From: Jonathan Gray <jsg@jsg.id.au>
Subject: Re: arm64 AI support
To: Mark Kettenis <mark.kettenis@xs4all.nl>
Cc: brad@comstyle.com, tech@openbsd.org
Date: Mon, 27 May 2024 12:54:06 +1000

On Sun, May 26, 2024 at 02:35:58PM +0200, Mark Kettenis wrote:
> Just kidding.
> 
> This adds support for recognizing a few more instruction set features,
> including the BF16 and I8MM features that are useful for implementing
> things like convolutional neural networks.
> 
> This also makes us export those features to userland through
> sysctl(2).  It is entirely up to userland to use these features or
> not.
> 
> ok?

the defines match the Arm ARM, but there is a mistake in the printfs
see below

> 
> Index: arch/arm64/include/armreg.h
> ===================================================================
> RCS file: /cvs/src/sys/arch/arm64/include/armreg.h,v
> retrieving revision 1.33
> diff -u -p -r1.33 armreg.h
> --- arch/arm64/include/armreg.h	18 Mar 2024 18:35:21 -0000	1.33
> +++ arch/arm64/include/armreg.h	26 May 2024 12:28:17 -0000
> @@ -337,7 +337,7 @@
>  #define	 ID_AA64ISAR0_RNDR_IMPL		(0x1ULL << ID_AA64ISAR0_RNDR_SHIFT)
>  
>  /* ID_AA64ISAR1_EL1 */
> -#define	ID_AA64ISAR1_MASK		0x00000fffffffffffULL
> +#define	ID_AA64ISAR1_MASK		0xffffffffffffffffULL
>  #define	ID_AA64ISAR1_DPB_SHIFT		0
>  #define	ID_AA64ISAR1_DPB_MASK		(0xfULL << ID_AA64ISAR1_DPB_SHIFT)
>  #define	ID_AA64ISAR1_DPB(x)		((x) & ID_AA64ISAR1_DPB_MASK)
> @@ -396,6 +396,34 @@
>  #define	ID_AA64ISAR1_SPECRES(x)		((x) & ID_AA64ISAR1_SPECRES_MASK)
>  #define	 ID_AA64ISAR1_SPECRES_NONE	(0x0ULL << ID_AA64ISAR1_SPECRES_SHIFT)
>  #define	 ID_AA64ISAR1_SPECRES_IMPL	(0x1ULL << ID_AA64ISAR1_SPECRES_SHIFT)
> +#define	ID_AA64ISAR1_BF16_SHIFT		44
> +#define	ID_AA64ISAR1_BF16_MASK		(0xfULL << ID_AA64ISAR1_BF16_SHIFT)
> +#define	ID_AA64ISAR1_BF16(x)		((x) & ID_AA64ISAR1_BF16_MASK)
> +#define	 ID_AA64ISAR1_BF16_NONE		(0x0ULL << ID_AA64ISAR1_BF16_SHIFT)
> +#define	 ID_AA64ISAR1_BF16_BASE		(0x1ULL << ID_AA64ISAR1_BF16_SHIFT)
> +#define	 ID_AA64ISAR1_BF16_EBF		(0x2ULL << ID_AA64ISAR1_BF16_SHIFT)
> +#define	ID_AA64ISAR1_DGH_SHIFT		48
> +#define	ID_AA64ISAR1_DGH_MASK		(0xfULL << ID_AA64ISAR1_DGH_SHIFT)
> +#define	ID_AA64ISAR1_DGH(x)		((x) & ID_AA64ISAR1_DGH_MASK)
> +#define	 ID_AA64ISAR1_DGH_NONE		(0x0ULL << ID_AA64ISAR1_DGH_SHIFT)
> +#define	 ID_AA64ISAR1_DGH_IMPL		(0x1ULL << ID_AA64ISAR1_DGH_SHIFT)
> +#define	ID_AA64ISAR1_I8MM_SHIFT		52
> +#define	ID_AA64ISAR1_I8MM_MASK		(0xfULL << ID_AA64ISAR1_I8MM_SHIFT)
> +#define	ID_AA64ISAR1_I8MM(x)		((x) & ID_AA64ISAR1_I8MM_MASK)
> +#define	 ID_AA64ISAR1_I8MM_NONE		(0x0ULL << ID_AA64ISAR1_I8MM_SHIFT)
> +#define	 ID_AA64ISAR1_I8MM_IMPL		(0x1ULL << ID_AA64ISAR1_I8MM_SHIFT)
> +#define	ID_AA64ISAR1_XS_SHIFT		56
> +#define	ID_AA64ISAR1_XS_MASK		(0xfULL << ID_AA64ISAR1_XS_SHIFT)
> +#define	ID_AA64ISAR1_XS(x)		((x) & ID_AA64ISAR1_XS_MASK)
> +#define	 ID_AA64ISAR1_XS_NONE		(0x0ULL << ID_AA64ISAR1_XS_SHIFT)
> +#define	 ID_AA64ISAR1_XS_IMPL		(0x1ULL << ID_AA64ISAR1_XS_SHIFT)
> +#define	ID_AA64ISAR1_LS64_SHIFT		60
> +#define	ID_AA64ISAR1_LS64_MASK		(0xfULL << ID_AA64ISAR1_LS64_SHIFT)
> +#define	ID_AA64ISAR1_LS64(x)		((x) & ID_AA64ISAR1_LS64_MASK)
> +#define	 ID_AA64ISAR1_LS64_NONE		(0x0ULL << ID_AA64ISAR1_LS64_SHIFT)
> +#define	 ID_AA64ISAR1_LS64_BASE		(0x1ULL << ID_AA64ISAR1_LS64_SHIFT)
> +#define	 ID_AA64ISAR1_LS64_V		(0x2ULL << ID_AA64ISAR1_LS64_SHIFT)
> +#define	 ID_AA64ISAR1_LS64_ACCDATA	(0x3ULL << ID_AA64ISAR1_LS64_SHIFT)
>  
>  /* ID_AA64ISAR2_EL1 */
>  #define	ID_AA64ISAR2_MASK		0x00000000f0000000ULL
> Index: arch/arm64/arm64/cpu.c
> ===================================================================
> RCS file: /cvs/src/sys/arch/arm64/arm64/cpu.c,v
> retrieving revision 1.114
> diff -u -p -r1.114 cpu.c
> --- arch/arm64/arm64/cpu.c	13 Apr 2024 14:19:39 -0000	1.114
> +++ arch/arm64/arm64/cpu.c	26 May 2024 12:28:17 -0000
> @@ -742,6 +742,37 @@ cpu_identify(struct cpu_info *ci)
>  	 */
>  	id = READ_SPECIALREG(id_aa64isar1_el1);
>  
> +	if (ID_AA64ISAR1_LS64(id) >= ID_AA64ISAR1_LS64_BASE) {
> +		printf("%sLS64", sep);
> +		sep = ",";
> +	}
> +	if (ID_AA64ISAR1_LS64(id) >= ID_AA64ISAR1_LS64_V)
> +		printf("+V");
> +	if (ID_AA64ISAR1_LS64(id) >= ID_AA64ISAR1_LS64_ACCDATA)
> +		printf("+ACCDATA");
> +
> +	if (ID_AA64ISAR1_XS(id) >= ID_AA64ISAR1_XS_IMPL) {
> +		printf("%sXS", sep);
> +		sep = ",";
> +	}
> +
> +	if (ID_AA64ISAR1_I8MM(id) >= ID_AA64ISAR1_I8MM_IMPL) {
> +		printf("%sI8MM", sep);
> +		sep = ",";
> +	}
> +
> +	if (ID_AA64ISAR1_DGH(id) >= ID_AA64ISAR1_DGH_IMPL) {
> +		printf("%sXS", sep);

should be DGH not XS.  ok jsg@ with that fixed

> +		sep = ",";
> +	}
> +
> +	if (ID_AA64ISAR1_BF16(id) >= ID_AA64ISAR1_BF16_BASE) {
> +		printf("%sBF16", sep);
> +		sep = ",";
> +	}
> +	if (ID_AA64ISAR1_BF16(id) >= ID_AA64ISAR1_BF16_EBF)
> +		printf("+EBF");
> +
>  	if (ID_AA64ISAR1_SPECRES(id) >= ID_AA64ISAR1_SPECRES_IMPL) {
>  		printf("%sSPECRES", sep);
>  		sep = ",";
> 
>