[FFmpeg-devel,3/4] aarch64: Add linux runtime cpu feature detection using getauxval(AT_HWCAP)

Based on code by Janne Grunau.

Using HWCAP_CPUID for user space access to the CPU feature registers. See
https://www.kernel.org/doc/html/latest/arm64/cpu-feature-registers.html.
---
 configure               |  2 ++
 libavutil/aarch64/cpu.c | 38 ++++++++++++++++++++++++++++++++++++++
 2 files changed, 40 insertions(+)

Le perjantaina 26. toukokuuta 2023, 11.03.14 EEST Martin Storsjö a écrit :
> Based on code by Janne Grunau.
> 
> Using HWCAP_CPUID for user space access to the CPU feature registers. See
> https://www.kernel.org/doc/html/latest/arm64/cpu-feature-registers.html.
> ---
>  configure               |  2 ++
>  libavutil/aarch64/cpu.c | 38 ++++++++++++++++++++++++++++++++++++++
>  2 files changed, 40 insertions(+)
> 
> diff --git a/configure b/configure
> index 3c7473efb2..b5357b8d27 100755
> --- a/configure
> +++ b/configure
> @@ -2207,6 +2207,7 @@ HAVE_LIST_PUB="
> 
>  HEADERS_LIST="
>      arpa_inet_h
> +    asm_hwcap_h
>      asm_types_h
>      cdio_paranoia_h
>      cdio_paranoia_paranoia_h
> @@ -6422,6 +6423,7 @@ check_headers io.h
>  enabled libdrm &&
>      check_headers linux/dma-buf.h
> 
> +check_headers asm/hwcap.h
>  check_headers linux/perf_event.h
>  check_headers libcrystalhd/libcrystalhd_if.h
>  check_headers malloc.h
> diff --git a/libavutil/aarch64/cpu.c b/libavutil/aarch64/cpu.c
> index 42b33e4a2d..34c838c2f5 100644
> --- a/libavutil/aarch64/cpu.c
> +++ b/libavutil/aarch64/cpu.c
> @@ -20,6 +20,42 @@
>  #include "libavutil/cpu_internal.h"
>  #include "config.h"
> 
> +#if (defined(__linux__) || defined(__ANDROID__)) && HAVE_GETAUXVAL &&
> HAVE_ASM_HWCAP_H +#include <stdint.h>
> +#include <asm/hwcap.h>
> +#include <sys/auxv.h>
> +
> +#define get_cpu_feature_reg(reg, val) \
> +        __asm__("mrs %0, " #reg : "=r" (val))

Strictly speaking, this can read any system register. One way to prevent that 
would be to include the ID_ prefix and _EL1 suffix in the macro. I would have 
used a pure static inline instead, but that's just a matter of taste.

> +
> +static int detect_flags(void)
> +{
> +    unsigned long ret = getauxval(AT_HWCAP);
> +    int flags = 0;
> +#if defined(HWCAP_CPUID)
> +    uint64_t tmp;
> +    if (!(ret & HWCAP_CPUID))
> +        return flags;
> +    get_cpu_feature_reg(ID_AA64ISAR0_EL1, tmp);
> +    if (((tmp >> 44) & 0xf) == 0x1)
> +        flags |= AV_CPU_FLAG_DOTPROD;
> +    get_cpu_feature_reg(ID_AA64ISAR1_EL1, tmp);
> +    if (((tmp >> 52) & 0xf) == 0x1)
> +        flags |= AV_CPU_FLAG_I8MM;
> +#endif

NEON detection could be added here, though I've yet to see an Armv8 
implementation without AdvSIMD.

FWIW, DotProd is exposed as HWCAP_ASIMDDP and I8MM is exposed via HWCAP2_I8MM, 
using trapped ID registers is not (yet) necessary.

> +
> +    return flags;
> +}
> +
> +#else
> +
> +static int detect_flags(void)
> +{
> +    return 0;
> +}
> +
> +#endif
> +
>  int ff_get_cpu_flags_aarch64(void)
>  {
>      int flags = AV_CPU_FLAG_ARMV8 * HAVE_ARMV8 |
> @@ -33,6 +69,8 @@ int ff_get_cpu_flags_aarch64(void)
>      flags |= AV_CPU_FLAG_I8MM;
>  #endif
> 
> +    flags |= detect_flags();
> +
>      return flags;
>  }

On Sat, 27 May 2023, Rémi Denis-Courmont wrote:

> Le perjantaina 26. toukokuuta 2023, 11.03.14 EEST Martin Storsjö a écrit :
>> Based on code by Janne Grunau.
>> 
>> Using HWCAP_CPUID for user space access to the CPU feature registers. See
>> https://www.kernel.org/doc/html/latest/arm64/cpu-feature-registers.html.
>> ---
>>  configure               |  2 ++
>>  libavutil/aarch64/cpu.c | 38 ++++++++++++++++++++++++++++++++++++++
>>  2 files changed, 40 insertions(+)
>> 
>> diff --git a/configure b/configure
>> index 3c7473efb2..b5357b8d27 100755
>> --- a/configure
>> +++ b/configure
>> @@ -2207,6 +2207,7 @@ HAVE_LIST_PUB="
>>
>>  HEADERS_LIST="
>>      arpa_inet_h
>> +    asm_hwcap_h
>>      asm_types_h
>>      cdio_paranoia_h
>>      cdio_paranoia_paranoia_h
>> @@ -6422,6 +6423,7 @@ check_headers io.h
>>  enabled libdrm &&
>>      check_headers linux/dma-buf.h
>> 
>> +check_headers asm/hwcap.h
>>  check_headers linux/perf_event.h
>>  check_headers libcrystalhd/libcrystalhd_if.h
>>  check_headers malloc.h
>> diff --git a/libavutil/aarch64/cpu.c b/libavutil/aarch64/cpu.c
>> index 42b33e4a2d..34c838c2f5 100644
>> --- a/libavutil/aarch64/cpu.c
>> +++ b/libavutil/aarch64/cpu.c
>> @@ -20,6 +20,42 @@
>>  #include "libavutil/cpu_internal.h"
>>  #include "config.h"
>> 
>> +#if (defined(__linux__) || defined(__ANDROID__)) && HAVE_GETAUXVAL &&
>> HAVE_ASM_HWCAP_H +#include <stdint.h>
>> +#include <asm/hwcap.h>
>> +#include <sys/auxv.h>
>> +
>> +#define get_cpu_feature_reg(reg, val) \
>> +        __asm__("mrs %0, " #reg : "=r" (val))
>
> Strictly speaking, this can read any system register. One way to prevent that 
> would be to include the ID_ prefix and _EL1 suffix in the macro. I would have 
> used a pure static inline instead, but that's just a matter of taste.
>
>> +
>> +static int detect_flags(void)
>> +{
>> +    unsigned long ret = getauxval(AT_HWCAP);
>> +    int flags = 0;
>> +#if defined(HWCAP_CPUID)
>> +    uint64_t tmp;
>> +    if (!(ret & HWCAP_CPUID))
>> +        return flags;
>> +    get_cpu_feature_reg(ID_AA64ISAR0_EL1, tmp);
>> +    if (((tmp >> 44) & 0xf) == 0x1)
>> +        flags |= AV_CPU_FLAG_DOTPROD;
>> +    get_cpu_feature_reg(ID_AA64ISAR1_EL1, tmp);
>> +    if (((tmp >> 52) & 0xf) == 0x1)
>> +        flags |= AV_CPU_FLAG_I8MM;
>> +#endif
>
> NEON detection could be added here, though I've yet to see an Armv8 
> implementation without AdvSIMD.

I guess we could, but as it's part of the require baseline for armv8-a I 
don't think there's much need for it? If configured with --disable-neon we 
don't return that cpuflag though.

> FWIW, DotProd is exposed as HWCAP_ASIMDDP and I8MM is exposed via 
> HWCAP2_I8MM, using trapped ID registers is not (yet) necessary.

Ah, I see. I guess using those would be more straightforward.

OTOH, HWCAP_CPUID is available much earlier than HWCAP_ASIMDDP or 
HWCAP2_I8MM (I do some amount of cross building with a fairly old 
sysroot). I'll think about it, whether it's worth complicating things to 
try both approaches, or if we should just go with the plain HWCAPs here.

// Martin

Le lauantaina 27. toukokuuta 2023, 23.35.14 EEST Martin Storsjö a écrit :
> > NEON detection could be added here, though I've yet to see an Armv8
> > implementation without AdvSIMD.
> 
> I guess we could, but as it's part of the require baseline for armv8-a I
> don't think there's much need for it?

I mean that if you have 95% of the code needed, you might as well add the last 
5%. I don't think it's used by much anything.

> If configured with --disable-neon we
> don't return that cpuflag though.
> 
> > FWIW, DotProd is exposed as HWCAP_ASIMDDP and I8MM is exposed via
> > HWCAP2_I8MM, using trapped ID registers is not (yet) necessary.
> 
> Ah, I see. I guess using those would be more straightforward.
> 
> OTOH, HWCAP_CPUID is available much earlier than HWCAP_ASIMDDP or
> HWCAP2_I8MM (I do some amount of cross building with a fairly old
> sysroot).

If the sysroot shares the not-too-old kernel of the host, then ID registers 
will work anyway. The age of the userspace toolchain affects availability of 
HWCAP constant definitions, not that of ID registers, which is a purely in-
kernel feature.

> I'll think about it, whether it's worth complicating things to
> try both approaches, or if we should just go with the plain HWCAPs here.

I don't really mind either way.