Update on "[ET-VK] Adding a common utility function to calculate 3d output position based on unique index."

trivedivivek · trivedivivek · commit b6d4072413da · 2025-01-06T22:12:47.000-08:00
This diff adds an indexing utils header file used in Vulkan backend of Executorch. The header file includes functions for converting a global index to u16 indices based on input sizes. Differential Revision: [D67821941](https://our.internmc.facebook.com/intern/diff/D67821941/) [ghstack-poisoned]
diff --git a/.lintrunner.toml b/.lintrunner.toml
@@ -302,8 +302,8 @@ include_patterns = [
     'profiler/**/*.py',
     'runtime/**/*.py',
     'scripts/**/*.py',
-    # 'test/**/*.py',
-    # 'util/**/*.py',
+    'test/**/*.py',
+    'util/**/*.py',
     '*.py',
 ]
 exclude_patterns = [
diff --git a/.mypy.ini b/.mypy.ini
@@ -21,10 +21,14 @@ files =
     profiler,
     runtime,
     scripts,
+    test,
     util
 
 mypy_path = executorch
 
+[mypy-executorch.backends.*]
+follow_untyped_imports = True
+
 [mypy-executorch.codegen.*]
 follow_untyped_imports = True
 
@@ -46,6 +50,12 @@ follow_untyped_imports = True
 [mypy-executorch.runtime.*]
 follow_untyped_imports = True
 
+[mypy-executorch.test.*]
+follow_untyped_imports = True
+
+[mypy-functorch.*]
+follow_untyped_imports = True
+
 [mypy-requests.*]
 follow_untyped_imports = True
 
@@ -80,4 +90,4 @@ ignore_missing_imports = True
 ignore_missing_imports = True
 
 [mypy-zstd]
-ignore_missing_imports = True
+ignore_missing_imports = True
diff --git a/backends/cadence/fusion_g3/operators/op_add.cpp b/backends/cadence/fusion_g3/operators/op_add.cpp
@@ -10,6 +10,7 @@
 
 #include <xa_nnlib_kernels_api.h>
 
+#include <executorch/backends/cadence/fusion_g3/operators/xt_macros.h>
 #include <executorch/kernels/portable/cpu/scalar_utils.h>
 #include <executorch/kernels/portable/cpu/util/elementwise_util.h>
 #include <executorch/kernels/portable/cpu/util/kernel_ops_util.h>
@@ -28,15 +29,6 @@ namespace impl {
 namespace G3 {
 namespace native {
 
-#define XT_KERNEL_CHECK(ctx, out, kernel, ...) \
-  const auto ret = kernel(__VA_ARGS__);        \
-  ET_KERNEL_CHECK_MSG(                         \
-      ctx,                                     \
-      ret == 0,                                \
-      InvalidArgument,                         \
-      out,                                     \
-      "Failed to run kernel: " #kernel "(" #__VA_ARGS__ ")");
-
 Tensor& add_out(
     KernelRuntimeContext& ctx,
     const Tensor& a,
diff --git a/backends/cadence/fusion_g3/operators/targets.bzl b/backends/cadence/fusion_g3/operators/targets.bzl
@@ -27,6 +27,7 @@ def define_operator(name: str, deps: list[str] | None = None) -> None:
         deps = deps + common_deps,
         exported_deps = [
             ":operators_header",
+            ":xt_macros",
         ],
     )
 
@@ -61,5 +62,17 @@ def define_common_targets():
         ],
     )
 
+    runtime.cxx_library(
+        name = "xt_macros",
+        exported_headers = ["xt_macros.h"],
+        visibility = [
+            "//executorch/backends/cadence/...",
+        ],
+        exported_deps = [
+            "//executorch/runtime/core/exec_aten:lib",
+            "//executorch/runtime/kernel:kernel_runtime_context",
+        ],
+    )
+
     for op in OPERATORS:
         define_operator(op)
diff --git a/backends/cadence/fusion_g3/operators/xt_macros.h b/backends/cadence/fusion_g3/operators/xt_macros.h
@@ -0,0 +1,20 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#include <executorch/runtime/core/exec_aten/util/tensor_util.h>
+
+#define XT_KERNEL_CHECK(ctx, out, kernel, ...) \
+  const auto ret = kernel(__VA_ARGS__);        \
+  ET_KERNEL_CHECK_MSG(                         \
+      ctx,                                     \
+      ret == 0,                                \
+      InvalidArgument,                         \
+      out,                                     \
+      "Failed to run kernel: " #kernel "(" #__VA_ARGS__ ")");
diff --git a/docs/TARGETS b/docs/TARGETS
@@ -1,7 +1,7 @@
 load("@fbcode_macros//build_defs:native_rules.bzl", "buck_filegroup", "buck_sh_test")
 load("@fbcode_macros//build_defs:python_binary.bzl", "python_binary")
 
-oncall("pytorch_r2p")
+oncall("executorch")
 
 python_binary(
     name = "sphinx",
diff --git a/extension/flat_tensor/serialize/TARGETS b/extension/flat_tensor/serialize/TARGETS
@@ -0,0 +1,16 @@
+load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
+load(":targets.bzl", "define_common_targets")
+
+oncall("executorch")
+
+define_common_targets()
+
+runtime.python_library(
+    name = "schema",
+    srcs = [
+        "flat_tensor_schema.py",
+    ],
+    visibility = [
+        "//executorch/...",
+    ],
+)
diff --git a/extension/flat_tensor/serialize/flat_tensor.fbs b/extension/flat_tensor/serialize/flat_tensor.fbs
@@ -13,7 +13,7 @@ table TensorMetadata {
   scalar_type: executorch_flatbuffer.ScalarType;
 
   // Size of each dimension.
-  dim_sizes: [int32];
+  sizes: [int32];
 
   // Specifies in what order the dimensions are laid out in memory (from outer
   // to inner).
diff --git a/extension/flat_tensor/serialize/flat_tensor_schema.py b/extension/flat_tensor/serialize/flat_tensor_schema.py
@@ -17,7 +17,7 @@
 class TensorMetadata:
     fully_qualified_name: str
     scalar_type: ScalarType
-    dim_sizes: List[int]
+    sizes: List[int]
     dim_order: List[bytes]
 
     segment_index: int
diff --git a/extension/flat_tensor/serialize/scalar_type.fbs b/extension/flat_tensor/serialize/scalar_type.fbs
diff --git a/extension/flat_tensor/serialize/targets.bzl b/extension/flat_tensor/serialize/targets.bzl
@@ -0,0 +1,36 @@
+load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
+
+def define_common_targets():
+    runtime.genrule(
+        name = "gen_schema",
+        srcs = [
+            "flat_tensor.fbs",
+            "scalar_type.fbs",
+        ],
+        outs = {
+            "schema_generated.h": ["flat_tensor_generated.h"],
+            "scalar_type_generated.h": ["scalar_type_generated.h"]
+        },
+        cmd = " ".join([
+            "$(exe {})".format(runtime.external_dep_location("flatc")),
+            "--cpp",
+            "--cpp-std c++11",
+            "--scoped-enums",
+            "-o ${OUT}",
+            "${SRCS}",
+        ]),
+        default_outs = ["."],
+    )
+
+    runtime.cxx_library(
+        name = "generated_headers",
+        srcs = [],
+        visibility = [
+            "//executorch/...",
+        ],
+        exported_headers = {
+            "schema_generated.h": ":gen_schema[schema_generated.h]",
+            "scalar_type_generated.h": ":gen_schema[scalar_type_generated.h]",
+        },
+        exported_external_deps = ["flatbuffers-api"],
+    )
diff --git a/test/end2end/exported_module.py b/test/end2end/exported_module.py
@@ -126,9 +126,7 @@ def return_wrapper():
         trace_inputs_method = "get_upper_bound_inputs"
         get_trace_inputs = get_inputs_adapter(
             (
-                # pyre-fixme[6]: For 1st argument expected `(...) -> Any` but got
-                #  `Union[Module, Tensor]`.
-                getattr(eager_module, trace_inputs_method)
+                getattr(eager_module, trace_inputs_method)  # type: ignore[arg-type]
                 if hasattr(eager_module, trace_inputs_method)
                 else eager_module.get_random_inputs
             ),
@@ -144,18 +142,14 @@ def return_wrapper():
         if hasattr(eager_module, "get_dynamic_shapes"):
             assert capture_config is not None
             assert capture_config.enable_aot is True
-            # pyre-fixme[29]: `Union[nn.modules.module.Module,
-            #  torch._tensor.Tensor]` is not a function.
-            trace_dynamic_shapes = eager_module.get_dynamic_shapes()
+            trace_dynamic_shapes = eager_module.get_dynamic_shapes()  # type: ignore[operator]
             method_name_to_dynamic_shapes = {}
             for method in methods:
                 method_name_to_dynamic_shapes[method] = trace_dynamic_shapes
 
         memory_planning_pass = MemoryPlanningPass()
         if hasattr(eager_module, "get_memory_planning_pass"):
-            # pyre-fixme[29]: `Union[nn.modules.module.Module,
-            #  torch._tensor.Tensor]` is not a function.
-            memory_planning_pass = eager_module.get_memory_planning_pass()
+            memory_planning_pass = eager_module.get_memory_planning_pass()  # type: ignore[operator]
 
         class WrapperModule(nn.Module):
             def __init__(self, method):
@@ -172,7 +166,7 @@ def __init__(self, method):
                 assert method_name == "forward"
                 ep = _export(
                     eager_module,
-                    method_input,
+                    method_input,  # type: ignore[arg-type]
                     dynamic_shapes=(
                         method_name_to_dynamic_shapes[method_name]
                         if method_name_to_dynamic_shapes
@@ -184,7 +178,7 @@ def __init__(self, method):
             else:
                 exported_methods[method_name] = export(
                     eager_module,
-                    method_input,
+                    method_input,  # type: ignore[arg-type]
                     dynamic_shapes=(
                         method_name_to_dynamic_shapes[method_name]
                         if method_name_to_dynamic_shapes
@@ -220,9 +214,7 @@ def __init__(self, method):
 
         # Get a function that creates random inputs appropriate for testing.
         get_random_inputs_fn = get_inputs_adapter(
-            # pyre-fixme[6]: For 1st argument expected `(...) -> Any` but got
-            #  `Union[Module, Tensor]`.
-            eager_module.get_random_inputs,
+            eager_module.get_random_inputs,  # type: ignore[arg-type]
             # all exported methods must have the same signature so just pick the first one.
             methods[0],
         )
diff --git a/test/end2end/test_end2end.py b/test/end2end/test_end2end.py
@@ -52,9 +52,7 @@
 kernel_mode = None  # either aten mode or lean mode
 try:
     from executorch.extension.pybindings.portable_lib import (
-        _load_bundled_program_from_buffer,
         _load_for_executorch_from_buffer,
-        _load_for_executorch_from_bundled_program,
     )
 
     kernel_mode = "lean"
@@ -63,10 +61,8 @@
     pass
 
 try:
-    from executorch.extension.pybindings.aten_lib import (
-        _load_bundled_program_from_buffer,
+    from executorch.extension.pybindings.aten_lib import (  # type: ignore[import-not-found]
         _load_for_executorch_from_buffer,
-        _load_for_executorch_from_bundled_program,
     )
 
     assert kernel_mode is None
diff --git a/test/models/export_delegated_program.py b/test/models/export_delegated_program.py
@@ -118,9 +118,7 @@ def export_module_to_program(
     eager_module = module_class().eval()
     inputs = ()
     if hasattr(eager_module, "get_random_inputs"):
-        # pyre-fixme[29]: `Union[nn.modules.module.Module, torch._tensor.Tensor]` is
-        #  not a function.
-        inputs = eager_module.get_random_inputs()
+        inputs = eager_module.get_random_inputs()  # type: ignore[operator]
 
     class WrapperModule(torch.nn.Module):
         def __init__(self, fn):
@@ -153,7 +151,7 @@ def forward(self, *args, **kwargs):
         ).to_executorch(config=et_config)
     else:
         edge: exir.EdgeProgramManager = to_edge(exported_program)
-        lowered_module = to_backend(
+        lowered_module = to_backend(  # type: ignore[call-arg]
             backend_id, edge.exported_program(), compile_specs=[]
         )
 
diff --git a/test/models/generate_linear_out_bundled_program.py b/test/models/generate_linear_out_bundled_program.py
@@ -27,7 +27,9 @@
 from executorch.exir.passes import MemoryPlanningPass, ToOutVarPass
 from executorch.exir.print_program import pretty_print
 
-from executorch.test.models.linear_model import LinearModel
+from executorch.test.models.linear_model import (  # type: ignore[import-not-found]
+    LinearModel,
+)
 from torch.export import export
 
 
diff --git a/util/activation_memory_profiler.py b/util/activation_memory_profiler.py
@@ -9,7 +9,7 @@
 import json
 import typing
 from dataclasses import dataclass, field
-from typing import List
+from typing import Any, Dict, List, Optional
 
 import executorch.exir.memory as memory
 import torch
@@ -52,7 +52,7 @@ def create_tensor_allocation_info(graph: torch.fx.Graph) -> List[MemoryTimeline]
     allocations at that timestep.
     """
     nodes = graph.nodes
-    memory_timeline = [None] * len(nodes)
+    memory_timeline: List[Optional[MemoryTimeline]] = [None for _ in range(len(nodes))]
     for _, node in enumerate(nodes):
         if node.op == "output":
             continue
@@ -72,11 +72,11 @@ def create_tensor_allocation_info(graph: torch.fx.Graph) -> List[MemoryTimeline]
             stack_trace = node.meta.get("stack_trace")
             fqn = _get_module_hierarchy(node)
             for j in range(start, end + 1):
-                if memory_timeline[j] is None:
-                    # pyre-ignore
-                    memory_timeline[j] = MemoryTimeline()
-                # pyre-ignore
-                memory_timeline[j].allocations.append(
+                memory_timeline_j = memory_timeline[j]
+                if memory_timeline_j is None:
+                    memory_timeline_j = MemoryTimeline()
+                assert memory_timeline_j
+                memory_timeline_j.allocations.append(
                     Allocation(
                         node.name,
                         node.target,
@@ -87,8 +87,7 @@ def create_tensor_allocation_info(graph: torch.fx.Graph) -> List[MemoryTimeline]
                         stack_trace,
                     )
                 )
-    # pyre-ignore
-    return memory_timeline
+    return memory_timeline  # type: ignore[return-value]
 
 
 def _validate_memory_planning_is_done(exported_program: ExportedProgram):
@@ -129,7 +128,7 @@ def generate_memory_trace(
 
     memory_timeline = create_tensor_allocation_info(exported_program.graph)
     root = {}
-    trace_events = []
+    trace_events: List[Dict[str, Any]] = []
     root["traceEvents"] = trace_events
 
     tid = 0
@@ -138,7 +137,7 @@ def generate_memory_trace(
         if memory_timeline_event is None:
             continue
         for allocation in memory_timeline_event.allocations:
-            e = {}
+            e: Dict[str, Any] = {}
             e["name"] = allocation.name
             e["cat"] = "memory_allocation"
             e["ph"] = "X"
diff --git a/util/python_profiler.py b/util/python_profiler.py
