Skip to content

Commit 1574bea

Browse files
zingozingo
authored
Merge branch 'main' into marlin-remove-unused-param
2 parents c22d879 + 9aedbeb commit 1574bea

File tree

106 files changed

+3987
-940
lines changed

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

106 files changed

+3987
-940
lines changed

‎.ci/scripts/build-qnn-sdk.sh

100644100755
Lines changed: 4 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -11,17 +11,12 @@ set -o xtrace
1111

1212
build_qnn_backend() {
1313
echo "Start building qnn backend."
14-
export ANDROID_NDK_ROOT=/opt/ndk
15-
export QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029
14+
export ANDROID_NDK_ROOT=${ANDROID_NDK_ROOT:-/opt/ndk}
15+
export QNN_SDK_ROOT=${QNN_SDK_ROOT:-/tmp/qnn/2.28.0.241029}
1616
export EXECUTORCH_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../.." && pwd)"
1717

18-
# Workaround to avoid issues around missing flatccrt library (depending on the
19-
# number of jobs used), see issue #7300:
20-
# Build twice (second time with `--no_clean`) to make sure libflatccrt.a is
21-
# available.
22-
# TODO: Remove this workaround once the underlying issue is fixed.
23-
bash backends/qualcomm/scripts/build.sh --skip_aarch64 --job_number 2 --release || \
24-
bash backends/qualcomm/scripts/build.sh --skip_aarch64 --job_number 2 --release --no_clean
18+
parallelism=$(( $(nproc) - 1 ))
19+
bash backends/qualcomm/scripts/build.sh --skip_aarch64 --job_number ${parallelism} --release
2520
}
2621

2722
set_up_aot() {

‎.github/workflows/build-presets.yml

Lines changed: 39 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -6,6 +6,8 @@ on:
66
branches:
77
- main
88
- release/*
9+
paths:
10+
- .github/workflows/build-presets.yml
911
workflow_dispatch:
1012

1113
concurrency:
@@ -16,15 +18,51 @@ jobs:
1618
apple:
1719
uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
1820
strategy:
21+
fail-fast: false
1922
matrix:
20-
preset: [macos-arm64]
23+
preset: [macos-arm64, pybind]
2124
with:
2225
job-name: build
26+
ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
2327
runner: macos-latest-xlarge
2428
python-version: 3.12
2529
submodules: recursive
30+
timeout: 90
2631
script: |
2732
set -eux
2833
${CONDA_RUN} ./install_requirements.sh > /dev/null
2934
${CONDA_RUN} cmake --preset ${{ matrix.preset }}
3035
${CONDA_RUN} cmake --build cmake-out --parallel
36+
37+
linux:
38+
uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
39+
strategy:
40+
fail-fast: false
41+
matrix:
42+
preset: [pybind]
43+
runner: [linux.2xlarge, linux.arm64.2xlarge]
44+
docker-image: [executorch-ubuntu-22.04-clang12, executorch-ubuntu-22.04-gcc11-aarch64]
45+
# Excluding specific runner + docker image combinations that don't make sense:
46+
# - Excluding the ARM64 gcc image on the x86 runner (linux.2xlarge)
47+
# - Excluding the x86 clang image on the ARM64 runner (linux.arm64.2xlarge)
48+
exclude:
49+
- runner: linux.2xlarge
50+
docker-image: executorch-ubuntu-22.04-gcc11-aarch64
51+
- runner: linux.arm64.2xlarge
52+
docker-image: executorch-ubuntu-22.04-clang12
53+
with:
54+
job-name: build
55+
ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
56+
runner: ${{ matrix.runner }}
57+
docker-image: ${{ matrix.docker-image }}
58+
submodules: recursive
59+
timeout: 90
60+
script: |
61+
set -eux
62+
# The generic Linux job chooses to use base env, not the one setup by the image
63+
CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
64+
conda activate "${CONDA_ENV}"
65+
66+
./install_requirements.sh > /dev/null
67+
cmake --preset ${{ matrix.preset }}
68+
cmake --build cmake-out --parallel

‎.github/workflows/pull.yml

Lines changed: 0 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -343,34 +343,6 @@ jobs:
343343
PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool "${BUILD_TOOL}"
344344
PYTHON_EXECUTABLE=python bash examples/xnnpack/quantization/test_quantize.sh "${BUILD_TOOL}" mv2
345345
346-
test-pybind-build-linux:
347-
name: test-pybind-build-linux
348-
uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
349-
permissions:
350-
id-token: write
351-
contents: read
352-
strategy:
353-
fail-fast: false
354-
with:
355-
runner: linux.2xlarge
356-
docker-image: executorch-ubuntu-22.04-clang12
357-
submodules: 'recursive'
358-
ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
359-
timeout: 90
360-
script: |
361-
# The generic Linux job chooses to use base env, not the one setup by the image
362-
CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
363-
conda activate "${CONDA_ENV}"
364-
365-
# build module for executorch.extension.pybindings.portable_lib
366-
BUILD_TOOL="cmake"
367-
PYTHON_EXECUTABLE=python \
368-
CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON" \
369-
bash .ci/scripts/setup-linux.sh --build-tool "${BUILD_TOOL}"
370-
371-
# see if we can import the module successfully
372-
python -c "from executorch.extension.pybindings import portable_lib; print('success!')"
373-
374346
test-binary-size-linux-gcc:
375347
name: test-binary-size-linux-gcc
376348
uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main

‎.github/workflows/trunk.yml

Lines changed: 0 additions & 24 deletions
Original file line numberDiff line numberDiff line change
@@ -289,30 +289,6 @@ jobs:
289289
# Build and test coreml delegate
290290
PYTHON_EXECUTABLE=python ${CONDA_RUN} bash backends/apple/coreml/scripts/build_all.sh
291291
292-
test-pybind-build-macos:
293-
name: test-pybind-build-macos
294-
uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
295-
strategy:
296-
matrix:
297-
include:
298-
- build-tool: cmake
299-
fail-fast: false
300-
with:
301-
runner: macos-m1-stable
302-
python-version: '3.11'
303-
submodules: 'recursive'
304-
ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
305-
timeout: 180
306-
script: |
307-
bash .ci/scripts/setup-conda.sh
308-
309-
# build module for executorch.extension.pybindings.portable_lib
310-
BUILD_TOOL=${{ matrix.build-tool }}
311-
CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON" PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/setup-macos.sh --build-tool "${BUILD_TOOL}"
312-
313-
# see if we can import the module successfully
314-
${CONDA_RUN} python -c "from executorch.extension.pybindings import portable_lib; print('success!')"
315-
316292
test-static-llama-ane:
317293
name: test-static-llama-ane
318294
uses: pytorch/test-infra/.github/workflows/macos_job.yml@main

‎CMakeLists.txt

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -582,6 +582,7 @@ if(EXECUTORCH_BUILD_PYBIND)
582582
${TORCH_PYTHON_LIBRARY}
583583
bundled_program
584584
etdump
585+
flatccrt
585586
executorch
586587
extension_data_loader
587588
util

‎CMakePresets.json

Lines changed: 15 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -15,6 +15,7 @@
1515
},
1616
{
1717
"name": "macos-arm64",
18+
"displayName": "Build everything buildable on macOS arm64",
1819
"inherits": ["common"],
1920
"generator": "Xcode",
2021
"cacheVariables": {
@@ -28,6 +29,20 @@
2829
"type": "equals",
2930
"rhs": "Darwin"
3031
}
32+
},
33+
{
34+
"name": "pybind",
35+
"displayName": "Build pybindings exported in the wheel",
36+
"inherits": ["common"],
37+
"cacheVariables": {
38+
"EXECUTORCH_BUILD_PRESET_FILE": "${sourceDir}/tools/cmake/preset/pybind.cmake",
39+
"CMAKE_OSX_DEPLOYMENT_TARGET": "10.15"
40+
},
41+
"condition": {
42+
"type": "inList",
43+
"string": "${hostSystemName}",
44+
"list": ["Darwin", "Linux", "Windows"]
45+
}
3146
}
3247
]
3348
}

‎backends/arm/_passes/__init__.py

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -8,6 +8,7 @@
88
from .annotate_channels_last_dim_order_pass import AnnotateChannelsLastDimOrder # noqa
99
from .annotate_decomposed_matmul import AnnotateDecomposedMatmulPass # noqa
1010
from .arm_pass import ArmPass # noqa
11+
from .broadcast_args_pass import BroadcastArgsPass # noqa
1112
from .cast_int64_pass import CastInt64BuffersToInt32Pass # noqa
1213
from .cast_to_int32_pass import CastToInt32Pass # noqa
1314
from .conv1d_unsqueeze_pass import Conv1dUnsqueezePass # noqa

‎backends/arm/_passes/arm_pass_manager.py

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -10,6 +10,7 @@
1010
from executorch.backends.arm._passes import (
1111
AnnotateChannelsLastDimOrder,
1212
AnnotateDecomposedMatmulPass,
13+
BroadcastArgsPass,
1314
CastInt64BuffersToInt32Pass,
1415
CastToInt32Pass,
1516
ComputeConstantOpsAOT,
@@ -104,6 +105,8 @@ def _tosa_080_BI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
104105
self.add_pass(RetraceFoldedDtypesPass())
105106
self.add_pass(UnsqueezeScalarPlaceholdersPass(exported_program))
106107
self.add_pass(MatchArgRanksPass(exported_program))
108+
if self.tosa_spec.is_U55_subset:
109+
self.add_pass(BroadcastArgsPass())
107110
self.add_pass(ComputeConstantOpsAOT(exported_program))
108111

109112
self.add_pass(RemoveClonePass())

‎backends/arm/_passes/broadcast_args_pass.py

Lines changed: 63 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,63 @@
1+
# Copyright 2025 Arm Limited and/or its affiliates.
2+
#
3+
# This source code is licensed under the BSD-style license found in the
4+
# LICENSE file in the root directory of this source tree.
5+
6+
from executorch.backends.arm._passes import ArmPass
7+
8+
from executorch.backends.arm._passes.arm_pass_utils import (
9+
create_node,
10+
get_first_fake_tensor,
11+
)
12+
13+
from executorch.exir.dialects._ops import ops as exir_ops
14+
15+
from executorch.exir.pass_base import PassResult
16+
from torch.fx import GraphModule, Node
17+
18+
19+
class BroadcastArgsPass(ArmPass):
20+
"""
21+
Pass to manually broadcast arguments by inserting repeats.
22+
This is done when more than one arg needs broadcasting.
23+
"""
24+
25+
targeted_ops = {
26+
exir_ops.edge.aten.add.Tensor,
27+
exir_ops.edge.aten.sub.Tensor,
28+
# mul is indirectly targeting div as div is decompsed to reciprocal + mul
29+
exir_ops.edge.aten.mul.Tensor,
30+
}
31+
32+
def call(self, graph_module: GraphModule) -> PassResult:
33+
for node in graph_module.graph.nodes:
34+
if node.op != "call_function" or node.target not in self.targeted_ops:
35+
continue
36+
37+
output_shape = get_first_fake_tensor(node).shape
38+
nbr_of_broacasts = 0
39+
for arg in node.args:
40+
if not isinstance(arg, Node):
41+
continue
42+
43+
shape = get_first_fake_tensor(arg).shape
44+
if shape != output_shape:
45+
nbr_of_broacasts += 1
46+
if nbr_of_broacasts > 1:
47+
multiples = [
48+
int(output_shape[d] / shape[d])
49+
for d in range(len(output_shape))
50+
]
51+
with graph_module.graph.inserting_before(node):
52+
repeat = create_node(
53+
graph_module.graph,
54+
exir_ops.edge.aten.repeat.default,
55+
args=(arg, multiples),
56+
kwargs={},
57+
from_node=node,
58+
)
59+
node.replace_input_with(arg, repeat)
60+
61+
graph_module.recompile()
62+
graph_module = super().call(graph_module).graph_module
63+
return PassResult(graph_module, True)

‎backends/arm/arm_vela.py

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -73,8 +73,8 @@ def vela_compile(tosa_flatbuffer: bytes, args: List[str], verbose: bool = False)
7373
np_path = os.path.join(tmpdir, "output", "out_vela.npz")
7474
else:
7575
np_path = os.path.join(tmpdir, "output", "out_sg0_vela.npz")
76-
blocks = b""
7776

77+
blocks = b""
7878
with np.load(np_path, allow_pickle=False) as data:
7979
# Construct our modified output_blocks with data in a form easily
8080
# digested on the device side
@@ -92,7 +92,7 @@ def vela_compile(tosa_flatbuffer: bytes, args: List[str], verbose: bool = False)
9292
if not isinstance(data["scratch_shape"][0], np.int64):
9393
raise RuntimeError("Expected scratch to be int64")
9494
block_length = int(data["scratch_shape"][0])
95-
bin_blocks["scratch_data"] = b"\x00" * block_length
95+
bin_blocks["scratch_size"] = struct.pack("<I", block_length)
9696

9797
# Capture inputs and outputs
9898
bin_blocks["inputs"] = vela_bin_pack_io("input", data)

‎backends/arm/operator_support/to_copy_support.py

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4,6 +4,7 @@
44
# LICENSE file in the root directory of this source tree.
55

66
# pyre-unsafe
7+
import copy
78
import logging
89

910
import torch
@@ -42,7 +43,9 @@ def _merge_supported_types(
4243
dtypes1: SupportedTypeDict,
4344
dtypes2: SupportedTypeDict,
4445
) -> SupportedTypeDict:
45-
merged_dtypes = dtypes1
46+
merged_dtypes = copy.deepcopy(
47+
dtypes1
48+
) # Use deepcopy to avoid unintentionally modifying SUPPORTED_INT_TYPES
4649
for k, v in dtypes2.items():
4750
merged_dtypes[k] = merged_dtypes.get(k, []) + v
4851
return merged_dtypes

‎backends/arm/operator_support/tosa_supported_operators.py

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -306,6 +306,7 @@ class CheckProperQuantization(OperatorSupportBase):
306306
exir_ops.edge.aten.sub.Tensor,
307307
exir_ops.edge.aten.upsample_bilinear2d.vec,
308308
exir_ops.edge.aten.upsample_nearest2d.vec,
309+
torch.ops.aten.scalar_tensor.default,
309310
*TableOps.included_ops(),
310311
)
311312

‎backends/arm/operators/op_abs.py

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -164,7 +164,7 @@ def define_node(
164164
scale_back = 1.0
165165
if inputs[0].dtype == ts.DType.INT8:
166166
rescaled_inputs, scale_back = tqutils.insert_rescale_ops_to_int32(
167-
tosa_graph, inputs, node, self.tosa_specs
167+
tosa_graph, inputs, node, self.tosa_spec
168168
) # type: ignore[possibly-undefined]
169169
else:
170170
# input[0].dtype == ts.DType.INT32
@@ -192,7 +192,7 @@ def define_node(
192192
# Scale output back to 8 bit
193193
# pyre-ignore
194194
tqutils.insert_rescale_op_to_int8(
195-
tosa_graph, abs_output, scale_back, node, self.tosa_specs
195+
tosa_graph, abs_output, scale_back, node, self.tosa_spec
196196
) # type: ignore[possibly-undefined]
197197

198198

‎backends/arm/operators/op_add.py

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -174,7 +174,7 @@ def define_node(
174174
scale_back = 1.0
175175
if inputs[0].dtype == ts.DType.INT8:
176176
rescaled_inputs, scale_back = tqutils.insert_rescale_ops_to_int32(
177-
tosa_graph, inputs, node, self.tosa_specs
177+
tosa_graph, inputs, node, self.tosa_spec
178178
)
179179
else:
180180
# input[0].dtype == ts.DType.INT32
@@ -202,7 +202,7 @@ def define_node(
202202
# Scale output back to 8 bit
203203
# pyre-ignore
204204
tqutils.insert_rescale_op_to_int8(
205-
tosa_graph, add_output, scale_back, node, self.tosa_specs
205+
tosa_graph, add_output, scale_back, node, self.tosa_spec
206206
) # type: ignore[possibly-undefined]
207207

208208

‎backends/arm/operators/op_eq.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -98,7 +98,7 @@ def define_node(
9898
if inputs[0].dtype == ts.DType.INT8:
9999
# Rescale inputs to 32 bit
100100
rescaled_inputs, _ = tqutils.insert_rescale_ops_to_int32(
101-
tosa_graph, inputs, node, self.tosa_specs
101+
tosa_graph, inputs, node, self.tosa_spec
102102
)
103103

104104
# Update IO

‎backends/arm/operators/op_ge.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -97,7 +97,7 @@ def define_node(
9797
if inputs[0].dtype == ts.DType.INT8:
9898
# Rescale inputs to 32 bit
9999
rescaled_inputs, _ = tqutils.insert_rescale_ops_to_int32(
100-
tosa_graph, inputs, node, self.tosa_specs
100+
tosa_graph, inputs, node, self.tosa_spec
101101
)
102102

103103
# Update IO

‎backends/arm/operators/op_gt.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -97,7 +97,7 @@ def define_node(
9797
if inputs[0].dtype == ts.DType.INT8:
9898
# Rescale inputs to 32 bit
9999
rescaled_inputs, _ = tqutils.insert_rescale_ops_to_int32(
100-
tosa_graph, inputs, node, self.tosa_specs
100+
tosa_graph, inputs, node, self.tosa_spec
101101
)
102102

103103
# Update IO

‎backends/arm/operators/op_le.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -97,7 +97,7 @@ def define_node(
9797
if inputs[0].dtype == ts.DType.INT8:
9898
# Rescale inputs to 32 bit
9999
rescaled_inputs, _ = tqutils.insert_rescale_ops_to_int32(
100-
tosa_graph, inputs, node, self.tosa_specs
100+
tosa_graph, inputs, node, self.tosa_spec
101101
)
102102

103103
# Update IO

0 commit comments

Comments
 (0)