Skip to content

Implement dual LUT scheme for int16 reciprocal/rsqrt operations #16396

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
3l1 wants to merge 1 commit into
base: main
Choose a base branch
from
+5 −3
Open

Implement dual LUT scheme for int16 reciprocal/rsqrt operations #16396

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions backends/arm/_passes/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@


from . import arm_pass_utils # noqa
from . import fb

Check failure on line 8 in backends/arm/_passes/__init__.py

View workflow job for this annotation

GitHub Actions / lintrunner-mypy / linux-job

MYPY attr-defined 

Module "backends.arm._passes" has no attribute "fb" 

To disable, use `  # type: ignore[attr-defined]`
from .arm_pass import ArmPass # noqa # usort: skip
from .annotate_decomposed_matmul import AnnotateDecomposedMatmulPass # noqa
from .annotate_output_dim_order_pass import AnnotateOutputDimOrderPass # noqa
Expand Down Expand Up @@ -80,6 +81,7 @@
)
from .decompose_var_pass import DecomposeVarPass # noqa
from .decorate_fp32_to_int32_casting_pass import DecorateFp32toInt32CastingPass # noqa
from .fb.reciprocal_dual_lut_pass import ReciprocalDualLutPass # noqa

Check failure on line 84 in backends/arm/_passes/__init__.py

View workflow job for this annotation

GitHub Actions / lintrunner-mypy / linux-job

MYPY import-not-found 

Cannot find implementation or library stub for module named "backends.arm._passes.fb.reciprocal_dual_lut_pass" 

To disable, use `  # type: ignore[import-not-found]`
from .fold_qdq_with_annotated_qparams_pass import ( # noqa
FoldAndAnnotateQParamsPass,
QuantizeClampArgumentsPass,
Expand Down
3 changes: 3 additions & 0 deletions backends/arm/_passes/arm_pass_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -107,6 +107,7 @@
RewriteConvPass,
RewriteMatmulPass,
RewriteUpsamplePass,
ReciprocalDualLutPass,
ScalarsToAttributePass,
SizeAdjustInputPass,
ToTosaMemoryFormatPass,
Expand Down Expand Up @@ -329,6 +330,8 @@ def _tosa_pipeline(
FuseViewCopyTransformPass(),
DecomposeConvWithInt16ActivationPass(),
DecomposeSumPass(),
# FIX: Fixed LUT generation, re-enabled dual LUT pass
ReciprocalDualLutPass(exported_program),
InsertTableOpsPass(exported_program),
]
)
Expand Down
3 changes: 0 additions & 3 deletions backends/arm/test/ops/test_rsqrt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -124,9 +124,6 @@ def test_rsqrt_tosa_INT_a16w8(test_tensor: torch.Tensor):

@common.parametrize("test_tensor", Rsqrt.test_parameters)
@common.XfailIfNoCorstone300
@pytest.mark.xfail(
reason="MLETORCH-707: AssertionError: Output 0 does not match reference output."
)
def test_rsqrt_16a8w_u55_INT(test_tensor: torch.Tensor):
"""Test rsqrt operation with int16 I/O quantization for U55"""
# Use wider tolerances for int16 I/O quantization on U55
Expand Down
Loading