Update

[ghstack-poisoned]

Author: GregoryComer

.ci/scripts/setup-vulkan-linux-deps.sh

@@ -23,6 +23,7 @@ install_swiftshader() {
 
   export VK_ICD_FILENAMES="${_swiftshader_dir}/swiftshader/build/Linux/vk_swiftshader_icd.json"
   export LD_LIBRARY_PATH="${_swiftshader_dir}/swiftshader/build/Linux/"
+  export ETVK_USING_SWIFTSHADER=1
 }
 
 install_vulkan_sdk() {

.github/workflows/pull.yml

@@ -864,6 +864,43 @@ jobs:
         PYTHON_EXECUTABLE=python bash examples/nxp/run_aot_example.sh
 
 
+  test-vulkan-models-linux:
+    name: test-vulkan-models-linux
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    with:
+      runner: linux.2xlarge
+      docker-image: ci-image:executorch-ubuntu-22.04-clang12
+      submodules: 'recursive'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
+      script: |
+        set -eux
+
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        # Setup swiftshader and Vulkan SDK which are required to build the Vulkan delegate
+        source .ci/scripts/setup-vulkan-linux-deps.sh
+
+        # Setup python
+        PYTHON_EXECUTABLE=python \
+        CMAKE_ARGS="-DEXECUTORCH_BUILD_VULKAN=ON" \
+        .ci/scripts/setup-linux.sh --build-tool "cmake"
+
+        PYTHON_EXECUTABLE=python bash backends/vulkan/test/scripts/test_model.sh --build
+
+        # Test models serially
+        models="mv2 mv3 edsr resnet18 resnet50 dl3"
+        for model in $models; do
+          python -m examples.vulkan.export --model_name=$model --test
+        done
+
+
+
   nxp-build-test:
     name: nxp-build-test
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main

.gitignore

@@ -24,6 +24,7 @@ pip-out/
 # Any exported models and profiling outputs
 *.bin
 *.model
+*.etdump
 tokenizer.json
 *.pte
 *.ptd
@@ -58,6 +59,7 @@ xcuserdata/
 /include/
 /share/
 /version.py
+*.csv
 
 # Android
 *.aar

CMakeLists.txt

@@ -757,16 +757,16 @@ if(EXECUTORCH_BUILD_PYBIND)
     list(APPEND _dep_libs openvino_backend)
   endif()
 
-  if(EXECUTORCH_BUILD_VULKAN)
-    list(APPEND _dep_libs vulkan_backend)
-  endif()
-
   if(EXECUTORCH_BUILD_XNNPACK)
     # need to explicitly specify XNNPACK and xnnpack-microkernels-prod here
     # otherwise uses XNNPACK and microkernel-prod symbols from libtorch_cpu
     list(APPEND _dep_libs xnnpack_backend XNNPACK xnnpack-microkernels-prod)
   endif()
 
+  if(EXECUTORCH_BUILD_VULKAN)
+    list(APPEND _dep_libs vulkan_backend)
+  endif()
+
   # compile options for pybind
   set(_pybind_compile_options -Wno-deprecated-declarations -fPIC -frtti
                               -fexceptions

backends/vulkan/CMakeLists.txt

@@ -101,7 +101,7 @@ set_target_properties(vulkan_schema PROPERTIES LINKER_LANGUAGE CXX)
 target_include_directories(
   vulkan_schema
   INTERFACE
-    ${SCHEMA_INCLUDE_DIR}
+    $<BUILD_INTERFACE:${SCHEMA_INCLUDE_DIR}>
     $<BUILD_INTERFACE:${EXECUTORCH_ROOT}/third-party/flatbuffers/include>
 )
 

backends/vulkan/cmake/ShaderLibrary.cmake

@@ -49,6 +49,15 @@ function(gen_vulkan_shader_lib_cpp shaders_path)
   set(VULKAN_SHADERGEN_ENV "")
   set(VULKAN_SHADERGEN_OUT_PATH ${CMAKE_BINARY_DIR}/vulkan_compute_shaders)
 
+  set(GEN_SPV_ARGS "--optimize")
+  if(DEFINED ENV{ETVK_USING_SWIFTSHADER})
+    if("$ENV{ETVK_USING_SWIFTSHADER}" STREQUAL "1"
+       OR "$ENV{ETVK_USING_SWIFTSHADER}" STREQUAL "True"
+    )
+      list(APPEND GEN_SPV_ARGS "--replace-u16vecn")
+    endif()
+  endif()
+
   add_custom_command(
     COMMENT "Generating Vulkan Compute Shaders"
     OUTPUT ${VULKAN_SHADERGEN_OUT_PATH}/spv.cpp
@@ -58,7 +67,7 @@ function(gen_vulkan_shader_lib_cpp shaders_path)
       ${shaders_path} --output-path ${VULKAN_SHADERGEN_OUT_PATH}
       --glslc-path=${GLSLC_PATH}
       --tmp-dir-path=${VULKAN_SHADERGEN_OUT_PATH}/shader_cache/ --env
-      ${VULKAN_GEN_ARG_ENV} --optimize
+      ${VULKAN_GEN_ARG_ENV} ${GEN_SPV_ARGS}
     DEPENDS ${shaders_path}/*
             ${EXECUTORCH_ROOT}/backends/vulkan/runtime/gen_vulkan_spv.py
   )

backends/vulkan/partitioner/vulkan_partitioner.py

@@ -7,7 +7,7 @@
 # pyre-strict
 
 import logging
-from typing import Any, Callable, Dict, final, List, Mapping, Optional, Tuple
+from typing import Any, Callable, Dict, final, List, Mapping, Optional, Set, Tuple
 
 import executorch.backends.vulkan.utils as utils
 
@@ -17,6 +17,7 @@
     get_op_features,
     has_impl,
     OpFeatures,
+    OpKey,
     vulkan_supported_ops,
 )
 
@@ -55,11 +56,17 @@ def __init__(
         texture_limits: utils.ImageExtents,
         buffer_limit: int,
         require_dynamic_shape: bool = False,
+        operator_blocklist: Optional[Set[OpKey]] = None,
+        operator_allowlist: Optional[Set[OpKey]] = None,
     ) -> None:
         super().__init__()
         self.texture_limits: utils.ImageExtents = texture_limits
         self.buffer_limit = buffer_limit
         self.require_dynamic_shapes = require_dynamic_shape
+        self.operator_blocklist: Set[OpKey] = (
+            operator_blocklist if operator_blocklist is not None else set()
+        )
+        self.operator_allowlist = operator_allowlist
 
     def op_node_is_compatible(  # noqa: C901: Function is too complex
         self, node: torch.fx.Node, features: Optional[OpFeatures] = None
@@ -77,6 +84,17 @@ def op_node_is_compatible(  # noqa: C901: Function is too complex
             assert isinstance(first_arg, torch._ops.OpOverload)
             target = first_arg.name()
 
+        # Operator allow list is only used for torch ops
+        if (
+            utils.is_torch_op_node(node)
+            and (self.operator_allowlist is not None)
+            and (target not in self.operator_allowlist)
+        ):
+            return False, "op is not in allowlist"
+
+        if target in self.operator_blocklist:
+            return False, "op is in blocklist"
+
         # Extract the features for the node's operator, if no override was provided
         if features is None:
             if not has_impl(target):
@@ -93,7 +111,7 @@ def op_node_is_compatible(  # noqa: C901: Function is too complex
         if op_repsets.any_is_empty():
             return (
                 False,
-                "No valid representations for a tensor in the operation",
+                f"no valid representations for op {utils.node_io_str(node)}",
             )
 
         return True, "Op is compatible"
@@ -277,6 +295,8 @@ class VulkanPartitioner(Partitioner):
     def __init__(
         self,
         compile_options: Optional[Dict[str, Any]] = None,
+        operator_blocklist: Optional[List[OpKey]] = None,
+        operator_allowlist: Optional[List[OpKey]] = None,
     ) -> None:
         self.options: Dict[str, Any] = {}
         if compile_options is not None:
@@ -285,6 +305,18 @@ def __init__(
         compile_spec = parse_compile_options(self.options)
         self.delegation_spec = DelegationSpec(VulkanBackend.__name__, compile_spec)
 
+        self.operator_blocklist: Set[OpKey] = set()
+        if operator_blocklist is not None:
+            for entry in operator_blocklist or []:
+                self.operator_blocklist.add(entry)
+
+        self.operator_allowlist: Optional[Set[OpKey]] = None
+        if operator_allowlist is not None:
+            self.operator_allowlist = set()
+            for entry in operator_allowlist:
+                assert self.operator_allowlist is not None
+                self.operator_allowlist.add(entry)
+
     def ops_to_not_decompose(
         self, ep: ExportedProgram
     ) -> Tuple[List[torch._ops.OpOverload], Optional[Callable[[torch.fx.Node], bool]]]:
@@ -308,6 +340,8 @@ def partition(self, exported_program: ExportedProgram) -> PartitionResult:
                 texture_limits,
                 buffer_limit,
                 require_dynamic_shape=self.options.get("require_dynamic_shapes", False),
+                operator_blocklist=self.operator_blocklist,
+                operator_allowlist=self.operator_allowlist,
             ),
             allows_single_node_partition=True,
         )

backends/vulkan/runtime/gen_vulkan_spv.py

@@ -1083,7 +1083,6 @@ def compile_spirv(shader_paths_pair) -> Tuple[str, str]:
             for spv_out_path, glsl_out_path in pool.map(
                 compile_spirv, self.output_file_map.items()
             ):
-                print(spv_to_glsl_map)
                 spv_to_glsl_map[spv_out_path] = glsl_out_path
 
         return spv_to_glsl_map