Fixing xnnpack+qnnpack test (#355)

Summary:

Fixing all qnnpack tests. In addition to the broken test linked by mergen, there were many tests set to expect to fail.

This diff fixes all tests.

Reviewed By: digantdesai, kirklandsign

Differential Revision: D49068312

Author: mcr229

backends/qnnpack/QNNPackBackend.cpp

@@ -140,7 +140,8 @@ class QnnpackBackend final : public PyTorchBackendInterface {
         weights_zp->buffer()->data(),
         ScalarType::QUInt8,
         runtime_allocator,
-        0,
+        pre_pad_bytes, // Not necessary to prepad but surpresses asan errors:
+                       // D42179009
         &zp_buf);
 
     // Create + copy Weight Scales Tensor
@@ -152,7 +153,8 @@ class QnnpackBackend final : public PyTorchBackendInterface {
         weights_scale->buffer()->data(),
         ScalarType::Float,
         runtime_allocator,
-        0,
+        pre_pad_bytes, // Not necessary to prepad but surpresses asan errors:
+                       // D42179009
         &scale_buf);
 
     // Create Quantized Input Tensor

backends/qnnpack/partition/qnnpack_partitioner.py

@@ -5,9 +5,7 @@
 # LICENSE file in the root directory of this source tree.
 
 import logging
-from typing import Callable, Dict, List, Optional, Union
-
-import torch
+from typing import Dict, List, Optional, Union
 
 from executorch.backends.qnnpack.partition.support_patterns import (
     get_dynamic_quant_addmm_with_view_copy_graph,
@@ -16,14 +14,13 @@
     get_dynamic_quant_mm_without_view_copy_graph,
 )
 from executorch.backends.qnnpack.qnnpack_preprocess import QnnpackBackend
-from executorch.backends.transforms.addmm_mm_to_linear import (
-    apply_addmm_mm_to_linear_transform,
-)
+from executorch.backends.transforms.addmm_mm_to_linear import AddmmToLinearTransform
 from executorch.exir.backend.partitioner import (
     DelegationSpec,
     Partitioner,
     PartitionResult,
 )
+from torch._export.pass_base import PassType
 from torch.export import ExportedProgram
 from torch.fx.passes.utils.matcher_utils import SubgraphMatcher
 
@@ -69,7 +66,7 @@ def __init__(
         self,
         delegate_name,
         patterns,
-        transforms: Optional[List[Callable[[torch.fx.Graph], torch.fx.Graph]]] = None,
+        transforms: Optional[List[PassType]] = None,
     ):
         """
         @param transforms: Optional list of transforms that will be applied to the graph before running the partitioner.
@@ -157,5 +154,5 @@ def __init__(self) -> None:
             get_dynamic_quant_mm_without_view_copy_graph(dynamic_shape=True),
         ]
         super().__init__(
-            QnnpackBackend.__name__, qnnp_patterns, [apply_addmm_mm_to_linear_transform]
+            QnnpackBackend.__name__, qnnp_patterns, [AddmmToLinearTransform()]
         )

backends/qnnpack/serialization/qnnpack_graph_serialize.py

@@ -21,7 +21,6 @@
 
 def convert_to_flatbuffer(qnn_dynamic_linear: QNNDynamicLinear) -> bytes:
     qnnpack_graph_json = json.dumps(qnn_dynamic_linear, cls=_DataclassEncoder)
-
     with tempfile.TemporaryDirectory() as d:
         schema_path = os.path.join(d, "schema.fbs")
         with open(schema_path, "wb") as schema_file:

backends/qnnpack/test/test_qnnpack.py

@@ -51,8 +51,7 @@
 
 EDGE_COMPILE_CONFIG = exir.EdgeCompileConfig(_check_ir_validity=False)
 
-# TODO(T158653285)
-@unittest.expectedFailure
+
 class TestQnnbackends(unittest.TestCase):
     k_dim = 5
     input_dims = (1, 4, k_dim)
@@ -89,7 +88,7 @@ def test_qnnpack_per_channel_dynamic_mm(self):
         ).check(
             "executorch_exir_dialects_edge__ops_quantized_decomposed_quantize_per_tensor_tensor"
         ).check(
-            "executorch_exir_dialects_edge__ops_aten_t_copy_default"
+            "executorch_exir_dialects_edge__ops_aten_permute_copy_default"
         ).check(
             "executorch_exir_dialects_edge__ops_aten_mm"
         ).run(
@@ -170,7 +169,7 @@ def test_qnnpack_per_channel_dynamic_qlinear(self):
         ).check(
             "aten_view_copy_default"
         ).check(
-            "aten_t_copy_default"
+            "aten_permute_copy_default"
         ).check(
             "aten_addmm_default"
         ).check(
@@ -245,7 +244,7 @@ def test_qnnpack_per_tensor_dynamic_mm(self):
         ).check(
             "executorch_exir_dialects_edge__ops_quantized_decomposed_quantize_per_tensor_tensor"
         ).check(
-            "executorch_exir_dialects_edge__ops_aten_t_copy_default"
+            "executorch_exir_dialects_edge__ops_aten_permute_copy_default"
         ).check(
             "executorch_exir_dialects_edge__ops_aten_mm"
         ).run(
@@ -326,7 +325,7 @@ def test_qnnpack_per_tensor_dynamic_qlinear(self):
         ).check(
             "aten_view_copy_default"
         ).check(
-            "aten_t_copy_default"
+            "aten_permute_copy_default"
         ).check(
             "aten_addmm_default"
         ).check(
@@ -400,7 +399,7 @@ def test_qnnpack_per_channel_dynamic_mm_with_dynamic_shape(self):
         ).check(
             "executorch_exir_dialects_edge__ops_quantized_decomposed_quantize_per_tensor_tensor"
         ).check(
-            "executorch_exir_dialects_edge__ops_aten_t_copy_default"
+            "executorch_exir_dialects_edge__ops_aten_permute_copy_default"
         ).check(
             "executorch_exir_dialects_edge__ops_aten_mm"
         ).run(
@@ -482,7 +481,7 @@ def test_qnnpack_per_channel_dynamic_qlinear_via_partitioner(self):
         ).check(
             "aten_view_copy_default"
         ).check(
-            "aten_t_copy_default"
+            "aten_permute_copy_default"
         ).check(
             "aten_addmm_default"
         ).check(

backends/qnnpack/test/test_qnnpack_partitioner.py

@@ -66,8 +66,6 @@ def get_actual_dyanmic_quantized_graph(
     return dynamic_quantized_exir_graph.graph
 
 
-# TODO(T158653285)
-@unittest.expectedFailure
 class TestQnnbackends(unittest.TestCase):
     def test_dynamic_quantize_addmm_with_view_copy_partitioner(self):
         example_inputs = (torch.rand(5, 1, 256),)

backends/transforms/TARGETS

@@ -19,6 +19,7 @@ python_library(
     srcs = ["addmm_mm_to_linear.py"],
     deps = [
         "//caffe2:torch",
+        "//executorch/exir:pass_base",
         "//executorch/exir:sym_util",
         "//executorch/exir/dialects:lib",
     ],

backends/transforms/addmm_mm_to_linear.py

@@ -6,6 +6,7 @@
 
 import torch
 from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.exir.pass_base import ExportPass, PassResult
 
 from executorch.exir.sym_util import eval_shape
 
@@ -105,7 +106,10 @@ def replace_addmm_mm_with_linear(graph: torch.fx.Graph) -> torch.fx.Graph:
             with graph.inserting_after(node):
                 if node.target == ops.aten.addmm.default:
                     weight_t_node = node.args[2]
-                    if weight_t_node.target != ops.aten.t_copy.default:
+                    if weight_t_node.target not in [
+                        ops.aten.t_copy.default,
+                        ops.aten.permute_copy.default,
+                    ]:
                         raise RuntimeError(
                             f"Weight input to addmm must be tranposed but found {weight_t_node}"
                         )
@@ -120,7 +124,10 @@ def replace_addmm_mm_with_linear(graph: torch.fx.Graph) -> torch.fx.Graph:
                     )
                 else:
                     weight_t_node = node.args[1]
-                    if weight_t_node.target != ops.aten.t_copy.default:
+                    if weight_t_node.target not in [
+                        ops.aten.t_copy.default,
+                        ops.aten.permute_copy.default,
+                    ]:
                         raise RuntimeError(
                             f"Weight input to addmm must be tranposed but found {weight_t_node}"
                         )
@@ -145,3 +152,9 @@ def apply_addmm_mm_to_linear_transform(graph: torch.fx.Graph) -> torch.fx.Graph:
     graph = replace_addmm_mm_with_linear(graph)
     graph = replace_linear_view_copy_input_output(graph)
     return graph
+
+
+class AddmmToLinearTransform(ExportPass):
+    def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
+        graph_module.graph = apply_addmm_mm_to_linear_transform(graph_module.graph)
+        return PassResult(graph_module, True)

exir/tests/test_memory_planning.py

@@ -457,8 +457,6 @@ def quantize(self, eager_model: nn.Module) -> nn.Module:
         )
         return quantized_model
 
-    # TODO(T158653285)
-    @unittest.expectedFailure
     def test_asr_joiner(self) -> None:
         eager_model = self.quantize(ASRJoiner())
         inputs = eager_model.get_random_inputs()