drivers: add generic build of SVE files
Subsystems (2)
RFC PATCH 1/1 drivers: add generic build of SVE files
For SVE, as for AVX2 and AVX-512, support building files for these ISAs from the top-level drivers/meson.build file, rather than having each driver re-implement it.
This removes the remaining build task for drivers in DPDK which is being done by a driver itself, rather than in the generic drivers' build rules.
--- drivers/meson.build | 27 +++++++++++++++++++++++++++ drivers/net/hns3/meson.build | 22 +--------------------- 2 files changed, 28 insertions(+), 21 deletions(-)
drivers/meson.build +27 -0
@@ -128,6 +128,7 @@ foreach subpath:subdirs sources = [] sources_avx2 = [] sources_avx512 = [] + sources_sve = [] headers = [] driver_sdk_headers = [] # public headers included by drivers objs = []
@@ -285,6 +286,32 @@ foreach subpath:subdirs endif endif + if (arch_subdir == 'arm' and sources_sve.length() > 0 + and cc.has_argument('-march=armv8.2-a+sve') + and cc.check_header('arm_sve.h')) + + if dpdk_conf.has('RTE_HAS_SVE_ACLE') + sources += sources_sve + else + cflags += ['-DRTE_HAS_SVE_ACLE=1'] + sve_cflags = ['-march=armv8.2-a+sve'] + foreach flag: cflags + if (flag.startswith('-march=') + or flag.startswith('-mcpu=') + or flag.startswith('-mtune=')) + continue + endif + sve_cflags += flag + endforeach + sve_lib = static_library(libname + '_sve_lib', + sources_sve, + dependencies: static_deps, + include_directories: includes, + c_args: sve_cflags) + objs += sve_lib.extract_objects(sources_sve) + endif + endif + # generate pmdinfo sources by building a temporary # lib and then running pmdinfogen on the contents of # that lib. The final lib reuses the object files and
drivers/net/hns3/meson.build +1 -21
@@ -43,25 +43,5 @@ cflags += no_wvla_cflag if arch_subdir == 'arm' and dpdk_conf.get('RTE_ARCH_64') sources += files('hns3_rxtx_vec.c') - - # compile SVE when: - # a. support SVE in minimum instruction set baseline - # b. it's not minimum instruction set, but compiler support - if dpdk_conf.has('RTE_HAS_SVE_ACLE') - sources += files('hns3_rxtx_vec_sve.c') - elif cc.has_argument('-march=armv8.2-a+sve') and cc.check_header('arm_sve.h') - cflags += ['-DRTE_HAS_SVE_ACLE=1'] - sve_cflags = [] - foreach flag: cflags - if not (flag.startswith('-march=') or flag.startswith('-mcpu=') or flag.startswith('-mtune=')) - sve_cflags += flag - endif - endforeach - hns3_sve_lib = static_library('hns3_sve_lib', - 'hns3_rxtx_vec_sve.c', - dependencies: [static_rte_ethdev], - include_directories: includes, - c_args: [sve_cflags, '-march=armv8.2-a+sve']) - objs += hns3_sve_lib.extract_objects('hns3_rxtx_vec_sve.c') - endif + sources_sve += files('hns3_rxtx_vec_sve.c') endif
--
2.45.2
Quoted 8 lines · Bruce Richardson
For SVE, as for AVX2 and AVX-512, support building files for these ISAs from the top-level
drivers/meson.buildfile, rather than having each driver re-implement it.This removes the remaining build task for drivers in DPDK which is being done by a driver itself, rather than in the generic drivers' build rules.
Quoted 6 lines
Signed-off-by Bruce Richardson <[email protected]>--- drivers/meson.build | 27 +++++++++++++++++++++++++++ drivers/net/hns3/meson.build | 22 +--------------------- 2 files changed, 28 insertions(+), 21 deletions(-)
As noted in my email in the other thread [1], this code has not been tested, and is being sent out here as a start point to hopefully enable someone more familiar with SVE to improve on it as-necessary. This support of SVE in hns3 driver is the last remaining instance of a driver doing a separate build of source code itself. All other drivers now simply pass all code files back to the top level meson.build file to have them built appropriately there.
/Bruce
[1] https://inbox.dpdk.org/dev/[email protected]/T/#m6614ca2fdef698e0448311f7255bf3be2438f101
Quoted 10 lines · Bruce Richardson
For SVE, as for AVX2 and AVX-512, support building files for these ISAs from the top-level
drivers/meson.buildfile, rather than having each driver re-implement it.This removes the remaining build task for drivers in DPDK which is being done by a driver itself, rather than in the generic drivers' build rules.
Signed-off-by Bruce Richardson <[email protected]>
Thanks for working on this additional cleanup.
Quoted 18 lines
--- drivers/meson.build | 27 +++++++++++++++++++++++++++ drivers/net/hns3/meson.build | 22 +--------------------- 2 files changed, 28 insertions(+), 21 deletions(-)drivers/meson.build +1
-0@@ -128,6 +128,7 @@ foreach subpath:subdirs sources = [] sources_avx2 = [] sources_avx512 = [] + sources_sve = [] headers = [] driver_sdk_headers = [] # public headers included by drivers objs = []
drivers/meson.build foreach subpath:subdirs · line 294
Quoted 9 lines
drivers/meson.build +6
-0@@ -285,6 +286,32 @@ foreach subpath:subdirs endif endif + if (arch_subdir == 'arm' and sources_sve.length() > 0 + and cc.has_argument('-march=armv8.2-a+sve') + and cc.check_header('arm_sve.h')) + + if dpdk_conf.has('RTE_HAS_SVE_ACLE') + sources += sources_sve
Do we need this special case?
Quoted 13 lines
drivers/meson.build +13
-0+ else + cflags += ['-DRTE_HAS_SVE_ACLE=1'] + sve_cflags = ['-march=armv8.2-a+sve'] + foreach flag: cflags + if (flag.startswith('-march=') + or flag.startswith('-mcpu=') + or flag.startswith('-mtune=')) + continue + endif + sve_cflags += flag + endforeach + sve_lib = static_library(libname + '_sve_lib', + sources_sve,
drivers/meson.build foreach subpath:subdirs · line 317
Quoted 10 lines
drivers/meson.build +7
-0+ dependencies: static_deps, + include_directories: includes, + c_args: sve_cflags) + objs += sve_lib.extract_objects(sources_sve) + endif + endif + # generate pmdinfo sources by building a temporary # lib and then running pmdinfogen on the contents of # that lib. The final lib reuses the object files and
On the principle, sources_sve must be passed to gen_version_map like other sources_XXX.
Signature · 1 line
David Marchand
Quoted 37 lines · David Marchand
Quoted 10 lines · Bruce Richardson
For SVE, as for AVX2 and AVX-512, support building files for these ISAs from the top-level
drivers/meson.buildfile, rather than having each driver re-implement it.This removes the remaining build task for drivers in DPDK which is being done by a driver itself, rather than in the generic drivers' build rules.
Signed-off-by Bruce Richardson <[email protected]>Thanks for working on this additional cleanup.
Quoted 21 lines
--- drivers/meson.build | 27 +++++++++++++++++++++++++++ drivers/net/hns3/meson.build | 22 +--------------------- 2 files changed, 28 insertions(+), 21 deletions(-)drivers/meson.build +1
-0@@ -128,6 +128,7 @@ foreach subpath:subdirs sources = [] sources_avx2 = [] sources_avx512 = [] + sources_sve = [] headers = [] driver_sdk_headers = [] # public headers included by drivers objs = []@@ -285,6 +286,32 @@ foreach subpath:subdirs endif endif
Quoted 9 lines
drivers/meson.build foreach subpath:subdirs · line 294
Quoted 6 lines
drivers/meson.build +6
-0+ if (arch_subdir == 'arm' and sources_sve.length() > 0 + and cc.has_argument('-march=armv8.2-a+sve') + and cc.check_header('arm_sve.h')) + + if dpdk_conf.has('RTE_HAS_SVE_ACLE') + sources += sources_sveDo we need this special case?
I put it there because it was in the original code. The handling of the march/mcpu flags below are also different to what is done on x86, and being unfamiliar with SVE compilation, I was very conservative in changing things. Again, hopefully an SVE expert (or hns3 driver expert), can rework this as necessary.
Quoted 25 lines
Quoted 14 lines
drivers/meson.build +13
-0+ else + cflags += ['-DRTE_HAS_SVE_ACLE=1'] + sve_cflags = ['-march=armv8.2-a+sve'] + foreach flag: cflags + if (flag.startswith('-march=') + or flag.startswith('-mcpu=') + or flag.startswith('-mtune=')) + continue + endif + sve_cflags += flag + endforeach + sve_lib = static_library(libname + '_sve_lib', + sources_sve,drivers/meson.build foreach subpath:subdirs · line 317
Quoted 10 lines
drivers/meson.build +7
-0+ dependencies: static_deps, + include_directories: includes, + c_args: sve_cflags) + objs += sve_lib.extract_objects(sources_sve) + endif + endif + # generate pmdinfo sources by building a temporary # lib and then running pmdinfogen on the contents of # that lib. The final lib reuses the object files and
Quoted 3 lines
On the principle,
sources_svemust be passed togen_version_maplike othersources_XXX.
Yep, missed that as I neglected to update from main the branch I was working on. :(
Quoted 46 lines
-----Original Message-----
From: David Marchand <[email protected]>
Sent: Monday, April 7, 2025 10:50 AM
To: Bruce Richardson <[email protected]>
Cc: [email protected]; Jie Hai <[email protected]>
Subject: Re: [RFC PATCH] drivers: add generic build of SVE filesQuoted 10 lines · Bruce Richardson
For SVE, as for AVX2 and AVX-512, support building files for these ISAs from the top-level
drivers/meson.buildfile, rather than having each driver re-implement it.This removes the remaining build task for drivers in DPDK which is being done by a driver itself, rather than in the generic drivers' build rules.
Signed-off-by Bruce Richardson <[email protected]>Thanks for working on this additional cleanup.
Quoted 21 lines
--- drivers/meson.build | 27 +++++++++++++++++++++++++++ drivers/net/hns3/meson.build | 22 +--------------------- 2 files changed, 28 insertions(+), 21 deletions(-)drivers/meson.build index renamed from drivers/meson.build +0
-0
b2d2537dc8..a6f0670a2f100644drivers/meson.build +1
-0@@ -128,6 +128,7 @@ foreach subpath:subdirs sources = [] sources_avx2 = [] sources_avx512 = [] + sources_sve = [] headers = [] driver_sdk_headers = [] # public headers included by drivers objs = []@@ -285,6 +286,32 @@ foreach subpath:subdirs endif endif
Quoted 9 lines
drivers/meson.build foreach subpath:subdirs · line 294
Quoted 6 lines
drivers/meson.build +6
-0+ if (arch_subdir == 'arm' and sources_sve.length() > 0 + and cc.has_argument('-march=armv8.2-a+sve') + and cc.check_header('arm_sve.h')) + + if dpdk_conf.has('RTE_HAS_SVE_ACLE') + sources += sources_sveDo we need this special case?
config/meson.build sets 'RTE_CPUFLAG_SVE for all SoCs with SVE support. For such SoCs RTE_HAS_SVE_ACLE is set to true if SVE ACLE is available. However, for some reason soc_cn10k sets RTE_HAS_SVE_ACLE to false even though SVE is available on that cnxk SoC.
--wathsala