD", commutative: true, typ: "Vec256", resultInArg0: false},
{name: "VPDPWSSDMasked256", argLength: 4, reg: fp3k1fp1, asm: "VPDPWSSD", commutative: false, typ: "Vec256", resultInArg0: true},
{name: "VPOPCNTDMasked256", argLength: 2, reg: fp1k1fp1, asm: "VPOPCNTD", commutative: false, typ: "Vec256", resultInArg0: false},
{name: "VPDPWSSDSMasked256", argLength: 4, reg: fp3k1fp1, asm: "VPDPWSSDS", commutative: false, typ: "Vec256", resultInArg0: true},
{name: "VPDPBUSDSMasked256", argLength: 4, reg: fp3k1fp1, asm: "VPDPBUSDS", commutative: false, typ: "Vec256", resultInArg0: true},
{name: "VPSUBDMasked256", argLength: 3, reg: fp2k1fp1, asm: "VPSUBD", commutative: false, typ: "Vec256", resultInArg0: false},
{name: "VPDPBUSDMasked256", argLength: 4, reg: fp3k1fp1, asm: "VPDPBUSD", commutative: false, typ: "Vec256", resultInArg0: true},
{name: "VPXORDMasked256", argLength: 3, reg: fp2k1fp1, asm: "VPXORD", commutative: true, typ: "Vec256", resultInArg0: false},
{name: "VPMAXSD256", argLength: 2, reg: fp21, asm: "VPMAXSD", commutative: true, typ: "Vec256", resultInArg0: false},
{name: "VPMINSD256", argLength: 2, reg: fp21, asm: "VPMINSD", commutative: true, typ: "Vec256", resultInArg0: false},
{name: "VPMULDQ256", argLength: 2, reg: fp21, asm: "VPMULDQ", commutative: true, typ: "Vec256", resultInArg0: false},
{name: "VPMULLD256", argLength: 2, reg: fp21, asm: "VPMULLD", commutative: true, typ: "Vec256", resultInArg0: false},
{name: "VPDPWSSD256", argLength: 3, reg: fp31, asm: "VPDPWSSD", commutative: false, typ: "Vec256", resultInArg0: true},
{name: "VPHADDD256", argLength: 2, reg: fp21, asm: "VPHADDD", commutative: false, typ: "Vec256", resultInArg0: false},
{name: "VPHSUBD256", argLength: 2, reg: fp21, asm: "VPHSUBD", commutative: false, typ: "Vec256", resultInArg0: false},
{name: "VPOPCNTD256", argLength: 1, reg: fp11, asm: "VPOPCNTD", commutative: false, typ: "Vec256", resultInArg0: false},
{name: "VPDPWSSDS256", argLength: 3, reg: fp31, asm: "VPDPWSSDS", commutative: false, typ: "Vec256", resultInArg0: true},
{name: "VPDPBUSDS256", argLength: 3, reg: fp31, asm: "VPDPBUSDS", commutative: false, typ: "Vec256", resultInArg0: true},
{name: "VPSIGND256", argLength: 2, reg: fp21, asm: "VPSIGND", commutative: false, typ: "Vec256", resultInArg0: false},
{name: "VPSUBD256", argLength: 2, reg: fp21, asm: "VPSUBD", commutative: false, typ: "Vec256", resultInArg0: false},
{name: "VPDPBUSD256", argLength: 3, reg: fp31, asm: "VPDPBUSD", commutative: false, typ: "Vec256", resultInArg0: true},
{name: "VPABSQ128", argLength: 1, reg: fp11, asm: "VPABSQ", commutative: false, typ: "Vec128", resultInArg0: false},
{name: "VPADDQ128", argLength: 2, reg: fp21, asm: "VPADDQ", commutative: true, typ: "Vec128", resultInArg0: false},
{name: "VPCMPEQQ128", argLength: 2, reg: fp21, asm: "VPCMPEQQ", commutative: true, typ: "Vec128", resultInArg0: false},