Update on "[ExecuTorch] Add broadcast support for optimized add op"

Summary:
This brings add op to feature parity, wrt, broadcasting, to mul op in
optimized kernels lib

Test Plan:
tests added

Reviewers:

Subscribers:

Tasks:

Tags:

cc larryliu0820 manuelcandales

Differential Revision: [D69491814](https://our.internmc.facebook.com/intern/diff/D69491814)

[ghstack-poisoned]

Author: kimishpatel

backends/xnnpack/partition/config/gemm_configs.py

@@ -334,6 +334,18 @@ def check_constraints(self, node: torch.fx.Node, ep: ExportedProgram) -> bool:
         is_transpose = node.args[6]
         groups = cast(int, node.args[8])
 
+        # XNNPack does not support non-zero output padding in transposed
+        # convolutions.
+        if is_transpose and any(
+            out_pad != 0 for out_pad in cast(List[int], node.args[7])
+        ):
+            why(
+                node,
+                "XNNPACK does not support transposed convolutions with"
+                "non-zero output padding",
+            )
+            return False
+
         if (
             is_transpose
             and weight_quant_params is not None

backends/xnnpack/test/ops/test_conv2d.py

@@ -657,3 +657,42 @@ def get_inputs(self):
                     quant_config=None,
                     conv_count=1,
                 )
+
+    def test_padded_output_tconv(self):
+        class TConv2d(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.conv = torch.nn.ConvTranspose2d(
+                    in_channels=2,
+                    out_channels=1,
+                    kernel_size=(3, 3),
+                    stride=(2, 2),
+                    padding=(1, 1),
+                    output_padding=(0, 1),
+                    dilation=(1, 1),
+                    groups=1,
+                    bias=True,
+                ).to(torch.float)
+
+            def forward(self, x):
+                return self.conv(x)
+
+        m = TConv2d()
+        inputs = (torch.randn(1, 2, 8, 8),)
+        tester = Tester(m.eval(), inputs)
+
+        conv_count: int = 1
+        op = "torch.ops.aten.conv_transpose2d"
+
+        (tester.export().check_count({op: conv_count}).to_edge_transform_and_lower())
+
+        # tconv should not be offloaded to XNNPack, since output padding is not
+        (
+            tester.check(
+                ["executorch_exir_dialects_edge__ops_aten_convolution_default"]
+            )
+            .check_not(["torch.ops.higher_order.executorch_call_delegate"])
+            .to_executorch()
+            .serialize()
+            .run_method_and_compare_outputs(qtol=1)
+        )

examples/xnnpack/README.md

@@ -31,7 +31,7 @@ Once we have the model binary (pte) file, then let's run it with ExecuTorch runt
 cd executorch
 
 # Get a clean cmake-out directory
-./install_requiements.sh --clean
+./install_executorch.sh --clean
 mkdir cmake-out
 
 # Configure cmake

extension/flat_tensor/serialize/CMakeLists.txt

@@ -4,7 +4,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-# Flatbuffer schema header lib. Please this file formatted by running:
+# Flatbuffer schema header lib. Please keep this file formatted by running:
 # ~~~
 # cmake-format -i CMakeLists.txt
 # ~~~

extension/flat_tensor/test/CMakeLists.txt

@@ -4,7 +4,6 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-# @generated by test/utils/generate_gtest_cmakelists.py
 #
 # This file should be formatted with
 # ~~~

runtime/executor/program.cpp

@@ -294,7 +294,12 @@ Result<const void*> Program::get_constant_buffer_data(
   // loaded during Program::load, or stored inside the flatbuffer data
   // (constant_buffer).
   if (constant_segment_data_.data() != nullptr) {
-    size_t num_elems = internal_program->constant_segment()->offsets()->size();
+    const auto* constant_segment = internal_program->constant_segment();
+    size_t num_elems = constant_segment == nullptr
+        ? 0
+        : (constant_segment->offsets() == nullptr
+               ? 0
+               : constant_segment->offsets()->size());
     ET_CHECK_OR_RETURN_ERROR(
         buffer_index < num_elems,
         InvalidArgument,
@@ -326,25 +331,27 @@ Result<const void*> Program::get_constant_buffer_data(
         offset);
   } else {
     // Otherwise, the constant data is stored inside Program.constant_buffer.
-    size_t num_elems = internal_program->constant_buffer()->size();
+    const auto* constant_buffer_ptr = internal_program->constant_buffer();
+    size_t num_elems =
+        constant_buffer_ptr == nullptr ? 0 : constant_buffer_ptr->size();
     ET_CHECK_OR_RETURN_ERROR(
         buffer_index < num_elems,
         InvalidArgument,
         "Constant buffer index %zu invalid for program constant buffer range %zu",
         buffer_index,
         num_elems);
 
-    const auto& constant_buffer = *internal_program->constant_buffer();
-
+    const auto& constant_buffer = *constant_buffer_ptr;
+    const auto* storage = constant_buffer[buffer_index]->storage();
+    auto storage_size = storage == nullptr ? 0 : storage->size();
     ET_CHECK_OR_RETURN_ERROR(
-        constant_buffer[buffer_index]->storage()->size() <= nbytes,
+        storage_size <= nbytes,
         InvalidArgument,
         "Constant buffer size %u larger than allocated nbytes %zu",
-        constant_buffer[buffer_index]->storage()->size(),
+        storage_size,
         nbytes);
 
-    return static_cast<const void*>(
-        constant_buffer[buffer_index]->storage()->data());
+    return storage->data();
   }
 }
 

test/utils/OSSTestConfig.json

@@ -18,13 +18,6 @@
             "../print_evalue.cpp"
         ]
     },
-    {
-        "directory": "extension/flat_tensor/test",
-        "sources": [
-            "flat_tensor_data_map_test.cpp",
-            "flat_tensor_header_test.cpp"
-        ]
-    },
     {
         "directory": "extension/kernel_util/test",
         "sources": [