[PATCH] x86: Support Intel AVX512 BF16

[Jan Beulich]

>>> On 08.04.19 at 20:17, <hjl.tools@gmail.com> wrote:
> On Mon, Apr 8, 2019 at 1:56 AM Jan Beulich <JBeulich@suse.com> wrote:
>>
>> >>> On 05.04.19 at 20:01, <hongjiu.lu@intel.com> wrote:
>> > --- a/opcodes/i386-opc.tbl
>> > +++ b/opcodes/i386-opc.tbl
>> > @@ -4710,3 +4710,33 @@ movdir64b, 2, 0x660f38f8, None, 3, 
> CpuMOVDIR64B|CpuNo64, Modrm|IgnoreSize|No_bSu
>> >  movdir64b, 2, 0x660f38f8, None, 3, CpuMOVDIR64B|Cpu64, 
> Modrm|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|NoRex64|AddrPrefixOpRe
> g, { Unspecified|BaseIndex, Reg32|Reg64 }
>> >
>> >  // MOVEDIR instructions end.
>> > +
>> > +// AVX512_BF16 instructions.
>> > +
>> > +vcvtne2ps2bf16, 3, 0xf272, None, 1, CpuAVX512_BF16|CpuAVX512VL, 
> Modrm|VexOpcode=1|EVex128|VexVVVV=1|Masking=3|VexW0|Broadcast|Disp8MemShift=4
> |IgnoreSize|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { 
> Dword|Xmmword|Unspecified|BaseIndex, RegXMM, RegXMM }
>> > +vcvtne2ps2bf16, 3, 0xf272, None, 1, CpuAVX512_BF16|CpuAVX512VL, 
> Modrm|VexOpcode=1|EVex256|VexVVVV=1|Masking=3|VexW0|Broadcast|Disp8MemShift=5
> |IgnoreSize|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { 
> Dword|Ymmword|Unspecified|BaseIndex, RegYMM, RegYMM }
>> > +vcvtne2ps2bf16, 3, 0xf272, None, 1, CpuAVX512_BF16, 
> Modrm|VexOpcode=1|EVex=1|VexVVVV=1|Masking=3|VexW0|Broadcast|Disp8MemShift=6|
> IgnoreSize|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { 
> Dword|Zmmword|Unspecified|BaseIndex, RegZMM, RegZMM }
>>
>> Here and below, can you please avoid introducing anew inefficient
>> table entries with partly irrelevant / redundant attributes:
>> - use Disp8ShiftVL (and no explicit EVex<NNN> / EVex=<N> nor
>>   CpuAVX512VL), thus folding the above three entries into one
>> - don't explicitly use [XYZ]mmWord (redundant with Reg[XYZ]MM,
>>   once folding with the non-broadcast forms is also done - see
>>   further down)
>> - presumably IgnoreSize is not needed
>> I hope I didn't forget further ones.
> 
> Done.

Much better, thanks!

> I am checking in this patch to update BF16 which was implemented by
> Xuepeng before he left Intel.

Oh, I see.

Jan