ARMv8 Tune reorg
cover ARMv8 Tune reorg
There is a large number of Arm Tune files located in meta/conf/machine/include/, and to support the current and upcoming Arm cores, more are needed. Adding more files is simply going to make it harder to find the relevant ones for an OE/YP developer/user. Also, there are problems with stale and erroneous configs (see my previous series), which will only be exacerbated by having more files.
I am proposing a reorganization of the existing tune files by including them in the generic family include file. For example, the tune-cortexa55.inc would be moved into arch-armv8-2a.inc, and tune-cortexa57.inc would be moved into arch-armv8a.inc. This reduces the number of files from 12 to 2 for ARMv8a, and that is excluding the 13 I am adding in this series that would otherwise be unique files.
To use, simply add
...
DEFAULTTUNE ?= "neoversen1"
require conf/machine/include/arm/arch-armv8-2a.inc
...
Which is arguably what should be done anyway (instead of taking the default of the tune include file).
See the qemuarm64 patch in the series for a working example.
Of course, by removing the existing tune files, current users are going to break. A simple script can be written to use sed (or similar) to replace the relevant parts for those users that would be affected (at least for those that are in the layer index and update regularly).
Thanks,
Jon
---
Originally sent as a RFC in
https://lists.openembedded.org/g/openembedded-core/message/142324
Given the generally positive feedback, sending as a patch series. Keeping the "hard fail" of the file removal (per Richards comment in
https://lists.openembedded.org/g/openembedded-core/message/142356).
Only difference of note is the removal of the "arm64: set BASE_LIB to lib64", as there needs to be more investigation (see
https://lists.openembedded.org/g/openembedded-core/message/142414).
Jon Mason (5):
arch-armv8-2a.inc: Add Cortex-A55 tunings arch-armv8a.inc: Add existing tunings
qemuarm64: change tuning
arch-armv8a.inc: Add tunes for supported ARMv8a cores arch-armv8-2a.inc: Add tunes for supported ARMv8.2a cores
.../machine/include/arm/arch-armv8-2a.inc | 175 ++++++++++++++- meta/conf/machine/include/arm/arch-armv8a.inc | 206 +++++++++++++++++- meta/conf/machine/include/tune-cortexa32.inc | 18 -- meta/conf/machine/include/tune-cortexa35.inc | 17 -- meta/conf/machine/include/tune-cortexa53.inc | 18 -- meta/conf/machine/include/tune-cortexa55.inc | 13 -- .../include/tune-cortexa57-cortexa53.inc | 15 -- meta/conf/machine/include/tune-cortexa57.inc | 17 -- .../include/tune-cortexa72-cortexa53.inc | 20 -- meta/conf/machine/include/tune-cortexa72.inc | 13 -- .../include/tune-cortexa73-cortexa53.inc | 20 -- meta/conf/machine/include/tune-thunderx.inc | 19 -- meta/conf/machine/qemuarm64.conf | 3 +- 13 files changed, 371 insertions(+), 183 deletions(-) delete mode 100644 meta/conf/machine/include/tune-cortexa32.inc delete mode 100644 meta/conf/machine/include/tune-cortexa35.inc delete mode 100644 meta/conf/machine/include/tune-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa55.inc delete mode 100644 meta/conf/machine/include/tune-cortexa57-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa57.inc delete mode 100644 meta/conf/machine/include/tune-cortexa72-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa72.inc delete mode 100644 meta/conf/machine/include/tune-cortexa73-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-thunderx.inc
Signature · 1 line
2.20.1
PATCH 1/5 arch-armv8-2a.inc: Add Cortex-A55 tunings
Migrate the settings in tune-cortexa55.inc to arch-armv8-2a.inc. This will allow for a single file to include all of the tunings of a family of processors. This will reduce the proliferation of unique files per processor currently occurring in conf/machine/include/
--- .../machine/include/arm/arch-armv8-2a.inc | 36 +++++++++++++------ meta/conf/machine/include/tune-cortexa55.inc | 13 ------- 2 files changed, 26 insertions(+), 23 deletions(-) delete mode 100644 meta/conf/machine/include/tune-cortexa55.inc
meta/conf/machine/include/arm/arch-armv8-2a.inc +26 -10
@@ -1,4 +1,20 @@ -DEFAULTTUNE ?= "armv8-2a" +# +# Tune Settings for Cortex-A55 +# +TUNEVALID[cortexa55] = "Enable Cortex-A55 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa55', ' -mcpu=cortex-a55', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa55" +ARMPKGARCH_tune-cortexa55 = "cortexa55" +TUNE_FEATURES_tune-cortexa55 = "aarch64 cortexa55 crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa55 = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} cortexa55" +BASE_LIB_tune-cortexa55 = "lib64" + +# +# Defaults for ARMv8-a +# +DEFAULTTUNE ?= "armv8-2a" TUNEVALID[armv8-2a] = "Enable instructions for ARMv8-a" TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'armv8-2a', ' -march=armv8.2-a', '', d)}"
@@ -8,12 +24,12 @@ MACHINEOVERRIDES =. "${@bb.utils.contains('TUNE_FEATURES', 'armv8-2a', 'armv8-2a require conf/machine/include/arm/arch-armv8a.inc # Little Endian base configs -AVAILTUNES += "armv8-2a armv8-2a-crypto" -ARMPKGARCH_tune-armv8-2a ?= "armv8-2a" -ARMPKGARCH_tune-armv8-2a-crypto ?= "armv8-2a" -TUNE_FEATURES_tune-armv8-2a = "aarch64 armv8-2a" -TUNE_FEATURES_tune-armv8-2a-crypto = "${TUNE_FEATURES_tune-armv8-2a} crypto" -PACKAGE_EXTRA_ARCHS_tune-armv8-2a = "${PACKAGE_EXTRA_ARCHS_tune-armv8a} armv8-2a" -PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a} armv8-2a-crypto" -BASE_LIB_tune-armv8-2a = "lib64" -BASE_LIB_tune-armv8-2a-crypto = "lib64" +AVAILTUNES += "armv8-2a armv8-2a-crypto" +ARMPKGARCH_tune-armv8-2a ?= "armv8-2a" +ARMPKGARCH_tune-armv8-2a-crypto ?= "armv8-2a" +TUNE_FEATURES_tune-armv8-2a = "aarch64 armv8-2a" +TUNE_FEATURES_tune-armv8-2a-crypto = "${TUNE_FEATURES_tune-armv8-2a} crypto" +PACKAGE_EXTRA_ARCHS_tune-armv8-2a = "${PACKAGE_EXTRA_ARCHS_tune-armv8a} armv8-2a" +PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a} armv8-2a-crypto" +BASE_LIB_tune-armv8-2a = "lib64" +BASE_LIB_tune-armv8-2a-crypto = "lib64"
meta/conf/machine/include/tune-cortexa55.inc deleted +0 -13
@@ -1,13 +0,0 @@ -DEFAULTTUNE ?= "cortexa55" - -TUNEVALID[cortexa55] = "Enable Cortex-A55 specific processor optimizations" -TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa55', ' -mcpu=cortex-a55', '', d)}" - -require conf/machine/include/arm/arch-armv8-2a.inc - -# Little Endian base configs -AVAILTUNES += "cortexa55" -ARMPKGARCH_tune-cortexa55 = "cortexa55" -TUNE_FEATURES_tune-cortexa55 = "aarch64 cortexa55 crypto" -PACKAGE_EXTRA_ARCHS_tune-cortexa55 = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} cortexa55" -BASE_LIB_tune-cortexa55 = "lib64"
Signature · 1 line
2.20.1
PATCH 2/5 arch-armv8a.inc: Add existing tunings
In reply to Jon MasonMigrate the existing tune settings to arch-armv8a.inc. This will allow for a single file to include all of the tunings of a family of processors. This will reduce the proliferation of unique files per processor currently occurring in conf/machine/include/
--- meta/conf/machine/include/arm/arch-armv8a.inc | 157 +++++++++++++++++- meta/conf/machine/include/tune-cortexa32.inc | 18 -- meta/conf/machine/include/tune-cortexa35.inc | 17 -- meta/conf/machine/include/tune-cortexa53.inc | 18 -- .../include/tune-cortexa57-cortexa53.inc | 15 -- meta/conf/machine/include/tune-cortexa57.inc | 17 -- .../include/tune-cortexa72-cortexa53.inc | 20 --- meta/conf/machine/include/tune-cortexa72.inc | 13 -- .../include/tune-cortexa73-cortexa53.inc | 20 --- meta/conf/machine/include/tune-thunderx.inc | 19 --- 10 files changed, 155 insertions(+), 159 deletions(-) delete mode 100644 meta/conf/machine/include/tune-cortexa32.inc delete mode 100644 meta/conf/machine/include/tune-cortexa35.inc delete mode 100644 meta/conf/machine/include/tune-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa57-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa57.inc delete mode 100644 meta/conf/machine/include/tune-cortexa72-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa72.inc delete mode 100644 meta/conf/machine/include/tune-cortexa73-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-thunderx.inc
meta/conf/machine/include/arm/arch-armv8a.inc +155 -2
@@ -1,4 +1,155 @@ -DEFAULTTUNE ?= "armv8a-crc" +# +# Tune Settings for Cortex-A32 (32bit only) +# +TUNEVALID[cortexa32] = "Enable Cortex-A32 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa32', ' -mcpu=cortex-a32', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa32 cortexa32-crypto" +ARMPKGARCH_tune-cortexa32 = "cortexa32" +ARMPKGARCH_tune-cortexa32-crypto = "cortexa32" +TUNE_FEATURES_tune-cortexa32 = "armv8a cortexa32 crc" +TUNE_FEATURES_tune-cortexa32-crypto = "${TUNE_FEATURES_tune-cortexa32} crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa32 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa32" +PACKAGE_EXTRA_ARCHS_tune-cortexa32-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa32 cortexa32-crypto" +BASE_LIB_tune-cortexa32 = "lib" +BASE_LIB_tune-cortexa32-crypto = "lib" + +# +# Tune Settings for Cortex-A35 +# +TUNEVALID[cortexa35] = "Enable Cortex-A35 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa35', ' -mcpu=cortex-a35', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa35 cortexa35-crypto" +ARMPKGARCH_tune-cortexa35 = "cortexa35" +ARMPKGARCH_tune-cortexa35-crypto = "cortexa35" +TUNE_FEATURES_tune-cortexa35 = "aarch64 cortexa35 crc" +TUNE_FEATURES_tune-cortexa35-crypto = "${TUNE_FEATURES_tune-cortexa35} crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa35 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa35" +PACKAGE_EXTRA_ARCHS_tune-cortexa35-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa35 cortexa35-crypto" +BASE_LIB_tune-cortexa35 = "lib64" +BASE_LIB_tune-cortexa35-crypto = "lib64" + +# +# Tune Settings for Cortex-A53 +# +TUNEVALID[cortexa53] = "Enable Cortex-A53 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa53', ' -mcpu=cortex-a53', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa53 cortexa53-crypto" +ARMPKGARCH_tune-cortexa53 = "cortexa53" +ARMPKGARCH_tune-cortexa53-crypto = "cortexa53-crypto" +TUNE_FEATURES_tune-cortexa53 = "aarch64 cortexa53 crc" +TUNE_FEATURES_tune-cortexa53-crypto = "${TUNE_FEATURES_tune-cortexa53} crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa53 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa53" +PACKAGE_EXTRA_ARCHS_tune-cortexa53-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa53 cortexa53-crypto" +BASE_LIB_tune-cortexa53 = "lib64" +BASE_LIB_tune-cortexa53-crypto = "lib64" + +# +# Tune Settings for Cortex-A57 +# +TUNEVALID[cortexa57] = "Enable Cortex-A57 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa57', ' -mcpu=cortex-a57', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa57 cortexa57-crypto" +ARMPKGARCH_tune-cortexa57 = "cortexa57" +ARMPKGARCH_tune-cortexa57-crypto = "cortexa57-crypto" +TUNE_FEATURES_tune-cortexa57 = "aarch64 cortexa57 crc" +TUNE_FEATURES_tune-cortexa57-crypto = "${TUNE_FEATURES_tune-cortexa57} crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa57 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa57" +PACKAGE_EXTRA_ARCHS_tune-cortexa57-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa57 cortexa57-crypto" +BASE_LIB_tune-cortexa57 = "lib64" +BASE_LIB_tune-cortexa57-crypto = "lib64" + +# +# Tune Settings for Cortex-A72 +# +TUNEVALID[cortexa72] = "Enable Cortex-A72 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa72', ' -mcpu=cortex-a72', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa72" +ARMPKGARCH_tune-cortexa72 = "cortexa72" +TUNE_FEATURES_tune-cortexa72 = "aarch64 cortexa72 crc crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa72 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa72" +BASE_LIB_tune-cortexa72 = "lib64" +BASE_LIB_tune-cortexa72-crypto = "lib64" + +# +# Tune Settings for ThunderX +# +TUNEVALID[thunderx] = "Enable instructions for Cavium ThunderX" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'thunderx', ' -mcpu=thunderx', '',d)}" + +AVAILTUNES += "thunderx thunderx_be" +ARMPKGARCH_tune-thunderx ?= "thunderx" +ARMPKGARCH_tune-thunderx_be ?= "thunderx_be" +TUNE_FEATURES_tune-thunderx = "${TUNE_FEATURES_tune-aarch64} thunderx" +TUNE_FEATURES_tune-thunderx_be = "${TUNE_FEATURES_tune-thunderx} bigendian" +PACKAGE_EXTRA_ARCHS_tune-thunderx = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} thunderx" +PACKAGE_EXTRA_ARCHS_tune-thunderx_be = "aarch64_be thunderx_be" +BASE_LIB_tune-thunderx = "lib64" +BASE_LIB_tune-thunderx_be = "lib64" + +# +# Tune Settings for big.LITTLE Cortex-A57 - Cortex-A53 +# +TUNEVALID[cortexa57-cortexa53] = "Enable big.LITTLE Cortex-A57.Cortex-A53 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains("TUNE_FEATURES", "cortexa57-cortexa53", " -mcpu=cortex-a57.cortex-a53", "", d)}" +MACHINEOVERRIDES =. "${@bb.utils.contains("TUNE_FEATURES", "cortexa57-cortexa53", "cortexa57-cortexa53:", "" ,d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa57-cortexa53" +ARMPKGARCH_tune-cortexa57-cortexa53 = "cortexa57-cortexa53" +TUNE_FEATURES_tune-cortexa57-cortexa53 = "aarch64 crc cortexa57-cortexa53" +PACKAGE_EXTRA_ARCHS_tune-cortexa57-cortexa53 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa57-cortexa53" +BASE_LIB_tune-cortexa57-cortexa53 = "lib64" + +# +# Tune Settings for big.LITTLE Cortex-A72 - Cortex-A53 +# +TUNEVALID[cortexa72-cortexa53] = "Enable big.LITTLE Cortex-A72.Cortex-A53 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains("TUNE_FEATURES", "cortexa72-cortexa53", " -mcpu=cortex-a72.cortex-a53", "", d)}" +MACHINEOVERRIDES =. "${@bb.utils.contains("TUNE_FEATURES", "cortexa72-cortexa53", "cortexa72-cortexa53:", "" ,d)}" + +# cortexa72.cortexa53 implies crc support +AVAILTUNES += "cortexa72-cortexa53 cortexa72-cortexa53-crypto" +ARMPKGARCH_tune-cortexa72-cortexa53 = "cortexa72-cortexa53" +ARMPKGARCH_tune-cortexa72-cortexa53-crypto = "cortexa72-cortexa53-crypto" +TUNE_FEATURES_tune-cortexa72-cortexa53 = "aarch64 crc cortexa72-cortexa53" +TUNE_FEATURES_tune-cortexa72-cortexa53-crypto = "${TUNE_FEATURES_tune-cortexa72-cortexa53} crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa72-cortexa53 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa72-cortexa53" +PACKAGE_EXTRA_ARCHS_tune-cortexa72-cortexa53-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa72-cortexa53 cortexa72-cortexa53-crypto" +BASE_LIB_tune-cortexa72-cortexa53 = "lib64" +BASE_LIB_tune-cortexa72-cortexa53-crypto = "lib64" + +# +# Tune Settings for big.LITTLE Cortex-A73 - Cortex-A53 +# +TUNEVALID[cortexa73-cortexa53] = "Enable big.LITTLE Cortex-A73.Cortex-A53 specific processor optimizations" +MACHINEOVERRIDES =. "${@bb.utils.contains("TUNE_FEATURES", "cortexa73-cortexa53", "cortexa73-cortexa53:", "" ,d)}" +TUNE_CCARGS .= "${@bb.utils.contains("TUNE_FEATURES", "cortexa73-cortexa53", " -mcpu=cortex-a73.cortex-a53", "", d)}" + +# cortexa73.cortexa53 implies crc support +AVAILTUNES += "cortexa73-cortexa53 cortexa73-cortexa53-crypto" +ARMPKGARCH_tune-cortexa73-cortexa53 = "cortexa73-cortexa53" +ARMPKGARCH_tune-cortexa73-cortexa53-crypto = "cortexa73-cortexa53-crypto" +TUNE_FEATURES_tune-cortexa73-cortexa53 = "aarch64 crc cortexa73-cortexa53" +TUNE_FEATURES_tune-cortexa73-cortexa53-crypto = "${TUNE_FEATURES_tune-cortexa73-cortexa53} crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa73-cortexa53 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa73-cortexa53" +PACKAGE_EXTRA_ARCHS_tune-cortexa73-cortexa53-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa73-cortexa53 cortexa73-cortexa53-crypto" +BASE_LIB_tune-cortexa73-cortexa53 = "lib64" +BASE_LIB_tune-cortexa73-cortexa53-crypto = "lib64" + +# +# Defaults for ARMv8-a +# +DEFAULTTUNE ?= "armv8a-crc" TUNEVALID[armv8a] = "Enable instructions for ARMv8-a" TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'armv8a', ' -march=armv8-a', '', d)}"
@@ -8,10 +159,12 @@ TUNEVALID[crypto] = "Enable instructions for ARMv8-a cryptographic" TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'crypto', '+crypto', '', d)}" MACHINEOVERRIDES =. "${@bb.utils.contains('TUNE_FEATURES', 'armv8a', 'armv8a:', '' ,d)}" +TUNECONFLICTS[aarch64] = "armv4 armv5 armv6 armv7 armv7a" + require conf/machine/include/arm/arch-arm64.inc # Little Endian base configs -AVAILTUNES += "armv8a armv8a-crc armv8a-crc-crypto armv8a-crypto" +AVAILTUNES += "armv8a armv8a-crc armv8a-crc-crypto armv8a-crypto" ARMPKGARCH_tune-armv8a ?= "armv8a" ARMPKGARCH_tune-armv8a-crc ?= "armv8a" ARMPKGARCH_tune-armv8a-crypto ?= "armv8a"
meta/conf/machine/include/tune-cortexa32.inc deleted +0 -18
@@ -1,18 +0,0 @@ -DEFAULTTUNE ?= "cortexa32" - - -TUNEVALID[cortexa32] = "Enable Cortex-A32 specific processor optimizations" -TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa32', ' -mcpu=cortex-a32', '', d)}" - -require conf/machine/include/arm/arch-armv8a.inc - -# Little Endian base configs -AVAILTUNES += "cortexa32 cortexa32-crypto" -ARMPKGARCH_tune-cortexa32 = "cortexa32" -ARMPKGARCH_tune-cortexa32-crypto = "cortexa32" -TUNE_FEATURES_tune-cortexa32 = "armv8a cortexa32 crc" -TUNE_FEATURES_tune-cortexa32-crypto = "${TUNE_FEATURES_tune-cortexa32} crypto" -PACKAGE_EXTRA_ARCHS_tune-cortexa32 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa32" -PACKAGE_EXTRA_ARCHS_tune-cortexa32-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa32 cortexa32-crypto" -BASE_LIB_tune-cortexa32 = "lib" -BASE_LIB_tune-cortexa32-crypto = "lib"
meta/conf/machine/include/tune-cortexa35.inc deleted +0 -17
@@ -1,17 +0,0 @@ -DEFAULTTUNE ?= "cortexa35" - -TUNEVALID[cortexa35] = "Enable Cortex-A35 specific processor optimizations" -TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa35', ' -mcpu=cortex-a35', '', d)}" - -require conf/machine/include/arm/arch-armv8a.inc - -# Little Endian base configs -AVAILTUNES += "cortexa35 cortexa35-crypto" -ARMPKGARCH_tune-cortexa35 = "cortexa35" -ARMPKGARCH_tune-cortexa35-crypto = "cortexa35" -TUNE_FEATURES_tune-cortexa35 = "aarch64 cortexa35 crc" -TUNE_FEATURES_tune-cortexa35-crypto = "${TUNE_FEATURES_tune-cortexa35} crypto" -PACKAGE_EXTRA_ARCHS_tune-cortexa35 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa35" -PACKAGE_EXTRA_ARCHS_tune-cortexa35-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa35 cortexa35-crypto" -BASE_LIB_tune-cortexa35 = "lib64" -BASE_LIB_tune-cortexa35-crypto = "lib64"
meta/conf/machine/include/tune-cortexa53.inc deleted +0 -18
@@ -1,18 +0,0 @@ -DEFAULTTUNE ?= "cortexa53" - -TUNEVALID[cortexa53] = "Enable Cortex-A53 specific processor optimizations" -TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa53', ' -mcpu=cortex-a53', '', d)}" - -require conf/machine/include/arm/arch-armv8a.inc - -# Little Endian base configs -AVAILTUNES += "cortexa53 cortexa53-crypto" -ARMPKGARCH_tune-cortexa53 = "cortexa53" -ARMPKGARCH_tune-cortexa53-crypto = "cortexa53-crypto" -TUNE_FEATURES_tune-cortexa53 = "aarch64 cortexa53 crc" -TUNE_FEATURES_tune-cortexa53-crypto = "${TUNE_FEATURES_tune-cortexa53} crypto" -PACKAGE_EXTRA_ARCHS_tune-cortexa53 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa53" -PACKAGE_EXTRA_ARCHS_tune-cortexa53-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa53 cortexa53-crypto" - -BASE_LIB_tune-cortexa53 = "lib64" -BASE_LIB_tune-cortexa53-crypto = "lib64"
meta/conf/machine/include/tune-cortexa57-cortexa53.inc deleted +0 -15
@@ -1,15 +0,0 @@ -DEFAULTTUNE ?= "cortexa57-cortexa53" - -TUNEVALID[cortexa57-cortexa53] = "Enable big.LITTLE Cortex-A57.Cortex-A53 specific processor optimizations" -TUNECONFLICTS[aarch64] = "armv4 armv5 armv6 armv7 armv7a" -TUNE_CCARGS .= "${@bb.utils.contains("TUNE_FEATURES", "cortexa57-cortexa53", " -mcpu=cortex-a57.cortex-a53", "", d)}" -MACHINEOVERRIDES =. "${@bb.utils.contains("TUNE_FEATURES", "cortexa57-cortexa53", "cortexa57-cortexa53:", "" ,d)}" - -require conf/machine/include/arm/arch-armv8a.inc - -# Little Endian base configs -AVAILTUNES += "cortexa57-cortexa53" -ARMPKGARCH_tune-cortexa57-cortexa53 = "cortexa57-cortexa53" -TUNE_FEATURES_tune-cortexa57-cortexa53 = "aarch64 crc cortexa57-cortexa53" -PACKAGE_EXTRA_ARCHS_tune-cortexa57-cortexa53 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa57-cortexa53" -BASE_LIB_tune-cortexa57-cortexa53 = "lib64"
meta/conf/machine/include/tune-cortexa57.inc deleted +0 -17
@@ -1,17 +0,0 @@ -DEFAULTTUNE ?= "cortexa57" - -TUNEVALID[cortexa57] = "Enable Cortex-A57 specific processor optimizations" -TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa57', ' -mcpu=cortex-a57', '', d)}" - -require conf/machine/include/arm/arch-armv8a.inc - -# Little Endian base configs -AVAILTUNES += "cortexa57 cortexa57-crypto" -ARMPKGARCH_tune-cortexa57 = "cortexa57" -ARMPKGARCH_tune-cortexa57-crypto = "cortexa57-crypto" -TUNE_FEATURES_tune-cortexa57 = "aarch64 cortexa57 crc" -TUNE_FEATURES_tune-cortexa57-crypto = "${TUNE_FEATURES_tune-cortexa57} crypto" -PACKAGE_EXTRA_ARCHS_tune-cortexa57 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa57" -PACKAGE_EXTRA_ARCHS_tune-cortexa57-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa57 cortexa57-crypto" -BASE_LIB_tune-cortexa57 = "lib64" -BASE_LIB_tune-cortexa57-crypto = "lib64"
meta/conf/machine/include/tune-cortexa72-cortexa53.inc deleted +0 -20
@@ -1,20 +0,0 @@ -DEFAULTTUNE ?= "cortexa72-cortexa53" - -TUNEVALID[cortexa72-cortexa53] = "Enable big.LITTLE Cortex-A72.Cortex-A53 specific processor optimizations" -TUNECONFLICTS[aarch64] = "armv4 armv5 armv6 armv7 armv7a" -TUNE_CCARGS .= "${@bb.utils.contains("TUNE_FEATURES", "cortexa72-cortexa53", " -mcpu=cortex-a72.cortex-a53", "", d)}" -MACHINEOVERRIDES =. "${@bb.utils.contains("TUNE_FEATURES", "cortexa72-cortexa53", "cortexa72-cortexa53:", "" ,d)}" - -require conf/machine/include/arm/arch-armv8a.inc - -# cortexa72.cortexa53 implies crc support -AVAILTUNES += "cortexa72-cortexa53 cortexa72-cortexa53-crypto" -ARMPKGARCH_tune-cortexa72-cortexa53 = "cortexa72-cortexa53" -ARMPKGARCH_tune-cortexa72-cortexa53-crypto = "cortexa72-cortexa53-crypto" -TUNE_FEATURES_tune-cortexa72-cortexa53 = "aarch64 crc cortexa72-cortexa53" -TUNE_FEATURES_tune-cortexa72-cortexa53-crypto = "${TUNE_FEATURES_tune-cortexa72-cortexa53} crypto" -PACKAGE_EXTRA_ARCHS_tune-cortexa72-cortexa53 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa72-cortexa53" -PACKAGE_EXTRA_ARCHS_tune-cortexa72-cortexa53-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa72-cortexa53 cortexa72-cortexa53-crypto" -BASE_LIB_tune-cortexa72-cortexa53 = "lib64" -BASE_LIB_tune-cortexa72-cortexa53-crypto = "lib64" -
meta/conf/machine/include/tune-cortexa72.inc deleted +0 -13
@@ -1,13 +0,0 @@ -DEFAULTTUNE ?= "cortexa72" - -TUNEVALID[cortexa72] = "Enable Cortex-A72 specific processor optimizations" -TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa72', ' -mcpu=cortex-a72', '', d)}" - -require conf/machine/include/arm/arch-armv8a.inc - -# Little Endian base configs -AVAILTUNES += "cortexa72" -ARMPKGARCH_tune-cortexa72 = "cortexa72" -TUNE_FEATURES_tune-cortexa72 = "aarch64 cortexa72 crc crypto" -PACKAGE_EXTRA_ARCHS_tune-cortexa72 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa72" -BASE_LIB_tune-cortexa72 = "lib64"
meta/conf/machine/include/tune-cortexa73-cortexa53.inc deleted +0 -20
@@ -1,20 +0,0 @@ -DEFAULTTUNE ?= "cortexa73-cortexa53" - -TUNEVALID[cortexa73-cortexa53] = "Enable big.LITTLE Cortex-A73.Cortex-A53 specific processor optimizations" -TUNECONFLICTS[aarch64] = "armv4 armv5 armv6 armv7 armv7a" -MACHINEOVERRIDES =. "${@bb.utils.contains("TUNE_FEATURES", "cortexa73-cortexa53", "cortexa73-cortexa53:", "" ,d)}" -TUNE_CCARGS .= "${@bb.utils.contains("TUNE_FEATURES", "cortexa73-cortexa53", " -mcpu=cortex-a73.cortex-a53", "", d)}" - -require conf/machine/include/arm/arch-armv8a.inc - -# cortexa73.cortexa53 implies crc support -AVAILTUNES += "cortexa73-cortexa53 cortexa73-cortexa53-crypto" -ARMPKGARCH_tune-cortexa73-cortexa53 = "cortexa73-cortexa53" -ARMPKGARCH_tune-cortexa73-cortexa53-crypto = "cortexa73-cortexa53-crypto" -TUNE_FEATURES_tune-cortexa73-cortexa53 = "aarch64 crc cortexa73-cortexa53" -TUNE_FEATURES_tune-cortexa73-cortexa53-crypto = "${TUNE_FEATURES_tune-cortexa73-cortexa53} crypto" -PACKAGE_EXTRA_ARCHS_tune-cortexa73-cortexa53 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa73-cortexa53" -PACKAGE_EXTRA_ARCHS_tune-cortexa73-cortexa53-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa73-cortexa53 cortexa73-cortexa53-crypto" -BASE_LIB_tune-cortexa73-cortexa53 = "lib64" -BASE_LIB_tune-cortexa73-cortexa53-crypto = "lib64" -
meta/conf/machine/include/tune-thunderx.inc deleted +0 -19
@@ -1,19 +0,0 @@ -DEFAULTTUNE ?= "thunderx" -AVAILTUNES += "thunderx thunderx_be" - -TUNEVALID[thunderx] = "Enable instructions for Cavium ThunderX" - -TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'thunderx', ' -mcpu=thunderx', '',d)}" - -require conf/machine/include/arm/arch-armv8a.inc - -ARMPKGARCH_tune-thunderx ?= "thunderx" -ARMPKGARCH_tune-thunderx_be ?= "thunderx_be" - -TUNE_FEATURES_tune-thunderx = "${TUNE_FEATURES_tune-aarch64} thunderx" -TUNE_FEATURES_tune-thunderx_be = "${TUNE_FEATURES_tune-thunderx} bigendian" -BASE_LIB_tune-thunderx = "lib64" -BASE_LIB_tune-thunderx_be = "lib64" - -PACKAGE_EXTRA_ARCHS_tune-thunderx = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} thunderx" -PACKAGE_EXTRA_ARCHS_tune-thunderx_be = "aarch64_be thunderx_be"
Signature · 1 line
2.20.1
PATCH 3/5 qemuarm64: change tuning
In reply to Jon MasonThe previous patch caused the tuning file referenced here to be removed. Use the new one with the new DEFAULTTUNE.
--- meta/conf/machine/qemuarm64.conf | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
meta/conf/machine/qemuarm64.conf +2 -1
@@ -2,7 +2,8 @@ #@NAME: QEMU ARMv8 machine #@DESCRIPTION: Machine configuration for running an ARMv8 system on QEMU -require conf/machine/include/tune-cortexa57.inc +DEFAULTTUNE ?= "cortexa57" +require conf/machine/include/arm/arch-armv8a.inc require conf/machine/include/qemu.inc KERNEL_IMAGETYPE = "Image"
Signature · 1 line
2.20.1
PATCH 4/5 arch-armv8a.inc: Add tunes for supported ARMv8a cores
In reply to Jon MasonAdd tunes for all the ARMv8a cores currently supported in GCC. This
is: Cortex-A34, Cortex-A73, and Cortex-A73-Cortex-A35.
--- meta/conf/machine/include/arm/arch-armv8a.inc | 49 +++++++++++++++++++ 1 file changed, 49 insertions(+)
meta/conf/machine/include/arm/arch-armv8a.inc +49 -0
@@ -15,6 +15,23 @@ PACKAGE_EXTRA_ARCHS_tune-cortexa32-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a- BASE_LIB_tune-cortexa32 = "lib" BASE_LIB_tune-cortexa32-crypto = "lib" +# +# Tune Settings for Cortex-A34 +# +TUNEVALID[cortexa34] = "Enable Cortex-A34 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa34', ' -mcpu=cortex-a34', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa34 cortexa34-crypto" +ARMPKGARCH_tune-cortexa34 = "cortexa34" +ARMPKGARCH_tune-cortexa34-crypto = "cortexa34" +TUNE_FEATURES_tune-cortexa34 = "aarch64 cortexa34 crc" +TUNE_FEATURES_tune-cortexa34-crypto = "${TUNE_FEATURES_tune-cortexa34} crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa34 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa34" +PACKAGE_EXTRA_ARCHS_tune-cortexa34-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa34 cortexa34-crypto" +BASE_LIB_tune-cortexa34 = "lib64" +BASE_LIB_tune-cortexa34-crypto = "lib64" + # # Tune Settings for Cortex-A35 #
@@ -80,6 +97,20 @@ PACKAGE_EXTRA_ARCHS_tune-cortexa72 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a- BASE_LIB_tune-cortexa72 = "lib64" BASE_LIB_tune-cortexa72-crypto = "lib64" +# +# Tune Settings for Cortex-A73 +# +TUNEVALID[cortexa73] = "Enable Cortex-A73 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa73', ' -mcpu=cortex-a73', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa73" +ARMPKGARCH_tune-cortexa73 = "cortexa73" +TUNE_FEATURES_tune-cortexa73 = "aarch64 cortexa73 crc crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa73 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa73" +BASE_LIB_tune-cortexa73 = "lib64" +BASE_LIB_tune-cortexa73-crypto = "lib64" + # # Tune Settings for ThunderX #
@@ -128,6 +159,24 @@ PACKAGE_EXTRA_ARCHS_tune-cortexa72-cortexa53-crypto = "${PACKAGE_EXTRA_ARCHS_tun BASE_LIB_tune-cortexa72-cortexa53 = "lib64" BASE_LIB_tune-cortexa72-cortexa53-crypto = "lib64" +# +# Tune Settings for big.LITTLE Cortex-A73 - Cortex-A35 +# +TUNEVALID[cortexa73-cortexa35] = "Enable big.LITTLE Cortex-A73.Cortex-A35 specific processor optimizations" +MACHINEOVERRIDES =. "${@bb.utils.contains("TUNE_FEATURES", "cortexa73-cortexa35", "cortexa73-cortexa35:", "" ,d)}" +TUNE_CCARGS .= "${@bb.utils.contains("TUNE_FEATURES", "cortexa73-cortexa35", " -mcpu=cortex-a73.cortex-a35", "", d)}" + +# cortexa73.cortexa35 implies crc support +AVAILTUNES += "cortexa73-cortexa35 cortexa73-cortexa35-crypto" +ARMPKGARCH_tune-cortexa73-cortexa35 = "cortexa73-cortexa35" +ARMPKGARCH_tune-cortexa73-cortexa35-crypto = "cortexa73-cortexa35-crypto" +TUNE_FEATURES_tune-cortexa73-cortexa35 = "aarch64 crc cortexa73-cortexa35" +TUNE_FEATURES_tune-cortexa73-cortexa35-crypto = "${TUNE_FEATURES_tune-cortexa73-cortexa35} crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa73-cortexa35 = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc} cortexa73-cortexa35" +PACKAGE_EXTRA_ARCHS_tune-cortexa73-cortexa35-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8a-crc-crypto} cortexa73-cortexa35 cortexa73-cortexa35-crypto" +BASE_LIB_tune-cortexa73-cortexa35 = "lib64" +BASE_LIB_tune-cortexa73-cortexa35-crypto = "lib64" + # # Tune Settings for big.LITTLE Cortex-A73 - Cortex-A53 #
Signature · 1 line
2.20.1
PATCH 5/5 arch-armv8-2a.inc: Add tunes for supported ARMv8.2a cores
In reply to Jon MasonAdd tunes for all the ARMv8.2a cores currently supported in GCC. This
is: Cortex-A65, Cortex-A65AE, Cortex-A75, Cortex-A76, Cortex-A76AE, Cortex-A77, Neoverse-E1, Neoverse-N1, Cortex-A75-Cortex-A55, and Cortex-A76-Cortex-A55.
--- .../machine/include/arm/arch-armv8-2a.inc | 139 ++++++++++++++++++ 1 file changed, 139 insertions(+)
meta/conf/machine/include/arm/arch-armv8-2a.inc +139 -0
@@ -11,6 +11,145 @@ TUNE_FEATURES_tune-cortexa55 = "aarch64 cortexa55 crypto" PACKAGE_EXTRA_ARCHS_tune-cortexa55 = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} cortexa55" BASE_LIB_tune-cortexa55 = "lib64" +# +# Tune Settings for Cortex-A65 +# +TUNEVALID[cortexa65] = "Enable Cortex-A65 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa65', ' -mcpu=cortex-a65', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa65" +ARMPKGARCH_tune-cortexa65 = "cortexa65" +TUNE_FEATURES_tune-cortexa65 = "aarch64 cortexa65 crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa65 = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} cortexa65" +BASE_LIB_tune-cortexa65 = "lib64" + +# +# Tune Settings for Cortex-A65AE +# +TUNEVALID[cortexa65ae] = "Enable Cortex-A65AE specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa65ae', ' -mcpu=cortex-a65ae', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa65ae" +ARMPKGARCH_tune-cortexa65ae = "cortexa65ae" +TUNE_FEATURES_tune-cortexa65ae = "aarch64 cortexa65ae crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa65ae = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} cortexa65ae" +BASE_LIB_tune-cortexa65ae = "lib64" + +# +# Tune Settings for Cortex-A75 +# +TUNEVALID[cortexa75] = "Enable Cortex-A75 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa75', ' -mcpu=cortex-a75', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa75" +ARMPKGARCH_tune-cortexa75 = "cortexa75" +TUNE_FEATURES_tune-cortexa75 = "aarch64 cortexa75 crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa75 = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} cortexa75" +BASE_LIB_tune-cortexa75 = "lib64" + +# +# Tune Settings for Cortex-A76 +# +TUNEVALID[cortexa76] = "Enable Cortex-A76 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa76', ' -mcpu=cortex-a76', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa76" +ARMPKGARCH_tune-cortexa76 = "cortexa76" +TUNE_FEATURES_tune-cortexa76 = "aarch64 cortexa76 crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa76 = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} cortexa76" +BASE_LIB_tune-cortexa76 = "lib64" + +# +# Tune Settings for Cortex-A76AE +# +TUNEVALID[cortexa76ae] = "Enable Cortex-A76AE specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa76ae', ' -mcpu=cortex-a76ae', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa76ae" +ARMPKGARCH_tune-cortexa76ae = "cortexa76ae" +TUNE_FEATURES_tune-cortexa76ae = "aarch64 cortexa76ae crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa76ae = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} cortexa76ae" +BASE_LIB_tune-cortexa76ae = "lib64" + +# +# Tune Settings for Cortex-A77 +# +TUNEVALID[cortexa77] = "Enable Cortex-A77 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa77', ' -mcpu=cortex-a77', '', d)}" + +# Little Endian base configs +AVAILTUNES += "cortexa77" +ARMPKGARCH_tune-cortexa77 = "cortexa77" +TUNE_FEATURES_tune-cortexa77 = "aarch64 cortexa77 crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa77 = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} cortexa77" +BASE_LIB_tune-cortexa77 = "lib64" + +# +# Tune Settings for Neoverse-E1 +# +TUNEVALID[neoversee1] = "Enable Neoverse-E1 specific processor optimizations" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'neoversee1', ' -mcpu=neoverse-e1', '', d)}" + +# Little Endian base configs +AVAILTUNES += "neoversee1" +ARMPKGARCH_tune-neoversee1 = "neoversee1" +TUNE_FEATURES_tune-neoversee1 = "aarch64 neoversee1 crypto" +PACKAGE_EXTRA_ARCHS_tune-neoversee1 = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} neoversee1" +BASE_LIB_tune-neoversee1 = "lib64" + +# +# Tune Settings for Neoverse-N1 +# +TUNEVALID[neoversen1] = "Enable Neoverse-N1 specific processor optimizations" +# Note: Neoverse was called Ares, and GCC will accept "ares" in place of "neoverse-n1" +TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'neoversen1', ' -mcpu=neoverse-n1', '', d)}" + +# Little Endian base configs +AVAILTUNES += "neoversen1" +ARMPKGARCH_tune-neoversen1 = "neoversen1" +TUNE_FEATURES_tune-neoversen1 = "aarch64 neoversen1 crypto" +PACKAGE_EXTRA_ARCHS_tune-neoversen1 = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} neoversen1" +BASE_LIB_tune-neoversen1 = "lib64" + +# +# Tune Settings for big.LITTLE Cortex-A75 - Cortex-A55 +# +TUNEVALID[cortexa75-cortexa55] = "Enable big.LITTLE Cortex-A75.Cortex-A55 specific processor optimizations" +MACHINEOVERRIDES =. "${@bb.utils.contains("TUNE_FEATURES", "cortexa75-cortexa55", "cortexa75-cortexa55:", "" ,d)}" +TUNE_CCARGS .= "${@bb.utils.contains("TUNE_FEATURES", "cortexa75-cortexa55", " -mcpu=cortex-a75.cortex-a55", "", d)}" + +AVAILTUNES += "cortexa75-cortexa55 cortexa75-cortexa55-crypto" +ARMPKGARCH_tune-cortexa75-cortexa55 = "cortexa75-cortexa55" +ARMPKGARCH_tune-cortexa75-cortexa55-crypto = "cortexa75-cortexa55-crypto" +TUNE_FEATURES_tune-cortexa75-cortexa55 = "aarch64 cortexa75-cortexa55" +TUNE_FEATURES_tune-cortexa75-cortexa55-crypto = "${TUNE_FEATURES_tune-cortexa75-cortexa55} crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa75-cortexa55 = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a} cortexa75-cortexa55" +PACKAGE_EXTRA_ARCHS_tune-cortexa75-cortexa55-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} cortexa75-cortexa55 cortexa75-cortexa55-crypto" +BASE_LIB_tune-cortexa75-cortexa55 = "lib64" +BASE_LIB_tune-cortexa75-cortexa55-crypto = "lib64" + +# +# Tune Settings for big.LITTLE Cortex-A76 - Cortex-A55 +# +TUNEVALID[cortexa76-cortexa55] = "Enable big.LITTLE Cortex-A76.Cortex-A55 specific processor optimizations" +MACHINEOVERRIDES =. "${@bb.utils.contains("TUNE_FEATURES", "cortexa76-cortexa55", "cortexa76-cortexa55:", "" ,d)}" +TUNE_CCARGS .= "${@bb.utils.contains("TUNE_FEATURES", "cortexa76-cortexa55", " -mcpu=cortex-a76.cortex-a55", "", d)}" + +AVAILTUNES += "cortexa76-cortexa55 cortexa76-cortexa55-crypto" +ARMPKGARCH_tune-cortexa76-cortexa55 = "cortexa76-cortexa55" +ARMPKGARCH_tune-cortexa76-cortexa55-crypto = "cortexa76-cortexa55-crypto" +TUNE_FEATURES_tune-cortexa76-cortexa55 = "aarch64 cortexa76-cortexa55" +TUNE_FEATURES_tune-cortexa76-cortexa55-crypto = "${TUNE_FEATURES_tune-cortexa76-cortexa55} crypto" +PACKAGE_EXTRA_ARCHS_tune-cortexa76-cortexa55 = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a} cortexa76-cortexa55" +PACKAGE_EXTRA_ARCHS_tune-cortexa76-cortexa55-crypto = "${PACKAGE_EXTRA_ARCHS_tune-armv8-2a-crypto} cortexa76-cortexa55 cortexa76-cortexa55-crypto" +BASE_LIB_tune-cortexa76-cortexa55 = "lib64" +BASE_LIB_tune-cortexa76-cortexa55-crypto = "lib64" + # # Defaults for ARMv8-a #
Signature · 1 line
2.20.1
Quoted 1 line
This reduces the number of files from 12 to 2 for ARMv8a, and that is
excluding the 13 I am adding in this series that would otherwise be unique files.
I don't have a strong opinion on this anymore, but is the number of the include files the issue here?
I think the issue is the number of possible combinations and using the same TUNE_PKGARCH for different tunes (unlike 32bit arm tune files which use different). Bundling all these combinations in fewer include files doesn't IMHO improve it.
With 32bit arm include files it was useful to be able to compare various include files to check that they follow the same "structure" (or to add new cortexa* one by copying existing and regex-replace).
Also AVAILTUNES will now contain all possible tunes from given family (which makes it almost useless for aarch64 tunes). So instead of BSP including .inc file corresponding with the core used in the MACHINE and getting some "sane" default DEFAULTTUNE (plus some other possible options listed in AVAILTUNES) we now force every BSP to set more specific DEFAULTTUNE and then let user figure out what other DEFAULTTUNEs might be compatible with it, e.g. when someone is doing multi MACHINE builds and wants to share TUNE_PKGARCH between them (to save build time, package feed size etc).
Cheers,
Quoted 82 lines · Jon Mason
There is a large number of Arm Tune files located in meta/conf/machine/include/, and to support the current and upcoming Arm cores, more are needed. Adding more files is simply going to make it harder to find the relevant ones for an OE/YP developer/user. Also, there are problems with stale and erroneous configs (see my previous series), which will only be exacerbated by having more files.
I am proposing a reorganization of the existing tune files by including them in the generic family include file. For example, the tune-cortexa55.inc would be moved into arch-armv8-2a.inc, and tune-cortexa57.inc would be moved into arch-armv8a.inc. This reduces the number of files from 12 to 2 for ARMv8a, and that is excluding the 13 I am adding in this series that would otherwise be unique files.
To use, simply add
...
DEFAULTTUNE ?= "neoversen1"
require conf/machine/include/arm/arch-armv8-2a.inc
...Which is arguably what should be done anyway (instead of taking the default of the tune include file).
See the qemuarm64 patch in the series for a working example.Of course, by removing the existing tune files, current users are going to break. A simple script can be written to use sed (or similar) to replace the relevant parts for those users that would be affected (at least for those that are in the layer index and update regularly).
Thanks,
Jon---
Originally sent as a RFC in
https://lists.openembedded.org/g/openembedded-core/message/142324Given the generally positive feedback, sending as a patch series. Keeping the "hard fail" of the file removal (per Richards comment in
https://lists.openembedded.org/g/openembedded-core/message/142356).
Only difference of note is the removal of the "arm64: setBASE_LIBto lib64", as there needs to be more investigation (see
https://lists.openembedded.org/g/openembedded-core/message/142414).Jon Mason (5):
arch-armv8-2a.inc: Add Cortex-A55 tunings arch-armv8a.inc: Add existing tunings
qemuarm64: change tuning
arch-armv8a.inc: Add tunes for supported ARMv8a cores arch-armv8-2a.inc: Add tunes for supported ARMv8.2a cores.../machine/include/arm/arch-armv8-2a.inc | 175 ++++++++++++++- meta/conf/machine/include/arm/arch-armv8a.inc | 206 +++++++++++++++++- meta/conf/machine/include/tune-cortexa32.inc | 18 -- meta/conf/machine/include/tune-cortexa35.inc | 17 -- meta/conf/machine/include/tune-cortexa53.inc | 18 -- meta/conf/machine/include/tune-cortexa55.inc | 13 -- .../include/tune-cortexa57-cortexa53.inc | 15 -- meta/conf/machine/include/tune-cortexa57.inc | 17 -- .../include/tune-cortexa72-cortexa53.inc | 20 -- meta/conf/machine/include/tune-cortexa72.inc | 13 -- .../include/tune-cortexa73-cortexa53.inc | 20 -- meta/conf/machine/include/tune-thunderx.inc | 19 -- meta/conf/machine/qemuarm64.conf | 3 +- 13 files changed, 371 insertions(+), 183 deletions(-) delete mode 100644 meta/conf/machine/include/tune-cortexa32.inc delete mode 100644 meta/conf/machine/include/tune-cortexa35.inc delete mode 100644 meta/conf/machine/include/tune-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa55.inc delete mode 100644 meta/conf/machine/include/tune-cortexa57-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa57.inc delete mode 100644 meta/conf/machine/include/tune-cortexa72-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa72.inc delete mode 100644 meta/conf/machine/include/tune-cortexa73-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-thunderx.inc--
2.20.1
Quoted 4 lines · Martin Jansa
Quoted 2 lines
This reduces the number of files from 12 to 2 for ARMv8a, and that is excluding the 13 I am adding in this series that would otherwise be unique files.
I don't have a strong opinion on this anymore, but is the number of the include files the issue here?
This is the reason why I'm doing it. Arm needs to support all the cortex-a and cortex-m cores. This would make the majority of the tune files for only Arm cores. We could move them into meta/conf/machine/include/arm and not make the level down directory so messy, but I think this is cleaner.
Quoted 1 line
I think the issue is the number of possible combinations and using the same
TUNE_PKGARCHfor different tunes (unlike 32bit arm tune files which use different). Bundling all these combinations in fewer include files doesn't IMHO improve it.
The multitude of combinations and redundancy code entries makes it very error prone. By putting them all in the same file, it makes it much easier to see the differences in entries and catch errors. Longer term, I'd like to see something that removes the redundancies completely, but it will take some work.
Quoted 3 lines
With 32bit arm include files it was useful to be able to compare various include files to check that they follow the same "structure" (or to add new cortexa* one by copying existing and regex-replace).
Also AVAILTUNES will now contain all possible tunes from given family (which makes it almost useless for aarch64 tunes). So instead of BSP including .inc file corresponding with the core used in the MACHINE and getting some "sane" default DEFAULTTUNE (plus some other possible options listed in AVAILTUNES) we now force every BSP to set more specific DEFAULTTUNE and then let user figure out what other DEFAULTTUNEs might be compatible with it, e.g. when someone is doing multi MACHINE builds and wants to share
TUNE_PKGARCHbetween them (to save build time, package feed size etc).
You can still set a generic default for all the machines of a given family. By including armv8a.inc, you still get 'DEFAULTTUNE ?= "armv8a-crc"'. This will allow for a generic for an entire family. I'm not seeing the issue.
Thanks,
Jon
Quoted 86 lines
Cheers,
Quoted 82 lines · Jon Mason
There is a large number of Arm Tune files located in meta/conf/machine/include/, and to support the current and upcoming Arm cores, more are needed. Adding more files is simply going to make it harder to find the relevant ones for an OE/YP developer/user. Also, there are problems with stale and erroneous configs (see my previous series), which will only be exacerbated by having more files.
I am proposing a reorganization of the existing tune files by including them in the generic family include file. For example, the tune-cortexa55.inc would be moved into arch-armv8-2a.inc, and tune-cortexa57.inc would be moved into arch-armv8a.inc. This reduces the number of files from 12 to 2 for ARMv8a, and that is excluding the 13 I am adding in this series that would otherwise be unique files.
To use, simply add
...
DEFAULTTUNE ?= "neoversen1"
require conf/machine/include/arm/arch-armv8-2a.inc
...Which is arguably what should be done anyway (instead of taking the default of the tune include file).
See the qemuarm64 patch in the series for a working example.Of course, by removing the existing tune files, current users are going to break. A simple script can be written to use sed (or similar) to replace the relevant parts for those users that would be affected (at least for those that are in the layer index and update regularly).
Thanks,
Jon---
Originally sent as a RFC in
https://lists.openembedded.org/g/openembedded-core/message/142324Given the generally positive feedback, sending as a patch series. Keeping the "hard fail" of the file removal (per Richards comment in
https://lists.openembedded.org/g/openembedded-core/message/142356).
Only difference of note is the removal of the "arm64: setBASE_LIBto lib64", as there needs to be more investigation (see
https://lists.openembedded.org/g/openembedded-core/message/142414).Jon Mason (5):
arch-armv8-2a.inc: Add Cortex-A55 tunings arch-armv8a.inc: Add existing tunings
qemuarm64: change tuning
arch-armv8a.inc: Add tunes for supported ARMv8a cores arch-armv8-2a.inc: Add tunes for supported ARMv8.2a cores.../machine/include/arm/arch-armv8-2a.inc | 175 ++++++++++++++- meta/conf/machine/include/arm/arch-armv8a.inc | 206 +++++++++++++++++- meta/conf/machine/include/tune-cortexa32.inc | 18 -- meta/conf/machine/include/tune-cortexa35.inc | 17 -- meta/conf/machine/include/tune-cortexa53.inc | 18 -- meta/conf/machine/include/tune-cortexa55.inc | 13 -- .../include/tune-cortexa57-cortexa53.inc | 15 -- meta/conf/machine/include/tune-cortexa57.inc | 17 -- .../include/tune-cortexa72-cortexa53.inc | 20 -- meta/conf/machine/include/tune-cortexa72.inc | 13 -- .../include/tune-cortexa73-cortexa53.inc | 20 -- meta/conf/machine/include/tune-thunderx.inc | 19 -- meta/conf/machine/qemuarm64.conf | 3 +- 13 files changed, 371 insertions(+), 183 deletions(-) delete mode 100644 meta/conf/machine/include/tune-cortexa32.inc delete mode 100644 meta/conf/machine/include/tune-cortexa35.inc delete mode 100644 meta/conf/machine/include/tune-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa55.inc delete mode 100644 meta/conf/machine/include/tune-cortexa57-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa57.inc delete mode 100644 meta/conf/machine/include/tune-cortexa72-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa72.inc delete mode 100644 meta/conf/machine/include/tune-cortexa73-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-thunderx.inc--
2.20.1
Quoted 12 lines · Jon Mason
Quoted 4 lines · Martin Jansa
Quoted 2 lines
This reduces the number of files from 12 to 2 for ARMv8a, and that is excluding the 13 I am adding in this series that would otherwise be unique files.
I don't have a strong opinion on this anymore, but is the number of the include files the issue here?
This is the reason why I'm doing it. Arm needs to support all the cortex-a and cortex-m cores. This would make the majority of the tune files for only Arm cores. We could move them into meta/conf/machine/include/arm and not make the level down directory so messy, but I think this is cleaner.
Quoted 5 lines
Quoted 1 line
I think the issue is the number of possible combinations and using the same
TUNE_PKGARCHfor different tunes (unlike 32bit arm tune files which use different). Bundling all these combinations in fewer include files doesn't IMHO improve it.The multitude of combinations and redundancy code entries makes it very error prone. By putting them all in the same file, it makes it much easier to see the differences in entries and catch errors.
It does make it easier? In my experience it's more difficult to compare very similar long sections in different parts of the same file, than comparing very similar files with diff.
When working on
https://git.openembedded.org/openembedded-core-contrib/log/?h=jansa/tune-test
https://git.openembedded.org/openembedded-core-contrib/log/?h=jansa/tune2-test
https://git.openembedded.org/openembedded-core-contrib/log/?h=jansa/tune3
https://git.openembedded.org/openembedded-core-contrib/log/?h=jansa/optdefaulttune
long time ago, I found it really convenient to compare all cortex*.inc files just by replacing the number with some placeholder and comparing the resulting files (which were in most cases identical) - doing the same inside single file just makes it more complicated, probably by first splitting the various sections back to separate files again.
Quoted 2 lines
Longer term, I'd like to see something that removes the redundancies completely, but it will take some work.
So now all BSPs need to adapt (by changing from the more specific .inc file corresponding to the actual CPU in MACHINE they implement to generic family .inc while setting one of the possible DEFAULTTUNE, instead of leaving the default from .inc file), but all this only for the dubious benefit of having fewer .inc files in oe-core, right?
Quoted 8 lines
Quoted 3 lines
With 32bit arm include files it was useful to be able to compare various include files to check that they follow the same "structure" (or to add new cortexa* one by copying existing and regex-replace).
Also AVAILTUNES will now contain all possible tunes from given family (which makes it almost useless for aarch64 tunes). So instead of BSP including .inc file corresponding with the core used in the MACHINE and getting some "sane" default DEFAULTTUNE (plus some other possible options listed in AVAILTUNES) we now force every BSP to set more specific DEFAULTTUNE and then let user figure out what other DEFAULTTUNEs might be compatible with it, e.g. when someone is doing multi MACHINE builds and wants to share
TUNE_PKGARCHbetween them (to save build time, package feed size etc).You can still set a generic default for all the machines of a given family. By including armv8a.inc, you still get 'DEFAULTTUNE ?= "armv8a-crc"'. This will allow for a generic for an entire family. I'm not seeing the issue.
OK, with e.g. raspberrypi4-64.conf from
https://github.com/agherzan/meta-raspberrypi/blob/a4c8118676ba8002edab29fc81b4e4edd9fad1f1/conf/machine/raspberrypi4-64.conf
which just
require conf/machine/include/tune-cortexa72.inc
Without your patches:
DEFAULTTUNE="cortexa72"
TUNE_PKGARCH="cortexa72"
TUNE_CCARGS=" -mcpu=cortex-a72+crc+crypto"
AVAILTUNES=" armv4 armv4t armv4b armv4tb armv5 armv5t armv5-vfp armv5t-vfp armv5hf-vfp armv5thf-vfp armv5b armv5tb armv5b-vfp armv5tb-vfp armv5hfb-vfp armv5thfb-vfp armv5e armv5te armv5e-vfp armv5te-vfp armv5ehf-vfp armv5tehf-vfp armv5eb armv5teb armv5eb-vfp armv5teb-vfp armv5ehfb-vfp armv5tehfb-vfp armv6-novfp armv6t-novfp armv6 armv6t armv6hf armv6thf armv6b-novfp armv6tb-novfp armv6b armv6tb armv6hfb armv6thfb armv7a armv7at armv7a-vfpv3d16 armv7at-vfpv3d16 armv7a-vfpv3 armv7at-vfpv3 armv7a-vfpv4d16 armv7at-vfpv4d16 armv7a-neon armv7at-neon armv7a-neon-vfpv4 armv7at-neon-vfpv4 armv7ahf armv7athf armv7ahf-vfpv3d16 armv7athf-vfpv3d16 armv7ahf-vfpv3 armv7athf-vfpv3 armv7ahf-vfpv4d16 armv7athf-vfpv4d16 armv7ahf-neon armv7athf-neon armv7ahf-neon-vfpv4 armv7athf-neon-vfpv4 armv7ab armv7atb armv7ab-vfpv3d16 armv7atb-vfpv3d16 armv7ab-vfpv3 armv7atb-vfpv3 armv7ab-vfpv4d16 armv7atb-vfpv4d16 armv7ab-neon armv7atb-neon armv7ab-neon-vfpv4 armv7atb-neon-vfpv4 armv7ahfb armv7athfb armv7ahfb-vfpv3d16 armv7athfb-vfpv3d16 armv7ahfb-vfpv3 armv7athfb-vfpv3 armv7ahfb-vfpv4d16 armv7athfb-vfpv4d16 armv7ahfb-neon armv7athfb-neon armv7ahfb-neon-vfpv4 armv7athfb-neon-vfpv4 armv7ve armv7vet armv7ve-vfpv3d16 armv7vet-vfpv3d16 armv7ve-vfpv3 armv7vet-vfpv3 armv7ve-vfpv4d16 armv7vet-vfpv4d16 armv7ve-neon armv7vet-neon armv7ve-neon-vfpv4 armv7vet-neon-vfpv4 armv7vehf armv7vethf armv7vehf-vfpv3d16 armv7vethf-vfpv3d16 armv7vehf-vfpv3 armv7vethf-vfpv3 armv7vehf-vfpv4d16 armv7vethf-vfpv4d16 armv7vehf-neon armv7vethf-neon armv7vehf-neon-vfpv4 armv7vethf-neon-vfpv4 armv7veb armv7vetb armv7veb-vfpv3d16 armv7vetb-vfpv3d16 armv7veb-vfpv3 armv7vetb-vfpv3 armv7veb-vfpv4d16 armv7vetb-vfpv4d16 armv7veb-neon armv7vetb-neon armv7veb-neon-vfpv4 armv7vetb-neon-vfpv4 armv7vehfb armv7vethfb armv7vehfb-vfpv3d16 armv7vethfb-vfpv3d16 armv7vehfb-vfpv3 armv7vethfb-vfpv3 armv7vehfb-vfpv4d16 armv7vethfb-vfpv4d16 armv7vehfb-neon armv7vethfb-neon armv7vehfb-neon-vfpv4 armv7vethfb-neon-vfpv4 aarch64 aarch64_be armv8a armv8a-crc armv8a-crc-crypto armv8a-crypto cortexa72"
With your patches:
ERROR: ParseError at /OE/build/oe-core/meta-raspberrypi/conf/machine/raspberrypi4-64.conf:13: Could not include required file conf/machine/include/tune-cortexa72.inc
After changing the require to conf/machine/include/arm/arch-armv8a.inc:
DEFAULTTUNE="armv8a-crc"
TUNE_PKGARCH="armv8a"
TUNE_CCARGS=" -march=armv8-a+crc"
AVAILTUNES=" cortexa32 cortexa32-crypto cortexa34 cortexa34-crypto cortexa35 cortexa35-crypto cortexa53 cortexa53-crypto cortexa57 cortexa57-crypto cortexa72 cortexa73 thunderx thunderx_be cortexa57-cortexa53 cortexa72-cortexa53 cortexa72-cortexa53-crypto cortexa73-cortexa35 cortexa73-cortexa35-crypto cortexa73-cortexa53 cortexa73-cortexa53-crypto armv4 armv4t armv4b armv4tb armv5 armv5t armv5-vfp armv5t-vfp armv5hf-vfp armv5thf-vfp armv5b armv5tb armv5b-vfp armv5tb-vfp armv5hfb-vfp armv5thfb-vfp armv5e armv5te armv5e-vfp armv5te-vfp armv5ehf-vfp armv5tehf-vfp armv5eb armv5teb armv5eb-vfp armv5teb-vfp armv5ehfb-vfp armv5tehfb-vfp armv6-novfp armv6t-novfp armv6 armv6t armv6hf armv6thf armv6b-novfp armv6tb-novfp armv6b armv6tb armv6hfb armv6thfb armv7a armv7at armv7a-vfpv3d16 armv7at-vfpv3d16 armv7a-vfpv3 armv7at-vfpv3 armv7a-vfpv4d16 armv7at-vfpv4d16 armv7a-neon armv7at-neon armv7a-neon-vfpv4 armv7at-neon-vfpv4 armv7ahf armv7athf armv7ahf-vfpv3d16 armv7athf-vfpv3d16 armv7ahf-vfpv3 armv7athf-vfpv3 armv7ahf-vfpv4d16 armv7athf-vfpv4d16 armv7ahf-neon armv7athf-neon armv7ahf-neon-vfpv4 armv7athf-neon-vfpv4 armv7ab armv7atb armv7ab-vfpv3d16 armv7atb-vfpv3d16 armv7ab-vfpv3 armv7atb-vfpv3 armv7ab-vfpv4d16 armv7atb-vfpv4d16 armv7ab-neon armv7atb-neon armv7ab-neon-vfpv4 armv7atb-neon-vfpv4 armv7ahfb armv7athfb armv7ahfb-vfpv3d16 armv7athfb-vfpv3d16 armv7ahfb-vfpv3 armv7athfb-vfpv3 armv7ahfb-vfpv4d16 armv7athfb-vfpv4d16 armv7ahfb-neon armv7athfb-neon armv7ahfb-neon-vfpv4 armv7athfb-neon-vfpv4 armv7ve armv7vet armv7ve-vfpv3d16 armv7vet-vfpv3d16 armv7ve-vfpv3 armv7vet-vfpv3 armv7ve-vfpv4d16 armv7vet-vfpv4d16 armv7ve-neon armv7vet-neon armv7ve-neon-vfpv4 armv7vet-neon-vfpv4 armv7vehf armv7vethf armv7vehf-vfpv3d16 armv7vethf-vfpv3d16 armv7vehf-vfpv3 armv7vethf-vfpv3 armv7vehf-vfpv4d16 armv7vethf-vfpv4d16 armv7vehf-neon armv7vethf-neon armv7vehf-neon-vfpv4 armv7vethf-neon-vfpv4 armv7veb armv7vetb armv7veb-vfpv3d16 armv7vetb-vfpv3d16 armv7veb-vfpv3 armv7vetb-vfpv3 armv7veb-vfpv4d16 armv7vetb-vfpv4d16 armv7veb-neon armv7vetb-neon armv7veb-neon-vfpv4 armv7vetb-neon-vfpv4 armv7vehfb armv7vethfb armv7vehfb-vfpv3d16 armv7vethfb-vfpv3d16 armv7vehfb-vfpv3 armv7vethfb-vfpv3 armv7vehfb-vfpv4d16 armv7vethfb-vfpv4d16 armv7vehfb-neon armv7vethfb-neon armv7vehfb-neon-vfpv4 armv7vethfb-neon-vfpv4 aarch64 aarch64_be armv8a armv8a-crc armv8a-crc-crypto armv8a-crypto"
Should really raspberrypi4-64 users see e.g. cortexa73-cortexa35-crypto in AVAILTUNES even when it doesn't make sense for their HW?
Quoted 83 lines
Quoted 83 lines · Jon Mason
Quoted 82 lines
There is a large number of Arm Tune files located in meta/conf/machine/include/, and to support the current and upcoming Arm cores, more are needed. Adding more files is simply going to make it harder to find the relevant ones for an OE/YP developer/user. Also, there are problems with stale and erroneous configs (see my previous series), which will only be exacerbated by having more files.
I am proposing a reorganization of the existing tune files by including them in the generic family include file. For example, the tune-cortexa55.inc would be moved into arch-armv8-2a.inc, and tune-cortexa57.inc would be moved into arch-armv8a.inc. This reduces the number of files from 12 to 2 for ARMv8a, and that is excluding the 13 I am adding in this series that would otherwise be unique files.
To use, simply add
...
DEFAULTTUNE ?= "neoversen1"
require conf/machine/include/arm/arch-armv8-2a.inc
...Which is arguably what should be done anyway (instead of taking the default of the tune include file).
See the qemuarm64 patch in the series for a working example.Of course, by removing the existing tune files, current users are going to break. A simple script can be written to use sed (or similar) to replace the relevant parts for those users that would be affected (at least for those that are in the layer index and update regularly).
Thanks,
Jon---
Originally sent as a RFC in
https://lists.openembedded.org/g/openembedded-core/message/142324Given the generally positive feedback, sending as a patch series. Keeping the "hard fail" of the file removal (per Richards comment in
https://lists.openembedded.org/g/openembedded-core/message/142356).
Only difference of note is the removal of the "arm64: setBASE_LIBto lib64", as there needs to be more investigation (see
https://lists.openembedded.org/g/openembedded-core/message/142414).Jon Mason (5):
arch-armv8-2a.inc: Add Cortex-A55 tunings arch-armv8a.inc: Add existing tunings
qemuarm64: change tuning
arch-armv8a.inc: Add tunes for supported ARMv8a cores arch-armv8-2a.inc: Add tunes for supported ARMv8.2a cores.../machine/include/arm/arch-armv8-2a.inc | 175 ++++++++++++++- meta/conf/machine/include/arm/arch-armv8a.inc | 206 +++++++++++++++++- meta/conf/machine/include/tune-cortexa32.inc | 18 -- meta/conf/machine/include/tune-cortexa35.inc | 17 -- meta/conf/machine/include/tune-cortexa53.inc | 18 -- meta/conf/machine/include/tune-cortexa55.inc | 13 -- .../include/tune-cortexa57-cortexa53.inc | 15 -- meta/conf/machine/include/tune-cortexa57.inc | 17 -- .../include/tune-cortexa72-cortexa53.inc | 20 -- meta/conf/machine/include/tune-cortexa72.inc | 13 -- .../include/tune-cortexa73-cortexa53.inc | 20 -- meta/conf/machine/include/tune-thunderx.inc | 19 -- meta/conf/machine/qemuarm64.conf | 3 +- 13 files changed, 371 insertions(+), 183 deletions(-) delete mode 100644 meta/conf/machine/include/tune-cortexa32.inc delete mode 100644 meta/conf/machine/include/tune-cortexa35.inc delete mode 100644 meta/conf/machine/include/tune-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa55.inc delete mode 100644 meta/conf/machine/include/tune-cortexa57-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa57.inc delete mode 100644 meta/conf/machine/include/tune-cortexa72-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa72.inc delete mode 100644 meta/conf/machine/include/tune-cortexa73-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-thunderx.inc--
2.20.1
Quoted 13 lines · Martin Jansa
Quoted 12 lines · Jon Mason
Quoted 4 lines · Martin Jansa
Quoted 2 lines
This reduces the number of files from 12 to 2 for ARMv8a, and that is excluding the 13 I am adding in this series that would otherwise be unique files.
I don't have a strong opinion on this anymore, but is the number of the include files the issue here?
This is the reason why I'm doing it. Arm needs to support all the cortex-a and cortex-m cores. This would make the majority of the tune files for only Arm cores. We could move them into meta/conf/machine/include/arm and not make the level down directory so messy, but I think this is cleaner.
Quoted 7 lines
Quoted 5 lines
Quoted 1 line
I think the issue is the number of possible combinations and using the same
TUNE_PKGARCHfor different tunes (unlike 32bit arm tune files which use different). Bundling all these combinations in fewer include files doesn't IMHO improve it.The multitude of combinations and redundancy code entries makes it very error prone. By putting them all in the same file, it makes it much easier to see the differences in entries and catch errors.
It does make it easier? In my experience it's more difficult to compare
For some of my usecases, it makes it significantly easier.
I need to build multilib configurations that have a variety of tunes in them. Before I had to include 5-7 different .inc files, and then ignore the warning about the same files being included multiple times.
With this approach, I can include just one file, the proper superset of them all, and then do my multilib configurations together.
(Yes, I realize this isn't the typical Linux use-case, but this seems to be fairly common to do things like this when working with baremetal configurations or trying to use the Yocto Project to build toolchains that are shared with others.)
Quoted 13 lines
very similar long sections in different parts of the same file, than comparing very similar files with diff.
When working on
https://git.openembedded.org/openembedded-core-contrib/log/?h=jansa/tune-test
https://git.openembedded.org/openembedded-core-contrib/log/?h=jansa/tune2-test
https://git.openembedded.org/openembedded-core-contrib/log/?h=jansa/tune3
https://git.openembedded.org/openembedded-core-contrib/log/?h=jansa/optdefaulttune
long time ago, I found it really convenient to compare all cortex*.inc files just by replacing the number with some placeholder and comparing the resulting files (which were in most cases identical) - doing the same inside single file just makes it more complicated, probably by first splitting the various sections back to separate files again.
Quoted 8 lines
Quoted 2 lines
Longer term, I'd like to see something that removes the redundancies completely, but it will take some work.
So now all BSPs need to adapt (by changing from the more specific .inc file corresponding to the actual CPU in MACHINE they implement to generic family .inc while setting one of the possible DEFAULTTUNE, instead of leaving the default from .inc file), but all this only for the dubious benefit of having fewer .inc files in oe-core, right?
I suggested how a transition could be done, but this was rejected in favor of make people move their BSPs to the new format.
--Mark
Quoted 184 lines
Quoted 8 lines
Quoted 3 lines
With 32bit arm include files it was useful to be able to compare various include files to check that they follow the same "structure" (or to add new cortexa* one by copying existing and regex-replace).
Also AVAILTUNES will now contain all possible tunes from given family (which makes it almost useless for aarch64 tunes). So instead of BSP including .inc file corresponding with the core used in the MACHINE and getting some "sane" default DEFAULTTUNE (plus some other possible options listed in AVAILTUNES) we now force every BSP to set more specific DEFAULTTUNE and then let user figure out what other DEFAULTTUNEs might be compatible with it, e.g. when someone is doing multi MACHINE builds and wants to share
TUNE_PKGARCHbetween them (to save build time, package feed size etc).You can still set a generic default for all the machines of a given family. By including armv8a.inc, you still get 'DEFAULTTUNE ?= "armv8a-crc"'. This will allow for a generic for an entire family. I'm not seeing the issue.
OK, with e.g. raspberrypi4-64.conf from
https://github.com/agherzan/meta-raspberrypi/blob/a4c8118676ba8002edab29fc81b4e4edd9fad1f1/conf/machine/raspberrypi4-64.conf
which just
require conf/machine/include/tune-cortexa72.incWithout your patches:
DEFAULTTUNE="cortexa72"
TUNE_PKGARCH="cortexa72"
TUNE_CCARGS=" -mcpu=cortex-a72+crc+crypto"
AVAILTUNES=" armv4 armv4t armv4b armv4tb armv5 armv5t armv5-vfp armv5t-vfp armv5hf-vfp armv5thf-vfp armv5b armv5tb armv5b-vfp armv5tb-vfp armv5hfb-vfp armv5thfb-vfp armv5e armv5te armv5e-vfp armv5te-vfp armv5ehf-vfp armv5tehf-vfp armv5eb armv5teb armv5eb-vfp armv5teb-vfp armv5ehfb-vfp armv5tehfb-vfp armv6-novfp armv6t-novfp armv6 armv6t armv6hf armv6thf armv6b-novfp armv6tb-novfp armv6b armv6tb armv6hfb armv6thfb armv7a armv7at armv7a-vfpv3d16 armv7at-vfpv3d16 armv7a-vfpv3 armv7at-vfpv3 armv7a-vfpv4d16 armv7at-vfpv4d16 armv7a-neon armv7at-neon armv7a-neon-vfpv4 armv7at-neon-vfpv4 armv7ahf armv7athf armv7ahf-vfpv3d16 armv7athf-vfpv3d16 armv7ahf-vfpv3 armv7athf-vfpv3 armv7ahf-vfpv4d16 armv7athf-vfpv4d16 armv7ahf-neon armv7athf-neon armv7ahf-neon-vfpv4 armv7athf-neon-vfpv4 armv7ab armv7atb armv7ab-vfpv3d16 armv7atb-vfpv3d16 armv7ab-vfpv3 armv7atb-vfpv3 armv7ab-vfpv4d16 armv7atb-vfpv4d16 armv7ab-neon armv7atb-neon armv7ab-neon-vfpv4 armv7atb-neon-vfpv4 armv7ahfb armv7athfb armv7ahfb-vfpv3d16 armv7athfb-vfpv3d16 armv7ahfb-vfpv3 armv7athfb-vfpv3 armv7ahfb-vfpv4d16 armv7athfb-vfpv4d16 armv7ahfb-neon armv7athfb-neon armv7ahfb-neon-vfpv4 armv7athfb-neon-vfpv4 armv7ve armv7vet armv7ve-vfpv3d16 armv7vet-vfpv3d16 armv7ve-vfpv3 armv7vet-vfpv3 armv7ve-vfpv4d16 armv7vet-vfpv4d16 armv7ve-neon armv7vet-neon armv7ve-neon-vfpv4 armv7vet-neon-vfpv4 armv7vehf armv7vethf armv7vehf-vfpv3d16 armv7vethf-vfpv3d16 armv7vehf-vfpv3 armv7vethf-vfpv3 armv7vehf-vfpv4d16 armv7vethf-vfpv4d16 armv7vehf-neon armv7vethf-neon armv7vehf-neon-vfpv4 armv7vethf-neon-vfpv4 armv7veb armv7vetb armv7veb-vfpv3d16 armv7vetb-vfpv3d16 armv7veb-vfpv3 armv7vetb-vfpv3 armv7veb-vfpv4d16 armv7vetb-vfpv4d16 armv7veb-neon armv7vetb-neon armv7veb-neon-vfpv4 armv7vetb-neon-vfpv4 armv7vehfb armv7vethfb armv7vehfb-vfpv3d16 armv7vethfb-vfpv3d16 armv7vehfb-vfpv3 armv7vethfb-vfpv3 armv7vehfb-vfpv4d16 armv7vethfb-vfpv4d16 armv7vehfb-neon armv7vethfb-neon armv7vehfb-neon-vfpv4 armv7vethfb-neon-vfpv4 aarch64aarch64_bearmv8a armv8a-crc armv8a-crc-crypto armv8a-crypto cortexa72"With your patches:
ERROR: ParseError at /OE/build/oe-core/meta-raspberrypi/conf/machine/raspberrypi4-64.conf:13: Could not include required file conf/machine/include/tune-cortexa72.incAfter changing the require to conf/machine/include/arm/arch-armv8a.inc:
DEFAULTTUNE="armv8a-crc"
TUNE_PKGARCH="armv8a"
TUNE_CCARGS=" -march=armv8-a+crc"
AVAILTUNES=" cortexa32 cortexa32-crypto cortexa34 cortexa34-crypto cortexa35 cortexa35-crypto cortexa53 cortexa53-crypto cortexa57 cortexa57-crypto cortexa72 cortexa73 thunderxthunderx_becortexa57-cortexa53 cortexa72-cortexa53 cortexa72-cortexa53-crypto cortexa73-cortexa35 cortexa73-cortexa35-crypto cortexa73-cortexa53 cortexa73-cortexa53-crypto armv4 armv4t armv4b armv4tb armv5 armv5t armv5-vfp armv5t-vfp armv5hf-vfp armv5thf-vfp armv5b armv5tb armv5b-vfp armv5tb-vfp armv5hfb-vfp armv5thfb-vfp armv5e armv5te armv5e-vfp armv5te-vfp armv5ehf-vfp armv5tehf-vfp armv5eb armv5teb armv5eb-vfp armv5teb-vfp armv5ehfb-vfp armv5tehfb-vfp armv6-novfp armv6t-novfp armv6 armv6t armv6hf armv6thf armv6b-novfp armv6tb-novfp armv6b armv6tb armv6hfb armv6thfb armv7a armv7at armv7a-vfpv3d16 armv7at-vfpv3d16 armv7a-vfpv3 armv7at-vfpv3 armv7a-vfpv4d16 armv7at-vfpv4d16 armv7a-neon armv7at-neon armv7a-neon-vfpv4 armv7at-neon-vfpv4 armv7ahf armv7athf armv7ahf-vfpv3d16 armv7athf-vfpv3d16 armv7ahf-vfpv3 armv7athf-vfpv3 armv7ahf-vfpv4d16 armv7athf-vfpv4d16 armv7ahf-neon armv7athf-neon armv7ahf-neon-vfpv4 armv7athf-neon-vfpv4 armv7ab armv7atb armv7ab-vfpv3d16 armv7atb-vfpv3d16 armv7ab-vfpv3 armv7atb-vfpv3 armv7ab-vfpv4d16 armv7atb-vfpv4d16 armv7ab-neon armv7atb-neon armv7ab-neon-vfpv4 armv7atb-neon-vfpv4 armv7ahfb armv7athfb armv7ahfb-vfpv3d16 armv7athfb-vfpv3d16 armv7ahfb-vfpv3 armv7athfb-vfpv3 armv7ahfb-vfpv4d16 armv7athfb-vfpv4d16 armv7ahfb-neon armv7athfb-neon armv7ahfb-neon-vfpv4 armv7athfb-neon-vfpv4 armv7ve armv7vet armv7ve-vfpv3d16 armv7vet-vfpv3d16 armv7ve-vfpv3 armv7vet-vfpv3 armv7ve-vfpv4d16 armv7vet-vfpv4d16 armv7ve-neon armv7vet-neon armv7ve-neon-vfpv4 armv7vet-neon-vfpv4 armv7vehf armv7vethf armv7vehf-vfpv3d16 armv7vethf-vfpv3d16 armv7vehf-vfpv3 armv7vethf-vfpv3 armv7vehf-vfpv4d16 armv7vethf-vfpv4d16 armv7vehf-neon armv7vethf-neon armv7vehf-neon-vfpv4 armv7vethf-neon-vfpv4 armv7veb armv7vetb armv7veb-vfpv3d16 armv7vetb-vfpv3d16 armv7veb-vfpv3 armv7vetb-vfpv3 armv7veb-vfpv4d16 armv7vetb-vfpv4d16 armv7veb-neon armv7vetb-neon armv7veb-neon-vfpv4 armv7vetb-neon-vfpv4 armv7vehfb armv7vethfb armv7vehfb-vfpv3d16 armv7vethfb-vfpv3d16 armv7vehfb-vfpv3 armv7vethfb-vfpv3 armv7vehfb-vfpv4d16 armv7vethfb-vfpv4d16 armv7vehfb-neon armv7vethfb-neon armv7vehfb-neon-vfpv4 armv7vethfb-neon-vfpv4 aarch64aarch64_bearmv8a armv8a-crc armv8a-crc-crypto armv8a-crypto"Should really raspberrypi4-64 users see e.g. cortexa73-cortexa35-crypto in AVAILTUNES even when it doesn't make sense for their HW?
Quoted 85 lines
Quoted 85 lines · Jon Mason
Quoted 84 lines
There is a large number of Arm Tune files located in meta/conf/machine/include/, and to support the current and upcoming Arm cores, more are needed. Adding more files is simply going to make it harder to find the relevant ones for an OE/YP developer/user. Also, there are problems with stale and erroneous configs (see my previous series), which will only be exacerbated by having more files.
I am proposing a reorganization of the existing tune files by including them in the generic family include file. For example, the tune-cortexa55.inc would be moved into arch-armv8-2a.inc, and tune-cortexa57.inc would be moved into arch-armv8a.inc. This reduces the number of files from 12 to 2 for ARMv8a, and that is excluding the 13 I am adding in this series that would otherwise be unique files.
To use, simply add
...
DEFAULTTUNE ?= "neoversen1"
require conf/machine/include/arm/arch-armv8-2a.inc
...Which is arguably what should be done anyway (instead of taking the default of the tune include file).
See the qemuarm64 patch in the series for a working example.Of course, by removing the existing tune files, current users are going to break. A simple script can be written to use sed (or similar) to replace the relevant parts for those users that would be affected (at least for those that are in the layer index and update regularly).
Thanks,
Jon---
Originally sent as a RFC in
https://lists.openembedded.org/g/openembedded-core/message/142324Given the generally positive feedback, sending as a patch series. Keeping the "hard fail" of the file removal (per Richards comment in
https://lists.openembedded.org/g/openembedded-core/message/142356).
Only difference of note is the removal of the "arm64: setBASE_LIBto lib64", as there needs to be more investigation (see
https://lists.openembedded.org/g/openembedded-core/message/142414).Jon Mason (5):
arch-armv8-2a.inc: Add Cortex-A55 tunings arch-armv8a.inc: Add existing tunings
qemuarm64: change tuning
arch-armv8a.inc: Add tunes for supported ARMv8a cores arch-armv8-2a.inc: Add tunes for supported ARMv8.2a cores.../machine/include/arm/arch-armv8-2a.inc | 175 ++++++++++++++- meta/conf/machine/include/arm/arch-armv8a.inc | 206 +++++++++++++++++- meta/conf/machine/include/tune-cortexa32.inc | 18 -- meta/conf/machine/include/tune-cortexa35.inc | 17 -- meta/conf/machine/include/tune-cortexa53.inc | 18 -- meta/conf/machine/include/tune-cortexa55.inc | 13 -- .../include/tune-cortexa57-cortexa53.inc | 15 -- meta/conf/machine/include/tune-cortexa57.inc | 17 -- .../include/tune-cortexa72-cortexa53.inc | 20 -- meta/conf/machine/include/tune-cortexa72.inc | 13 -- .../include/tune-cortexa73-cortexa53.inc | 20 -- meta/conf/machine/include/tune-thunderx.inc | 19 -- meta/conf/machine/qemuarm64.conf | 3 +- 13 files changed, 371 insertions(+), 183 deletions(-) delete mode 100644 meta/conf/machine/include/tune-cortexa32.inc delete mode 100644 meta/conf/machine/include/tune-cortexa35.inc delete mode 100644 meta/conf/machine/include/tune-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa55.inc delete mode 100644 meta/conf/machine/include/tune-cortexa57-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa57.inc delete mode 100644 meta/conf/machine/include/tune-cortexa72-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-cortexa72.inc delete mode 100644 meta/conf/machine/include/tune-cortexa73-cortexa53.inc delete mode 100644 meta/conf/machine/include/tune-thunderx.inc--
2.20.1
Hi Jon,
That's not really a comment on the reorganization of compiler tunes, but more like "Do they actually do something meaningful?"
I posted here[1] some benchmarks and at least with the benchmarks I tried on the chips I tried there is no obvious impact.
i.mx6q:
TUNE_FEATURES = "arm armv7a vfp thumb callconvention-hard"TARGET_FPU = "hard"
vs.
TUNE_FEATURES = "arm vfp cortexa9 neon thumb callconvention-hard"TARGET_FPU = "hard"
i.m8mm:
TUNE_FEATURES = "aarch64 cortexa53 crc crypto"TARGET_FPU = ""
vs.
TUNE_FEATURES = "aarch64 armv8a crc crypto"TARGET_FPU = ""
[1] https://yoctoproject.blogspot.com/2020/09/compiler-tunes-benchmarks-with-yocto.html
Should we expect so see differences?
If so can you suggest benchmarks which show those differences?
Regards,
Robert
Quoted 27 lines · Robert Berger
Hi Jon,
That's not really a comment on the reorganization of compiler tunes, but more like "Do they actually do something meaningful?"
I posted here[1] some benchmarks and at least with the benchmarks I tried on the chips I tried there is no obvious impact.
i.mx6q:
TUNE_FEATURES= "arm armv7a vfp thumb callconvention-hard"TARGET_FPU= "hard"vs.
TUNE_FEATURES= "arm vfp cortexa9 neon thumb callconvention-hard"TARGET_FPU= "hard"i.m8mm:
TUNE_FEATURES= "aarch64 cortexa53 crc crypto"TARGET_FPU= ""vs.
Quoted 8 lines
TUNE_FEATURES= "aarch64 armv8a crc crypto"TARGET_FPU= ""[1]
https://yoctoproject.blogspot.com/2020/09/compiler-tunes-benchmarks-with-yocto.htmlShould we expect so see differences?
There are more things at play here than simply performance. But speaking of performance, there might not be much of a benefit between a generic ARMv8.0 and a ARMv8 based core (like A53), but I do expect to see a performance bump for a ARMv8.2 based core (like A76). The delta between the former is much smaller than the latter.
The differences are not only performance. Tuning for A76 versus a more generic armv8a allows for security features like branch-protection to be enabled (as it isn't supported in older versions). You get these kind of things "by default" when tuning for the specific model.
Thanks,
Jon
Quoted 6 lines
If so can you suggest benchmarks which show those differences?
Regards,
Robert
Hi,
Please see my comments in-line.
Quoted 6 lines · Jon Mason
Quoted 4 lines
[1]
https://yoctoproject.blogspot.com/2020/09/compiler-tunes-benchmarks-with-yocto.htmlShould we expect so see differences?
There are more things at play here than simply performance.
Hmm interesting. This did not cross my mind at all;)
Quoted 5 lines
But
speaking of performance, there might not be much of a benefit between a generic ARMv8.0 and a ARMv8 based core (like A53), but I do expect to see a performance bump for a ARMv8.2 based core (like A76). The delta between the former is much smaller than the latter.
I guess you mean between ARMv8.0/A53 and ARMv8.2/A76. I'll need some toys to run tests/benchmarks there;)
Quoted 6 lines
The differences are not only performance. Tuning for A76 versus a more generic armv8a allows for security features like branch-protection to be enabled (as it isn't supported in older versions). You get these kind of things "by default" when tuning for the specific model.
OK - This is a very good point I completely ignored so far. I guess this will be tricky to test.
Quoted 10 lines
Thanks,
JonQuoted 6 lines
If so can you suggest benchmarks which show those differences?
Regards,
Robert
Thanks - I updated my Blog post.
Regards,
Robert
Quoted 20 lines · Jon Mason
There is a large number of Arm Tune files located in meta/conf/machine/include/, and to support the current and upcoming Arm cores, more are needed. Adding more files is simply going to make it harder to find the relevant ones for an OE/YP developer/user. Also, there are problems with stale and erroneous configs (see my previous series), which will only be exacerbated by having more files.
I am proposing a reorganization of the existing tune files by including them in the generic family include file. For example, the tune-cortexa55.inc would be moved into arch-armv8-2a.inc, and tune-cortexa57.inc would be moved into arch-armv8a.inc. This reduces the number of files from 12 to 2 for ARMv8a, and that is excluding the 13 I am adding in this series that would otherwise be unique files.
To use, simply add
...
DEFAULTTUNE ?= "neoversen1"
require conf/machine/include/arm/arch-armv8-2a.inc
...
Quoted 8 lines
Which is arguably what should be done anyway (instead of taking the default of the tune include file).
See the qemuarm64 patch in the series for a working example.Of course, by removing the existing tune files, current users are going to break. A simple script can be written to use sed (or similar) to replace the relevant parts for those users that would be affected (at least for those that are in the layer index and update regularly).
I've just looked at this in a bit more details and I'm worried.
The intent is the BSP includes the processor/core tune file that its based upon. The BSP should usually know which one is present.
That tune then presents the possible options which the BSPs selects from but the end user or distro can override.
With this change, you get all tunes and you don't know which are compatible with a given core/processor. I'm not sure that is an improvement.
I appreciate not wanting "lots of files" but we need to consider the usability too.
As others have mentioned, the errors are fairly obvious with diff between the files (or a GUI like meld).
What are the advantages this brings other than fewer files? Am I missing something?
Cheers,
Richard
Quoted 36 lines · Richard Purdie
Quoted 20 lines · Jon Mason
There is a large number of Arm Tune files located in meta/conf/machine/include/, and to support the current and upcoming Arm cores, more are needed. Adding more files is simply going to make it harder to find the relevant ones for an OE/YP developer/user. Also, there are problems with stale and erroneous configs (see my previous series), which will only be exacerbated by having more files.
I am proposing a reorganization of the existing tune files by including them in the generic family include file. For example, the tune-cortexa55.inc would be moved into arch-armv8-2a.inc, and tune-cortexa57.inc would be moved into arch-armv8a.inc. This reduces the number of files from 12 to 2 for ARMv8a, and that is excluding the 13 I am adding in this series that would otherwise be unique files.
To use, simply add
...
DEFAULTTUNE ?= "neoversen1"
require conf/machine/include/arm/arch-armv8-2a.inc
...Quoted 8 lines
Which is arguably what should be done anyway (instead of taking the default of the tune include file).
See the qemuarm64 patch in the series for a working example.Of course, by removing the existing tune files, current users are going to break. A simple script can be written to use sed (or similar) to replace the relevant parts for those users that would be affected (at least for those that are in the layer index and update regularly).
I've just looked at this in a bit more details and I'm worried.
The intent is the BSP includes the processor/core tune file that its based upon. The BSP should usually know which one is present.
Quoted 6 lines
That tune then presents the possible options which the BSPs selects from but the end user or distro can override.
With this change, you get all tunes and you don't know which are compatible with a given core/processor. I'm not sure that is an improvement.
Before you were selecting an A75 inc file (for example), now you are specifying an ARMv8.2 inc file and saying it is an A75 from the available list. Either way, you needed to know which core you were running.
In addition to this, you have the ability to run the generic default for the family, which will work for all of them. And for ARMv8, even the more generic armv8 will work for all of them. So, there isn't an incompatibility problem.
Quoted 8 lines
I appreciate not wanting "lots of files" but we need to consider the usability too.
As others have mentioned, the errors are fairly obvious with diff between the files (or a GUI like meld).
What are the advantages this brings other than fewer files? Am I missing something?
This is the main benefit of it, TBH. If I add 25 more arm tunes, it's going to get ugly. And if Arm keeps adding cores at this same pace every year, it's going to get even uglier. However, I'm fine to send out those patches if it only bothers me to have so many files.
Thanks,
Jon
Quoted 7 lines
Cheers,
Richard
Quoted 63 lines · Jon Mason
Quoted 49 lines · Richard Purdie
Quoted 33 lines · Jon Mason
There is a large number of Arm Tune files located in meta/conf/machine/include/, and to support the current and upcoming Arm
cores, more are needed. Adding more files is simply going to make it
harder to find the relevant ones for an OE/YP developer/user. Also,
there are problems with stale and erroneous configs (see my previous
series), which will only be exacerbated by having more files.I am proposing a reorganization of the existing tune files by including
them in the generic family include file. For example, the tune-cortexa55.inc would be moved into arch-armv8-2a.inc, and tune-cortexa57.inc would be moved into arch-armv8a.inc. This reduces
the number of files from 12 to 2 for ARMv8a, and that is excluding the
13 I am adding in this series that would otherwise be unique files.To use, simply add
...
DEFAULTTUNE ?= "neoversen1"
require conf/machine/include/arm/arch-armv8-2a.inc
...Which is arguably what should be done anyway (instead of taking the
default of the tune include file).
See the qemuarm64 patch in the series for a working example.Quoted 8 lines
Of course, by removing the existing tune files, current users are going
to break. A simple script can be written to use sed (or similar) to
replace the relevant parts for those users that would be affected (at
least for those that are in the layer index and update regularly).I've just looked at this in a bit more details and I'm worried.
The intent is the BSP includes the processor/core tune file that its
based upon. The BSP should usually know which one is present.Quoted 6 lines
That tune then presents the possible options which the BSPs selects from but the end user or distro can override.
With this change, you get all tunes and you don't know which are compatible with a given core/processor. I'm not sure that is an improvement.
Before you were selecting an A75 inc file (for example), now you are specifying an ARMv8.2 inc file and saying it is an A75 from the available list. Either way, you needed to know which core you were running.
Quoted 4 lines
In addition to this, you have the ability to run the generic default for the family, which will work for all of them. And for ARMv8, even the more generic armv8 will work for all of them. So, there isn't an incompatibility problem.
You should have the ability to access the generic tune with either approach?
Quoted 7 lines
Quoted 7 lines
I appreciate not wanting "lots of files" but we need to consider the
usability too.As others have mentioned, the errors are fairly obvious with diff between the files (or a GUI like meld).
Quoted 7 lines
Quoted 2 lines
What are the advantages this brings other than fewer files? Am I missing something?
This is the main benefit of it, TBH. If I add 25 more arm tunes, it's going to get ugly. And if Arm keeps adding cores at this same pace every year, it's going to get even uglier. However, I'm fine to send out those patches if it only bothers me to have so many files.
My worry is that the current system:
a) tells BSP developers to select their processor, not an arch
b) only shows them tunes that work on that processor
So we're changing the model, but only for armv8XXX. This is going to be confusing. I do like only showing people things they can use too.
How about we split the difference and add the new tune files into subdirs by architecture?
Cheers,
Richard
Quoted 90 lines · Richard Purdie
Quoted 63 lines · Jon Mason
Quoted 49 lines · Richard Purdie
Quoted 33 lines · Jon Mason
There is a large number of Arm Tune files located in meta/conf/machine/include/, and to support the current and upcoming Arm
cores, more are needed. Adding more files is simply going to make it
harder to find the relevant ones for an OE/YP developer/user. Also,
there are problems with stale and erroneous configs (see my previous
series), which will only be exacerbated by having more files.I am proposing a reorganization of the existing tune files by including
them in the generic family include file. For example, the tune-cortexa55.inc would be moved into arch-armv8-2a.inc, and tune-cortexa57.inc would be moved into arch-armv8a.inc. This reduces
the number of files from 12 to 2 for ARMv8a, and that is excluding the
13 I am adding in this series that would otherwise be unique files.To use, simply add
...
DEFAULTTUNE ?= "neoversen1"
require conf/machine/include/arm/arch-armv8-2a.inc
...Which is arguably what should be done anyway (instead of taking the
default of the tune include file).
See the qemuarm64 patch in the series for a working example.Quoted 8 lines
Of course, by removing the existing tune files, current users are going
to break. A simple script can be written to use sed (or similar) to
replace the relevant parts for those users that would be affected (at
least for those that are in the layer index and update regularly).I've just looked at this in a bit more details and I'm worried.
The intent is the BSP includes the processor/core tune file that its
based upon. The BSP should usually know which one is present.Quoted 6 lines
That tune then presents the possible options which the BSPs selects from but the end user or distro can override.
With this change, you get all tunes and you don't know which are compatible with a given core/processor. I'm not sure that is an improvement.
Before you were selecting an A75 inc file (for example), now you are specifying an ARMv8.2 inc file and saying it is an A75 from the available list. Either way, you needed to know which core you were running.
Quoted 4 lines
In addition to this, you have the ability to run the generic default for the family, which will work for all of them. And for ARMv8, even the more generic armv8 will work for all of them. So, there isn't an incompatibility problem.
You should have the ability to access the generic tune with either approach?
Quoted 7 lines
Quoted 7 lines
I appreciate not wanting "lots of files" but we need to consider the
usability too.As others have mentioned, the errors are fairly obvious with diff between the files (or a GUI like meld).
Quoted 7 lines
Quoted 2 lines
What are the advantages this brings other than fewer files? Am I missing something?
This is the main benefit of it, TBH. If I add 25 more arm tunes, it's going to get ugly. And if Arm keeps adding cores at this same pace every year, it's going to get even uglier. However, I'm fine to send out those patches if it only bothers me to have so many files.
My worry is that the current system:
Quoted 8 lines
a) tells BSP developers to select their processor, not an arch
b) only shows them tunes that work on that processorSo we're changing the model, but only for armv8XXX. This is going to be confusing. I do like only showing people things they can use too.
How about we split the difference and add the new tune files into subdirs by architecture?
So it would look something like
meta/conf/machine/arm/armv8.0/tune-cortexa57.inc
meta/conf/machine/arm/armv8.2/tune-cortexa75.inc
And inside of those, it would reference meta/conf/machine/include/arm/arch-armv8.inc or meta/conf/machine/include/arm/arch-armv8-2a.inc (just as it does now). Correct?
Assuming so, does it make sense to try and match this with all the other CPU architectures (e.g., x86, ppc, mips, armv7, etc)?
Thanks,
Jon
Quoted 4 lines
Cheers,
Richard
Quoted 98 lines · Jon Mason
Quoted 96 lines · Richard Purdie
Quoted 63 lines · Jon Mason
Quoted 49 lines · Richard Purdie
Quoted 33 lines · Jon Mason
There is a large number of Arm Tune files located in meta/conf/machine/include/, and to support the current and upcoming Arm
cores, more are needed. Adding more files is simply going to make it
harder to find the relevant ones for an OE/YP developer/user. Also,
there are problems with stale and erroneous configs (see my previous
series), which will only be exacerbated by having more files.I am proposing a reorganization of the existing tune files by including
them in the generic family include file. For example, the tune-cortexa55.inc would be moved into arch-armv8-2a.inc, and tune-cortexa57.inc would be moved into arch-armv8a.inc. This reduces
the number of files from 12 to 2 for ARMv8a, and that is excluding the
13 I am adding in this series that would otherwise be unique files.To use, simply add
...
DEFAULTTUNE ?= "neoversen1"
require conf/machine/include/arm/arch-armv8-2a.inc
...Which is arguably what should be done anyway (instead of taking the
default of the tune include file).
See the qemuarm64 patch in the series for a working example.Quoted 8 lines
Of course, by removing the existing tune files, current users are going
to break. A simple script can be written to use sed (or similar) to
replace the relevant parts for those users that would be affected (at
least for those that are in the layer index and update regularly).I've just looked at this in a bit more details and I'm worried.
The intent is the BSP includes the processor/core tune file that its
based upon. The BSP should usually know which one is present.Quoted 6 lines
That tune then presents the possible options which the BSPs selects from but the end user or distro can override.
With this change, you get all tunes and you don't know which are compatible with a given core/processor. I'm not sure that is an improvement.
Before you were selecting an A75 inc file (for example), now you are specifying an ARMv8.2 inc file and saying it is an A75 from the available list. Either way, you needed to know which core you were running.
Quoted 4 lines
In addition to this, you have the ability to run the generic default for the family, which will work for all of them. And for ARMv8, even the more generic armv8 will work for all of them. So, there isn't an incompatibility problem.
You should have the ability to access the generic tune with either approach?
Quoted 7 lines
Quoted 7 lines
I appreciate not wanting "lots of files" but we need to consider the
usability too.As others have mentioned, the errors are fairly obvious with diff between the files (or a GUI like meld).
Quoted 7 lines
Quoted 2 lines
What are the advantages this brings other than fewer files? Am I missing something?
This is the main benefit of it, TBH. If I add 25 more arm tunes, it's going to get ugly. And if Arm keeps adding cores at this same pace every year, it's going to get even uglier. However, I'm fine to send out those patches if it only bothers me to have so many files.
My worry is that the current system:
a) tells BSP developers to select their processor, not an arch
b) only shows them tunes that work on that processorSo we're changing the model, but only for armv8XXX. This is going to be confusing. I do like only showing people things they can use too.
Quoted 6 lines
Quoted 2 lines
How about we split the difference and add the new tune files into subdirs by architecture?
So it would look something like
meta/conf/machine/arm/armv8.0/tune-cortexa57.inc
meta/conf/machine/arm/armv8.2/tune-cortexa75.inc
This additional level will still force all BSPs to update the include/require line in MACHINE configs.
Quoted 4 lines
And inside of those, it would reference meta/conf/machine/include/arm/arch-armv8.inc or meta/conf/machine/include/arm/arch-armv8-2a.inc (just as it does now). Correct?
Isn't this enough to see which tune-coretex* file belongs to which family? (e.g. git grep arch-armv8-2a.inc to see all available tune files from armv8.2a family instead of armv8.2 subdirectory)
Quoted 13 lines
Assuming so, does it make sense to try and match this with all the other CPU architectures (e.g., x86, ppc, mips, armv7, etc)?
Thanks,
JonQuoted 4 lines
Cheers,
Richard
Quoted 116 lines · Jon Mason
Quoted 114 lines · Richard Purdie
Quoted 72 lines · Jon Mason
Quoted 54 lines · Richard Purdie
Quoted 34 lines · Jon Mason
There is a large number of Arm Tune files located in meta/conf/machine/include/, and to support the current and upcoming Arm
cores, more are needed. Adding more files is simply going to make it
harder to find the relevant ones for an OE/YP developer/user. Also,
there are problems with stale and erroneous configs (see my previous
series), which will only be exacerbated by having more files.I am proposing a reorganization of the existing tune files by including
them in the generic family include file. For example, the tune-cortexa55.inc would be moved into arch-armv8-2a.inc, and tune-cortexa57.inc would be moved into arch-armv8a.inc. This reduces
the number of files from 12 to 2 for ARMv8a, and that is excluding the
13 I am adding in this series that would otherwise be unique files.To use, simply add
...
DEFAULTTUNE ?= "neoversen1"
require conf/machine/include/arm/arch-armv8-2a.inc
...Which is arguably what should be done anyway (instead of taking
the
default of the tune include file).
See the qemuarm64 patch in the series for a working example.Quoted 11 lines
Of course, by removing the existing tune files, current users are
going
to break. A simple script can be written to use sed (or similar)
to
replace the relevant parts for those users that would be affected
(at
least for those that are in the layer index and update regularly).I've just looked at this in a bit more details and I'm worried.
The intent is the BSP includes the processor/core tune file that
its
based upon. The BSP should usually know which one is present.Quoted 8 lines
That tune then presents the possible options which the BSPs selects
from but the end user or distro can override.With this change, you get all tunes and you don't know which are
compatible with a given core/processor. I'm not sure that is an improvement.Before you were selecting an A75 inc file (for example), now you are
specifying an ARMv8.2 inc file and saying it is an A75 from the available list. Either way, you needed to know which core you were
running.Quoted 7 lines
In addition to this, you have the ability to run the generic default
for the family, which will work for all of them. And for ARMv8, even
the more generic armv8 will work for all of them. So, there isn't an
incompatibility problem.You should have the ability to access the generic tune with either approach?
Quoted 13 lines
Quoted 5 lines
I appreciate not wanting "lots of files" but we need to consider
the
usability too.Quoted 7 lines
As others have mentioned, the errors are fairly obvious with diff
between the files (or a GUI like meld).What are the advantages this brings other than fewer files? Am I
missing something?Quoted 7 lines
This is the main benefit of it, TBH. If I add 25 more arm tunes, it's going to get ugly. And if Arm keeps adding cores at this same
pace every year, it's going to get even uglier. However, I'm fine to
send out those patches if it only bothers me to have so many files.My worry is that the current system:
a) tells BSP developers to select their processor, not an arch
b) only shows them tunes that work on that processorSo we're changing the model, but only for armv8XXX. This is going to be
confusing. I do like only showing people things they can use too.
Quoted 6 lines
Quoted 2 lines
How about we split the difference and add the new tune files into subdirs by architecture?
So it would look something like
meta/conf/machine/arm/armv8.0/tune-cortexa57.inc
meta/conf/machine/arm/armv8.2/tune-cortexa75.inc
Yes. With an open question on whether we move any existing files. Maybe just the armv8.0+ ones?
Quoted 4 lines
And inside of those, it would reference meta/conf/machine/include/arm/arch-armv8.inc or meta/conf/machine/include/arm/arch-armv8-2a.inc (just as it does now). Correct?
Yes.
Quoted 2 lines
Assuming so, does it make sense to try and match this with all the other CPU architectures (e.g., x86, ppc, mips, armv7, etc)?
I'd say not worth it, just do this either for new ones or the armv8* onwards ones.
Cheers,
Richard