From: "Chang S. Bae" <chang.seok.bae@xxxxxxxxx> 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 instruction 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: Chang S. Bae <chang.seok.bae@xxxxxxxxx> Signed-off-by: Jiaxi Chen <jiaxi.chen@xxxxxxxxxxxxxxx> --- arch/x86/include/asm/cpufeatures.h | 1 + arch/x86/kvm/cpuid.c | 2 +- 2 files changed, 2 insertions(+), 1 deletion(-) diff --git a/arch/x86/include/asm/cpufeatures.h b/arch/x86/include/asm/cpufeatures.h index 19db3940f262..df4a7f7505a9 100644 --- a/arch/x86/include/asm/cpufeatures.h +++ b/arch/x86/include/asm/cpufeatures.h @@ -309,6 +309,7 @@ #define X86_FEATURE_AVX_VNNI (12*32+ 4) /* AVX VNNI instructions */ #define X86_FEATURE_AVX512_BF16 (12*32+ 5) /* AVX512 BFLOAT16 instructions */ #define X86_FEATURE_CMPCCXADD (12*32+ 7) /* CMPccXADD instructions */ +#define X86_FEATURE_AMX_FP16 (12*32+21) /* AMX fp16 Support */ /* AMD-defined CPU features, CPUID level 0x80000008 (EBX), word 13 */ #define X86_FEATURE_CLZERO (13*32+ 0) /* CLZERO instruction */ diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c index 7ab7cc717b1c..2a334d4cd04e 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_mask(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, -- 2.27.0
