Merge branch 'main' into wasm-bindings-js

Author: Conarnar

.github/workflows/trunk.yml

@@ -302,6 +302,37 @@ jobs:
           exit 1
         fi
 
+  test-arm-ootb-linux:
+    name: test-arm-ootb-linux
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    with:
+      runner: linux.2xlarge
+      docker-image: executorch-ubuntu-22.04-arm-sdk
+      submodules: 'recursive'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        # Follow the steps required before running the notebooks
+        # Try to mirror these as closely as possible
+        source .ci/scripts/utils.sh
+        install_executorch "--use-pt-pinned-commit"
+
+        .ci/scripts/setup-arm-baremetal-tools.sh
+        source examples/arm/ethos-u-scratch/setup_path.sh
+
+        # Install requirements for converting notebooks
+        pip install notebook
+
+        # Run OOTB tests
+        backends/arm/test/test_arm_ootb.sh
+
   test-coreml-delegate:
     name: test-coreml-delegate
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main

backends/arm/_passes/decompose_asin_pass.py

@@ -85,12 +85,11 @@ def _build_polynomial(
         return result
 
     def call_operator(self, op, args, kwargs, meta):
+        if op not in edge_asin_op:
+            return super().call_operator(op, args, kwargs, meta)
         logging.info(
             f"Approximating asin. This may introduce small numerical errors. For details, see {__file__}."
         )
-        if op not in edge_asin_op:
-            return super().call_operator(op, args, kwargs, meta)
-
         x = args[0]
         half = 0.5
         one = 1.0

backends/arm/test/ops/test_add.py

@@ -205,6 +205,5 @@ def test_add_tensor_vgf_int(test_data: input_t1):
         aten_op,
         exir_op,
         tosa_version="TOSA-1.0+INT",
-        symmetric_io_quantization=True,
     )
     pipeline.run()

backends/arm/test/test_arm_ootb.sh

@@ -0,0 +1,18 @@
+#!/usr/bin/env bash
+
+# Copyright 2025 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+set -e
+
+run_ootb_tests_ethos_u() {
+    echo "$FUNCNAME: Running out-of-the-box tests for Arm Ethos-U"
+    jupyter nbconvert \
+        --to notebook \
+        --execute examples/arm/ethos_u_minimal_example.ipynb
+    echo "${FUNCNAME}: PASS"
+}
+
+run_ootb_tests_ethos_u

backends/arm/test/tester/arm_tester.py

@@ -43,6 +43,7 @@
     EthosUQuantizer,
     get_symmetric_quantization_config,
     TOSAQuantizer,
+    VgfQuantizer,
 )
 from executorch.backends.arm.test.runner_utils import (
     dbg_tosa_fb_to_json,
@@ -332,6 +333,8 @@ def quantize(
                 quantizer = TOSAQuantizer(tosa_spec)
             elif is_ethosu(self.compile_spec):
                 quantizer = EthosUQuantizer(self.compile_spec)
+            elif is_vgf(self.compile_spec):
+                quantizer = VgfQuantizer(self.compile_spec)
             quantize_stage = tester.Quantize(
                 quantizer,
                 get_symmetric_quantization_config(),

backends/arm/test/tester/test_pipeline.py

@@ -861,18 +861,15 @@ def __init__(
         rtol: float = 1e-03,
         qtol: int = 1,
         dynamic_shapes: Optional[Tuple[Any]] = None,
+        transform_passes: Optional[
+            Union[Sequence[PassType], Dict[str, Sequence[PassType]]]
+        ] = None,
     ):
 
         if (
             symmetric_io_quantization or per_channel_quantization
         ) and tosa_version == "TOSA-1.0+FP":
             raise ValueError("Dont configure quantization with FP TOSA profile.")
-        if (
-            symmetric_io_quantization is False
-            and per_channel_quantization is False
-            and tosa_version == "TOSA-1.0+INT"
-        ):
-            raise ValueError("Missing quantization options for INT TOSA profile.")
 
         tosa_profile = TosaSpecification.create_from_string(tosa_version)
         compile_spec = common.get_vgf_compile_spec(
@@ -887,6 +884,7 @@ def __init__(
             exir_op,
             use_to_edge_transform_and_lower,
             dynamic_shapes,
+            transform_passes=transform_passes,
         )
 
         if symmetric_io_quantization or per_channel_quantization:
@@ -900,7 +898,7 @@ def __init__(
         else:
             quant_stage = None
 
-        if quant_stage:
+        if "INT" in tosa_version:
             self.add_stage(self.tester.quantize, quant_stage, pos=0)
 
             self.add_stage_after(

backends/cadence/aot/tests/test_fusion_ops_passes.py

@@ -12,7 +12,6 @@
 
 import executorch.backends.cadence.aot.ops_registrations  # noqa
 import torch
-from executorch.backends.cadence.aot import compiler
 from executorch.backends.cadence.aot.fuse_ops import (
     FuseCascadedTransposeOrPermuteOps,
     FuseCascadedViewOps,
@@ -30,7 +29,6 @@
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.dialects.edge._ops import EdgeOpOverload
 from executorch.exir.pass_base import PassResult, ProxyValue
-from torch import nn
 
 
 class TestFusionPassesBase(unittest.TestCase):
@@ -178,43 +176,6 @@ def test_keep_mm_add_with_multiple_users(self) -> None:
         self.assertEqual(count_node(converted_graph, exir_ops.edge.aten.mm.default), 1)
         self.assertEqual(count_node(converted_graph, exir_ops.edge.aten.add.Tensor), 3)
 
-    # TODO(matthiascremon) -> None: enable that pass with new flow
-    @torch.no_grad()
-    @unittest.expectedFailure
-    def test_legacy_conv_bn_fusion(self) -> None:
-        class ModelConvBN(torch.nn.Module):
-            def __init__(
-                self, in_features: int, out_features: int, kernel_size: int
-            ) -> None:
-                super().__init__()
-                self.conv1d = nn.Conv1d(in_features, out_features, kernel_size)
-                self.bn = nn.BatchNorm1d(out_features)
-
-            def forward(self, x: torch.Tensor) -> torch.Tensor:
-                y = self.conv1d(x)
-                return self.bn(y)
-
-        model = ModelConvBN(64, 1, 2)
-        x = torch.randn(1, 64, 4)
-
-        graph_module = (
-            compiler.export_to_executorch_gen_etrecord(model.eval(), (x,))
-            .exported_program()
-            .graph_module
-        )
-        # Assert that after running the fusion passes, batchnorm was fused with conv1d
-        self.assertEqual(
-            count_node(graph_module, torch.ops.aten.linear.out)
-            + count_node(graph_module, torch.ops.cadence.convolution.out),
-            1,
-        )
-        self.assertEqual(
-            count_node(
-                graph_module, torch.ops.aten._native_batch_norm_legit_no_training.out
-            ),
-            0,
-        )
-
     def test_permute_transpose_fusion(self) -> None:
         builder = GraphBuilder()
         x = builder.placeholder("x", torch.randn(3, 1, 3, 1, 4, dtype=torch.float32))

backends/nxp/tests/ir/converter/node_converter/test_max_pool_2d_converter.py

@@ -61,7 +61,8 @@ def test_max_pool_2d_conversion(input_shape, padding):
     # Otherwise, we get violation that this op is not part of ATen Core ops.
     edge_program._verifiers = [
         EXIREdgeDialectVerifier(
-            class_only=True, exception_list=[torch.ops.aten.max_pool2d.default]
+            class_only=True,
+            core_aten_ops_exception_list=[torch.ops.aten.max_pool2d.default],
         )
     ]
 

backends/nxp/tests/test_node_format_inference.py

@@ -71,7 +71,8 @@ def test_maxpool2d():
     # Otherwise, we get violation that this op is not part of ATen Core ops.
     edge_program._verifiers = [
         EXIREdgeDialectVerifier(
-            class_only=True, exception_list=[torch.ops.aten.max_pool2d.default]
+            class_only=True,
+            core_aten_ops_exception_list=[torch.ops.aten.max_pool2d.default],
         )
     ]
 

backends/vulkan/runtime/api/Context.cpp

@@ -39,6 +39,7 @@ Context::Context(vkapi::Adapter* adapter, const ContextConfig& config)
       // Command buffer submission
       cmd_mutex_{},
       cmd_(VK_NULL_HANDLE, VK_NULL_HANDLE, 0u),
+      prev_semaphore_(VK_NULL_HANDLE),
       submit_count_{0u},
       // Memory Management
       buffer_clearlist_mutex_{},
@@ -195,10 +196,21 @@ void Context::register_blit(
 }
 
 void Context::submit_cmd_to_gpu(VkFence fence_handle, const bool final_use) {
+  // Wait semaphore would be previous command buffer's signal semaphore
+  VkSemaphore wait_semaphore = prev_semaphore_;
+  // Signal semaphore for the the current command buffer
+  VkSemaphore signal_semaphore = cmd_.get_signal_semaphore();
+  // Next command buffer would wait on this command buffer's signal semaphore
+  prev_semaphore_ = signal_semaphore;
+
   if (cmd_) {
     cmd_.end();
     adapter_p_->submit_cmd(
-        queue_, cmd_.get_submit_handle(final_use), fence_handle);
+        queue_,
+        cmd_.get_submit_handle(final_use),
+        fence_handle,
+        wait_semaphore,
+        signal_semaphore);
 
     submit_count_ = 0u;
   }
@@ -214,6 +226,8 @@ void Context::flush() {
   if (cmd_) {
     cmd_.invalidate();
   }
+  // Reset previous command buffer semaphore
+  prev_semaphore_ = VK_NULL_HANDLE;
 
   std::lock_guard<std::mutex> bufferlist_lock(buffer_clearlist_mutex_);
   std::lock_guard<std::mutex> imagelist_lock(image_clearlist_mutex_);