Skip to content

Commit 6f0487f

Browse files
metascroymetascroy
authored
Merge branch 'main' into fix_assets
2 parents 54c8298 + c78b0fd commit 6f0487f

Some content is hidden

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

50 files changed

+1637
-221
lines changed

‎.ci/scripts/test_backend_linux.sh‎

Lines changed: 31 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -10,6 +10,8 @@ SUITE=$1
1010
FLOW=$2
1111
ARTIFACT_DIR=$3
1212

13+
REPORT_FILE="$ARTIFACT_DIR/test-report-$FLOW-$SUITE.csv"
14+
1315
echo "Running backend test job for suite $SUITE, flow $FLOW."
1416
echo "Saving job artifacts to $ARTIFACT_DIR."
1517

@@ -18,13 +20,38 @@ eval "$(conda shell.bash hook)"
1820
CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
1921
conda activate "${CONDA_ENV}"
2022

21-
# Setup swiftshader and Vulkan SDK which are required to build the Vulkan delegate
22-
source .ci/scripts/setup-vulkan-linux-deps.sh
23+
export PYTHON_EXECUTABLE=python
2324

2425
# CMake options to use, in addition to the defaults.
25-
EXTRA_BUILD_ARGS="-DEXECUTORCH_BUILD_VULKAN=ON"
26+
EXTRA_BUILD_ARGS=""
27+
28+
if [[ "$FLOW" == *qnn* ]]; then
29+
# Setup QNN sdk and deps - note that this is a bit hacky due to the nature of the
30+
# Qualcomm build. TODO (gjcomer) Clean this up once the QNN pybinding integration is
31+
# cleaned up.
32+
PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool cmake
33+
PYTHON_EXECUTABLE=python bash .ci/scripts/setup-qnn-deps.sh
34+
PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
35+
QNN_X86_LIB_DIR=`realpath build-x86/lib/`
36+
QNN_SDK_ROOT="/tmp/qnn/2.28.0.241029"
37+
export LD_LIBRARY_PATH"=$QNN_X86_LIB_DIR:$QNN_SDK_ROOT/lib/x86_64-linux-clang/:${LD_LIBRARY_PATH:-}"
38+
39+
# TODO Get SDK root from install scripts
40+
EXTRA_BUILD_ARGS+=" -DEXECUTORCH_BUILD_QNN=ON -DQNN_SDK_ROOT=$QNN_SDK_ROOT"
41+
fi
42+
43+
if [[ "$FLOW" == *vulkan* ]]; then
44+
# Setup swiftshader and Vulkan SDK which are required to build the Vulkan delegate
45+
source .ci/scripts/setup-vulkan-linux-deps.sh
46+
47+
EXTRA_BUILD_ARGS+=" -DEXECUTORCH_BUILD_VULKAN=ON"
48+
fi
2649

2750
# We need the runner to test the built library.
2851
PYTHON_EXECUTABLE=python CMAKE_ARGS="$EXTRA_BUILD_ARGS" .ci/scripts/setup-linux.sh --build-tool cmake --build-mode Release --editable true
2952

30-
python -m executorch.backends.test.suite.runner $SUITE --flow $FLOW --report "$ARTIFACT_DIR/test_results.csv"
53+
EXIT_CODE=0
54+
python -m executorch.backends.test.suite.runner $SUITE --flow $FLOW --report "$REPORT_FILE" || EXIT_CODE=$?
55+
56+
# Generate markdown summary.
57+
python -m executorch.backends.test.suite.generate_markdown_summary "$REPORT_FILE" > ${GITHUB_STEP_SUMMARY:-"step_summary.md"} --exit-code $EXIT_CODE

‎.ci/scripts/test_backend_macos.sh‎

Lines changed: 7 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -10,6 +10,8 @@ SUITE=$1
1010
FLOW=$2
1111
ARTIFACT_DIR=$3
1212

13+
REPORT_FILE="$ARTIFACT_DIR/test-report-$FLOW-$SUITE.csv"
14+
1315
echo "Running backend test job for suite $SUITE, flow $FLOW."
1416
echo "Saving job artifacts to $ARTIFACT_DIR."
1517

@@ -21,4 +23,8 @@ eval "$(conda shell.bash hook)"
2123
PYTHON_EXECUTABLE=python
2224
${CONDA_RUN} --no-capture-output .ci/scripts/setup-macos.sh --build-tool cmake --build-mode Release
2325

24-
${CONDA_RUN} --no-capture-output python -m executorch.backends.test.suite.runner $SUITE --flow $FLOW --report "$ARTIFACT_DIR/test_results.csv"
26+
EXIT_CODE=0
27+
${CONDA_RUN} --no-capture-output python -m executorch.backends.test.suite.runner $SUITE --flow $FLOW --report "$REPORT_FILE" || EXIT_CODE=$?
28+
29+
# Generate markdown summary.
30+
${CONDA_RUN} --no-capture-output python -m executorch.backends.test.suite.generate_markdown_summary "$REPORT_FILE" > ${GITHUB_STEP_SUMMARY:-"step_summary.md"} --exit-code $EXIT_CODE

‎.github/workflows/nightly.yml‎

Lines changed: 8 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -42,7 +42,11 @@ jobs:
4242
strategy:
4343
fail-fast: false
4444
matrix:
45-
flow: [vulkan, xnnpack, xnnpack_static_int8_per_channel]
45+
flow: [
46+
qnn, qnn_16a16w, qnn_16a8w, qnn_16a4w, qnn_16a4w_block, qnn_8a8w,
47+
vulkan, vulkan_static_int8_per_channel,
48+
xnnpack, xnnpack_dynamic_int8_per_channel, xnnpack_static_int8_per_channel, xnnpack_static_int8_per_tensor
49+
]
4650
suite: [models, operators]
4751
with:
4852
ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
@@ -53,11 +57,8 @@ jobs:
5357
upload-artifact: test-report-${{ matrix.flow }}-${{ matrix.suite }}
5458
script: |
5559
set -eux
56-
# Intentionally suppressing exit code for now.
57-
# TODO (gjcomer) Remove this when jobs are stable.
58-
EXIT_CODE=0
59-
.ci/scripts/test_backend_linux.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}" || EXIT_CODE=$?
60-
echo "Test run complete with exit code $EXIT_CODE."
60+
61+
source .ci/scripts/test_backend_linux.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"
6162
6263
backend-test-macos:
6364
uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
@@ -82,6 +83,4 @@ jobs:
8283
# This is needed to get the prebuilt PyTorch wheel from S3
8384
${CONDA_RUN} --no-capture-output pip install awscli==1.37.21
8485
85-
EXIT_CODE=0
86-
.ci/scripts/test_backend_macos.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}" || EXIT_CODE=$?
87-
echo "Test run complete with exit code $EXIT_CODE."
86+
source .ci/scripts/test_backend_macos.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"

‎backends/cadence/aot/functions.yaml‎

Lines changed: 15 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -249,6 +249,21 @@
249249
- arg_meta: null
250250
kernel_name: impl::reference::quantized_relu_asym8u_asym8u_per_tensor_out
251251

252+
- func: cadence::quantized_add.per_tensor_out(Tensor X, float X_scale, int X_zero_point, Tensor Y, float Y_scale, int Y_zero_point, float out_scale, int out_zero_point, *, Tensor(a!) out) -> Tensor(a!)
253+
kernels:
254+
- arg_meta: null
255+
kernel_name: impl::reference::quantized_add_per_tensor_out
256+
257+
- func: cadence::quantized_add_asym8sxasym8s_asym8s.per_tensor_out(Tensor X, float X_scale, int X_zero_point, Tensor Y, float Y_scale, int Y_zero_point, float out_scale, int out_zero_point, *, Tensor(a!) out) -> Tensor(a!)
258+
kernels:
259+
- arg_meta: null
260+
kernel_name: impl::reference::quantized_add_asym8sxasym8s_asym8s_per_tensor_out
261+
262+
- func: cadence::quantized_add_asym8uxasym8u_asym8u.per_tensor_out(Tensor X, float X_scale, int X_zero_point, Tensor Y, float Y_scale, int Y_zero_point, float out_scale, int out_zero_point, *, Tensor(a!) out) -> Tensor(a!)
263+
kernels:
264+
- arg_meta: null
265+
kernel_name: impl::reference::quantized_add_asym8uxasym8u_asym8u_per_tensor_out
266+
252267
- func: cadence::quantized_matmul.out(Tensor X, int X_zero_point, Tensor Y, int Y_zero_point, Tensor? bias, int out_multiplier, int out_shift, int out_zero_point, bool transposed, *, Tensor(a!) out) -> Tensor(a!)
253268
kernels:
254269
- arg_meta: null

‎backends/cadence/aot/functions_hifi.yaml‎

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -404,6 +404,16 @@
404404
- arg_meta: null
405405
kernel_name: cadence::impl::HiFi::quantized_relu_asym8u_asym8u_per_tensor_out
406406

407+
- func: cadence::quantized_add_asym8sxasym8s_asym8s.per_tensor_out(Tensor X, float X_scale, int X_zero_point, Tensor Y, float Y_scale, int Y_zero_point, float out_scale, int out_zero_point, *, Tensor(a!) out) -> Tensor(a!)
408+
kernels:
409+
- arg_meta: null
410+
kernel_name: cadence::impl::HiFi::quantized_add_asym8sxasym8s_asym8s_per_tensor_out
411+
412+
- func: cadence::quantized_add_asym8uxasym8u_asym8u.per_tensor_out(Tensor X, float X_scale, int X_zero_point, Tensor Y, float Y_scale, int Y_zero_point, float out_scale, int out_zero_point, *, Tensor(a!) out) -> Tensor(a!)
413+
kernels:
414+
- arg_meta: null
415+
kernel_name: cadence::impl::HiFi::quantized_add_asym8uxasym8u_asym8u_per_tensor_out
416+
407417
- func: cadence::quantized_matmul.out(Tensor X, int X_zero_point, Tensor Y, int Y_zero_point, Tensor? bias, int out_multiplier, int out_shift, int out_zero_point, bool transposed, *, Tensor(a!) out) -> Tensor(a!)
408418
kernels:
409419
- arg_meta: null

‎backends/cadence/aot/memory_planning.py‎

Lines changed: 8 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -116,6 +116,9 @@ def plan_spec(
116116
Greedily place the spec in the first memory that can fit it.
117117
"""
118118
for spec.mem_id in range(1, self.get_num_memories()):
119+
if placement_constraints.is_mem_id_in_blocklist(spec, spec.mem_id):
120+
# Skip placement for blocked memory id.
121+
continue
119122
prev_offset, smallest_gap = 0, float("inf")
120123
for allocated_spec in state.allocated_buffers[spec.mem_id]:
121124
if not Verifier.lifetime_overlap(spec, allocated_spec):
@@ -141,11 +144,11 @@ def plan_spec(
141144
)
142145
if spec.mem_offset is None:
143146
spec.mem_offset = prev_offset
144-
if not self.is_valid_placement(spec, placement_constraints):
145-
spec.mem_offset = None
146-
continue
147-
else:
148-
spec.mem_offset = prev_offset
147+
148+
if not self.is_valid_placement(spec, placement_constraints):
149+
# Skip placement for invalid memory id.
150+
spec.mem_offset = None
151+
continue
149152

150153
state.place_spec(spec)
151154
# A data structure used for maintaining the tensor order

‎backends/cadence/aot/memory_planning_algo.py‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -204,7 +204,7 @@ def _place_memory_id_pinned_specs(
204204
for spec, c in spec_with_abs_constraint.items()
205205
if c is not None and c.pinned_memory_id == mem_id and c.offset is None
206206
}
207-
logging.error(f"Placing specs {mem_id_pinned_specs} for {mem_id=}")
207+
logging.debug(f"Placing specs {mem_id_pinned_specs} for {mem_id=}")
208208

209209
with self.block_memories_except(mem_id):
210210
self.plan(
@@ -220,7 +220,7 @@ def _place_memory_id_pinned_specs(
220220
if constraint is None:
221221
continue
222222

223-
logging.error(f"Placing spec {spec} with {constraint}")
223+
logging.debug(f"Placing spec {spec} with {constraint}")
224224

225225
if not state.is_placed(spec):
226226
raise MemoryError(

‎backends/cadence/aot/ops_registrations.py‎

Lines changed: 46 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -325,6 +325,22 @@
325325
"quantized_add.per_tensor_out(Tensor X, float X_scale, int X_zero_point, Tensor Y, float Y_scale, "
326326
"int Y_zero_point, float out_scale, int out_zero_point, *, Tensor(a!) out) -> Tensor(a!)"
327327
)
328+
lib.define(
329+
"quantized_add_asym8sxasym8s_asym8s.per_tensor(Tensor X, float X_scale, int X_zero_point, Tensor Y, float Y_scale, "
330+
"int Y_zero_point, float out_scale, int out_zero_point) -> Tensor"
331+
)
332+
lib.define(
333+
"quantized_add_asym8sxasym8s_asym8s.per_tensor_out(Tensor X, float X_scale, int X_zero_point, Tensor Y, float Y_scale, "
334+
"int Y_zero_point, float out_scale, int out_zero_point, *, Tensor(a!) out) -> Tensor(a!)"
335+
)
336+
lib.define(
337+
"quantized_add_asym8uxasym8u_asym8u.per_tensor(Tensor X, float X_scale, int X_zero_point, Tensor Y, float Y_scale, "
338+
"int Y_zero_point, float out_scale, int out_zero_point) -> Tensor"
339+
)
340+
lib.define(
341+
"quantized_add_asym8uxasym8u_asym8u.per_tensor_out(Tensor X, float X_scale, int X_zero_point, Tensor Y, float Y_scale, "
342+
"int Y_zero_point, float out_scale, int out_zero_point, *, Tensor(a!) out) -> Tensor(a!)"
343+
)
328344
lib.define(
329345
"quantized_mul.out(Tensor X, Tensor X_scale, Tensor X_zero_point, Tensor Y, Tensor Y_scale, "
330346
"Tensor Y_zero_point, float out_scale, int out_zero_point, *, Tensor(a!) out) -> Tensor(a!)"
@@ -503,6 +519,36 @@ def quantized_add_per_tensor_meta(
503519
return X.new_empty(out_size, dtype=X.dtype)
504520

505521

522+
@register_fake("cadence::quantized_add_asym8sxasym8s_asym8s.per_tensor")
523+
def quantized_add_asym8sxasym8s_asym8s_per_tensor_meta(
524+
X: torch.Tensor,
525+
X_scale: float,
526+
X_zero_point: int,
527+
Y: torch.Tensor,
528+
Y_scale: float,
529+
Y_zero_point: int,
530+
out_scale: float,
531+
out_zero_point: int,
532+
) -> torch.Tensor:
533+
out_size = torch.broadcast_shapes(X.size(), Y.size())
534+
return X.new_empty(out_size, dtype=X.dtype)
535+
536+
537+
@register_fake("cadence::quantized_add_asym8uxasym8u_asym8u.per_tensor")
538+
def quantized_add_asym8uxasym8u_asym8u_per_tensor_meta(
539+
X: torch.Tensor,
540+
X_scale: float,
541+
X_zero_point: int,
542+
Y: torch.Tensor,
543+
Y_scale: float,
544+
Y_zero_point: int,
545+
out_scale: float,
546+
out_zero_point: int,
547+
) -> torch.Tensor:
548+
out_size = torch.broadcast_shapes(X.size(), Y.size())
549+
return X.new_empty(out_size, dtype=X.dtype)
550+
551+
506552
@register_fake("cadence::quantized_linear")
507553
def quantized_linear_meta(
508554
src: torch.Tensor,

‎backends/cadence/aot/tests/test_memory_passes.py‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1044,7 +1044,7 @@ class DummyMemIdBlockConstraintGen(PassBase):
10441044
mul: blocks 1, 3
10451045
"""
10461046

1047-
def __init__(self, memory_constraints: MemoryConfig):
1047+
def __init__(self, memory_constraints: MemConstraints):
10481048
self.memory_constraints = memory_constraints
10491049

10501050
def call(self, graph_module: torch.fx.GraphModule) -> PassResult:

‎backends/cadence/aot/tests/test_type_dispatch_passes.py‎

Lines changed: 50 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -445,3 +445,53 @@ def test_uint8_dispatch_quantized_conv_nhwc_dilated(self) -> None:
445445
),
446446
1,
447447
)
448+
449+
def test_int8_dispatch_quantized_add(self) -> None:
450+
"""Test int8 x int8 inputs should dispatch to asym8sxasym8s_asym8s variant for quantized_add"""
451+
x = torch.randint(-128, 127, (2, 3), dtype=torch.int8)
452+
y = torch.randint(-128, 127, (2, 3), dtype=torch.int8)
453+
gm = single_op_builder(
454+
placeholders=(x, y),
455+
op=exir_ops.edge.cadence.quantized_add.per_tensor,
456+
args=(x, 1.0, 0, y, 1.0, 0, 1.0, 0),
457+
)
458+
p = CompileTimeTypeDispatchPass()
459+
gm = cast(PassResult, p(gm)).graph_module
460+
# Original op should be replaced
461+
self.assertEqual(
462+
count_node(gm, exir_ops.edge.cadence.quantized_add.per_tensor),
463+
0,
464+
)
465+
# Should be replaced with int8 specific variant
466+
self.assertEqual(
467+
count_node(
468+
gm,
469+
exir_ops.edge.cadence.quantized_add_asym8sxasym8s_asym8s.per_tensor,
470+
),
471+
1,
472+
)
473+
474+
def test_uint8_dispatch_quantized_add(self) -> None:
475+
"""Test uint8 x uint8 inputs should dispatch to asym8uxasym8u_asym8u variant for quantized_add"""
476+
x = torch.randint(0, 255, (2, 3), dtype=torch.uint8)
477+
y = torch.randint(0, 255, (2, 3), dtype=torch.uint8)
478+
gm = single_op_builder(
479+
placeholders=(x, y),
480+
op=exir_ops.edge.cadence.quantized_add.per_tensor,
481+
args=(x, 1.0, 0, y, 1.0, 0, 1.0, 0),
482+
)
483+
p = CompileTimeTypeDispatchPass()
484+
gm = cast(PassResult, p(gm)).graph_module
485+
# Original op should be replaced
486+
self.assertEqual(
487+
count_node(gm, exir_ops.edge.cadence.quantized_add.per_tensor),
488+
0,
489+
)
490+
# Should be replaced with uint8 specific variant
491+
self.assertEqual(
492+
count_node(
493+
gm,
494+
exir_ops.edge.cadence.quantized_add_asym8uxasym8u_asym8u.per_tensor,
495+
),
496+
1,
497+
)

0 commit comments

Comments
 (0)