Merge branch 'pytorch:main' into main

Author: jakmro

backends/arm/test/conftest.py

@@ -11,7 +11,6 @@
 import shutil
 import subprocess
 import sys
-from enum import auto, Enum
 from typing import Any
 
 import pytest
@@ -22,30 +21,24 @@
 """
 
 
-class arm_test_options(Enum):
-    quantize_io = auto()
-    corstone_fvp = auto()
-    fast_fvp = auto()
-
-
-_test_options: dict[arm_test_options, Any] = {}
-
 # ==== Pytest hooks ====
 
 
 def pytest_configure(config):
+    pytest._test_options = {}
+
     if config.option.arm_quantize_io:
         _load_libquantized_ops_aot_lib()
-        _test_options[arm_test_options.quantize_io] = True
+        pytest._test_options["quantize_io"] = True
     if config.option.arm_run_corstoneFVP:
         corstone300_exists = shutil.which("FVP_Corstone_SSE-300_Ethos-U55")
         corstone320_exists = shutil.which("FVP_Corstone_SSE-320")
         if not (corstone300_exists and corstone320_exists):
             raise RuntimeError(
                 "Tests are run with --arm_run_corstoneFVP but corstone FVP is not installed."
             )
-        _test_options[arm_test_options.corstone_fvp] = True
-    _test_options[arm_test_options.fast_fvp] = config.option.fast_fvp
+        pytest._test_options["corstone_fvp"] = True
+    pytest._test_options["fast_fvp"] = config.option.fast_fvp
     logging.basicConfig(level=logging.INFO, stream=sys.stdout)
 
 
@@ -131,9 +124,7 @@ def expectedFailureOnFVP(test_item):
 # ==== End of Custom Pytest decorators =====
 
 
-def is_option_enabled(
-    option: str | arm_test_options, fail_if_not_enabled: bool = False
-) -> bool:
+def is_option_enabled(option: str, fail_if_not_enabled: bool = False) -> bool:
     """
     Returns whether an option is successfully enabled, i.e. if the flag was
     given to pytest and the necessary requirements are available.
@@ -144,10 +135,8 @@ def is_option_enabled(
     The optional parameter 'fail_if_not_enabled' makes the function raise
       a RuntimeError instead of returning False.
     """
-    if isinstance(option, str):
-        option = arm_test_options[option.lower()]
 
-    if option in _test_options and _test_options[option]:
+    if option in pytest._test_options and pytest._test_options[option]:
         return True
     else:
         if fail_if_not_enabled:
@@ -156,15 +145,15 @@ def is_option_enabled(
             return False
 
 
-def get_option(option: arm_test_options) -> Any | None:
+def get_option(option: str) -> Any | None:
     """
     Returns the value of an pytest option if it is set, otherwise None.
 
     Args:
-        option (arm_test_options): The option to check for.
+        option (str): The option to check for.
     """
-    if option in _test_options:
-        return _test_options[option]
+    if option in pytest._test_options:
+        return pytest._test_options[option]
     return None
 
 

backends/arm/test/ops/test_depthwise_conv.py

@@ -156,14 +156,11 @@
     ("two_dw_conv2d", two_dw_conv2d),
 ]
 
-testsuite_conv2d_u85 = [
+testsuite_conv2d_u85_xfails = [
     ("2x2_1x6x4x4_gp6_st1", dw_conv2d_2x2_1x6x4x4_gp6_st1),
     ("3x3_1x3x256x256_gp3_st1", dw_conv2d_3x3_1x3x256x256_gp3_st1),
     ("3x3_1x4x256x256_gp4_st1", dw_conv2d_3x3_1x4x256x256_gp4_st1),
     ("3x3_1x4x256x256_gp4_nobias", dw_conv2d_3x3_1x4x256x256_gp4_nobias),
-]
-
-testsuite_conv2d_u85_xfails = [
     ("3x3_2x8x198x198_gp8_st3", dw_conv2d_3x3_2x8x198x198_gp8_st3),
     ("two_dw_conv2d", two_dw_conv2d),
 ]
@@ -260,6 +257,7 @@ def test_dw_conv_tosa_BI(self, test_name: str, model: torch.nn.Module):
     )  # Works
 
     @parameterized.expand(testsuite_conv2d, skip_on_empty=True)
+    @unittest.expectedFailure
     def test_dw_conv2d_u55_BI(
         self, test_name: str, model: torch.nn.Module, set_quantize_io: bool = False
     ):
@@ -286,7 +284,7 @@ def test_dw_conv1d_u55_BI(
             model.get_inputs(),
         )
 
-    @parameterized.expand(testsuite_conv1d + testsuite_conv2d_u85)
+    @parameterized.expand(testsuite_conv1d[2:])
     def test_dw_conv_u85_BI(
         self, test_name: str, model: torch.nn.Module, set_quantize_io: bool = False
     ):
@@ -298,8 +296,12 @@ def test_dw_conv_u85_BI(
             model.get_inputs(),
         )
 
+    testsuite_conv2d_u85_xfails.remove(
+        ("3x3_1x3x256x256_gp3_st1", dw_conv2d_3x3_1x3x256x256_gp3_st1)
+    )  # Works
+
     # All test cases except 3x3_1x3x256x256_gp3_st1 have numerical issues on FVP. MLETORCH-520
-    @parameterized.expand(testsuite_conv2d_u85_xfails)
+    @parameterized.expand(testsuite_conv2d_u85_xfails + testsuite_conv1d[:2])
     @conftest.expectedFailureOnFVP
     def test_dw_conv_u85_BI_xfails(
         self, test_name: str, model: torch.nn.Module, set_quantize_io: bool = False

backends/arm/test/ops/test_div.py

@@ -183,21 +183,8 @@ def test_div_tosa_BI(
         test_data = (input_, other_)
         self._test_div_tosa_BI_pipeline(self.Div(), test_data)
 
-    @parameterized.expand(test_data_suite[:2])
-    def test_div_u55_BI(
-        self,
-        test_name: str,
-        input_: Union[torch.Tensor, torch.types.Number],
-        other_: Union[torch.Tensor, torch.types.Number],
-        rounding_mode: Optional[str] = None,
-    ):
-        test_data = (input_, other_)
-        self._test_div_ethos_BI_pipeline(
-            self.Div(), common.get_u55_compile_spec(), test_data
-        )
-
     # Numerical issues on FVP likely due to mul op, MLETORCH-521
-    @parameterized.expand(test_data_suite[2:])
+    @parameterized.expand(test_data_suite)
     @conftest.expectedFailureOnFVP
     def test_div_u55_BI_xfails(
         self,
@@ -211,21 +198,8 @@ def test_div_u55_BI_xfails(
             self.Div(), common.get_u55_compile_spec(), test_data
         )
 
-    @parameterized.expand(test_data_suite[:2])
-    def test_div_u85_BI(
-        self,
-        test_name: str,
-        input_: Union[torch.Tensor, torch.types.Number],
-        other_: Union[torch.Tensor, torch.types.Number],
-        rounding_mode: Optional[str] = None,
-    ):
-        test_data = (input_, other_)
-        self._test_div_ethos_BI_pipeline(
-            self.Div(), common.get_u85_compile_spec(), test_data
-        )
-
     # Numerical issues on FVP likely due to mul op, MLETORCH-521
-    @parameterized.expand(test_data_suite[2:])
+    @parameterized.expand(test_data_suite)
     @conftest.expectedFailureOnFVP
     def test_div_u85_BI_xfails(
         self,

backends/arm/test/ops/test_mul.py

@@ -152,7 +152,9 @@ def test_mul_tosa_BI(
         test_data = (input_, other_)
         self._test_mul_tosa_BI_pipeline(self.Mul(), test_data)
 
+    # Numerical issues on FVP, MLETORCH-521
     @parameterized.expand(test_data_sute)
+    @conftest.expectedFailureOnFVP
     def test_mul_u55_BI(
         self,
         test_name: str,
@@ -164,7 +166,10 @@ def test_mul_u55_BI(
             common.get_u55_compile_spec(), self.Mul(), test_data
         )
 
-    @parameterized.expand(test_data_sute)
+    # Numerical issues on FVP, MLETORCH-521
+    # test_data_sute[0] works on U85
+    @parameterized.expand(test_data_sute[1:])
+    @conftest.expectedFailureOnFVP
     def test_mul_u85_BI(
         self,
         test_name: str,

backends/arm/test/runner_utils.py

@@ -17,7 +17,7 @@
 import numpy as np
 import torch
 
-from executorch.backends.arm.test.conftest import arm_test_options, is_option_enabled
+from executorch.backends.arm.test.conftest import is_option_enabled
 
 from torch.export import ExportedProgram
 from torch.fx.node import Node
@@ -251,7 +251,7 @@ def run_corstone(
             cmd_line += f" -i {input_path}"
 
         ethos_u_extra_args = ""
-        if is_option_enabled(arm_test_options.fast_fvp):
+        if is_option_enabled("fast_fvp"):
             ethos_u_extra_args = ethos_u_extra_args + "--fast"
 
         command_args = {

backends/cadence/aot/ops_registrations.py

@@ -146,7 +146,10 @@
     "quantized_fully_connected(Tensor src, Tensor weight, Tensor bias, int src_zero_point, "
     "Tensor weight_zero_point, Tensor out_multiplier, Tensor out_shift, int out_zero_point, Tensor? offset) -> (Tensor Z)"
 )
-
+lib.define(
+    "quantized_fully_connected.per_tensor(Tensor src, Tensor weight, Tensor bias, int src_zero_point, "
+    "int weight_zero_point, int out_multiplier, int out_shift, int out_zero_point, Tensor? offset) -> (Tensor Z)"
+)
 
 # ------------------------------------ #
 #   Migrated from custom_ops.ymal      #
@@ -192,6 +195,10 @@
     "quantized_fully_connected.out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, "
     "Tensor weight_zero_point, Tensor out_multiplier, Tensor out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)"
 )
+lib.define(
+    "quantized_fully_connected.per_tensor_out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, "
+    "int weight_zero_point, int out_multiplier, int out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)"
+)
 lib.define(
     "quantized_embedding_byte.out(Tensor weight, Tensor weight_scales, Tensor weight_zero_points, "
     "Tensor indices, bool pruned_weights=False, *, Tensor(a!) out) -> Tensor(a!)"
@@ -595,6 +602,28 @@ def quantized_fully_connected_meta(
     bias: torch.Tensor,
     in_zero_point: int,
     weight_zero_point: torch.Tensor,
+    out_multiplier: torch.Tensor,
+    out_shift: torch.Tensor,
+    out_zero_point: int,
+    offset: Optional[torch.Tensor],
+) -> torch.Tensor:
+    # src comes in shape [leading_dims, in_dim]
+    # weight comes in shape [out_dim, in_dim]
+    # output comes in empty with shape [leading_dims, out_dim]
+    out_size = list(src.size())
+    weight_size = list(weight.size())
+    assert len(weight_size) == 2
+    out_size[-1] = weight_size[0]
+    return src.new_empty(out_size, dtype=src.dtype)
+
+
+@register_fake("cadence::quantized_fully_connected.per_tensor")
+def quantized_fully_connected_per_tensor_meta(
+    src: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    in_zero_point: int,
+    weight_zero_point: int,
     out_multiplier: int,
     out_shift: int,
     out_zero_point: int,
@@ -607,7 +636,7 @@ def quantized_fully_connected_meta(
     weight_size = list(weight.size())
     assert len(weight_size) == 2
     out_size[-1] = weight_size[0]
-    return src.new_empty(out_size, dtype=torch.uint8)
+    return src.new_empty(out_size, dtype=src.dtype)
 
 
 @register_fake("cadence::convolution")

backends/cadence/aot/replace_ops.py

@@ -9,6 +9,8 @@
 # 3. functions that replace an ATen op with another semantically equivalent ATen op.
 # 4. functions that concretize optional args.
 
+# pyre-unsafe
+
 import math
 from operator import neg
 from typing import cast, Dict, Iterable, Sequence, Set, Tuple
@@ -1698,12 +1700,6 @@ def call_operator(self, op, args, kwargs, meta):
         if leading_dims != 1:
             return super().call_operator(op, args, kwargs, meta)
 
-        # If the op is quantized::linear, but per-channel quantized, bail.
-        if op == exir_ops.edge.cadence.quantized_linear.default:
-            weight = args[1].to_tensor() if isinstance(args[1], ProxyValue) else args[1]
-            if weight.shape != [1]:
-                return super().call_operator(op, args, kwargs, meta)
-
         # Replace the linear with fully connected op
         return super().call_operator(
             self.linear_to_fc_op[op],
@@ -1893,6 +1889,10 @@ class ReplaceSingleElementTensorArgumentsFromFullOpWithScalarPass(ExportPass):
             exir_ops.edge.cadence.quantized_conv.per_tensor,
             [8, 9, 12, 13],
         ),
+        exir_ops.edge.cadence.quantized_fully_connected: (
+            exir_ops.edge.cadence.quantized_fully_connected.per_tensor,
+            [4, 5, 6],
+        ),
         exir_ops.edge.cadence.quantized_layer_norm: (
             exir_ops.edge.cadence.quantized_layer_norm.per_tensor,
             [1, 2],