kvm: x86: AVX512_BF16 feature support
RFC PATCH 1/1 kvm: x86: AVX512_BF16 feature support
The patch focuses on a new instruction AVX512_BF16 support for kvm guest, defined as CPUID.(EAX=7,ECX=1):EAX[bit 5], see spec[1].
The kvm implementation depends on kernel patch[2] which is in lkml discussion.
References:
[1] https://software.intel.com/sites/default/files/managed/c5/15/\
architecture-instruction-set-extensions-programming-reference.pdf
[2] https://lkml.org/lkml/2019/6/19/912
Jing Liu (1):
kvm: x86: Expose AVX512_BF16 feature to guest
arch/x86/kvm/cpuid.c | 16 ++++++++++++++-- 1 file changed, 14 insertions(+), 2 deletions(-)
Signature · 1 line
1.8.3.1
RFC PATCH 1/1 kvm: x86: Expose AVX512_BF16 feature to guest
AVX512 BFLOAT16 instructions support 16-bit BFLOAT16 floating-point format (BF16) for deep learning optimization.
Intel adds AVX512 BFLOAT16 feature in CooperLake, which is CPUID.7.1.EAX[5].
Detailed information of the CPUID bit can be found here,
https://software.intel.com/sites/default/files/managed/c5/15/\
architecture-instruction-set-extensions-programming-reference.pdf.
--- arch/x86/kvm/cpuid.c | 16 ++++++++++++++-- 1 file changed, 14 insertions(+), 2 deletions(-)
arch/x86/kvm/cpuid.c +14 -2
@@ -484,6 +484,7 @@ static inline int __do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 function, entry->edx =0 ;break ;case 7 : { +int i, times = entry->eax; entry->flags |=KVM_CPUID_FLAG_SIGNIFCANT_INDEX ;/* Mask ebx against host capability word 9 */ if (index ==0 ) {
@@ -507,12 +508,23 @@ static inline int __do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 function,* if the host doesn't support it. */ entry->edx |=F (ARCH_CAPABILITIES ); - }else { + }else if (index > times) { + entry->eax =0 ; entry->ebx =0 ; entry->ecx =0 ; entry->edx =0 ; } - entry->eax =0 ; +for (i =1 ; i <= times; i++) { +if (*nent >= maxnent) + goto out; +do_cpuid_1_ent (&entry[i], function, i); + entry[i].eax &=F (AVX512_BF16 ); + entry[i].ebx =0 ; + entry[i].ecx =0 ; + entry[i].edx =0 ; + entry[i].flags |=KVM_CPUID_FLAG_SIGNIFCANT_INDEX ; + ++*nent; + }break ; }case 9 :
Signature · 1 line
1.8.3.1
Paolo Bonzini
maintainerRe: [PATCH RFC] kvm: x86: Expose AVX512_BF16 feature to guest
arch/x86/kvm/cpuid.c static inline int __do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 function, · line 526
Quoted 10 lines · Jing Liu
arch/x86/kvm/cpuid.c +10
-0+for (i =1 ; i <= times; i++) { +if (*nent >= maxnent) + goto out; +do_cpuid_1_ent (&entry[i], function, i); + entry[i].eax &=F (AVX512_BF16 ); + entry[i].ebx =0 ; + entry[i].ecx =0 ; + entry[i].edx =0 ; + entry[i].flags |=KVM_CPUID_FLAG_SIGNIFCANT_INDEX ; + ++*nent;
This woud be wrong for i > 1, so instead make this
if (entry->eax >= 1)
and define F(AVX512_BF16) as a new constant kvm_cpuid_7_1_eax_features.
Paolo
Re: [PATCH RFC] kvm: x86: Expose AVX512_BF16 feature to guest
Hi Paolo,
Quoted 14 lines · Paolo Bonzini
arch/x86/kvm/cpuid.c static inline int __do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 function, · line 526
Quoted 10 lines · Jing Liu
arch/x86/kvm/cpuid.c +10
-0+for (i =1 ; i <= times; i++) { +if (*nent >= maxnent) + goto out; +do_cpuid_1_ent (&entry[i], function, i); + entry[i].eax &=F (AVX512_BF16 ); + entry[i].ebx =0 ; + entry[i].ecx =0 ; + entry[i].edx =0 ; + entry[i].flags |=KVM_CPUID_FLAG_SIGNIFCANT_INDEX ; + ++*nent;This woud be wrong for i > 1, so instead make this
Quoted 2 lines
if (entry->eax >= 1)
I am confused about the @index parameter. @index seems not used for every case except 0x07. Since the caller function only has @index=0, so all other cases except 0x07 put cpuid info from subleaf=0 to max subleaf.
What do you think about @index in current function? Does it mean, we need put cpuid from index to max subleaf to @entry[i]? If so, the logic seems as follows,
if (index == 0) {
// Put subleaf 0 into @entry
// Put subleaf 1 into @entry[1]
} else if (index < entry->eax) {
// Put subleaf 1 into @entry
} else {
// Put all zero into @entry
}But this seems not identical with other cases, for current caller function. Or we can simply ignore @index in 0x07 and just put all possible subleaf info back?
Quoted 2 lines
and define
F(AVX512_BF16)as a new constantkvm_cpuid_7_1_eax_features.
Got it.
Thanks,
Jing
Quoted 2 lines
Paolo
Re: [PATCH RFC] kvm: x86: Expose AVX512_BF16 feature to guest
Hi Paolo,
After thinking more, I found way to satisfy all cases in a easy way. How about things like this?
@@ -507,12 +510,26 @@ static inline int __do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 fu
* if the host doesn't support it.
*/
entry->edx |= F(ARCH_CAPABILITIES);
+ } else if (index == 1) {
+ entry->eax &= kvm_cpuid_7_1_eax_x86_features;
+ entry->ebx = 0;
+ entry->ecx = 0;
+ entry->edx = 0;
} else {
+ entry->eax = 0;
entry->ebx = 0;
entry->ecx = 0;
entry->edx = 0;
}
- entry->eax = 0;
+
+ if (index == 0 && entry->eax >= 1) {
+ entry[1].eax &= kvm_cpuid_7_1_eax_x86_features;
+ entry[1].ebx = 0;
+ entry[1].ecx = 0;
+ entry[1].edx = 0;
+ entry[1].flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
+ ++*nent;
+ }
break;
}Or you prefer that I update this into another version later?
Thanks!
Jing
Quoted 51 lines · On 6/20/2019 11:09 PM, Liu, Jing2 wrote:
Hi Paolo,
Quoted 12 lines · Paolo Bonzini
Quoted 10 lines · Jing Liu
+ for (i = 1; i <= times; i++) {
+ if (*nent >= maxnent)
+ goto out;
+do_cpuid_1_ent(&entry[i], function, i); + entry[i].eax &=F(AVX512_BF16);
+ entry[i].ebx = 0;
+ entry[i].ecx = 0;
+ entry[i].edx = 0;
+ entry[i].flags |=KVM_CPUID_FLAG_SIGNIFCANT_INDEX; + ++*nent;Quoted 4 lines
This woud be wrong for i > 1, so instead make this
if (
entry->eax>= 1)I am confused about the @index parameter. @index seems not used for every case except
0x07. Since the caller function only has @index=0, so all other cases except0x07put cpuid info from subleaf=0 to max subleaf.What do you think about @index in current function? Does it mean, we need put cpuid from index to max subleaf to @entry[i]? If so, the logic seems as follows,
if (index == 0) {
// Put subleaf 0 into @entry
// Put subleaf 1 into @entry[1]
} else if (index <entry->eax) {
// Put subleaf 1 into @entry
} else {
// Put all zero into @entry
}But this seems not identical with other cases, for current caller function. Or we can simply ignore @index in
0x07and just put all possible subleaf info back?Quoted 2 lines
and define
F(AVX512_BF16)as a new constantkvm_cpuid_7_1_eax_features.Got it.
Thanks,
JingQuoted 2 lines
Paolo
Paolo Bonzini
maintainerRe: [PATCH RFC] kvm: x86: Expose AVX512_BF16 feature to guest
Quoted 13 lines · Jing Liu
Quoted 13 lines
What do you think about @index in current function? Does it mean, we need put cpuid from index to max subleaf to @entry[i]? If so, the logic seems as follows,
if (index == 0) {
// Put subleaf 0 into @entry
// Put subleaf 1 into @entry[1]
} else if (index <entry->eax) {
// Put subleaf 1 into @entry
} else {
// Put all zero into @entry
}
Quoted 3 lines
Quoted 3 lines
But this seems not identical with other cases, for current caller function. Or we can simply ignore @index in
0x07and just put all possible subleaf info back?
There are indeed quite some cleanups to be made there. Let me post a series as soon as possible, and you can base your work on it.
Paolo
Re: [PATCH RFC] kvm: x86: Expose AVX512_BF16 feature to guest
Hi Paolo,
Quoted 14 lines · Paolo Bonzini
Quoted 13 lines · Jing Liu
Quoted 13 lines
What do you think about @index in current function? Does it mean, we need put cpuid from index to max subleaf to @entry[i]? If so, the logic seems as follows,
if (index == 0) {
// Put subleaf 0 into @entry
// Put subleaf 1 into @entry[1]
} else if (index <entry->eax) {
// Put subleaf 1 into @entry
} else {
// Put all zero into @entry
}
Quoted 7 lines
Quoted 3 lines
Quoted 3 lines
But this seems not identical with other cases, for current caller function. Or we can simply ignore @index in
0x07and just put all possible subleaf info back?There are indeed quite some cleanups to be made there. Let me post a series as soon as possible, and you can base your work on it.
Thanks. I just had another mail (replying you in this serial) appending some codes to deal with case 7. If you prefer to firstly cleanup, I can wait for the patch then. :)
Thanks,
Jing
Quoted 2 lines
Paolo