Re: [dpdk-dev] [PATCH 1/3] eal: introduce rte_vect_* abstractions

[Jan Viktorin <viktorin@rehivetech.com>]

On Mon, 30 Nov 2015 22:54:11 +0530
Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote:

> introduce rte_vect_* abstractions to remove SSE/AVX specific
> code in the common code(i.e the test applications)
> 
> The patch does not provide any functional change for IA, the goal is to

Does IA mean Intel Architecture?

> have infrastructure to reuse the common vector-based test code across
> all the architectures.
> 
> Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
> ---
>  lib/librte_eal/common/include/arch/arm/rte_vect.h | 17 ++++++++++++++++-
>  lib/librte_eal/common/include/arch/x86/rte_vect.h |  8 ++++++++
>  2 files changed, 24 insertions(+), 1 deletion(-)
> 
> diff --git a/lib/librte_eal/common/include/arch/arm/rte_vect.h b/lib/librte_eal/common/include/arch/arm/rte_vect.h
> index 21cdb4d..d300951 100644
> --- a/lib/librte_eal/common/include/arch/arm/rte_vect.h
> +++ b/lib/librte_eal/common/include/arch/arm/rte_vect.h
> @@ -33,13 +33,14 @@
>  #ifndef _RTE_VECT_ARM_H_
>  #define _RTE_VECT_ARM_H_
>  
> -#include "arm_neon.h"
> +#include <arm_neon.h>
>  
>  #ifdef __cplusplus
>  extern "C" {
>  #endif
>  
>  typedef int32x4_t xmm_t;
> +typedef int32x4_t __m128i;

As Jianbo pointed out recently, the __m128i type should be refactored in
a general rte_vect API too. If we do something like

#if SSE
typedef __m128i rte_128i;
#elif NEON
typedef int32x4_y rte_128i;
#endif

does it make somebody angry? I am afraid that it will influence a lot of
code. However, from the ABI point of view, it is OK, isn't it?

>  
>  #define	XMM_SIZE	(sizeof(xmm_t))
>  #define	XMM_MASK	(XMM_SIZE - 1)
> @@ -53,6 +54,20 @@ typedef union rte_xmm {
>  	double   pd[XMM_SIZE / sizeof(double)];
>  } __attribute__((aligned(16))) rte_xmm_t;
>  
> +/* rte_vect_* abstraction implementation using NEON */
> +
> +/* loads the __m128i value from address p(does not need to be 16-byte aligned)*/
> +#define rte_vect_loadu_sil128(p) vld1q_s32((const int32_t *)p)
> +
> +/* sets the 4 signed 32-bit integer values and returns the __m128i variable */
> +static inline __m128i  __attribute__((always_inline))
> +rte_vect_set_epi32(int i3, int i2, int i1, int i0)
> +{
> +	int32_t data[4] = {i0, i1, i2, i3};
> +
> +	return vld1q_s32(data);
> +}
> +
>  #ifdef __cplusplus
>  }
>  #endif
> 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 b698797..91c6523 100644
> --- a/lib/librte_eal/common/include/arch/x86/rte_vect.h
> +++ b/lib/librte_eal/common/include/arch/x86/rte_vect.h
> @@ -125,6 +125,14 @@ typedef union rte_ymm {
>  })
>  #endif /* (defined(__ICC) && __ICC < 1210) */
>  
> +/* rte_vect_* abstraction implementation using SSE */
> +
> +/* loads the __m128i value from address p(does not need to be 16-byte aligned)*/
> +#define rte_vect_loadu_sil128(p) _mm_loadu_si128(p)
> +
> +/* sets the 4 signed 32-bit integer values and returns the __m128i variable */
> +#define rte_vect_set_epi32(i3, i2, i1, i0) _mm_set_epi32(i3, i2, i1, i0)
> +
>  #ifdef __cplusplus
>  }
>  #endif

I like this approach. It is a question whether to inherit names from
SSE. However, why to reinvent the wheel...

We probably need other people to give their ideas about such
generalization of the API.

I think, there should be an autotest of the rte_vect API. Is it
possible to create one?

Regards
Jan

-- 
   Jan Viktorin                  E-mail: Viktorin@RehiveTech.com
   System Architect              Web:    www.RehiveTech.com
   RehiveTech
   Brno, Czech Republic