Re: [dpdk-dev] [PATCH v5 16/17] net: add checks for max SIMD bitwidth

["Ananyev, Konstantin" <konstantin.ananyev@intel.com>]

> >
> > > When choosing a vector path to take, an extra condition must be
> > > satisfied to ensure the max SIMD bitwidth allows for the CPU enabled
> > > path.
> > >
> > > The vector path was initially chosen in RTE_INIT, however this is no
> > > longer suitable as we cannot check the max SIMD bitwidth at that time.
> > > Default handlers are now chosen in RTE_INIT, these default handlers
> > > are used the first time the crc calc is called, and they set the suitable
> > > handlers to be used going forward.
> > >
> > > Suggested-by: Jasvinder Singh <jasvinder.singh@intel.com>
> > > Suggested-by: Olivier Matz <olivier.matz@6wind.com>
> > >
> > > Signed-off-by: Ciara Power <ciara.power@intel.com>
> > >
> > > ---
> > > v4:
> > >   - Added default handlers to be set at RTE_INIT time, rather than
> > >     choosing scalar handlers.
> > >   - Modified logging.
> > >   - Updated enum name.
> > > v3:
> > >   - Moved choosing vector paths out of RTE_INIT.
> > >   - Moved checking max_simd_bitwidth into the set_alg function.
> > > ---
> > >  lib/librte_net/rte_net_crc.c | 75 ++++++++++++++++++++++++++++++------
> > >  lib/librte_net/rte_net_crc.h |  8 ++++
> > >  2 files changed, 72 insertions(+), 11 deletions(-)
> > >
> > > diff --git a/lib/librte_net/rte_net_crc.c b/lib/librte_net/rte_net_crc.c
> > > index 4f5b9e8286..11d0161a32 100644
> > > --- a/lib/librte_net/rte_net_crc.c
> > > +++ b/lib/librte_net/rte_net_crc.c
> > > @@ -9,6 +9,7 @@
> > >  #include <rte_cpuflags.h>
> > >  #include <rte_common.h>
> > >  #include <rte_net_crc.h>
> > > +#include <rte_eal.h>
> > >
> > >  #if defined(RTE_ARCH_X86_64) && defined(__PCLMUL__)
> > >  #define X86_64_SSE42_PCLMULQDQ     1
> > > @@ -32,6 +33,12 @@
> > >  static uint32_t crc32_eth_lut[CRC_LUT_SIZE];
> > >  static uint32_t crc16_ccitt_lut[CRC_LUT_SIZE];
> > >
> > > +static uint32_t
> > > +rte_crc16_ccitt_default_handler(const uint8_t *data, uint32_t data_len);
> > > +
> > > +static uint32_t
> > > +rte_crc32_eth_default_handler(const uint8_t *data, uint32_t data_len);
> > > +
> > >  static uint32_t
> > >  rte_crc16_ccitt_handler(const uint8_t *data, uint32_t data_len);
> > >
> > > @@ -41,7 +48,12 @@ rte_crc32_eth_handler(const uint8_t *data, uint32_t data_len);
> > >  typedef uint32_t
> > >  (*rte_net_crc_handler)(const uint8_t *data, uint32_t data_len);
> > >
> > > -static rte_net_crc_handler *handlers;
> > > +static rte_net_crc_handler handlers_default[] = {
> > > +	[RTE_NET_CRC16_CCITT] = rte_crc16_ccitt_default_handler,
> > > +	[RTE_NET_CRC32_ETH] = rte_crc32_eth_default_handler,
> > > +};
> > > +
> > > +static rte_net_crc_handler *handlers = handlers_default;
> > >
> > >  static rte_net_crc_handler handlers_scalar[] = {
> > >  	[RTE_NET_CRC16_CCITT] = rte_crc16_ccitt_handler,
> > > @@ -60,6 +72,9 @@ static rte_net_crc_handler handlers_neon[] = {
> > >  };
> > >  #endif
> > >
> > > +static uint16_t max_simd_bitwidth;
> > > +RTE_LOG_REGISTER(libnet_logtype, lib.net, INFO);
> > > +
> > >  /**
> > >   * Reflect the bits about the middle
> > >   *
> > > @@ -112,6 +127,42 @@ crc32_eth_calc_lut(const uint8_t *data,
> > >  	return crc;
> > >  }
> > >
> > > +static uint32_t
> > > +rte_crc16_ccitt_default_handler(const uint8_t *data, uint32_t data_len)
> > > +{
> > > +	if (max_simd_bitwidth == 0)
> > > +		max_simd_bitwidth = rte_get_max_simd_bitwidth();
> > > +	handlers = handlers_scalar;
> > > +#ifdef X86_64_SSE42_PCLMULQDQ
> > > +	if (max_simd_bitwidth >= RTE_SIMD_128)
> > > +		handlers = handlers_sse42;
> > > +#endif
> > > +#ifdef ARM64_NEON_PMULL
> > > +	if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_PMULL) &&
> > > +			max_simd_bitwidth >= RTE_SIMD_128) {
> > > +		handlers = handlers_neon;
> > > +#endif
> >
> > You probably don't want to make all these checks for *every* invocation
> > of that function. I think it would be better:
> > if (ma_simd_bitwidth == 0) {....}
> > return handlers[..](...);
> 
> As another thougth - it is probably a bit safer to update max_simd_bitwidht
> after handler value update.
> 
> handler = ...; rte_smp_wmb(); max_simd_width = ...;
> 
> >
> > BTW, while it allows us to use best possible handler,
> > such approach means extra indirect call(/jump) anyway.

Ah sorry, that would only happen once, please ignore that one. 

> > Hard to say off-hand would it affect performance,
> > and if yes how significantly.
> > Couldn't find any perf tests in our UT for it...
> >
> > > +	return handlers[RTE_NET_CRC16_CCITT](data, data_len);
> > > +}
> > > +
> > > +static uint32_t
> > > +rte_crc32_eth_default_handler(const uint8_t *data, uint32_t data_len)
> > > +{
> > > +	if (max_simd_bitwidth == 0)
> > > +		max_simd_bitwidth = rte_get_max_simd_bitwidth();
> > > +	handlers = handlers_scalar;
> > > +#ifdef X86_64_SSE42_PCLMULQDQ
> > > +	if (max_simd_bitwidth >= RTE_SIMD_128)
> > > +		handlers = handlers_sse42;
> > > +#endif
> > > +#ifdef ARM64_NEON_PMULL
> > > +	if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_PMULL) &&
> > > +			max_simd_bitwidth >= RTE_SIMD_128) {
> > > +		handlers = handlers_neon;
> > > +#endif
> > > +	return handlers[RTE_NET_CRC32_ETH](data, data_len);
> > > +}
> > > +
> > >  static void
> > >  rte_net_crc_scalar_init(void)
> > >  {
> > > @@ -145,18 +196,26 @@ rte_crc32_eth_handler(const uint8_t *data, uint32_t data_len)
> > >  void
> > >  rte_net_crc_set_alg(enum rte_net_crc_alg alg)
> > >  {
> > > +	if (max_simd_bitwidth == 0)
> > > +		max_simd_bitwidth = rte_get_max_simd_bitwidth();
> > > +
> > >  	switch (alg) {
> > >  #ifdef X86_64_SSE42_PCLMULQDQ
> > >  	case RTE_NET_CRC_SSE42:
> > > -		handlers = handlers_sse42;
> > > -		break;
> > > +		if (max_simd_bitwidth >= RTE_SIMD_128) {
> > > +			handlers = handlers_sse42;
> > > +			return;
> > > +		}
> > > +		NET_LOG(INFO, "Max SIMD Bitwidth too low, can't use SSE\n");
> > >  #elif defined ARM64_NEON_PMULL
> > >  		/* fall-through */
> > >  	case RTE_NET_CRC_NEON:
> > > -		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_PMULL)) {
> > > +		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_PMULL) &&
> > > +				max_simd_bitwidth >= RTE_SIMD_128) {
> > >  			handlers = handlers_neon;
> > > -			break;
> > > +			return;
> > >  		}
> > > +		NET_LOG(INFO, "Max SIMD Bitwidth too low or CPU flag not enabled, can't use NEON\n");
> > >  #endif
> > >  		/* fall-through */
> > >  	case RTE_NET_CRC_SCALAR:
> > > @@ -184,19 +243,13 @@ rte_net_crc_calc(const void *data,
> > >  /* Select highest available crc algorithm as default one */
> > >  RTE_INIT(rte_net_crc_init)
> > >  {
> > > -	enum rte_net_crc_alg alg = RTE_NET_CRC_SCALAR;
> > > -
> > >  	rte_net_crc_scalar_init();
> > >
> > >  #ifdef X86_64_SSE42_PCLMULQDQ
> > > -	alg = RTE_NET_CRC_SSE42;
> > >  	rte_net_crc_sse42_init();
> > >  #elif defined ARM64_NEON_PMULL
> > >  	if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_PMULL)) {
> > > -		alg = RTE_NET_CRC_NEON;
> > >  		rte_net_crc_neon_init();
> > >  	}
> > >  #endif
> > > -
> > > -	rte_net_crc_set_alg(alg);
> > >  }
> > > diff --git a/lib/librte_net/rte_net_crc.h b/lib/librte_net/rte_net_crc.h
> > > index 16e85ca970..c942865ecf 100644
> > > --- a/lib/librte_net/rte_net_crc.h
> > > +++ b/lib/librte_net/rte_net_crc.h
> > > @@ -7,6 +7,8 @@
> > >
> > >  #include <stdint.h>
> > >
> > > +#include <rte_log.h>
> > > +
> > >  #ifdef __cplusplus
> > >  extern "C" {
> > >  #endif
> > > @@ -25,6 +27,12 @@ enum rte_net_crc_alg {
> > >  	RTE_NET_CRC_NEON,
> > >  };
> > >
> > > +extern int libnet_logtype;
> > > +
> > > +#define NET_LOG(level, fmt, args...)					\
> > > +	rte_log(RTE_LOG_ ## level, libnet_logtype, "%s(): " fmt "\n",	\
> > > +		__func__, ## args)
> > > +
> > >  /**
> > >   * This API set the CRC computation algorithm (i.e. scalar version,
> > >   * x86 64-bit sse4.2 intrinsic version, etc.) and internal data
> > > --
> > > 2.22.0