[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [PATCH v2] x86: Intel AVX512_BF16 feature enabling
From: |
Paolo Bonzini |
Subject: |
Re: [Qemu-devel] [PATCH v2] x86: Intel AVX512_BF16 feature enabling |
Date: |
Mon, 19 Aug 2019 16:11:48 +0200 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101 Thunderbird/60.8.0 |
On 25/07/19 08:14, Jing Liu wrote:
> Intel CooperLake cpu adds AVX512_BF16 instruction, defining as
> CPUID.(EAX=7,ECX=1):EAX[bit 05].
>
> The patch adds a property for setting the subleaf of CPUID leaf 7 in
> case that people would like to specify it.
>
> The release spec link as follows,
> https://software.intel.com/sites/default/files/managed/c5/15/\
> architecture-instruction-set-extensions-programming-reference.pdf
>
> Signed-off-by: Jing Liu <address@hidden>
> ---
> target/i386/cpu.c | 39 ++++++++++++++++++++++++++++++++++++++-
> target/i386/cpu.h | 7 +++++++
> target/i386/kvm.c | 3 ++-
> 3 files changed, 47 insertions(+), 2 deletions(-)
>
> diff --git a/target/i386/cpu.c b/target/i386/cpu.c
> index 805ce95..517dedb 100644
> --- a/target/i386/cpu.c
> +++ b/target/i386/cpu.c
> @@ -770,6 +770,7 @@ static void x86_cpu_vendor_words2str(char *dst, uint32_t
> vendor1,
> /* CPUID_7_0_ECX_OSPKE is dynamic */ \
> CPUID_7_0_ECX_LA57)
> #define TCG_7_0_EDX_FEATURES 0
> +#define TCG_7_1_EAX_FEATURES 0
> #define TCG_APM_FEATURES 0
> #define TCG_6_EAX_FEATURES CPUID_6_EAX_ARAT
> #define TCG_XSAVE_FEATURES (CPUID_XSAVE_XSAVEOPT | CPUID_XSAVE_XGETBV1)
> @@ -1095,6 +1096,25 @@ static FeatureWordInfo
> feature_word_info[FEATURE_WORDS] = {
> },
> .tcg_features = TCG_7_0_EDX_FEATURES,
> },
> + [FEAT_7_1_EAX] = {
> + .type = CPUID_FEATURE_WORD,
> + .feat_names = {
> + NULL, NULL, NULL, NULL,
> + NULL, "avx512-bf16", NULL, NULL,
> + NULL, NULL, NULL, NULL,
> + NULL, NULL, NULL, NULL,
> + NULL, NULL, NULL, NULL,
> + NULL, NULL, NULL, NULL,
> + NULL, NULL, NULL, NULL,
> + NULL, NULL, NULL, NULL,
> + },
> + .cpuid = {
> + .eax = 7,
> + .needs_ecx = true, .ecx = 1,
> + .reg = R_EAX,
> + },
> + .tcg_features = TCG_7_1_EAX_FEATURES,
> + },
> [FEAT_8000_0007_EDX] = {
> .type = CPUID_FEATURE_WORD,
> .feat_names = {
> @@ -4293,13 +4313,19 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index,
> uint32_t count,
> case 7:
> /* Structured Extended Feature Flags Enumeration Leaf */
> if (count == 0) {
> - *eax = 0; /* Maximum ECX value for sub-leaves */
> + /* Maximum ECX value for sub-leaves */
> + *eax = env->cpuid_level_func7;
> *ebx = env->features[FEAT_7_0_EBX]; /* Feature flags */
> *ecx = env->features[FEAT_7_0_ECX]; /* Feature flags */
> if ((*ecx & CPUID_7_0_ECX_PKU) && env->cr[4] & CR4_PKE_MASK) {
> *ecx |= CPUID_7_0_ECX_OSPKE;
> }
> *edx = env->features[FEAT_7_0_EDX]; /* Feature flags */
> + } else if (count == 1) {
> + *eax = env->features[FEAT_7_1_EAX];
> + *ebx = 0;
> + *ecx = 0;
> + *edx = 0;
> } else {
> *eax = 0;
> *ebx = 0;
> @@ -4949,6 +4975,11 @@ static void x86_cpu_adjust_feat_level(X86CPU *cpu,
> FeatureWord w)
> x86_cpu_adjust_level(cpu, &env->cpuid_min_xlevel2, eax);
> break;
> }
> +
> + if (eax == 7) {
> + x86_cpu_adjust_level(cpu, &env->cpuid_min_level_func7,
> + fi->cpuid.ecx);
> + }
> }
>
> /* Calculate XSAVE components based on the configured CPU feature flags */
> @@ -5067,6 +5098,7 @@ static void x86_cpu_expand_features(X86CPU *cpu, Error
> **errp)
> x86_cpu_adjust_feat_level(cpu, FEAT_1_ECX);
> x86_cpu_adjust_feat_level(cpu, FEAT_6_EAX);
> x86_cpu_adjust_feat_level(cpu, FEAT_7_0_ECX);
> + x86_cpu_adjust_feat_level(cpu, FEAT_7_1_EAX);
> x86_cpu_adjust_feat_level(cpu, FEAT_8000_0001_EDX);
> x86_cpu_adjust_feat_level(cpu, FEAT_8000_0001_ECX);
> x86_cpu_adjust_feat_level(cpu, FEAT_8000_0007_EDX);
> @@ -5098,6 +5130,9 @@ static void x86_cpu_expand_features(X86CPU *cpu, Error
> **errp)
> }
>
> /* Set cpuid_*level* based on cpuid_min_*level, if not explicitly set */
> + if (env->cpuid_level_func7 == UINT32_MAX) {
> + env->cpuid_level_func7 = env->cpuid_min_level_func7;
> + }
> if (env->cpuid_level == UINT32_MAX) {
> env->cpuid_level = env->cpuid_min_level;
> }
> @@ -5869,6 +5904,8 @@ static Property x86_cpu_properties[] = {
> DEFINE_PROP_BOOL("host-phys-bits", X86CPU, host_phys_bits, false),
> DEFINE_PROP_UINT8("host-phys-bits-limit", X86CPU, host_phys_bits_limit,
> 0),
> DEFINE_PROP_BOOL("fill-mtrr-mask", X86CPU, fill_mtrr_mask, true),
> + DEFINE_PROP_UINT32("level-func7", X86CPU, env.cpuid_level_func7,
> + UINT32_MAX),
> DEFINE_PROP_UINT32("level", X86CPU, env.cpuid_level, UINT32_MAX),
> DEFINE_PROP_UINT32("xlevel", X86CPU, env.cpuid_xlevel, UINT32_MAX),
> DEFINE_PROP_UINT32("xlevel2", X86CPU, env.cpuid_xlevel2, UINT32_MAX),
> diff --git a/target/i386/cpu.h b/target/i386/cpu.h
> index 05393cf..df9106f 100644
> --- a/target/i386/cpu.h
> +++ b/target/i386/cpu.h
> @@ -479,6 +479,7 @@ typedef enum FeatureWord {
> FEAT_7_0_EBX, /* CPUID[EAX=7,ECX=0].EBX */
> FEAT_7_0_ECX, /* CPUID[EAX=7,ECX=0].ECX */
> FEAT_7_0_EDX, /* CPUID[EAX=7,ECX=0].EDX */
> + FEAT_7_1_EAX, /* CPUID[EAX=7,ECX=1].EAX */
> FEAT_8000_0001_EDX, /* CPUID[8000_0001].EDX */
> FEAT_8000_0001_ECX, /* CPUID[8000_0001].ECX */
> FEAT_8000_0007_EDX, /* CPUID[8000_0007].EDX */
> @@ -692,6 +693,8 @@ typedef uint32_t FeatureWordArray[FEATURE_WORDS];
> #define CPUID_7_0_EDX_CORE_CAPABILITY (1U << 30) /*Core Capability*/
> #define CPUID_7_0_EDX_SPEC_CTRL_SSBD (1U << 31) /* Speculative Store Bypass
> Disable */
>
> +#define CPUID_7_1_EAX_AVX512_BF16 (1U << 5) /* AVX512 BFloat16 Instruction */
> +
> #define CPUID_8000_0008_EBX_WBNOINVD (1U << 9) /* Write back and
>
> do not invalidate cache */
> #define CPUID_8000_0008_EBX_IBPB (1U << 12) /* Indirect Branch Prediction
> Barrier */
> @@ -1322,6 +1325,10 @@ typedef struct CPUX86State {
> /* Fields after this point are preserved across CPU reset. */
>
> /* processor features (e.g. for CPUID insn) */
> + /* Minimum cpuid leaf 7 value */
> + uint32_t cpuid_level_func7;
> + /* Actual cpuid leaf 7 value */
> + uint32_t cpuid_min_level_func7;
> /* Minimum level/xlevel/xlevel2, based on CPU model + features */
> uint32_t cpuid_min_level, cpuid_min_xlevel, cpuid_min_xlevel2;
> /* Maximum level/xlevel/xlevel2 value for auto-assignment: */
> diff --git a/target/i386/kvm.c b/target/i386/kvm.c
> index ec7870c..fd0a447 100644
> --- a/target/i386/kvm.c
> +++ b/target/i386/kvm.c
> @@ -1493,6 +1493,7 @@ int kvm_arch_init_vcpu(CPUState *cs)
> c = &cpuid_data.entries[cpuid_i++];
> }
> break;
> + case 0x7:
> case 0x14: {
> uint32_t times;
>
> @@ -1505,7 +1506,7 @@ int kvm_arch_init_vcpu(CPUState *cs)
> for (j = 1; j <= times; ++j) {
> if (cpuid_i == KVM_MAX_CPUID_ENTRIES) {
> fprintf(stderr, "cpuid_data is full, no space for "
> - "cpuid(eax:0x14,ecx:0x%x)\n", j);
> + "cpuid(eax:0x%x,ecx:0x%x)\n", i, j);
> abort();
> }
> c = &cpuid_data.entries[cpuid_i++];
>
Queued, thanks (I was waiting for QEMU 4.1 to be released).
Paolo