@@ -321,3 +321,33 @@ following command::
``scalar`` can be selected on core ``3`` by the following command::
$ ovs-appctl dpif-netdev/miniflow-parser-set -pmd 3 scalar
+
+
+Actions Implementations (Experimental)
+--------------------------------------
+
+Actions describe what processing or modification should be performed on a
+packet when it matches a given flow. Similar to the datapath interface,
+DPCLS and MFEX (see above), the implementation of these actions can be
+accelerated using SIMD instructions, resulting in improved performance.
+
+OVS provides multiple implementations of the actions, however some
+implementations requiring a CPU capable of executing the required SIMD
+instructions.
+
+Available implementations can be listed with the following command::
+
+ $ ovs-appctl odp-execute/action-impl-show
+ Available Actions implementations:
+ scalar (available: Yes, active: Yes)
+ autovalidator (available: Yes, active: No)
+ avx512 (available: Yes, active: No)
+
+By default, ``scalar`` is used. Implementations can be selected by
+name::
+
+ $ ovs-appctl odp-execute/action-impl-set avx512
+ Action implementation set to avx512.
+
+ $ ovs-appctl odp-execute/action-impl-set scalar
+ Action implementation set to scalar.
@@ -361,12 +361,12 @@ testsuite.
Userspace datapath: Testing and Validation of CPU-specific Optimizations
''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''
-As multiple versions of the datapath classifier and packet parsing functions
-can co-exist, each with different CPU ISA optimizations, it is important to
-validate that they all give the exact same results. To easily test all the
-implementations, an ``autovalidator`` implementation of them exists. This
-implementation runs all other available implementations, and verifies that the
-results are identical.
+As multiple versions of the datapath classifier, packet parsing functions and
+actions can co-exist, each with different CPU ISA optimizations, it is
+important to validate that they all give the exact same results. To easily
+test all the implementations, an ``autovalidator`` implementation of them
+exists. This implementation runs all other available implementations, and
+verifies that the results are identical.
Running the OVS unit tests with the autovalidator enabled ensures all
implementations provide the same results. Note that the performance of the
@@ -382,18 +382,26 @@ To set the autovalidator for the packet parser, use this command::
$ ovs-appctl dpif-netdev/miniflow-parser-set autovalidator
+To set the autovalidator for actions, use this command::
+
+ $ ovs-appctl odp-execute/action-impl-set autovalidator
+
To run the OVS unit test suite with the autovalidator as the default
implementation, it is required to recompile OVS. During the recompilation,
the default priority of the `autovalidator` implementation is set to the
-maximum priority, ensuring every test will be run with every implementation::
+maximum priority, ensuring every test will be run with every implementation.
+Priority is only related to mfex autovalidator and not the actions
+autovalidator.::
- $ ./configure --enable-autovalidator --enable-mfex-default-autovalidator
+ $ ./configure --enable-autovalidator --enable-mfex-default-autovalidator \
+ --enable-actions-default-autovalidator
The following line should be seen in the configuration log when the above
options are used::
checking whether DPCLS Autovalidator is default implementation... yes
checking whether MFEX Autovalidator is default implementation... yes
+ checking whether actions Autovalidator is default implementation... yes
Compile OVS in debug mode to have `ovs_assert` statements error out if
there is a mis-match in the datapath classifier lookup or packet parser
@@ -61,6 +61,7 @@ Post-v2.17.0
implementations available at run time.
* Add build time configure command to enable auto-validator as default
actions implementation at build time.
+ * Add AVX512 implementation of actions.
- Linux datapath:
* Add offloading meter tc police.
* Add support for offloading the check_pkt_len action.
@@ -96,6 +96,7 @@ dnl
dnl Checks if compiler and binutils supports various AVX512 ISA.
AC_DEFUN([OVS_CHECK_AVX512], [
OVS_CHECK_BINUTILS_AVX512
+ OVS_CHECK_GCC_AVX512VL
OVS_CONDITIONAL_CC_OPTION_DEFINE([-mavx512f], [HAVE_AVX512F])
OVS_CONDITIONAL_CC_OPTION_DEFINE([-mavx512bw], [HAVE_AVX512BW])
OVS_CONDITIONAL_CC_OPTION_DEFINE([-mavx512vl], [HAVE_AVX512VL])
@@ -53,6 +53,7 @@ X86_ISA(X86_EXT_FEATURES_LEAF, EBX, 16, OVS_CPU_ISA_X86_AVX512F)
X86_ISA(X86_EXT_FEATURES_LEAF, EBX, 30, OVS_CPU_ISA_X86_AVX512BW)
X86_ISA(X86_EXT_FEATURES_LEAF, ECX, 1, OVS_CPU_ISA_X86_AVX512VBMI)
X86_ISA(X86_EXT_FEATURES_LEAF, ECX, 14, OVS_CPU_ISA_X86_VPOPCNTDQ)
+X86_ISA(X86_EXT_FEATURES_LEAF, EBX, 31, OVS_CPU_ISA_X86_AVX512VL)
#endif
bool
@@ -25,6 +25,7 @@ enum ovs_cpu_isa {
OVS_CPU_ISA_X86_AVX512F,
OVS_CPU_ISA_X86_AVX512BW,
OVS_CPU_ISA_X86_AVX512VBMI,
+ OVS_CPU_ISA_X86_AVX512VL,
OVS_CPU_ISA_X86_VPOPCNTDQ,
OVS_CPU_ISA_X86_LAST = OVS_CPU_ISA_X86_VPOPCNTDQ,
};
@@ -41,6 +41,14 @@ static struct odp_execute_action_impl action_impls[] = {
.name = "scalar",
.init_func = odp_action_scalar_init,
},
+
+#if ACTION_IMPL_AVX512_CHECK
+ [ACTION_IMPL_AVX512] = {
+ .available = false,
+ .name = "avx512",
+ .init_func = NULL,
+ },
+#endif
};
static void
@@ -22,6 +22,14 @@
#include "odp-netlink.h"
#include "ovs-atomic.h"
+/* Combine all required ISA and Linker checks into a single #define
+ * for readability and simplicity where the checks are needed. Note
+ * that it is always #defined, so code must use the #if preprocesor
+ * directive (not #ifdef). */
+#define ACTION_IMPL_AVX512_CHECK (__x86_64__ && HAVE_AVX512F \
+ && HAVE_LD_AVX512_GOOD && __SSE4_2__ && HAVE_AVX512BW && HAVE_AVX512VL \
+ && HAVE_GCC_AVX512VL_GOOD)
+
/* Forward declaration for typedef. */
struct odp_execute_action_impl;
@@ -56,6 +64,10 @@ enum odp_execute_action_impl_idx {
* Do not change the autovalidator position in this list without updating
* the define below. */
+#if ACTION_IMPL_AVX512_CHECK
+ ACTION_IMPL_AVX512,
+#endif
+
ACTION_IMPL_MAX,
};
@@ -421,6 +421,35 @@ AC_DEFUN([OVS_CHECK_SPHINX],
AC_ARG_VAR([SPHINXBUILD])
AM_CONDITIONAL([HAVE_SPHINX], [test "$SPHINXBUILD" != none])])
+
+dnl Checks for compiler correctly emitting AVX512-VL vpermd instruction.
+dnl GCC5 says it exports AVX512-VL, but it doesn't implement "vpermd" instruction
+dnl resulting in compilation failures. To workaround this "reported vs actual"
+dnl mismatch, we compile a small snippet, and conditionally enable AVX512-VL.
+AC_DEFUN([OVS_CHECK_GCC_AVX512VL], [
+ AC_MSG_CHECKING([whether compiler correctly emits AVX512-VL])
+ AC_COMPILE_IFELSE(
+ [AC_LANG_PROGRAM([#include <immintrin.h>
+ static void __attribute__((__target__("avx512vl")))
+ check_permutexvar(void)
+ {
+ __m256i v_swap32a = _mm256_setr_epi32(0x0, 0x4, 0xF,
+ 0xF, 0xF, 0xF,
+ 0xF, 0xF);
+ v_swap32a = _mm256_permutexvar_epi32(v_swap32a,
+ v_swap32a);
+ }],[])],
+ [AC_MSG_RESULT([yes])
+ ovs_cv_gcc_avx512vl_good=yes],
+ [AC_MSG_RESULT([no])
+ ovs_cv_gcc_avx512vl_good=no])
+ if test "$ovs_cv_gcc_avx512vl_good" = yes; then
+ AC_DEFINE([HAVE_GCC_AVX512VL_GOOD], [1],
+ [Define to 1 if gcc implements the vpermd instruction.])
+ fi
+ AM_CONDITIONAL([HAVE_GCC_AVX512VL_GOOD],
+ [test "$ovs_cv_gcc_avx512vl_good" = yes])])
+
dnl Checks for binutils/assembler known issue with AVX512.
dnl Due to backports, we probe assembling a reproducer instead of checking
dnl binutils version string. More details, including ASM dumps and debug here: