[dpdk-dev] [PATCH] acl: If build does not support sse4.2, emulate missing instructions with C code

[Ananyev, Konstantin]

Hi Neil,

> From: Neil Horman [mailto:nhorman at tuxdriver.com]
> Sent: Monday, August 04, 2014 4:36 PM
> To: dev at dpdk.org
> Cc: Neil Horman; Thomas Monjalon; Ananyev, Konstantin; Richardson, Bruce
> Subject: [PATCH] acl: If build does not support sse4.2, emulate missing instructions with C code
> 
> The ACL library makes extensive use of some SSE4.2 instructions, which means the
> default build can't compile this library.  Work around the problem by testing
> the __SSE42__ definition in the acl_vects.h file and defining the macros there
> as intrinsics or c-level equivalants.  Note this is a minimal patch, adjusting
> only the definitions that are currently used in the ACL library.
> 

My comments about actual implementations of c-level equivalents below.
None of them look correct to me.
Of course it could be fixed.
Though I am not sure that is a right way to proceed:
At first I really doubt that these equivalents will provide similar performance.
As you probably note - we do have a scalar version of  rte_acl_classify():  rte_acl_classify_scalar().
So I think it might be faster than vector one with 'emulated' instrincts. 
Unfortunately it is all mixed right now into one file and 'scalar' version could use few sse4 instrincts through resolve_priority().  
Another thing - we consider to add another version of rte_acl_classify() that will use avx2 instrincts.
If we go the way you suggest - I am afraid will soon have to provide scalar equivalents for several AVX2 instrincts too.
So in summary that way (providing our own scalar equivalents of SIMD instrincts) seems to me slow, hard to maintain and error prone.

What porbably can be done instead: rework acl_run.c a bit, so it provide:
rte_acl_classify_scalar() - could be build and used on all systems.
rte_acl_classify_sse() - could be build and used only on systems with sse4.2 and upper, return ENOTSUP on lower arch.
In future: rte_acl_classify_avx2 - could be build and used only on systems with avx2 and upper, return ENOTSUP on lower arch.

I am looking at rte_acl right now anyway.
So will try to come up with something workable.

> Only compile tested so far, but I wanted to post it for early review so that
> others could aid in unit testing.

Could you please stop submitting untested patches.
It is common (and good) practice to do at least some minimal testing before submitting your code changes.
At the end, it will save your own and other people time.
For ACL there is a simple UT, that could be run as:
./x86_64-native-linuxapp-gcc/app/test ...
RTE>>acl_autotest
It takes just few seconds to run.
Also if you plan further development for ACL, we probably can provide you with extra test-cases that we use for functional testing.
Thanks
Konstantin

> 
> Signed-off-by: Neil Horman <nhorman at tuxdriver.com>
> CC: Thomas Monjalon <thomas.monjalon at 6wind.com>
> CC: "Konstantin Ananyev" <konstantin.ananyev at intel.com>
> CC: Bruce Richardson <bruce.richardson at intel.com>
> ---
>  lib/librte_acl/acl_bld.c  |   3 +-
>  lib/librte_acl/acl_vect.h | 102 ++++++++++++++++++++++++++++++++++++++++++++--
>  2 files changed, 100 insertions(+), 5 deletions(-)
> 
> diff --git a/lib/librte_acl/acl_bld.c b/lib/librte_acl/acl_bld.c
> index 873447b..de974a4 100644
> --- a/lib/librte_acl/acl_bld.c
> +++ b/lib/librte_acl/acl_bld.c
> @@ -31,7 +31,6 @@
>   *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
>   */
> 
> -#include <nmmintrin.h>
>  #include <rte_acl.h>
>  #include "tb_mem.h"
>  #include "acl.h"
> @@ -1481,7 +1480,7 @@ acl_calc_wildness(struct rte_acl_build_rule *head,
>  			switch (rule->config->defs[n].type) {
>  			case RTE_ACL_FIELD_TYPE_BITMASK:
>  				wild = (size -
> -					_mm_popcnt_u32(fld->mask_range.u8)) /
> +					__builtin_popcountl(fld->mask_range.u8)) /
>  					size;
>  				break;
> 
> diff --git a/lib/librte_acl/acl_vect.h b/lib/librte_acl/acl_vect.h
> index d813600..e5f391b 100644
> --- a/lib/librte_acl/acl_vect.h
> +++ b/lib/librte_acl/acl_vect.h
> @@ -34,6 +34,10 @@
>  #ifndef _RTE_ACL_VECT_H_
>  #define _RTE_ACL_VECT_H_
> 
> +#ifdef __SSE4_1__
> +#include <smmintrin.h>
> +#endif
> +
>  /**
>   * @file
>   *
> @@ -44,12 +48,12 @@
>  extern "C" {
>  #endif
> 
> +
>  #define	MM_ADD16(a, b)		_mm_add_epi16(a, b)
>  #define	MM_ADD32(a, b)		_mm_add_epi32(a, b)
>  #define	MM_ALIGNR8(a, b, c)	_mm_alignr_epi8(a, b, c)
>  #define	MM_AND(a, b)		_mm_and_si128(a, b)
>  #define MM_ANDNOT(a, b)		_mm_andnot_si128(a, b)
> -#define MM_BLENDV8(a, b, c)	_mm_blendv_epi8(a, b, c)
>  #define MM_CMPEQ16(a, b)	_mm_cmpeq_epi16(a, b)
>  #define MM_CMPEQ32(a, b)	_mm_cmpeq_epi32(a, b)
>  #define	MM_CMPEQ8(a, b)		_mm_cmpeq_epi8(a, b)
> @@ -59,7 +63,6 @@ extern "C" {
>  #define	MM_CVT32(a)		_mm_cvtsi128_si32(a)
>  #define MM_CVTU32(a)		_mm_cvtsi32_si128(a)
>  #define	MM_INSERT16(a, c, b)	_mm_insert_epi16(a, c, b)
> -#define	MM_INSERT32(a, c, b)	_mm_insert_epi32(a, c, b)
>  #define	MM_LOAD(a)		_mm_load_si128(a)
>  #define	MM_LOADH_PI(a, b)	_mm_loadh_pi(a, b)
>  #define	MM_LOADU(a)		_mm_loadu_si128(a)
> @@ -82,7 +85,6 @@ extern "C" {
>  #define	MM_SRL32(a, b)		_mm_srli_epi32(a, b)
>  #define	MM_STORE(a, b)		_mm_store_si128(a, b)
>  #define	MM_STOREU(a, b)		_mm_storeu_si128(a, b)
> -#define	MM_TESTZ(a, b)		_mm_testz_si128(a, b)
>  #define	MM_XOR(a, b)		_mm_xor_si128(a, b)
> 
>  #define	MM_SET16(a, b, c, d, e, f, g, h)	\
> @@ -93,6 +95,100 @@ extern "C" {
>  	_mm_set_epi8(c0, c1, c2, c3, c4, c5, c6, c7,	\
>  		c8, c9, cA, cB, cC, cD, cE, cF)
> 
> +
> +#ifndef __SSE4_1__
> +static inline xmm_t pblendvb(xmm_t dst, xmm_t src, xmm_t mask)
> +{
> +	unsigned char tmpd[16], tmps[16], tmpm[16];
> +	int i;
> +
> +	MM_STOREU((xmm_t *)&tmpd, dst);
> +	MM_STOREU((xmm_t *)&tmps, src);
> +	MM_STOREU((xmm_t *)&tmpm, mask);
> +

Generic comment:
You don't need to use char temp[16] or other stuff each time you need to swap between __m128i and char[16];
Specially for that purpose inside rte_common_vect.h we have:
typedef union rte_xmm {
        xmm_t    m;
        uint8_t  u8[XMM_SIZE / sizeof(uint8_t)];
        uint16_t u16[XMM_SIZE / sizeof(uint16_t)];
        uint32_t u32[XMM_SIZE / sizeof(uint32_t)];
        uint64_t u64[XMM_SIZE / sizeof(uint64_t)];
        double   pd[XMM_SIZE / sizeof(double)];
} rte_xmm_t;


> +	for (i = 0; i < 16; i++)
> +		if (mask[i] & 0x8)

Should be tmpm, not mask.
Wonder how did it compile without warnings?

> +			dst[i] = src[i];
> +


Reading Intel Architecture Manual:
MASK <-- SRC3
IF (MASK[7] = 1) THEN DEST[7:0]  SRC2[7:0];
ELSE DEST[7:0] <-- SRC1[7:0];
IF (MASK[15] = 1) THEN DEST[15:8]  SRC2[15:8];
ELSE DEST[15:8] <-- SRC1[15:8];
...
So it should be 'tmpm[i] & 0x80'

> +	dst = MM_LOADU((xmm_t *)&tmpd);
> +
> +	return dst;
> +}

Taking into the account all of the above, I think it should be something like that:

static inline xmm_t
pblendvb(xmm_t dst, xmm_t src, xmm_t mask)
{
        rte_xmm_t d, m, s;
        int32_t i;

        d.m = dst;
        s.m = src;
        m.m = mask;

        for (i = 0; i != RTE_DIM(m.u8); i++)
                if ((m.u8[i] & INT8_MIN) != 0)
                        d.u8[i] = s.u8[i];

        return (d.m);
}

> +
> +#define MM_BLENDV8(a, b, c)	pblendvb(a, b, c)
> +
> +
> +static inline int ptestz(xmm_t a, xmm_t b)
> +{
> +	unsigned long long tmpa[2], tmpb[2];
> +
> +	MM_STOREU((xmm_t *)&tmpa, a);
> +	MM_STOREU((xmm_t *)&tmpb, b);
> +
> +	if (tmpa[0] & tmpb[0])
> +		return 1;
> +	if (tmpa[1] & tmpb[1])
> +		return 1;
> +
> +	return 0;
> +}

Again from IA manual:
(V)PTEST (128-bit version)
IF (SRC[127:0] BITWISE AND DEST[127:0] = 0)
THEN ZF <-- 1;
ELSE ZF <-- 0;
IF (SRC[127:0] BITWISE AND NOT DEST[127:0] = 0)
THEN CF <-- 1;
ELSE CF <-- 0;

So _mm_testz_si128 (__m128i s1, __m128i s2) returns 1 if the bitwise AND operation on s1 and s2 results in all zeros, else returns 0.
Yours one doing opposite stuff.
Should be:

static inline int
ptestz(xmm_t a, xmm_t b)
{
        rte_xmm_t x, y;

        x.m = a;
        y.m = b;

        return ((x.u64[0] & y.u64[0]) == 0 && (x.u64[1] & y.u64[1]) == 0);
}

> +
> +#define	MM_TESTZ(a, b)		ptestz(a, b)
> +
> +static inline xmm_t pinsrd(xmm_t dst, int32_t val, char off)
> +{
> +	unsigned long long tmpa[2];
> +	unsigned long long mask;
> +	int32_t tmp;
> +
> +	MM_STOREU((xmm_t *)&tmpa, dst);
> +
> +	/*
> +	 * Inserting a dword is a bit odd as it can cross a word boundary
> +	 */
> +
> +	if (off > 32) {
> +		/*
> +		 * If the offset is more than 32, then part of the
> +		 * inserted word will appear in the upper half of the xmm
> +		 * register.  Grab the part of the value that crosses the 64 bit
> +		 * boundary.
> +		 */
> +		tmp = val >> (off - 32);
> +
> +		/*
> +		 * Mask off the least significant bits of the upper longword
> +		 */
> +		mask = ~((1 << (off - 32)) - 1);
> +		tmpa[1] &= mask;
> +
> +		/*
> +		 * and insert the new value
> +		 */
> +		tmpa[1] |= tmp;
> +	}
> +	if (off < 64) {
> +		/*
> +		 * If the offset is less than 64 bits, we also need to mask and
> +		 * assign the lower longword
> +		 */
> +		mask = (1 << off) - 1;
> +		tmpa[0] &= mask;
> +		tmpa[0] |= (val << off);
> +	}
> +
> +	dst = MM_LOADU((xmm_t *)&tmpa);
> +	return dst;
> +}

Again from IA manual:
PINSRD: SEL <-- COUNT[1:0];
MASK <-- (0FFFFFFFFH << (SEL * 32));
TEMP <-- (((SRC << (SEL *32)) AND MASK) ;
DEST <-- ((DEST AND NOT MASK) OR TEMP);

So basically, all we need is to treat dst as array of 4 integers.
Off - is just an index [0-3] where to insert val:

Static inline xmm_t
pinsrd(xmm_t dst, int32_t val, uint32_t off)
{
        rte_xmm_t x;
        int32_t idx;

        x.m = dst;
        idx = off % RTE_DIM(x.u32);
        x.u32[idx] = val;
        return x.m;
}

> +
> +#define	MM_INSERT32(a, c, b)	pinsrd(a, c, b)
> +
> +#else
> +#define	MM_BLENDV8(a, b, c)	_mm_blendv_epi8(a, b, c)
> +#define	MM_TESTZ(a, b)		_mm_testz_si128(a, b)
> +#define	MM_INSERT32(a, c, b)	_mm_insert_epi32(a, c, b)
> +#endif
> +
>  #ifdef RTE_ARCH_X86_64
> 
>  #define	MM_CVT64(a)		_mm_cvtsi128_si64(a)
> --
> 1.8.3.1