Latest Intel platform Granite Rapids has introduced a new instruction -
AMX-FP16, which performs dot-products of two FP16 tiles and accumulates
the results into a packed single precision tile. This instrucion needs
no additional enabling on top of the existing kernel AMX enabling.
AMX-FP16 adds FP16 capability and also allows a FP16 GPU trained model
to run faster without loss of accuracy or added SW overhead.
The bit definition:
CPUID.(EAX=7,ECX=1):EAX[bit 21]
This CPUID is exposed to user space. Besides, there is no other VMX
control for this instruction.
Signed-off-by: Jiaxi Chen <jiaxi.chen@xxxxxxxxxxxxxxx>
---
arch/x86/kvm/cpuid.c | 2 +-
arch/x86/kvm/reverse_cpuid.h | 1 +
2 files changed, 2 insertions(+), 1 deletion(-)
diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c
index b388ef52f8c8..19ef02d5b11b 100644
--- a/arch/x86/kvm/cpuid.c
+++ b/arch/x86/kvm/cpuid.c
@@ -657,7 +657,7 @@ void kvm_set_cpu_caps(void)
kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
kvm_cpu_cap_init_scattered(CPUID_7_1_EAX,
- F(AVX_VNNI) | F(AVX512_BF16) | F(CMPCCXADD)
+ F(AVX_VNNI) | F(AVX512_BF16) | F(CMPCCXADD) | F(AMX_FP16)
);
kvm_cpu_cap_mask(CPUID_D_1_EAX,
diff --git a/arch/x86/kvm/reverse_cpuid.h b/arch/x86/kvm/reverse_cpuid.h
index 24f570ddb225..05fd43ebd226 100644
--- a/arch/x86/kvm/reverse_cpuid.h
+++ b/arch/x86/kvm/reverse_cpuid.h
@@ -29,6 +29,7 @@ enum kvm_only_cpuid_leafs {
#define X86_FEATURE_AVX_VNNI KVM_X86_FEATURE(CPUID_7_1_EAX, 4)
#define X86_FEATURE_AVX512_BF16 KVM_X86_FEATURE(CPUID_7_1_EAX, 5)
#define X86_FEATURE_CMPCCXADD KVM_X86_FEATURE(CPUID_7_1_EAX, 7)
+#define X86_FEATURE_AMX_FP16 KVM_X86_FEATURE(CPUID_7_1_EAX, 21)
struct cpuid_reg {
u32 function;
--
2.27.0
