[v4,1/8] eal: introduce zmm type for AVX 512-bit
Checks
Commit Message
New data type to manipulate 512 bit AVX values.
Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
---
lib/librte_eal/x86/include/rte_vect.h | 21 +++++++++++++++++++++
1 file changed, 21 insertions(+)
Comments
On Wed, Jul 8, 2020 at 10:17 PM Vladimir Medvedkin
<vladimir.medvedkin@intel.com> wrote:
>
> New data type to manipulate 512 bit AVX values.
The title mentions a "zmm" type that is not added by this patch.
Maybe instead, "eal/x86: introduce AVX 512-bit type"
>
> Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
> Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
> ---
> lib/librte_eal/x86/include/rte_vect.h | 21 +++++++++++++++++++++
> 1 file changed, 21 insertions(+)
>
> diff --git a/lib/librte_eal/x86/include/rte_vect.h b/lib/librte_eal/x86/include/rte_vect.h
> index df5a60762..ae59126bc 100644
> --- a/lib/librte_eal/x86/include/rte_vect.h
> +++ b/lib/librte_eal/x86/include/rte_vect.h
> @@ -13,6 +13,7 @@
>
> #include <stdint.h>
> #include <rte_config.h>
> +#include <rte_common.h>
> #include "generic/rte_vect.h"
>
> #if (defined(__ICC) || \
> @@ -90,6 +91,26 @@ __extension__ ({ \
> })
> #endif /* (defined(__ICC) && __ICC < 1210) */
>
> +#ifdef __AVX512F__
> +
> +typedef __m512i __x86_zmm_t;
We don't need this interim type, using the native __m512 is enough afaics.
Looking at the whole applied series:
$ git grep -lw __x86_zmm_t
lib/librte_eal/x86/include/rte_vect.h
> +
> +#define ZMM_SIZE (sizeof(__x86_zmm_t))
> +#define ZMM_MASK (ZMM_SIZE - 1)
Macros in a public header need a RTE_ prefix + this is x86 specific,
then RTE_X86_.
Looking at the whole applied series:
$ git grep -lw ZMM_SIZE
lib/librte_eal/x86/include/rte_vect.h
$ git grep -lw ZMM_MASK
lib/librte_eal/x86/include/rte_vect.h
So I wonder if we need to export it or we can instead just #undef
after the struct definition.
> +
> +typedef union __rte_x86_zmm {
> + __x86_zmm_t z;
> + ymm_t y[ZMM_SIZE / sizeof(ymm_t)];
> + xmm_t x[ZMM_SIZE / sizeof(xmm_t)];
> + uint8_t u8[ZMM_SIZE / sizeof(uint8_t)];
> + uint16_t u16[ZMM_SIZE / sizeof(uint16_t)];
> + uint32_t u32[ZMM_SIZE / sizeof(uint32_t)];
> + uint64_t u64[ZMM_SIZE / sizeof(uint64_t)];
> + double pd[ZMM_SIZE / sizeof(double)];
> +} __rte_aligned(ZMM_SIZE) __rte_x86_zmm_t;
I don't understand this forced alignment statement.
Would not natural alignment be enough, since all fields in this union
have the same size?
Hi David,
Thanks for review
On 09/07/2020 14:48, David Marchand wrote:
> On Wed, Jul 8, 2020 at 10:17 PM Vladimir Medvedkin
> <vladimir.medvedkin@intel.com> wrote:
>>
>> New data type to manipulate 512 bit AVX values.
>
> The title mentions a "zmm" type that is not added by this patch.
>
> Maybe instead, "eal/x86: introduce AVX 512-bit type"
>
Agree
>
>>
>> Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
>> Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
>> ---
>> lib/librte_eal/x86/include/rte_vect.h | 21 +++++++++++++++++++++
>> 1 file changed, 21 insertions(+)
>>
>> diff --git a/lib/librte_eal/x86/include/rte_vect.h b/lib/librte_eal/x86/include/rte_vect.h
>> index df5a60762..ae59126bc 100644
>> --- a/lib/librte_eal/x86/include/rte_vect.h
>> +++ b/lib/librte_eal/x86/include/rte_vect.h
>> @@ -13,6 +13,7 @@
>>
>> #include <stdint.h>
>> #include <rte_config.h>
>> +#include <rte_common.h>
>> #include "generic/rte_vect.h"
>>
>> #if (defined(__ICC) || \
>> @@ -90,6 +91,26 @@ __extension__ ({ \
>> })
>> #endif /* (defined(__ICC) && __ICC < 1210) */
>>
>> +#ifdef __AVX512F__
>> +
>> +typedef __m512i __x86_zmm_t;
>
> We don't need this interim type, using the native __m512 is enough afaics.
>
Agree
> Looking at the whole applied series:
> $ git grep -lw __x86_zmm_t
> lib/librte_eal/x86/include/rte_vect.h
>
>
>> +
>> +#define ZMM_SIZE (sizeof(__x86_zmm_t))
>> +#define ZMM_MASK (ZMM_SIZE - 1)
>
> Macros in a public header need a RTE_ prefix + this is x86 specific,
> then RTE_X86_.
>
> Looking at the whole applied series:
> $ git grep -lw ZMM_SIZE
> lib/librte_eal/x86/include/rte_vect.h
> $ git grep -lw ZMM_MASK
> lib/librte_eal/x86/include/rte_vect.h
>
> So I wonder if we need to export it or we can instead just #undef
> after the struct definition.
I think it's better to undef it
>
>
>> +
>> +typedef union __rte_x86_zmm {
>> + __x86_zmm_t z;
>> + ymm_t y[ZMM_SIZE / sizeof(ymm_t)];
>> + xmm_t x[ZMM_SIZE / sizeof(xmm_t)];
>> + uint8_t u8[ZMM_SIZE / sizeof(uint8_t)];
>> + uint16_t u16[ZMM_SIZE / sizeof(uint16_t)];
>> + uint32_t u32[ZMM_SIZE / sizeof(uint32_t)];
>> + uint64_t u64[ZMM_SIZE / sizeof(uint64_t)];
>> + double pd[ZMM_SIZE / sizeof(double)];
>> +} __rte_aligned(ZMM_SIZE) __rte_x86_zmm_t;
>
> I don't understand this forced alignment statement.
> Would not natural alignment be enough, since all fields in this union
> have the same size?
>
Some compilers won't align this union
https://mails.dpdk.org/archives/dev/2020-March/159591.html
>
On Thu, Jul 9, 2020 at 4:52 PM Medvedkin, Vladimir
<vladimir.medvedkin@intel.com> wrote:
> >> +
> >> +#define ZMM_SIZE (sizeof(__x86_zmm_t))
> >> +#define ZMM_MASK (ZMM_SIZE - 1)
> >
> > Macros in a public header need a RTE_ prefix + this is x86 specific,
> > then RTE_X86_.
> >
> > Looking at the whole applied series:
> > $ git grep -lw ZMM_SIZE
> > lib/librte_eal/x86/include/rte_vect.h
> > $ git grep -lw ZMM_MASK
> > lib/librte_eal/x86/include/rte_vect.h
> >
> > So I wonder if we need to export it or we can instead just #undef
> > after the struct definition.
>
> I think it's better to undef it
Even if you undef the macro, please still prefix it.
This is to avoid conflicts with macros defined before including this
rte_vect.h header.
>
> >
> >
> >> +
> >> +typedef union __rte_x86_zmm {
> >> + __x86_zmm_t z;
> >> + ymm_t y[ZMM_SIZE / sizeof(ymm_t)];
> >> + xmm_t x[ZMM_SIZE / sizeof(xmm_t)];
> >> + uint8_t u8[ZMM_SIZE / sizeof(uint8_t)];
> >> + uint16_t u16[ZMM_SIZE / sizeof(uint16_t)];
> >> + uint32_t u32[ZMM_SIZE / sizeof(uint32_t)];
> >> + uint64_t u64[ZMM_SIZE / sizeof(uint64_t)];
> >> + double pd[ZMM_SIZE / sizeof(double)];
> >> +} __rte_aligned(ZMM_SIZE) __rte_x86_zmm_t;
> >
> > I don't understand this forced alignment statement.
> > Would not natural alignment be enough, since all fields in this union
> > have the same size?
> >
>
> Some compilers won't align this union
> https://mails.dpdk.org/archives/dev/2020-March/159591.html
Ok, interesting, I will try to keep in mind.
@@ -13,6 +13,7 @@
#include <stdint.h>
#include <rte_config.h>
+#include <rte_common.h>
#include "generic/rte_vect.h"
#if (defined(__ICC) || \
@@ -90,6 +91,26 @@ __extension__ ({ \
})
#endif /* (defined(__ICC) && __ICC < 1210) */
+#ifdef __AVX512F__
+
+typedef __m512i __x86_zmm_t;
+
+#define ZMM_SIZE (sizeof(__x86_zmm_t))
+#define ZMM_MASK (ZMM_SIZE - 1)
+
+typedef union __rte_x86_zmm {
+ __x86_zmm_t z;
+ ymm_t y[ZMM_SIZE / sizeof(ymm_t)];
+ xmm_t x[ZMM_SIZE / sizeof(xmm_t)];
+ uint8_t u8[ZMM_SIZE / sizeof(uint8_t)];
+ uint16_t u16[ZMM_SIZE / sizeof(uint16_t)];
+ uint32_t u32[ZMM_SIZE / sizeof(uint32_t)];
+ uint64_t u64[ZMM_SIZE / sizeof(uint64_t)];
+ double pd[ZMM_SIZE / sizeof(double)];
+} __rte_aligned(ZMM_SIZE) __rte_x86_zmm_t;
+
+#endif /* __AVX512F__ */
+
#ifdef __cplusplus
}
#endif