[dpdk-dev] [PATCH v2 00/17] ACL: New AVX2 classify method and several other enhancements.
Thomas Monjalon
thomas.monjalon at 6wind.com
Mon Jan 19 18:16:02 CET 2015
2015-01-14 13:39, Neil Horman:
> On Mon, Jan 12, 2015 at 07:16:04PM +0000, Konstantin Ananyev wrote:
> > v2 changes:
> > - When build with the compilers that don't support AVX2 instructions,
> > make rte_acl_classify_avx2() do nothing and return an error.
> > - Remove unneeded 'ifdef __AVX2__' in acl_run_avx2.*.
> > - Reorder order of patches in the set, to keep RTE_LIBRTE_ACL_STANDALONE=y
> > always buildable.
> >
> > This patch series contain several fixes and enhancements for ACL library.
> > See complete list below.
> > Two main changes that are externally visible:
> > - Introduce new classify method: RTE_ACL_CLASSIFY_AVX2.
> > It uses AVX2 instructions and 256 bit wide data types
> > to perform internal trie traversal.
> > That helps to increase classify() throughput.
> > This method is selected as default one on CPUs that supports AVX2.
> > - Introduce new field in the build config structure: max_size.
> > It specifies maximum size that internal RT structure for given context
> > can reach.
> > The purpose of that is to allow user to decide about space/performance trade-off
> > (faster classify() vs less space for RT internal structures)
> > for each given set of rules.
> >
> > Konstantin Ananyev (17):
> > fix fix compilation issues with RTE_LIBRTE_ACL_STANDALONE=y
> > app/test: few small fixes fot test_acl.c
> > librte_acl: make data_indexes long enough to survive idle transitions.
> > librte_acl: remove build phase heuristsic with negative perfomance
> > effect.
> > librte_acl: fix a bug at build phase that can cause matches beeing
> > overwirtten.
> > librte_acl: introduce DFA nodes compression (group64) for identical
> > entries.
> > librte_acl: build/gen phase - simplify the way match nodes are
> > allocated.
> > librte_acl: make scalar RT code to be more similar to vector one.
> > librte_acl: a bit of RT code deduplication.
> > EAL: introduce rte_ymm and relatives in rte_common_vect.h.
> > librte_acl: add AVX2 as new rte_acl_classify() method
> > test-acl: add ability to manually select RT method.
> > librte_acl: Remove search_sse_2 and relatives.
> > libter_acl: move lo/hi dwords shuffle out from calc_addr
> > libte_acl: make calc_addr a define to deduplicate the code.
> > libte_acl: introduce max_size into rte_acl_config.
> > libte_acl: remove unused macros.
> >
> > app/test-acl/main.c | 126 +++--
> > app/test/test_acl.c | 8 +-
> > examples/l3fwd-acl/main.c | 3 +-
> > examples/l3fwd/main.c | 2 +-
> > lib/librte_acl/Makefile | 18 +
> > lib/librte_acl/acl.h | 58 ++-
> > lib/librte_acl/acl_bld.c | 392 +++++++---------
> > lib/librte_acl/acl_gen.c | 268 +++++++----
> > lib/librte_acl/acl_run.h | 7 +-
> > lib/librte_acl/acl_run_avx2.c | 54 +++
> > lib/librte_acl/acl_run_avx2.h | 284 ++++++++++++
> > lib/librte_acl/acl_run_scalar.c | 65 ++-
> > lib/librte_acl/acl_run_sse.c | 585 +-----------------------
> > lib/librte_acl/acl_run_sse.h | 357 +++++++++++++++
> > lib/librte_acl/acl_vect.h | 132 +++---
> > lib/librte_acl/rte_acl.c | 47 +-
> > lib/librte_acl/rte_acl.h | 4 +
> > lib/librte_acl/rte_acl_osdep_alone.h | 47 +-
> > lib/librte_eal/common/include/rte_common_vect.h | 39 +-
> > lib/librte_lpm/rte_lpm.h | 2 +-
> > 20 files changed, 1444 insertions(+), 1054 deletions(-)
> > create mode 100644 lib/librte_acl/acl_run_avx2.c
> > create mode 100644 lib/librte_acl/acl_run_avx2.h
> > create mode 100644 lib/librte_acl/acl_run_sse.h
> >
> Series
> Acked-by: Neil Horman <nhorman at tuxdriver.com>
Are you sure there is nothing to change or add in the documentation?
Maybe that explaining the space/performance trade-off would be a good idea.
> Nice work
Yes, great work!
--
Thomas
More information about the dev
mailing list