fixed empty output from runtime executor in CPU flow (#6184)

Summary:
Pull Request resolved: pytorch/executorch#6184

- add missing dependency dtype_util.cpp
- it was missing etrecord. And dump-output is not toggled in executor.
- Also enable runtime logging

Reviewed By: Vysarat, mcremon-meta

Differential Revision: D64272605

fbshipit-source-id: 9299fafeb6575b392d26b10caddef2192cae68e5

Author: zonglinpeng

backends/cadence/aot/TARGETS

@@ -43,6 +43,7 @@ python_library(
         "//executorch/backends/transforms:decompose_sdpa",
         "//executorch/backends/transforms:remove_clone_ops",
         "//executorch/exir:lib",
+        "//executorch/devtools:lib",
     ],
 )
 

backends/cadence/aot/compiler.py

@@ -7,6 +7,7 @@
 # pyre-strict
 
 import logging
+from pathlib import Path
 from typing import Optional
 
 import torch
@@ -29,7 +30,13 @@
     DecomposeScaledDotProductAttention,
 )
 from executorch.backends.transforms.remove_clone_ops import RemoveCloneOpsTransform
-from executorch.exir import EdgeCompileConfig, EdgeProgramManager, to_edge
+from executorch.devtools import generate_etrecord
+from executorch.exir import (
+    EdgeCompileConfig,
+    EdgeProgramManager,
+    ExecutorchProgramManager,
+    to_edge,
+)
 from torch.ao.quantization.pt2e.export_utils import model_is_exported
 from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
 
@@ -197,11 +204,12 @@ def export_to_edge(
 # Export the model and lower it to an EdgeProgramManager (in edge IR), and
 # apply passes specific to Cadence DSP execution. Return both to print the
 # differences.
-def export_to_cadence(
+def export_to_cadence_edge_executorch(
     model: torch.nn.Module,
     inputs: tuple[object, ...],
     dump_graphs: bool = False,
-) -> EdgeProgramManager:
+    output_dir: Optional[str] = None,
+) -> ExecutorchProgramManager:
     edge_prog_manager = export_to_edge(model, inputs)
 
     # Run a couple required passes for quant/dequant ops
@@ -225,4 +233,29 @@ def export_to_cadence(
         cadence_prog_manager.exported_program().graph_module,
     )
 
-    return cadence_prog_manager
+    # Get executorch program after Cadence specific passes
+    exec_prog: ExecutorchProgramManager = cadence_prog_manager.to_executorch()
+    if output_dir:
+        _gen_etrecord(edge_prog_manager, exec_prog, Path(output_dir))
+    else:
+        logging.warning("No output directory provided, skipping ETRecord generation")
+
+    return exec_prog
+
+
+def _gen_etrecord(
+    edge_program: EdgeProgramManager,
+    et_program: ExecutorchProgramManager,
+    output_dir: Path,
+) -> None:
+    etrec_path = output_dir / "etrecord.bin"
+    try:
+        generate_etrecord(
+            et_record=etrec_path,
+            edge_dialect_program=edge_program,
+            executorch_program=et_program,
+        )
+        logging.info(f"Generated ETRecord at {etrec_path}")
+    except Exception:
+        # Any errors here shouldn't block the rest of the flow
+        logging.exception("Encountered exception while generating ETRecord")

backends/cadence/aot/export_example.py

@@ -14,7 +14,7 @@
 
 from executorch.backends.cadence.aot.compiler import (
     convert_pt2,
-    export_to_cadence,
+    export_to_cadence_edge_executorch,
     fuse_pt2,
 )
 from executorch.backends.cadence.aot.quantizer.quantizer import CadenceQuantizer
@@ -53,10 +53,9 @@ def export_model(
     quantized_model = fuse_pt2(converted_model, quantizer)
 
     # Get edge program after Cadence specific passes
-    cadence_prog_manager = export_to_cadence(quantized_model, example_inputs)
-
-    # Get executorch program after Cadence specific passes
-    exec_prog: ExecutorchProgramManager = cadence_prog_manager.to_executorch()
+    exec_prog: ExecutorchProgramManager = export_to_cadence_edge_executorch(
+        quantized_model, example_inputs, working_dir
+    )
 
     logging.info("Final exported graph:\n")
     exec_prog.exported_program().graph_module.graph.print_tabular()

backends/cadence/build_cadence_runner.sh

@@ -25,6 +25,7 @@ main() {
     -DCMAKE_BUILD_TYPE=Release \
     -DEXECUTORCH_BUILD_DEVTOOLS=ON \
     -DEXECUTORCH_ENABLE_EVENT_TRACER=ON \
+    -DEXECUTORCH_ENABLE_LOGGING=ON \
     -Bcmake-out .
   cmake --build cmake-out --target install --config Release -j16
 
@@ -35,6 +36,7 @@ main() {
   cmake -DCMAKE_PREFIX_PATH="${cmake_prefix_path}" \
     -DCMAKE_BUILD_TYPE=Release \
     -DEXECUTORCH_CADENCE_CPU_RUNNER=ON \
+    -DEXECUTORCH_ENABLE_LOGGING=ON \
     -B"${build_dir}" \
     "${example_dir}"
   cmake --build "${build_dir}" --config Release -j16

backends/cadence/reference/operators/CMakeLists.txt

@@ -27,6 +27,7 @@ set(_aten_ops__srcs
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/activation_ops_util.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/copy_ops_util.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/broadcast_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/dtype_util.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/index_util.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/kernel_ops_util.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/matmul_ops_util.cpp"

backends/cadence/runtime/executor.py

@@ -123,6 +123,7 @@ def __call__(self) -> None:
             ),
             "etdump_path": os.path.join(self.working_dir, "etdump.etdp"),
             "debug_output_path": os.path.join(self.working_dir, "debug_output.bin"),
+            "dump_outputs": "true",
         }
         args = self.get_bash_command(self.execute_runner, cmd_args)
         logging.info(f"\33[33m{' '.join(args)}\33[0m")