Summary: Fix CI job: test-arm-backend (test_run_ethosu_fvp) / linux-job.

Add 'qdq_fusion_pass' flag to run.sh & aot_arm_compiler. This
allows certain CI jobs not to fail as we conditionally allow fusion op
to be enabled / disabled with this flag. Note that since the fusion op
passes are still WIP and not all ops are yet supported, it is good to
rely on this flag until we have full fledged support for fusion ops.

Test Plan:
1. examples/arm/run.sh --et_build_root=arm_test/test_run --target=ethos-u55-128 --model_name=qops --bundleio --no_delegate --portable_kernels="aten::sub.out,aten::add.out,aten::mul.out" --qdq_fusion_op=true

// The above fails as 'QuantOpTest' has mixed binary operations.

2. examples/arm/run.sh --et_build_root=arm_test/test_run
   --target=ethos-u55-128 --model_name=qops --bundleio --no_delegate
--portable_kernels="aten::sub.out,aten::add.out,aten::mul.out"  -->
passes (as default flag value is set to false)

Reviewers:

Subscribers:

Tasks:

Tags:

Author: Github Executorch

examples/arm/aot_arm_compiler.py

@@ -600,6 +600,12 @@ def get_args():
         action="store_false",
         help="Disable strict checking while exporting models.",
     )
+    parser.add_argument(
+        "--enable_qdq_fusion_pass",
+        action="store",
+        default=False,
+        help="Skip the QuantizedOpFusionPass fusion step (default: False)",
+    )
     args = parser.parse_args()
 
     if args.evaluate and (
@@ -791,14 +797,20 @@ def to_edge_no_delegate(exported_program, args, model: torch.nn.Module, example_
     return model_int8, edge
 
 
-def transform_for_cortex_m_backend(edge):
+def transform_for_cortex_m_backend(edge, args):
     # Let's make sure we are using optimized Cortex M backend
     # NB: If we can't find and replace ops those are expected to be replaced,
     # bad things will happen at runtime, like "missing operator" errors!
-    # Instantiate the pass
-    replace_quant_pass = ReplaceQuantNodesPass()
-    quantized_op_fusion_pass = QuantizedOpFusionPass()
-    edge = edge.transform([replace_quant_pass, quantized_op_fusion_pass])
+
+    # Instantiate the mandatory ReplaceQuantNodesPass
+    passes = [ReplaceQuantNodesPass()]
+
+    # Conditionally add the QuantizedOpFusionPass
+    if args.enable_qdq_fusion_pass.lower() == "true":
+        passes.append(QuantizedOpFusionPass())
+
+    # Apply the passes
+    edge = edge.transform(passes)
 
     return edge
 
@@ -835,7 +847,7 @@ def transform_for_cortex_m_backend(edge):
         )
 
     # Transform so we can use ops from the Cortex M backend
-    edge = transform_for_cortex_m_backend(edge)
+    edge = transform_for_cortex_m_backend(edge, args)
 
     dump_delegation_info(edge, args.intermediates)
 

examples/arm/run.sh

@@ -40,6 +40,7 @@ ethos_u_scratch_dir=${script_dir}/ethos-u-scratch
 scratch_dir_set=false
 toolchain=arm-none-eabi-gcc
 select_ops_list="aten::_softmax.out"
+qdq_fusion_op=false
 
 function help() {
     echo "Usage: $(basename $0) [options]"
@@ -69,6 +70,7 @@ function help() {
     echo "  --pte_placement=<elf|ADDR>             Ethos-U: Control if runtime has PTE baked into the elf or if its placed in memory outside of the elf, defaults to ${pte_placement}"
     echo "  --et_build_root=<FOLDER>               Executorch build output root folder to use, defaults to ${et_build_root}"
     echo "  --scratch-dir=<FOLDER>                 Path to your Ethos-U scrach dir if you not using default ${ethos_u_scratch_dir}"
+    echo "  --qdq_fusion_op=<true/false>           Enable/Disable QDQ fusion op"
     exit 0
 }
 
@@ -96,6 +98,7 @@ for arg in "$@"; do
       --pte_placement=*) pte_placement="${arg#*=}";;
       --et_build_root=*) et_build_root="${arg#*=}";;
       --scratch-dir=*) ethos_u_scratch_dir="${arg#*=}" ; scratch_dir_set=true ;;
+      --qdq_fusion_op=*) qdq_fusion_op="${arg#*=}";;
       *)
       ;;
     esac
@@ -275,7 +278,7 @@ for i in "${!test_model[@]}"; do
         model_compiler_flags="${model_compiler_flags} --model_input=${model_input}"
     fi
 
-    ARM_AOT_CMD="python3 -m examples.arm.aot_arm_compiler --model_name=${model} --target=${target} ${model_compiler_flags} --intermediate=${output_folder} --output=${pte_file} --system_config=${system_config} --memory_mode=${memory_mode} $bundleio_flag ${etrecord_flag} --config=${config}"
+    ARM_AOT_CMD="python3 -m examples.arm.aot_arm_compiler --model_name=${model} --target=${target} ${model_compiler_flags} --intermediate=${output_folder} --output=${pte_file} --system_config=${system_config} --memory_mode=${memory_mode} $bundleio_flag ${etrecord_flag} --config=${config} --enable_qdq_fusion_pass=${qdq_fusion_op}"
     echo "CALL ${ARM_AOT_CMD}" >&2
     ${ARM_AOT_CMD} 1>&2