Merge branch 'main' into gh/gasoonjia/61/orig

Author: Gasoonjia

CONTRIBUTING.md

@@ -33,7 +33,7 @@ executorch
 │   ├── <a href="backends/openvino">openvino</a> - OpenVINO backend for Intel hardware.
 │   ├── <a href="backends/qualcomm">qualcomm</a> - Qualcomm-specific backends. See <a href="docs/source/backends-qualcomm.md">doc</a>.
 │   ├── <a href="backends/transforms">transforms</a> - Transformations for backend optimization.
-│   ├── <a href="backends/vulkan">vulkan</a> - Vulkan backend for cross-platform GPU support. See <a href="docs/source/backends-vulkan.md">doc</a>.
+│   ├── <a href="backends/vulkan">vulkan</a> - Vulkan backend for cross-platform GPU support. See <a href="docs/source/backends/vulkan/vulkan-overview.md">doc</a>.
 │   └── <a href="backends/xnnpack">xnnpack</a> - XNNPACK backend for optimized neural network operations. See <a href="docs/source/backends/xnnpack/xnnpack-overview.md">doc</a>.
 ├── <a href="codegen">codegen</a> - Tooling to autogenerate bindings between kernels and the runtime.
 ├── <a href="configurations">configurations</a> - Configuration files.

backends/arm/test/models/test_nn_modules.py

@@ -17,32 +17,91 @@
 - Transformer
 """
 
+from typing import Callable
+
 import torch
 from executorch.backends.arm.test.common import parametrize
 from executorch.backends.arm.test.tester.test_pipeline import (
     TosaPipelineFP,
     TosaPipelineINT,
 )
 
+
+def make_module_wrapper(
+    name: str, module_factory: Callable[[], torch.nn.Module]
+) -> torch.nn.Module:
+    class ModuleWrapper(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self._module = module_factory()
+
+        def forward(self, *args, **kwargs):
+            return self._module(*args, **kwargs)
+
+    ModuleWrapper.__name__ = name
+    ModuleWrapper.__qualname__ = name
+    return ModuleWrapper()
+
+
 example_input = torch.rand(1, 6, 16, 16)
 
 module_tests = [
-    (torch.nn.Embedding(10, 10), (torch.LongTensor([[1, 2, 4, 5], [4, 3, 2, 9]]),)),
-    (torch.nn.LeakyReLU(), (example_input,)),
-    (torch.nn.BatchNorm1d(16), (torch.rand(6, 16, 16),)),
-    (torch.nn.AdaptiveAvgPool2d((12, 12)), (example_input,)),
-    (torch.nn.ConvTranspose2d(6, 3, 2), (example_input,)),
-    (torch.nn.GRU(10, 20, 2), (torch.randn(5, 3, 10), torch.randn(2, 3, 20))),
-    (torch.nn.GroupNorm(2, 6), (example_input,)),
-    (torch.nn.InstanceNorm2d(16), (example_input,)),
-    (torch.nn.PReLU(), (example_input,)),
     (
-        torch.nn.Transformer(
-            d_model=64,
-            nhead=1,
-            num_encoder_layers=1,
-            num_decoder_layers=1,
-            dtype=torch.float32,
+        make_module_wrapper(
+            "EmbeddingModule",
+            lambda: torch.nn.Embedding(10, 10),
+        ),
+        (torch.LongTensor([[1, 2, 4, 5], [4, 3, 2, 9]]),),
+    ),
+    (
+        make_module_wrapper("LeakyReLUModule", torch.nn.LeakyReLU),
+        (example_input,),
+    ),
+    (
+        make_module_wrapper("BatchNorm1dModule", lambda: torch.nn.BatchNorm1d(16)),
+        (torch.rand(6, 16, 16),),
+    ),
+    (
+        make_module_wrapper(
+            "AdaptiveAvgPool2dModule",
+            lambda: torch.nn.AdaptiveAvgPool2d((12, 12)),
+        ),
+        (example_input,),
+    ),
+    (
+        make_module_wrapper(
+            "ConvTranspose2dModule", lambda: torch.nn.ConvTranspose2d(6, 3, 2)
+        ),
+        (example_input,),
+    ),
+    (
+        make_module_wrapper("GRUModule", lambda: torch.nn.GRU(10, 20, 2)),
+        (torch.randn(5, 3, 10), torch.randn(2, 3, 20)),
+    ),
+    (
+        make_module_wrapper("GroupNormModule", lambda: torch.nn.GroupNorm(2, 6)),
+        (example_input,),
+    ),
+    (
+        make_module_wrapper(
+            "InstanceNorm2dModule", lambda: torch.nn.InstanceNorm2d(16)
+        ),
+        (example_input,),
+    ),
+    (
+        make_module_wrapper("PReLUModule", torch.nn.PReLU),
+        (example_input,),
+    ),
+    (
+        make_module_wrapper(
+            "TransformerModule",
+            lambda: torch.nn.Transformer(
+                d_model=64,
+                nhead=1,
+                num_encoder_layers=1,
+                num_decoder_layers=1,
+                dtype=torch.float32,
+            ),
         ),
         (torch.rand((10, 32, 64)), torch.rand((20, 32, 64))),
     ),
@@ -78,9 +137,9 @@ def test_nn_Modules_FP(test_data):
     "test_data",
     test_parameters,
     xfails={
-        "GRU": "RuntimeError: Node aten_linear_default with op <EdgeOpOverload: aten.linear[...]> was not decomposed or delegated.",
-        "PReLU": "RuntimeError: mul(): functions with out=... arguments don't support automatic differentiation, but one of the arguments requires grad.",
-        "Transformer": "AssertionError: Output 0 does not match reference output.",
+        "GRUModule": "RuntimeError: Node aten_linear_default with op <EdgeOpOverload: aten.linear[...]> was not decomposed or delegated.",
+        "PReLUModule": "RuntimeError: mul(): functions with out=... arguments don't support automatic differentiation, but one of the arguments requires grad.",
+        "TransformerModule": "AssertionError: Output 0 does not match reference output.",
     },
 )
 def test_nn_Modules_INT(test_data):

backends/arm/test/models/test_resnet18.py

@@ -79,9 +79,6 @@ def test_resnet_u55_INT(per_channel_quantization):
 
 
 @pytest.mark.slow
-@pytest.mark.xfail(
-    reason="For resnet18 for Ethos-U85, the SRAM memory footprint is very high. The compiler team is investigating."
-)
 @common.XfailIfNoCorstone320
 @common.parametrize("per_channel_quantization", quant_test_data)
 def test_resnet_u85_INT(per_channel_quantization):

backends/arm/test/models/test_torch_functions.py

@@ -128,7 +128,6 @@ def test_torch_fns_FP(test_data):
         "Requires dynamic output shape.",
         "topk": "NotImplementedError: No registered serialization name for <class 'torch.return_types.topk'> found",
         "sort": "NotImplementedError: No registered serialization name for <class 'torch.return_types.sort'> found",
-        "t": "MLETORCH-855: Issue with Quantization folding.",
     },
     strict=False,
 )

backends/arm/test/ops/test_conv2d.py

@@ -117,26 +117,26 @@ def forward(self, x):
         return x
 
 
-conv2d_2x2_3x2x40x40_nobias = Conv2d(
+conv2d_2x2_3x2x14x14_nobias = Conv2d(
     in_channels=2,
     out_channels=3,
     kernel_size=(2, 2),
     stride=1,
     bias=False,
     padding=0,
-    width=40,
-    height=40,
-    batches=3,
+    width=14,
+    height=14,
+    batches=2,
 )
 
-conv2d_3x3_1x3x256x256_st1 = Conv2d(
+conv2d_3x3_1x3x24x24_st1 = Conv2d(
     in_channels=3,
     out_channels=10,
     kernel_size=(3, 3),
     stride=1,
     padding=0,
-    width=256,
-    height=256,
+    width=24,
+    height=24,
     batches=1,
 )
 
@@ -151,14 +151,14 @@ def forward(self, x):
     batches=1,
 )
 
-conv2d_1x1_1x2x128x128_st1 = Conv2d(
+conv2d_1x1_1x2x16x16_st1 = Conv2d(
     in_channels=2,
     out_channels=1,
     kernel_size=(1, 1),
     stride=1,
     padding=0,
-    width=128,
-    height=128,
+    width=16,
+    height=16,
     batches=1,
 )
 
@@ -173,25 +173,25 @@ def forward(self, x):
     batches=1,
 )
 
-conv2d_5x5_3x2x128x128_st1 = Conv2d(
+conv2d_5x5_3x2x24x24_st1 = Conv2d(
     in_channels=2,
     out_channels=3,
     kernel_size=(5, 5),
     stride=1,
     padding=0,
-    width=128,
-    height=128,
-    batches=3,
+    width=24,
+    height=24,
+    batches=2,
 )
 
-conv2d_3x3_1x3x224x224_st2_pd1 = Conv2d(
+conv2d_3x3_1x3x28x28_st2_pd1 = Conv2d(
     in_channels=3,
     out_channels=16,
     kernel_size=(3, 3),
     stride=2,
     padding=1,
-    width=224,
-    height=224,
+    width=28,
+    height=28,
     batches=1,
 )
 
@@ -304,8 +304,8 @@ def forward(self, x):
 
 two_conv2d_nobias = Conv2d(
     nbr_conv=2,
-    width=256,
-    height=256,
+    width=32,
+    height=32,
     in_channels=[3, 10],
     out_channels=[10, 15],
     kernel_size=[(5, 5), (5, 5)],
@@ -317,8 +317,8 @@ def forward(self, x):
 
 two_conv2d = Conv2d(
     nbr_conv=2,
-    width=256,
-    height=256,
+    width=32,
+    height=32,
     in_channels=[3, 10],
     out_channels=[10, 15],
     kernel_size=[(5, 5), (5, 5)],
@@ -359,10 +359,10 @@ def forward(self, x):
 # Shenanigan to get a nicer output when test fails. With unittest it looks like:
 # FAIL: test_convolution_2d_tosa_INT_2_3x3_1x3x12x12_st2_pd1
 test_data_FP = {
-    "2x2_3x2x40x40_nobias": lambda: conv2d_2x2_3x2x40x40_nobias,
-    "3x3_1x3x256x256_st1": lambda: conv2d_3x3_1x3x256x256_st1,
+    "2x2_3x2x14x14_nobias": lambda: conv2d_2x2_3x2x14x14_nobias,
+    "3x3_1x3x24x24_st1": lambda: conv2d_3x3_1x3x24x24_st1,
     "3x3_1x3x12x12_st2_pd1": lambda: conv2d_3x3_1x3x12x12_st2_pd1,
-    "1x1_1x2x128x128_st1": lambda: conv2d_1x1_1x2x128x128_st1,
+    "1x1_1x2x16x16_st1": lambda: conv2d_1x1_1x2x16x16_st1,
     "2x2_1x1x14x13_st2_needs_adjust_pass": lambda: conv2d_2x2_1x1x14x13_st2,
     "5x5_1x3x14x15_st3_pd1_needs_adjust_pass": lambda: conv2d_5x5_1x3x14x15_st3_pd1,
     "7x7_1x3x16x16_st2_pd1_dl2_needs_adjust_pass": lambda: conv2d_7x7_1x3x16x16_st2_pd1_dl2,
@@ -373,8 +373,8 @@ def forward(self, x):
     "3x3_1x3x8x9_st3_pd0_dl1_needs_adjust_pass": lambda: conv2d_3x3_1x3x8x9_st3_pd0_dl1,
     "3x4_1x3x7x7_st3_pd0_dl1_needs_adjust_pass": lambda: conv2d_3x4_1x3x7x7_st3_pd0_dl1,
     "4x3_1x3x7x7_st3_pd0_dl1_needs_adjust_pass": lambda: conv2d_4x3_1x3x7x7_st3_pd0_dl1,
-    "5x5_3x2x128x128_st1": lambda: conv2d_5x5_3x2x128x128_st1,
-    "3x3_1x3x224x224_st2_pd1": lambda: conv2d_3x3_1x3x224x224_st2_pd1,
+    "5x5_3x2x24x24_st1": lambda: conv2d_5x5_3x2x24x24_st1,
+    "3x3_1x3x28x28_st2_pd1": lambda: conv2d_3x3_1x3x28x28_st2_pd1,
     "two_conv2d_nobias": lambda: two_conv2d_nobias,
     "two_conv2d": lambda: two_conv2d,
     "groups": lambda: conv2d_groups,

backends/arm/test/ops/test_conv_combos.py

@@ -48,7 +48,7 @@ def __init__(self):
         # 1. 1x1 CONV2d + ReLU6 (Pointwise)
         self.pointwise_conv2d = torch.nn.Conv2d(
             in_channels=16, out_channels=96, kernel_size=1, stride=1, groups=1
-        )  ## (1, 128, 81, 81)
+        )  ## Example output shape (1, 96, 33, 33)
         self.batch_norm2d_16 = torch.nn.BatchNorm2d(96, affine=False)
         self.relu6 = torch.nn.ReLU6()
 
@@ -60,15 +60,15 @@ def __init__(self):
             padding=1,
             stride=1,
             groups=96,
-        )  ## (1, 128, H, W)
+        )  ## Example output shape (1, 96, H, W)
 
         # 3. Linear 1x1 Conv2d
         self.pointwise_conv2d_linear = torch.nn.Conv2d(
             in_channels=96, out_channels=16, kernel_size=1, stride=1, groups=1
-        )  ## (1, 32, 81, 81)
+        )  ## Example output shape (1, 16, 33, 33)
 
     def get_inputs(self) -> Tuple[torch.Tensor]:
-        return (torch.randn(1, 16, 81, 81),)
+        return (torch.randn(1, 16, 33, 33),)
 
     def forward(self, x):
         input = x
@@ -106,7 +106,7 @@ def __init__(self):
         self.adaptive_avg_pool2d = torch.nn.AdaptiveAvgPool2d((1, 1))
 
     def get_inputs(self) -> Tuple[torch.Tensor]:
-        return (torch.randn(1, 3, 128, 128),)
+        return (torch.randn(1, 3, 48, 48),)
 
     def forward(self, x):
         x = self.conv2d(x)
@@ -145,7 +145,7 @@ def __init__(self, affine: bool):
         self.relu6 = torch.nn.ReLU6()
 
     def get_inputs(self) -> Tuple[torch.Tensor]:
-        return (torch.randn(1, 3, 256, 256),)
+        return (torch.randn(1, 3, 64, 64),)
 
     def forward(self, x):
         x = self.conv2d(x)
@@ -161,11 +161,11 @@ class ComboConvRelu6(torch.nn.Module):
     ]
 
     test_data_FP = {
-        "combo_conv_relu_2_x_4d": lambda: (2 * torch.randn(1, 3, 256, 256),),
-        "combo_conv_relu_0_5_x_4d": lambda: (0.5 * torch.randn(1, 3, 256, 256),),
-        "combo_conv_relu_4d": lambda: (torch.randn(1, 3, 256, 256),),
-        "combo_conv_relu_neg_0_5_x_4d": lambda: (-0.5 * torch.randn(1, 3, 256, 256),),
-        "combo_conv_relu_neg_2_x_4d": lambda: (-2 * torch.randn(1, 3, 256, 256),),
+        "combo_conv_relu_2_x_4d": lambda: (2 * torch.randn(1, 3, 64, 64),),
+        "combo_conv_relu_0_5_x_4d": lambda: (0.5 * torch.randn(1, 3, 64, 64),),
+        "combo_conv_relu_4d": lambda: (torch.randn(1, 3, 64, 64),),
+        "combo_conv_relu_neg_0_5_x_4d": lambda: (-0.5 * torch.randn(1, 3, 64, 64),),
+        "combo_conv_relu_neg_2_x_4d": lambda: (-2 * torch.randn(1, 3, 64, 64),),
     }
 
     # Generate a new test set paired with per_channel_quant=True/False.
@@ -196,10 +196,10 @@ class ComboConvAvgPool2d(torch.nn.Module):
     ]
 
     test_data_FP = {
-        "combo_conv_avgpool_20_x_4d": lambda: (20 * torch.randn(1, 3, 64, 32),),
-        "combo_conv_avgpool_4d": lambda: (torch.randn(1, 3, 100, 200),),
-        "combo_conv_avgpool_5_x_4d_randn": lambda: (5 * torch.randn(1, 3, 256, 256),),
-        "combo_conv_avgpool_2_x_4d": lambda: (torch.rand(1, 3, 512, 128),),
+        "combo_conv_avgpool_20_x_4d": lambda: (20 * torch.randn(1, 3, 48, 24),),
+        "combo_conv_avgpool_4d": lambda: (torch.randn(1, 3, 60, 120),),
+        "combo_conv_avgpool_5_x_4d_randn": lambda: (5 * torch.randn(1, 3, 64, 64),),
+        "combo_conv_avgpool_2_x_4d": lambda: (torch.rand(1, 3, 96, 32),),
     }
 
     # Generate a new test set paired with per_channel_quant=True/False.

backends/arm/test/ops/test_cosh.py

@@ -76,9 +76,6 @@ def test_cosh_u55_INT(test_data: Tuple):
 @common.parametrize(
     "test_data",
     test_data_suite,
-    xfails={
-        "ones_4D": "MLBEDSW-11046 - Incorrect output for TABLE followed by RESHAPE"
-    },
     strict=False,
 )
 def test_cosh_u85_INT(test_data: Tuple):