Skip to content

Commit 1052f90

Browse files
Sentimentronxnnpack-bot
Sentimentron
authored and
xnnpack-bot
committed
[gn] Add support for building/testing AArch32
Although AArch32 is not really the current focus for Chrome performance, as we've got the kernels around, there's no harm in having it. Also modifies the GN test script to support testing on x64 hosts. PiperOrigin-RevId: 916281274
1 parent 8da42ae commit 1052f90

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")
@@ -117,7 +120,8 @@ declare_args() {
117120
xnnpack_enable_arm_dotprod = false
118121
}
119122

120-
assert(target_cpu == "arm64" || target_cpu == "x64" || target_cpu == "x86",
123+
assert(target_cpu == "arm64" || target_cpu == "x64" || target_cpu == "x86" ||
124+
target_cpu == "arm",
121125
"Unsupported target_cpu '${target_cpu}'.")
122126

123127
assert(!xnnpack_enable_arm_kleidiai || target_cpu == "arm64",
@@ -194,6 +198,14 @@ config("xnnpack_private_config") {
194198
} else {
195199
defines += [ "XNN_ENABLE_ARM_SME2=0" ]
196200
}
201+
} else if (current_cpu == "arm") {
202+
defines += [
203+
"XNN_ENABLE_ARM_BF16=1",
204+
"XNN_ENABLE_ARM_DOTPROD=1",
205+
"XNN_ENABLE_ARM_FP16_SCALAR=1",
206+
"XNN_ENABLE_ARM_FP16_VECTOR=1",
207+
"XNN_ENABLE_ASSEMBLY=1",
208+
]
197209
} else if (current_cpu == "x64" || current_cpu == "x86") {
198210
# x64 support baseline, which is SSSE3 and below to match Chromium.
199211
defines += [
@@ -361,20 +373,6 @@ xnnpack_source_set("operators") {
361373
}
362374
}
363375

364-
xnnpack_source_set("subgraph_matchers") {
365-
sources = [
366-
"test/subgraph/rewrites/subgraph_matcher.cc",
367-
"test/subgraph/rewrites/subgraph_matcher.h",
368-
]
369-
370-
public_deps = [
371-
":xnnpack_headers",
372-
"//third_party/googletest:gtest",
373-
]
374-
375-
testonly = true
376-
}
377-
378376
xnnpack_source_set("subgraph") {
379377
deps = [
380378
":microkernel_headers",
@@ -573,6 +571,223 @@ if (current_cpu == "arm64") {
573571
}
574572
}
575573

574+
if (current_cpu == "arm") {
575+
config("armsimd32_config") {
576+
cflags = [
577+
"-marm",
578+
"-mfpu=vfp",
579+
"-munaligned-access",
580+
]
581+
}
582+
583+
config("aarch32_asm_config") {
584+
asmflags = [
585+
"-marm",
586+
"-march=armv8.2-a+dotprod+fp16",
587+
"-mfpu=neon-fp-armv8",
588+
]
589+
}
590+
591+
config("neon_config") {
592+
cflags = [
593+
"-marm",
594+
"-march=armv7-a",
595+
"-mfpu=neon",
596+
]
597+
}
598+
599+
config("neonfp16_config") {
600+
cflags = [
601+
"-marm",
602+
"-march=armv7-a",
603+
"-mfpu=neon-fp16",
604+
]
605+
}
606+
607+
config("neonfma_config") {
608+
cflags = [
609+
"-marm",
610+
"-march=armv7-a",
611+
"-mfpu=neon-vfpv4",
612+
]
613+
}
614+
615+
config("neonv8_config") {
616+
cflags = [
617+
"-marm",
618+
"-march=armv8-a",
619+
"-mfpu=neon-fp-armv8",
620+
]
621+
}
622+
623+
config("fp16arith_config") {
624+
cflags = [
625+
"-marm",
626+
"-march=armv8.2-a+fp16",
627+
"-mfpu=neon-fp-armv8",
628+
"-Xclang",
629+
"-target-feature",
630+
"-Xclang",
631+
"+fullfp16",
632+
]
633+
}
634+
635+
config("neonfp16arith_config") {
636+
cflags = [
637+
"-marm",
638+
"-march=armv8.2-a+fp16",
639+
"-mfpu=neon-fp-armv8",
640+
"-Xclang",
641+
"-target-feature",
642+
"-Xclang",
643+
"+fullfp16",
644+
]
645+
}
646+
647+
config("neon_dotprod_config") {
648+
cflags = [
649+
"-marm",
650+
"-march=armv8.2-a+dotprod",
651+
"-mfpu=neon-fp-armv8",
652+
"-Xclang",
653+
"-target-feature",
654+
"-Xclang",
655+
"+dotprod",
656+
]
657+
}
658+
659+
config("neon_dotprodfp16arith_config") {
660+
cflags = [
661+
"-marm",
662+
"-march=armv8.2-a+dotprod+fp16",
663+
"-mfpu=neon-fp-armv8",
664+
"-Xclang",
665+
"-target-feature",
666+
"-Xclang",
667+
"+dotprod",
668+
"-Xclang",
669+
"-target-feature",
670+
"-Xclang",
671+
"+fullfp16",
672+
]
673+
}
674+
675+
config("neon_bf16_config") {
676+
cflags = [
677+
"-marm",
678+
"-march=armv8.2-a+bf16",
679+
"-mfpu=neon-fp-armv8",
680+
"-Xclang",
681+
"-target-feature",
682+
"-Xclang",
683+
"+dotprod",
684+
"-Xclang",
685+
"-target-feature",
686+
"-Xclang",
687+
"+fullfp16",
688+
]
689+
}
690+
691+
xnnpack_source_set("armsimd32_microkernels") {
692+
deps = [
693+
":microkernel_defs",
694+
":microkernel_headers",
695+
]
696+
sources = ALL_ARMSIMD32_MICROKERNEL_SRCS
697+
configs = [ ":armsimd32_config" ]
698+
}
699+
700+
xnnpack_source_set("aarch32_asm_microkernels") {
701+
deps = [
702+
":microkernel_defs",
703+
":microkernel_headers",
704+
]
705+
sources = AARCH32_ASM_MICROKERNEL_SRCS
706+
configs = [ ":aarch32_asm_config" ]
707+
}
708+
709+
xnnpack_source_set("neon_microkernels") {
710+
deps = [
711+
":microkernel_defs",
712+
":microkernel_headers",
713+
]
714+
sources = ALL_NEON_MICROKERNEL_SRCS
715+
configs = [ ":neon_config" ]
716+
}
717+
718+
xnnpack_source_set("neonfp16_microkernels") {
719+
deps = [
720+
":microkernel_defs",
721+
":microkernel_headers",
722+
]
723+
sources = ALL_NEONFP16_MICROKERNEL_SRCS
724+
configs = [ ":neonfp16_config" ]
725+
}
726+
727+
xnnpack_source_set("neonfma_microkernels") {
728+
deps = [
729+
":microkernel_defs",
730+
":microkernel_headers",
731+
]
732+
sources = ALL_NEONFMA_MICROKERNEL_SRCS
733+
configs = [ ":neonfma_config" ]
734+
}
735+
736+
xnnpack_source_set("neonv8_microkernels") {
737+
deps = [
738+
":microkernel_defs",
739+
":microkernel_headers",
740+
]
741+
sources = ALL_NEONV8_MICROKERNEL_SRCS
742+
configs = [ ":neonv8_config" ]
743+
}
744+
745+
xnnpack_source_set("fp16arith_microkernels") {
746+
deps = [
747+
":microkernel_defs",
748+
":microkernel_headers",
749+
]
750+
sources = ALL_FP16ARITH_MICROKERNEL_SRCS
751+
configs = [ ":fp16arith_config" ]
752+
}
753+
754+
xnnpack_source_set("neonfp16arith_microkernels") {
755+
deps = [
756+
":microkernel_defs",
757+
":microkernel_headers",
758+
]
759+
sources = ALL_NEONFP16ARITH_MICROKERNEL_SRCS
760+
configs = [ ":neonfp16arith_config" ]
761+
}
762+
763+
xnnpack_source_set("neon_dotprod_microkernels") {
764+
deps = [
765+
":microkernel_defs",
766+
":microkernel_headers",
767+
]
768+
sources = ALL_NEONDOT_MICROKERNEL_SRCS
769+
configs = [ ":neon_dotprod_config" ]
770+
}
771+
772+
xnnpack_source_set("neon_dotprodfp16arith_microkernels") {
773+
deps = [
774+
":microkernel_defs",
775+
":microkernel_headers",
776+
]
777+
sources = ALL_NEONDOTFP16ARITH_MICROKERNEL_SRCS
778+
configs = [ ":neon_dotprodfp16arith_config" ]
779+
}
780+
781+
xnnpack_source_set("neon_bf16_microkernels") {
782+
deps = [
783+
":microkernel_defs",
784+
":microkernel_headers",
785+
]
786+
sources = ALL_NEONBF16_MICROKERNEL_SRCS
787+
configs = [ ":neon_bf16_config" ]
788+
}
789+
}
790+
576791
if (current_cpu == "x64" || current_cpu == "x86") {
577792
xnnpack_source_set("avxvnni_microkernels") {
578793
deps = [
@@ -766,6 +981,20 @@ xnnpack_source_set("xnnpack") {
766981
if (xnnpack_enable_arm_sme2) {
767982
deps += [ ":sme2_microkernels" ]
768983
}
984+
} else if (current_cpu == "arm") {
985+
deps += [
986+
":aarch32_asm_microkernels",
987+
":armsimd32_microkernels",
988+
":fp16arith_microkernels",
989+
":neon_bf16_microkernels",
990+
":neon_dotprod_microkernels",
991+
":neon_dotprodfp16arith_microkernels",
992+
":neon_microkernels",
993+
":neonfma_microkernels",
994+
":neonfp16_microkernels",
995+
":neonfp16arith_microkernels",
996+
":neonv8_microkernels",
997+
]
769998
} else if (current_cpu == "x64" || current_cpu == "x86") {
770999
if (xnnpack_enable_avx) {
7711000
deps += [ ":avx_microkernels" ]

0 commit comments

Comments
 (0)