Skip to content

Commit b1efb7d

Browse files
Sentimentronxnnpack-bot
Sentimentron
authored and
xnnpack-bot
committed
[gn] Add support for building/testing AArch32
Although AArch32 is legacy at this point and not really the focus for Chrome performance, as we've got the kernels around there's probably no harm in having it. PiperOrigin-RevId: 916180875
1 parent c3ac56a commit b1efb7d

3 files changed

Lines changed: 348 additions & 17 deletions

File tree

‎.github/workflows/build.yml‎

Lines changed: 63 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -335,6 +335,69 @@ jobs:
335335
run: cmake --build build/ios/arm64 --parallel $(sysctl -n hw.ncpu) -- -quiet
336336
working-directory: ${{ github.workspace }}
337337

338+
gn-linux-arm:
339+
runs-on: 'ubuntu-24.04-8core'
340+
timeout-minutes: 15
341+
env:
342+
CCACHE_SLOPPINESS: time_macros,modules
343+
CCACHE_DIR: /home/runner/.cache/ccache
344+
CCACHE_DEPEND: 1
345+
CCACHE_MAX_SIZE: 5G
346+
CCACHE_COMPILERCHECK: content
347+
steps:
348+
- name: Restore ccache
349+
id: cache-ccache-restore
350+
uses: actions/cache@v5
351+
with:
352+
path: |
353+
/home/runner/.cache/ccache
354+
key: ccache-linux-arm-${{ github.ref_name }}-${{ github.sha }}
355+
restore-keys: |
356+
ccache-linux-arm-${{ github.ref_name }}-
357+
ccache-linux-arm-master-
358+
- name: Install system dependencies
359+
run: |
360+
sudo apt-get install -y ccache qemu-user qemu-user-static g++-arm-linux-gnueabihf
361+
- name: Install Depot Tools
362+
uses: newkdev/setup-depot-tools@v1.0.1
363+
- name: Write .gclient
364+
run: |
365+
cat <<EOF > .gclient
366+
solutions = [
367+
{ "name" : 'xnnpack',
368+
"url" : 'https://github.com/google/XNNPACK',
369+
"deps_file" : 'DEPS',
370+
"managed" : False,
371+
"custom_deps" : {
372+
},
373+
"custom_vars": {},
374+
},
375+
]
376+
EOF
377+
- name: Sync to commit and run hooks
378+
run: gclient sync --reset --force --upstream --revision $GITHUB_SHA
379+
- name: Install sysroot
380+
run: ./build/linux/sysroot_scripts/install-sysroot.py --arch=armhf
381+
working-directory: ${{ github.workspace }}/xnnpack
382+
- name: Check GN formatting
383+
run: |
384+
git cl format --presubmit --no-clang-format --no-python --diff
385+
working-directory: ${{ github.workspace }}/xnnpack
386+
- name: Generate build files
387+
run: |
388+
gn gen --check --args="is_debug=false clang_use_chrome_plugins=false dcheck_always_on=true cc_wrapper=\"ccache\" symbol_level=0 target_cpu=\"arm\" use_siso=true" out/arm.dchecks
389+
working-directory: ${{ github.workspace }}/xnnpack
390+
- name: Build all targets (Release + debug checks)
391+
run: |
392+
autoninja -C out/arm.dchecks
393+
working-directory: ${{ github.workspace }}/xnnpack
394+
- name: Print ccache stats
395+
run: ccache -s
396+
- name: Run tests
397+
run: |
398+
python3 scripts/run-gn-tests.py out/arm.dchecks --sysroot=/usr/arm-linux-gnueabihf --architecture=arm --shards=32
399+
working-directory: ${{ github.workspace }}/xnnpack
400+
338401
gn-linux-x86_64:
339402
runs-on: 'ubuntu-24.04-8core'
340403
timeout-minutes: 15

‎BUILD.gn‎

Lines changed: 244 additions & 15 deletions
Original file line numberDiff line numberDiff line change
@@ -28,7 +28,10 @@ if (is_clang) {
2828
# can be read with GN too.
2929
import("gen/scalar_microkernels.bzl")
3030

31+
import("gen/aarch32_microkernels.bzl")
3132
import("gen/aarch64_microkernels.bzl")
33+
import("gen/armsimd32_microkernels.bzl")
34+
import("gen/fp16arith_microkernels.bzl")
3235
import("gen/neon_aarch64_microkernels.bzl")
3336
import("gen/neon_microkernels.bzl")
3437
import("gen/neonbf16_microkernels.bzl")
@@ -116,7 +119,8 @@ declare_args() {
116119
xnnpack_enable_arm_dotprod = false
117120
}
118121

119-
assert(target_cpu == "arm64" || target_cpu == "x64" || target_cpu == "x86",
122+
assert(target_cpu == "arm64" || target_cpu == "x64" || target_cpu == "x86" ||
123+
target_cpu == "arm",
120124
"Unsupported target_cpu '${target_cpu}'.")
121125

122126
assert(!xnnpack_enable_arm_kleidiai || target_cpu == "arm64",
@@ -193,6 +197,14 @@ config("xnnpack_private_config") {
193197
} else {
194198
defines += [ "XNN_ENABLE_ARM_SME2=0" ]
195199
}
200+
} else if (current_cpu == "arm") {
201+
defines += [
202+
"XNN_ENABLE_ARM_BF16=1",
203+
"XNN_ENABLE_ARM_DOTPROD=1",
204+
"XNN_ENABLE_ARM_FP16_SCALAR=1",
205+
"XNN_ENABLE_ARM_FP16_VECTOR=1",
206+
"XNN_ENABLE_ASSEMBLY=1",
207+
]
196208
} else if (current_cpu == "x64" || current_cpu == "x86") {
197209
# x64 support baseline, which is SSSE3 and below to match Chromium.
198210
defines += [
@@ -360,20 +372,6 @@ xnnpack_source_set("operators") {
360372
}
361373
}
362374

363-
xnnpack_source_set("subgraph_matchers") {
364-
sources = [
365-
"test/subgraph/rewrites/subgraph_matcher.cc",
366-
"test/subgraph/rewrites/subgraph_matcher.h",
367-
]
368-
369-
public_deps = [
370-
":xnnpack_headers",
371-
"//third_party/googletest:gtest",
372-
]
373-
374-
testonly = true
375-
}
376-
377375
xnnpack_source_set("subgraph") {
378376
deps = [
379377
":microkernel_headers",
@@ -572,6 +570,223 @@ if (current_cpu == "arm64") {
572570
}
573571
}
574572

573+
if (current_cpu == "arm") {
574+
config("armsimd32_config") {
575+
cflags = [
576+
"-marm",
577+
"-mfpu=vfp",
578+
"-munaligned-access",
579+
]
580+
}
581+
582+
config("aarch32_asm_config") {
583+
asmflags = [
584+
"-marm",
585+
"-march=armv8.2-a+dotprod+fp16",
586+
"-mfpu=neon-fp-armv8",
587+
]
588+
}
589+
590+
config("neon_config") {
591+
cflags = [
592+
"-marm",
593+
"-march=armv7-a",
594+
"-mfpu=neon",
595+
]
596+
}
597+
598+
config("neonfp16_config") {
599+
cflags = [
600+
"-marm",
601+
"-march=armv7-a",
602+
"-mfpu=neon-fp16",
603+
]
604+
}
605+
606+
config("neonfma_config") {
607+
cflags = [
608+
"-marm",
609+
"-march=armv7-a",
610+
"-mfpu=neon-vfpv4",
611+
]
612+
}
613+
614+
config("neonv8_config") {
615+
cflags = [
616+
"-marm",
617+
"-march=armv8-a",
618+
"-mfpu=neon-fp-armv8",
619+
]
620+
}
621+
622+
config("fp16arith_config") {
623+
cflags = [
624+
"-marm",
625+
"-march=armv8.2-a+fp16",
626+
"-mfpu=neon-fp-armv8",
627+
"-Xclang",
628+
"-target-feature",
629+
"-Xclang",
630+
"+fullfp16",
631+
]
632+
}
633+
634+
config("neonfp16arith_config") {
635+
cflags = [
636+
"-marm",
637+
"-march=armv8.2-a+fp16",
638+
"-mfpu=neon-fp-armv8",
639+
"-Xclang",
640+
"-target-feature",
641+
"-Xclang",
642+
"+fullfp16",
643+
]
644+
}
645+
646+
config("neon_dotprod_config") {
647+
cflags = [
648+
"-marm",
649+
"-march=armv8.2-a+dotprod",
650+
"-mfpu=neon-fp-armv8",
651+
"-Xclang",
652+
"-target-feature",
653+
"-Xclang",
654+
"+dotprod",
655+
]
656+
}
657+
658+
config("neon_dotprodfp16arith_config") {
659+
cflags = [
660+
"-marm",
661+
"-march=armv8.2-a+dotprod+fp16",
662+
"-mfpu=neon-fp-armv8",
663+
"-Xclang",
664+
"-target-feature",
665+
"-Xclang",
666+
"+dotprod",
667+
"-Xclang",
668+
"-target-feature",
669+
"-Xclang",
670+
"+fullfp16",
671+
]
672+
}
673+
674+
config("neon_bf16_config") {
675+
cflags = [
676+
"-marm",
677+
"-march=armv8.2-a+bf16",
678+
"-mfpu=neon-fp-armv8",
679+
"-Xclang",
680+
"-target-feature",
681+
"-Xclang",
682+
"+dotprod",
683+
"-Xclang",
684+
"-target-feature",
685+
"-Xclang",
686+
"+fullfp16",
687+
]
688+
}
689+
690+
xnnpack_source_set("armsimd32_microkernels") {
691+
deps = [
692+
":microkernel_defs",
693+
":microkernel_headers",
694+
]
695+
sources = ALL_ARMSIMD32_MICROKERNEL_SRCS
696+
configs = [ ":armsimd32_config" ]
697+
}
698+
699+
xnnpack_source_set("aarch32_asm_microkernels") {
700+
deps = [
701+
":microkernel_defs",
702+
":microkernel_headers",
703+
]
704+
sources = AARCH32_ASM_MICROKERNEL_SRCS
705+
configs = [ ":aarch32_asm_config" ]
706+
}
707+
708+
xnnpack_source_set("neon_microkernels") {
709+
deps = [
710+
":microkernel_defs",
711+
":microkernel_headers",
712+
]
713+
sources = ALL_NEON_MICROKERNEL_SRCS
714+
configs = [ ":neon_config" ]
715+
}
716+
717+
xnnpack_source_set("neonfp16_microkernels") {
718+
deps = [
719+
":microkernel_defs",
720+
":microkernel_headers",
721+
]
722+
sources = ALL_NEONFP16_MICROKERNEL_SRCS
723+
configs = [ ":neonfp16_config" ]
724+
}
725+
726+
xnnpack_source_set("neonfma_microkernels") {
727+
deps = [
728+
":microkernel_defs",
729+
":microkernel_headers",
730+
]
731+
sources = ALL_NEONFMA_MICROKERNEL_SRCS
732+
configs = [ ":neonfma_config" ]
733+
}
734+
735+
xnnpack_source_set("neonv8_microkernels") {
736+
deps = [
737+
":microkernel_defs",
738+
":microkernel_headers",
739+
]
740+
sources = ALL_NEONV8_MICROKERNEL_SRCS
741+
configs = [ ":neonv8_config" ]
742+
}
743+
744+
xnnpack_source_set("fp16arith_microkernels") {
745+
deps = [
746+
":microkernel_defs",
747+
":microkernel_headers",
748+
]
749+
sources = ALL_FP16ARITH_MICROKERNEL_SRCS
750+
configs = [ ":fp16arith_config" ]
751+
}
752+
753+
xnnpack_source_set("neonfp16arith_microkernels") {
754+
deps = [
755+
":microkernel_defs",
756+
":microkernel_headers",
757+
]
758+
sources = ALL_NEONFP16ARITH_MICROKERNEL_SRCS
759+
configs = [ ":neonfp16arith_config" ]
760+
}
761+
762+
xnnpack_source_set("neon_dotprod_microkernels") {
763+
deps = [
764+
":microkernel_defs",
765+
":microkernel_headers",
766+
]
767+
sources = ALL_NEONDOT_MICROKERNEL_SRCS
768+
configs = [ ":neon_dotprod_config" ]
769+
}
770+
771+
xnnpack_source_set("neon_dotprodfp16arith_microkernels") {
772+
deps = [
773+
":microkernel_defs",
774+
":microkernel_headers",
775+
]
776+
sources = ALL_NEONDOTFP16ARITH_MICROKERNEL_SRCS
777+
configs = [ ":neon_dotprodfp16arith_config" ]
778+
}
779+
780+
xnnpack_source_set("neon_bf16_microkernels") {
781+
deps = [
782+
":microkernel_defs",
783+
":microkernel_headers",
784+
]
785+
sources = ALL_NEONBF16_MICROKERNEL_SRCS
786+
configs = [ ":neon_bf16_config" ]
787+
}
788+
}
789+
575790
if (current_cpu == "x64" || current_cpu == "x86") {
576791
xnnpack_source_set("avxvnni_microkernels") {
577792
deps = [
@@ -771,6 +986,20 @@ xnnpack_source_set("xnnpack") {
771986
if (xnnpack_enable_arm_sme2) {
772987
deps += [ ":sme2_microkernels" ]
773988
}
989+
} else if (current_cpu == "arm") {
990+
deps += [
991+
":aarch32_asm_microkernels",
992+
":armsimd32_microkernels",
993+
":fp16arith_microkernels",
994+
":neon_bf16_microkernels",
995+
":neon_dotprod_microkernels",
996+
":neon_dotprodfp16arith_microkernels",
997+
":neon_microkernels",
998+
":neonfma_microkernels",
999+
":neonfp16_microkernels",
1000+
":neonfp16arith_microkernels",
1001+
":neonv8_microkernels",
1002+
]
7741003
} else if (current_cpu == "x64" || current_cpu == "x86") {
7751004
if (xnnpack_enable_avx) {
7761005
deps += [ ":avx_microkernels" ]

0 commit comments

Comments
 (0)