Update base for Update on "[ET-VK] Minor tuning for conv2d pw op to improve performance."

The diff introduces minor tuning for the Conv2d pointwise (PW) operation in the Vulkan backend to improve performance.
Conv 2d pw now issues a 2D dispatch instead of 1D, where dispatch axis y is now sized based on output texture's batch size.

Differential Revision: [D75251145](https://our.internmc.facebook.com/intern/diff/D75251145/)

[ghstack-poisoned]

Author: trivedivivek

.github/workflows/build-presets.yml

@@ -20,7 +20,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        preset: [macos-arm64, pybind, llm]
+        preset: [macos, ios, ios-simulator, pybind, llm]
     with:
       job-name: build
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
@@ -39,7 +39,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        preset: [pybind, llm]
+        preset: [linux, pybind, llm]
         runner: [linux.2xlarge, linux.arm64.2xlarge]
         docker-image: [executorch-ubuntu-22.04-clang12, executorch-ubuntu-22.04-gcc11-aarch64]
         # Excluding specific runner + docker image combinations that don't make sense:

.lintrunner.toml

@@ -390,7 +390,6 @@ exclude_patterns = [
     "backends/arm/test/ops/**",
     "backends/vulkan/quantizer/**",
     "backends/vulkan/test/**",
-    "backends/cadence/aot/quantizer/**",
     "backends/qualcomm/quantizer/**",
     "examples/qualcomm/**",
     "backends/xnnpack/quantizer/**",

CMakePresets.json

@@ -7,13 +7,13 @@
       "binaryDir": "${sourceDir}/cmake-out"
     },
     {
-      "name": "macos-arm64",
-      "displayName": "Build everything buildable on macOS arm64",
+      "name": "macos",
+      "displayName": "Build everything buildable on macOS",
       "inherits": ["common"],
       "generator": "Xcode",
       "cacheVariables": {
         "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/third-party/ios-cmake/ios.toolchain.cmake",
-        "EXECUTORCH_BUILD_PRESET_FILE": "${sourceDir}/tools/cmake/preset/macos-arm64.cmake",
+        "EXECUTORCH_BUILD_PRESET_FILE": "${sourceDir}/tools/cmake/preset/macos.cmake",
         "PLATFORM": "MAC_ARM64",
         "DEPLOYMENT_TARGET": "10.15"
       },
@@ -23,6 +23,54 @@
         "rhs": "Darwin"
       }
     },
+    {
+      "name": "ios",
+      "displayName": "Build everything buildable on iOS",
+      "inherits": ["common"],
+      "generator": "Xcode",
+      "cacheVariables": {
+        "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/third-party/ios-cmake/ios.toolchain.cmake",
+        "EXECUTORCH_BUILD_PRESET_FILE": "${sourceDir}/tools/cmake/preset/ios.cmake",
+        "PLATFORM": "OS64",
+        "DEPLOYMENT_TARGET": "17.0"
+      },
+      "condition": {
+        "lhs": "${hostSystemName}",
+        "type": "equals",
+        "rhs": "Darwin"
+      }
+    },
+    {
+      "name": "ios-simulator",
+      "displayName": "Build everything buildable on iOS simulator",
+      "inherits": ["common"],
+      "generator": "Xcode",
+      "cacheVariables": {
+        "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/third-party/ios-cmake/ios.toolchain.cmake",
+        "EXECUTORCH_BUILD_PRESET_FILE": "${sourceDir}/tools/cmake/preset/ios.cmake",
+        "PLATFORM": "SIMULATORARM64",
+        "DEPLOYMENT_TARGET": "17.0"
+      },
+      "condition": {
+        "lhs": "${hostSystemName}",
+        "type": "equals",
+        "rhs": "Darwin"
+      }
+    },
+    {
+      "name": "linux",
+      "displayName": "Build everything buildable on Linux",
+      "inherits": ["common"],
+      "cacheVariables": {
+        "CMAKE_SYSTEM_NAME": "Linux",
+        "EXECUTORCH_BUILD_PRESET_FILE": "${sourceDir}/tools/cmake/preset/linux.cmake"
+      },
+      "condition": {
+        "lhs": "${hostSystemName}",
+        "type": "equals",
+        "rhs": "Linux"
+      }
+    },
     {
       "name": "pybind",
       "displayName": "Build pybindings exported in the wheel",

backends/cadence/aot/compiler.py

@@ -123,7 +123,7 @@ def prepare_and_convert_pt2(
     assert isinstance(model_gm, torch.fx.GraphModule)
 
     # Prepare
-    prepared_model = prepare_pt2e(model_gm, quantizer)  # pyre-ignore[6]
+    prepared_model = prepare_pt2e(model_gm, quantizer)
 
     # Calibrate
     # If no calibration data is provided, use the inputs

backends/cadence/aot/quantizer/TARGETS

@@ -9,6 +9,7 @@ python_library(
     ],
     deps = [
         "//caffe2:torch",
+        "//pytorch/ao:torchao",
     ],
 )
 
@@ -34,7 +35,6 @@ python_library(
         ":patterns",
         ":utils",
         "//caffe2:torch",
-        "//executorch/backends/xnnpack/quantizer:xnnpack_quantizer_utils",
     ],
 )
 

backends/cadence/aot/quantizer/patterns.py

@@ -15,7 +15,7 @@
 
 from torch import fx
 from torch._ops import OpOverload
-from torch.ao.quantization.quantizer import (
+from torchao.quantization.pt2e.quantizer import (
     DerivedQuantizationSpec,
     SharedQuantizationSpec,
 )

backends/cadence/aot/quantizer/quantizer.py

@@ -29,19 +29,20 @@
     is_annotated,
     no_outside_users,
 )
-from executorch.backends.xnnpack.quantizer.xnnpack_quantizer_utils import (
+
+from torch import fx
+
+from torchao.quantization.pt2e import HistogramObserver, MinMaxObserver
+from torchao.quantization.pt2e.quantizer import (
+    ComposableQuantizer,
+    DerivedQuantizationSpec,
     OperatorConfig,
     QuantizationAnnotation,
     QuantizationConfig,
     QuantizationSpec,
+    Quantizer,
 )
 
-from torch import fx
-
-from torch.ao.quantization.observer import HistogramObserver, MinMaxObserver
-from torch.ao.quantization.quantizer import DerivedQuantizationSpec, Quantizer
-from torch.ao.quantization.quantizer.composable_quantizer import ComposableQuantizer
-
 
 act_qspec_asym8s = QuantizationSpec(
     dtype=torch.int8,

backends/cadence/aot/quantizer/utils.py

@@ -14,13 +14,13 @@
 import torch
 from torch import fx
 from torch._ops import OpOverload
-from torch.ao.quantization import ObserverOrFakeQuantize
 
 from torch.fx import GraphModule
 from torch.fx.passes.utils.source_matcher_utils import (
     check_subgraphs_connected,
     SourcePartition,
 )
+from torchao.quantization.pt2e import ObserverOrFakeQuantize
 
 
 def quantize_tensor_multiplier(

backends/cadence/aot/remove_ops.py

@@ -235,10 +235,7 @@ def call_operator(
         kwargs: dict[str, Argument],
         meta: NodeMetadata,
     ) -> ProxyValue:
-        if op not in {
-            exir_ops.edge.aten.linalg_vector_norm.default,
-            exir_ops.edge.cadence.linalg_vector_norm.default,
-        }:
+        if op is not exir_ops.edge.aten.linalg_vector_norm.default:
             return super().call_operator(op, args, kwargs, meta)
 
         # If the op has three args or less, it can't be a nop

backends/cadence/aot/tests/test_remove_ops_passes.py

@@ -467,10 +467,7 @@ def forward(self, x: torch.Tensor):
 
         # Expect the linalg_vector_norm op to be removed by the pass
         self.assertEqual(
-            count_node(graph_module, exir_ops.edge.aten.linalg_vector_norm.default)
-            + count_node(
-                graph_module, exir_ops.edge.cadence.linalg_vector_norm.default
-            ),
+            count_node(graph_module, exir_ops.edge.aten.linalg_vector_norm.default),
             0,
         )