Update on "[ET-VK] Store unique ptr to Tensor in Value instead of inlined tensor object, to reduce Value struct size from 448 to 80 bytes."

This diff aims to reduce the size of the Value struct in the Executorch Vulkan runtime by storing a unique pointer to the Tensor object instead of an inlined tensor object. This change reduces the size of the Value struct from 448 bytes to 80 bytes, which can improve performance and reduce memory usage.

Differential Revision: [D66655991](https://our.internmc.facebook.com/intern/diff/D66655991/)

[ghstack-poisoned]

Author: trivedivivek

.github/scripts/extract_benchmark_results.py

@@ -310,6 +310,7 @@ def transform(
     workflow_run_attempt: int,
     job_name: str,
     job_id: int,
+    schema_version: str,
 ) -> List:
     """
     Transform the benchmark results into the format writable into the benchmark database
@@ -319,45 +320,91 @@ def transform(
     for r in benchmark_results:
         r["deviceInfo"]["device"] = job_name
 
-    # TODO (huydhn): This is the current schema of the database oss_ci_benchmark_v2,
-    # and I'm trying to fit ET benchmark results into it, which is kind of awkward.
-    # However, the schema is going to be updated soon
-    return [
-        {
-            # GH-info to identify where the benchmark is run
-            "repo": repo,
-            "head_branch": head_branch,
-            "workflow_id": workflow_run_id,
-            "run_attempt": workflow_run_attempt,
-            "job_id": job_id,
-            # The model
-            "name": f"{r['benchmarkModel']['name']} {r['benchmarkModel'].get('backend', '')}".strip(),
-            "dtype": (
-                r["benchmarkModel"]["quantization"]
-                if r["benchmarkModel"]["quantization"]
-                else "unknown"
-            ),
-            # The metric value
-            "metric": r["metric"],
-            "actual": r["actualValue"],
-            "target": r["targetValue"],
-            # The device
-            "device": r["deviceInfo"]["device"],
-            "arch": r["deviceInfo"].get("os", ""),
-            # Not used here, just set it to something unique here
-            "filename": workflow_name,
-            "test_name": app_type,
-            "runner": job_name,
-        }
-        for r in benchmark_results
-    ]
+    if schema_version == "v2":
+        # TODO (huydhn): Clean up this branch after ExecuTorch dashboard migrates to v3
+        return [
+            {
+                # GH-info to identify where the benchmark is run
+                "repo": repo,
+                "head_branch": head_branch,
+                "workflow_id": workflow_run_id,
+                "run_attempt": workflow_run_attempt,
+                "job_id": job_id,
+                # The model
+                "name": f"{r['benchmarkModel']['name']} {r['benchmarkModel'].get('backend', '')}".strip(),
+                "dtype": (
+                    r["benchmarkModel"]["quantization"]
+                    if r["benchmarkModel"]["quantization"]
+                    else "unknown"
+                ),
+                # The metric value
+                "metric": r["metric"],
+                "actual": r["actualValue"],
+                "target": r["targetValue"],
+                # The device
+                "device": r["deviceInfo"]["device"],
+                "arch": r["deviceInfo"].get("os", ""),
+                # Not used here, just set it to something unique here
+                "filename": workflow_name,
+                "test_name": app_type,
+                "runner": job_name,
+            }
+            for r in benchmark_results
+        ]
+    elif schema_version == "v3":
+        quantization = (
+            r["benchmarkModel"]["quantization"]
+            if r["benchmarkModel"]["quantization"]
+            else "unknown"
+        )
+        # From https://github.com/pytorch/pytorch/wiki/How-to-integrate-with-PyTorch-OSS-benchmark-database
+        return [
+            {
+                "benchmark": {
+                    "name": "ExecuTorch",
+                    "mode": "inference",
+                    "dtype": quantization,
+                    "extra_info": {
+                        "app_type": app_type,
+                    },
+                },
+                "model": {
+                    "name": r["benchmarkModel"]["name"],
+                    "type": "OSS model",
+                    "backend": r["benchmarkModel"].get("backend", ""),
+                    "extra_info": {
+                        "quantization": quantization,
+                    },
+                },
+                "metric": {
+                    "name": r["metric"],
+                    "benchmark_values": [r["actualValue"]],
+                    "target_value": r["targetValue"],
+                    "extra_info": {
+                        "method": r.get("method", ""),
+                    },
+                },
+                "runners": [
+                    {
+                        "name": r["deviceInfo"]["device"],
+                        "type": r["deviceInfo"]["os"],
+                        "avail_mem_in_gb": r["deviceInfo"].get("availMem", ""),
+                        "total_mem_in_gb": r["deviceInfo"].get("totalMem", ""),
+                    }
+                ],
+            }
+            for r in benchmark_results
+        ]
 
 
 def main() -> None:
     args = parse_args()
 
-    # Across all devices
-    all_benchmark_results = []
+    # Across all devices, keeping both schemas for now until ExecuTorch dashboard migrates to v3
+    all_benchmark_results = {
+        "v2": [],
+        "v3": [],
+    }
 
     with open(args.artifacts) as f:
         for artifact in json.load(f):
@@ -384,23 +431,31 @@ def main() -> None:
                 )
 
             if benchmark_results:
-                benchmark_results = transform(
-                    app_type,
-                    benchmark_results,
-                    args.repo,
-                    args.head_branch,
-                    args.workflow_name,
-                    args.workflow_run_id,
-                    args.workflow_run_attempt,
-                    job_name,
-                    extract_job_id(args.artifacts),
-                )
-                all_benchmark_results.extend(benchmark_results)
+                for schema in all_benchmark_results.keys():
+                    results = transform(
+                        app_type,
+                        benchmark_results,
+                        args.repo,
+                        args.head_branch,
+                        args.workflow_name,
+                        args.workflow_run_id,
+                        args.workflow_run_attempt,
+                        job_name,
+                        extract_job_id(args.artifacts),
+                        schema,
+                    )
+                    all_benchmark_results[schema].extend(results)
+
+    for schema in all_benchmark_results.keys():
+        if not all_benchmark_results.get(schema):
+            continue
+
+        output_dir = os.path.join(args.output_dir, schema)
+        os.mkdir(output_dir)
 
-    if all_benchmark_results:
         output_file = os.path.basename(args.artifacts)
-        with open(f"{args.output_dir}/{output_file}", "w") as f:
-            json.dump(all_benchmark_results, f)
+        with open(f"{output_dir}/{output_file}", "w") as f:
+            json.dump(all_benchmark_results[schema], f)
 
 
 if __name__ == "__main__":

.github/workflows/android-perf.yml

@@ -298,15 +298,25 @@ jobs:
               --workflow-run-attempt ${{ github.run_attempt }}
           done
 
-          ls -lah benchmark-results
-
-          for BENCHMARK_RESULTS in benchmark-results/*.json; do
-            cat "${BENCHMARK_RESULTS}"
-            echo
+          for SCHEMA in v2 v3; do
+            for BENCHMARK_RESULTS in benchmark-results/"${SCHEMA}"/*.json; do
+              cat "${BENCHMARK_RESULTS}"
+              echo
+            done
           done
 
-      - name: Upload the benchmark results
+      # TODO (huydhn): Remove v2 schema once the benchmark dashboard finishes the migration
+      - name: Upload the benchmark results (v2)
+        uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
+        with:
+          benchmark-results-dir: benchmark-results/v2
+          dry-run: false
+          schema-version: v2
+
+      - name: Upload the benchmark results (v3)
         uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
         with:
-          benchmark-results-dir: 'benchmark-results'
+          benchmark-results-dir: benchmark-results/v3
           dry-run: false
+          schema-version: v3
+          github-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/apple-perf.yml

@@ -372,15 +372,25 @@ jobs:
               --workflow-run-attempt ${{ github.run_attempt }}
           done
 
-          ls -lah benchmark-results
-
-          for BENCHMARK_RESULTS in benchmark-results/*.json; do
-            cat "${BENCHMARK_RESULTS}"
-            echo
+          for SCHEMA in v2 v3; do
+            for BENCHMARK_RESULTS in benchmark-results/"${SCHEMA}"/*.json; do
+              cat "${BENCHMARK_RESULTS}"
+              echo
+            done
           done
 
-      - name: Upload the benchmark results
+      # TODO (huydhn): Remove v2 schema once the benchmark dashboard finishes the migration
+      - name: Upload the benchmark results (v2)
+        uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
+        with:
+          benchmark-results-dir: benchmark-results/v2
+          dry-run: false
+          schema-version: v2
+
+      - name: Upload the benchmark results (v3)
         uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
         with:
-          benchmark-results-dir: 'benchmark-results'
+          benchmark-results-dir: benchmark-results/v3
           dry-run: false
+          schema-version: v3
+          github-token: ${{ secrets.GITHUB_TOKEN }}

CMakeLists.txt

@@ -56,6 +56,21 @@ if(NOT CMAKE_BUILD_TYPE)
   set(CMAKE_BUILD_TYPE Debug)
 endif()
 
+# Setup RPATH.
+# See https://gitlab.kitware.com/cmake/community/-/wikis/doc/cmake/RPATH-handling
+# Use separate rpaths during build and install phases
+set(CMAKE_SKIP_BUILD_RPATH OFF)
+# Don't use the install-rpath during the build phase
+set(CMAKE_BUILD_WITH_INSTALL_RPATH ON)
+# Automatically add all linked folders that are NOT in the build directory to
+# the rpath (per library?)
+# TODO: Doesn't work for us right now because we are not installing .so's into the
+# correct locations. For example we have libcustom_ops_aot_lib.so depending on
+# _portable_lib.so, which was eventually put under <site-packages>/executorch/extension/pybindings/
+# but this rpath is not automatically added because at build time it seems `portable_lib`
+# is being built under the same directory, so no extra rpath is being added. To
+# properly fix this we need to install `portable_lib` into the correct path.
+set(CMAKE_INSTALL_RPATH_USE_LINK_PATH ON)
 # ------------------------------ OPTIONS -------------------------------------
 # WARNING: Please don't add example specific options in this CMakeLists.txt.
 # Instead please use `find_package(executorch REQUIRED)` in the example
@@ -682,22 +697,6 @@ if(EXECUTORCH_BUILD_PTHREADPOOL
 endif()
 
 if(EXECUTORCH_BUILD_PYBIND)
-  # Setup RPATH.
-  # See https://gitlab.kitware.com/cmake/community/-/wikis/doc/cmake/RPATH-handling
-  if(APPLE)
-    set(CMAKE_MACOSX_RPATH ON)
-    set(_rpath_portable_origin "@loader_path")
-  else()
-    set(_rpath_portable_origin $ORIGIN)
-  endif(APPLE)
-  # Use separate rpaths during build and install phases
-  set(CMAKE_SKIP_BUILD_RPATH  FALSE)
-  # Don't use the install-rpath during the build phase
-  set(CMAKE_BUILD_WITH_INSTALL_RPATH FALSE)
-  set(CMAKE_INSTALL_RPATH "${_rpath_portable_origin}")
-  # Automatically add all linked folders that are NOT in the build directory to
-  # the rpath (per library?)
-  set(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/third-party/pybind11)
 
   if(NOT EXECUTORCH_BUILD_EXTENSION_DATA_LOADER)

backends/cadence/aot/compiler.py

@@ -28,6 +28,7 @@
     to_edge,
 )
 from executorch.exir.pass_base import PassResult
+from torch._inductor.decomposition import remove_decompositions
 from torch.ao.quantization.pt2e.export_utils import model_is_exported
 from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
 
@@ -58,16 +59,33 @@ def convert_pt2(
     Returns a GraphModule with the converted model.
     """
 
+    # Get default decompositions
+    decomp_table = torch.export.default_decompositions()
+    # Select ops to keep
+    ops_to_keep = [
+        torch.ops.aten.conv1d.default,
+        torch.ops.aten.conv2d.default,
+        torch.ops.aten.layer_norm.default,
+        torch.ops.aten.linear.default,
+        torch.ops.aten.matmul.default,
+    ]
+    # Remove decompositions for the ops we want to keep
+    # pyre-fixme[6]: For 1st argument expected `Dict[typing.Callable[..., typing.Any
+    remove_decompositions(decomp_table, ops_to_keep)
     # Export with dynamo
-    model_gm = torch.export.export_for_training(model, inputs).module()
+    model_gm = (
+        torch.export.export_for_training(model, inputs)
+        .run_decompositions(decomp_table)
+        .module()
+    )
 
-    if model_gm_has_SDPA(model_gm):  # pyre-fixme[6]
+    if model_gm_has_SDPA(model_gm):
         # Decompose SDPA
-        DecomposeScaledDotProductAttention(False)(model_gm)  # pyre-fixme[6]
+        DecomposeScaledDotProductAttention(False)(model_gm)
 
         # Swap _safe_softmax with _softmax (see https://github.com/pytorch/pytorch/pull/133882
         # for details).
-        result = ReplaceSafeSoftmaxWithSoftmax()(model_gm)  # pyre-fixme[6]
+        result = ReplaceSafeSoftmaxWithSoftmax()(model_gm)
         assert result is not None
         model_gm = result.graph_module
 

backends/cadence/hifi/kernels/kernels.h

@@ -92,9 +92,9 @@ extern "C" WORD32 xa_nn_elm_mul_broadcast_4D_f32xf32_f32(
     const WORD32* const p_inp2_shape);
 
 extern "C" void xa_nn_elm_pow_f32(
-    FLOAT32* restrict z,
-    const FLOAT32* restrict x,
-    const FLOAT32* restrict y,
+    FLOAT32* __restrict__ z,
+    const FLOAT32* __restrict__ x,
+    const FLOAT32* __restrict__ y,
     WORD32 N);
 
 extern "C" WORD32 xa_nn_elm_where_f32xf32_f32(

backends/cadence/hifi/third-party/nnlib/xa_nn_elm_minimum_maximum_f32.c

@@ -19,12 +19,12 @@
 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 
 ******************************************************************************/
-#include "nnlib-hifi4/xa_nnlib/include/xa_type_def.h"
-#include "nnlib-hifi4/xa_nnlib/algo/common/include/xa_nnlib_common_fpu.h"
-#include "nnlib-hifi4/xa_nnlib/algo/common/include/xa_nn_common.h"
-#include "nnlib-hifi4/xa_nnlib/algo/common/include/xa_nnlib_err_chk.h"
-#include "nnlib-hifi4/xa_nnlib/algo/kernels/basic/hifi4/xa_nn_basic_state.h"
-#include "nnlib-hifi4/xa_nnlib/include/nnlib/xa_nnlib_kernels_api.h"
+#include "xa_type_def.h"
+#include "xa_nnlib_common_fpu.h"
+#include "xa_nn_common.h"
+#include "xa_nnlib_err_chk.h"
+// #include "xa_nn_basic_state.h"
+#include "xa_nnlib_kernels_api.h"
 
 #if !HAVE_VFPU
 DISCARD_FUN_FOR_NONVOID_RETURN(
@@ -844,4 +844,4 @@ WORD32 xa_nn_elm_minimum_broadcast_4D_f32xf32_f32(FLOAT32 * __restrict__ p_out,
   }
   return 0;
 }
-#endif
+#endif

backends/test/README.md

@@ -0,0 +1,8 @@
+# Any targets that should be shared between fbcode and xplat must be defined in
+# targets.bzl. This file can contain fbcode-only targets.
+
+load(":targets.bzl", "define_common_targets")
+
+oncall("executorch")
+
+define_common_targets(is_fbcode = True)