[1/6] 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>
---
lib/librte_eal/common/include/arch/x86/rte_vect.h | 20 ++++++++++++++++++++
1 file changed, 20 insertions(+)
Comments
On Mon, Mar 9, 2020 at 6:14 PM Vladimir Medvedkin
<vladimir.medvedkin@intel.com> wrote:
>
> New data type to manipulate 512 bit AVX values.
>
> Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
> ---
> lib/librte_eal/common/include/arch/x86/rte_vect.h | 20 ++++++++++++++++++++
> 1 file changed, 20 insertions(+)
>
> diff --git a/lib/librte_eal/common/include/arch/x86/rte_vect.h b/lib/librte_eal/common/include/arch/x86/rte_vect.h
> index df5a607..09f30e6 100644
> --- a/lib/librte_eal/common/include/arch/x86/rte_vect.h
> +++ b/lib/librte_eal/common/include/arch/x86/rte_vect.h
> @@ -90,6 +90,26 @@ __extension__ ({ \
> })
> #endif /* (defined(__ICC) && __ICC < 1210) */
>
> +#ifdef __AVX512F__
> +
> +typedef __m512i zmm_t;
> +
> +#define ZMM_SIZE (sizeof(zmm_t))
> +#define ZMM_MASK (ZMM_SIZE - 1)
> +
> +typedef union rte_zmm {
> + 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)];
Are we missing __attribute__((aligned(64))) here?
> +} rte_zmm_t;
IMO, Due to legacy reason, we have selected rte_xmm_t, rte_ymm_t for
128 and 256 operations in public APIs[1]
# Not sure where xmm_t and ymm_t and new zmm_t come from? Is this name
x86 arch-specific? If so,
why not give the more generic name rte_512i_t or something?
# Currently, In every arch file, we are repeating the definition for
rte_xmm_t, Why not make, this generic definition
in common file. ie. rte_zmm_t or rte_512i_t definition in common
file(./lib/librte_eal/common/include/generic/rte_vect.h)
# Currently ./lib/librte_eal/common/include/generic/rte_vect.h has
defintion for rte_vXsY_t for vector representation, would that
be enough for public API? Do we need to new type?
[1]
rte_lpm_lookupx4(const struct rte_lpm *lpm, xmm_t ip, uint32_t hop[4],
uint32_t defv)
> +
> +#endif /* __AVX512F__ */
> +
> #ifdef __cplusplus
> }
> #endif
> --
> 2.7.4
>
Hi Jerin,
On 09/03/2020 16:39, Jerin Jacob wrote:
> On Mon, Mar 9, 2020 at 6:14 PM Vladimir Medvedkin
> <vladimir.medvedkin@intel.com> wrote:
>> New data type to manipulate 512 bit AVX values.
>>
>> Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
>> ---
>> lib/librte_eal/common/include/arch/x86/rte_vect.h | 20 ++++++++++++++++++++
>> 1 file changed, 20 insertions(+)
>>
>> diff --git a/lib/librte_eal/common/include/arch/x86/rte_vect.h b/lib/librte_eal/common/include/arch/x86/rte_vect.h
>> index df5a607..09f30e6 100644
>> --- a/lib/librte_eal/common/include/arch/x86/rte_vect.h
>> +++ b/lib/librte_eal/common/include/arch/x86/rte_vect.h
>> @@ -90,6 +90,26 @@ __extension__ ({ \
>> })
>> #endif /* (defined(__ICC) && __ICC < 1210) */
>>
>> +#ifdef __AVX512F__
>> +
>> +typedef __m512i zmm_t;
>> +
>> +#define ZMM_SIZE (sizeof(zmm_t))
>> +#define ZMM_MASK (ZMM_SIZE - 1)
>> +
>> +typedef union rte_zmm {
>> + 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)];
> Are we missing __attribute__((aligned(64))) here?
Agree. While modern compilers align __m512i by default, some old could
failure to align. Please correct me if I'm wrong.
>
>> +} rte_zmm_t;
> IMO, Due to legacy reason, we have selected rte_xmm_t, rte_ymm_t for
> 128 and 256 operations in public APIs[1]
As for me, since these functions are inlined, prototype should be
changed to uint32_t ip[4] instead of passing vector type as an argument.
> # Not sure where xmm_t and ymm_t and new zmm_t come from? Is this name
> x86 arch-specific?
Yes, that's why they are in arch/x86/rte_vect.h
> If so,
> why not give the more generic name rte_512i_t or something?
> # Currently, In every arch file, we are repeating the definition for
> rte_xmm_t, Why not make, this generic definition
> in common file. ie. rte_zmm_t or rte_512i_t definition in common
> file(./lib/librte_eal/common/include/generic/rte_vect.h)
I think there could be some arch specific thing that prevents it from
being generic.
> # Currently ./lib/librte_eal/common/include/generic/rte_vect.h has
> defintion for rte_vXsY_t for vector representation, would that
> be enough for public API? Do we need to new type?
Definitions for rte_vXsY_tare almost the same as compiler's
__m[128,256,512]i apart from alignment.
Union types such as rte_zmm_t are very useful because of the ability to
access parts of a wide vector register with an arbitrary granularity.
For example, some old compiler don't support
_mm512_set_epi8()/_mm512_set_epi16() intrinsics, so accessing ".u8[]" of
".u16[]" solves the problem.
>
>
> [1]
> rte_lpm_lookupx4(const struct rte_lpm *lpm, xmm_t ip, uint32_t hop[4],
> uint32_t defv)
>
>
>> +
>> +#endif /* __AVX512F__ */
>> +
>> #ifdef __cplusplus
>> }
>> #endif
>> --
>> 2.7.4
>>
On Tue, Mar 10, 2020 at 8:14 PM Medvedkin, Vladimir
<vladimir.medvedkin@intel.com> wrote:
>
> Hi Jerin,
Hi Vladimir,
>
> Are we missing __attribute__((aligned(64))) here?
>
> Agree. While modern compilers align __m512i by default, some old could failure to align. Please correct me if I'm wrong.
Yes.
>
> +} rte_zmm_t;
>
> IMO, Due to legacy reason, we have selected rte_xmm_t, rte_ymm_t for
> 128 and 256 operations in public APIs[1]
>
> As for me, since these functions are inlined, prototype should be changed to uint32_t ip[4] instead of passing vector type as an argument.
OK. Makes sense.
> # Not sure where xmm_t and ymm_t and new zmm_t come from? Is this name
> x86 arch-specific?
>
> Yes, that's why they are in arch/x86/rte_vect.h
See the last comment.
>
> If so,
> why not give the more generic name rte_512i_t or something?
> # Currently, In every arch file, we are repeating the definition for
> rte_xmm_t, Why not make, this generic definition
> in common file. ie. rte_zmm_t or rte_512i_t definition in common
> file(./lib/librte_eal/common/include/generic/rte_vect.h)
>
> I think there could be some arch specific thing that prevents it from being generic.
>
> # Currently ./lib/librte_eal/common/include/generic/rte_vect.h has
> defintion for rte_vXsY_t for vector representation, would that
> be enough for public API? Do we need to new type?
>
> Definitions for rte_vXsY_tare almost the same as compiler's __m[128,256,512]i apart from alignment.
> Union types such as rte_zmm_t are very useful because of the ability to access parts of a wide vector register with an arbitrary granularity. For example, some old compiler don't support _mm512_set_epi8()/_mm512_set_epi16() intrinsics, so accessing ".u8[]" of ".u16[]" solves the problem.
Yes. We are on the same page.
I think, the only difference in thought is, the x86 specific
definition(rte_zmm_t) name should be something
it needs to be reflected as internal or arch-specific. Earlier APIs
such rte_lpm_lookupx4 has leaked
the xmm_t definition to public API.
To avoid that danger, please make rte_zmm_t as internal/arch-specific.
Something __rte_x86_zmm_t or
so that denotes it is not a public symbol.
@@ -90,6 +90,26 @@ __extension__ ({ \
})
#endif /* (defined(__ICC) && __ICC < 1210) */
+#ifdef __AVX512F__
+
+typedef __m512i zmm_t;
+
+#define ZMM_SIZE (sizeof(zmm_t))
+#define ZMM_MASK (ZMM_SIZE - 1)
+
+typedef union rte_zmm {
+ 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_zmm_t;
+
+#endif /* __AVX512F__ */
+
#ifdef __cplusplus
}
#endif