Update on "Skip annotate boolean input"

Differential Revision: [D55946526](https://our.internmc.facebook.com/intern/diff/D55946526/)

[ghstack-poisoned]

Author: cccclai

CMakeLists.txt

@@ -352,23 +352,27 @@ add_subdirectory(schema)
 # Only contains primitive operators; does not contain portable kernels or other
 # full operators. Does not contain any backends.
 #
-
-add_library(executorch ${_executorch__srcs})
-target_link_libraries(executorch PRIVATE program_schema)
-target_link_options_shared_lib(executorch)
+add_library(executorch_no_prim_ops ${_executorch_no_prim_ops__srcs})
+target_link_libraries(executorch_no_prim_ops PRIVATE program_schema)
 # Check if dl exists for this toolchain and only then link it.
 find_library(DL_LIBRARY_EXISTS NAMES dl)
 # Check if the library was found
 if(DL_LIBRARY_EXISTS)
-  target_link_libraries(executorch PRIVATE dl) # For dladdr()
+  target_link_libraries(executorch_no_prim_ops PRIVATE dl) # For dladdr()
 endif()
-target_include_directories(executorch PUBLIC ${_common_include_directories})
-target_compile_options(executorch PUBLIC ${_common_compile_options})
+target_include_directories(executorch_no_prim_ops PUBLIC ${_common_include_directories})
+target_compile_options(executorch_no_prim_ops PUBLIC ${_common_compile_options})
 if(MAX_KERNEL_NUM)
-  target_compile_definitions(executorch
+  target_compile_definitions(executorch_no_prim_ops
                              PRIVATE MAX_KERNEL_NUM=${MAX_KERNEL_NUM})
 endif()
 
+add_library(executorch ${_executorch__srcs})
+target_link_libraries(executorch PRIVATE executorch_no_prim_ops)
+target_include_directories(executorch PUBLIC ${_common_include_directories})
+target_compile_options(executorch PUBLIC ${_common_compile_options})
+target_link_options_shared_lib(executorch)
+
 #
 # portable_ops_lib: A library to register core ATen ops using portable kernels,
 # see kernels/portable/CMakeLists.txt.
@@ -406,7 +410,7 @@ endif()
 # Install `executorch` library as well as `executorch-config.cmake` under
 # ${CMAKE_INSTALL_PREFIX}/
 install(
-  TARGETS executorch
+  TARGETS executorch executorch_no_prim_ops
   DESTINATION lib
   INCLUDES
   DESTINATION ${_common_include_directories})

backends/apple/coreml/runtime/workspace/executorchcoreml.xcodeproj/project.pbxproj

@@ -100,6 +100,7 @@
 		C9E7D7952AB3F9BF00CCAE5D /* ETCoreMLModelManagerTests.mm in Sources */ = {isa = PBXBuildFile; fileRef = C9E7D78D2AB3F9BF00CCAE5D /* ETCoreMLModelManagerTests.mm */; };
 		C9E7D7962AB3F9BF00CCAE5D /* KeyValueStoreTests.mm in Sources */ = {isa = PBXBuildFile; fileRef = C9E7D78E2AB3F9BF00CCAE5D /* KeyValueStoreTests.mm */; };
 		C9E7D7A22AB3FBB200CCAE5D /* CoreMLBackendDelegateTests.mm in Sources */ = {isa = PBXBuildFile; fileRef = C9E7D7A12AB3FBB200CCAE5D /* CoreMLBackendDelegateTests.mm */; };
+		F24817E52BC655E100E80D98 /* libexecutorch_no_prim_ops.a in Frameworks */ = {isa = PBXBuildFile; fileRef = F24817E42BC655E100E80D98 /* libexecutorch_no_prim_ops.a */; };
 /* End PBXBuildFile section */
 
 /* Begin PBXCopyFilesBuildPhase section */
@@ -297,6 +298,7 @@
 		C9EA3DB22B71A2B200B7D7BD /* CoreML.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = CoreML.framework; path = System/Library/Frameworks/CoreML.framework; sourceTree = SDKROOT; };
 		C9EA3FDE2B73EEA000B7D7BD /* libsqlite3.tbd */ = {isa = PBXFileReference; lastKnownFileType = "sourcecode.text-based-dylib-definition"; name = libsqlite3.tbd; path = usr/lib/libsqlite3.tbd; sourceTree = SDKROOT; };
 		C9EA3FE52B73EF6300B7D7BD /* Accelerate.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Accelerate.framework; path = System/Library/Frameworks/Accelerate.framework; sourceTree = SDKROOT; };
+		F24817E42BC655E100E80D98 /* libexecutorch_no_prim_ops.a */ = {isa = PBXFileReference; lastKnownFileType = archive.ar; name = libexecutorch_no_prim_ops.a; path = ../libraries/libexecutorch_no_prim_ops.a; sourceTree = "<group>"; };
 /* End PBXFileReference section */
 
 /* Begin PBXFrameworksBuildPhase section */
@@ -305,6 +307,7 @@
 			buildActionMask = 2147483647;
 			files = (
 				C94D510F2ABDF87500AF47FD /* Accelerate.framework in Frameworks */,
+				F24817E52BC655E100E80D98 /* libexecutorch_no_prim_ops.a in Frameworks */,
 				C94D510E2ABDF86800AF47FD /* libsqlite3.tbd in Frameworks */,
 				C94D50D92ABD7B2400AF47FD /* CoreML.framework in Frameworks */,
 				C99883862B95AD7D000953A3 /* libprotobuf-lite.a in Frameworks */,
@@ -523,6 +526,7 @@
 				C96560942AABFDCE005F8126 /* libsqlite3.tbd */,
 				C96560922AABF992005F8126 /* CoreML.framework */,
 				C96560902AABF982005F8126 /* Accelerate.framework */,
+				F24817E42BC655E100E80D98 /* libexecutorch_no_prim_ops.a */,
 				C965608D2AABF72A005F8126 /* libexecutorch.a */,
 			);
 			name = "Recovered References";

backends/apple/coreml/scripts/build_tests.sh

@@ -59,6 +59,7 @@ cmake --build "$CMAKE_PROTOBUF_BUILD_DIR_PATH"  -j9 -t libprotobuf-lite
 echo "ExecuTorch: Copying libraries"
 mkdir "$LIBRARIES_DIR_PATH"
 cp -f "$CMAKE_EXECUTORCH_BUILD_DIR_PATH/libexecutorch.a" "$LIBRARIES_DIR_PATH"
+cp -f "$CMAKE_EXECUTORCH_BUILD_DIR_PATH/libexecutorch_no_prim_ops.a" "$LIBRARIES_DIR_PATH"
 cp -f "$CMAKE_PROTOBUF_BUILD_DIR_PATH/libprotobuf-lite.a" "$LIBRARIES_DIR_PATH"
 
 #Copy ExecuTorch headers

backends/qualcomm/quantizer/utils.py

@@ -42,13 +42,13 @@ def decorator(annotator: Callable):
 
     return decorator
 
-def _is_input_non_float_tensor(node: Node):
+def _is_input_float_tensor(node: Node):
     """Check if the input is not a float tensor, so that we can skip quantization for the node
     since observers only works with float Tensors
     """
     if "val" not in node.meta or not isinstance(node.meta["val"], FakeTensor):
         return True
-    return node.meta["val"].dtype != torch.float32
+    return node.meta["val"].dtype == torch.float32
 
 
 def _is_annotated(nodes: List[Node]):
@@ -132,11 +132,11 @@ def annotate_binary(node: Node, quantization_config: QuantizationConfig) -> None
 
     input_qspec_map = {}
     input_act0 = node.args[0]
-    if isinstance(input_act0, Node) and not _is_input_non_float_tensor(input_act0):
+    if isinstance(input_act0, Node) and _is_input_float_tensor(input_act0):
         input_qspec_map[input_act0] = input_act_qspec
 
     input_act1 = node.args[1]
-    if isinstance(input_act1, Node) and not _is_input_non_float_tensor(input_act1):
+    if isinstance(input_act1, Node) and _is_input_float_tensor(input_act1):
         input_qspec_map[input_act1] = input_act_qspec
 
     node.meta[QUANT_ANNOTATION_KEY] = QuantizationAnnotation(

backends/vulkan/runtime/VulkanBackend.cpp

@@ -77,26 +77,26 @@ api::StorageType get_storage_type(
     const vkgraph::VkStorageType& vk_storage_type) {
   switch (vk_storage_type) {
     case vkgraph::VkStorageType::BUFFER:
-      return api::StorageType::BUFFER;
+      return api::kBuffer;
     case vkgraph::VkStorageType::TEXTURE_3D:
-      return api::StorageType::TEXTURE_3D;
+      return api::kTexture3D;
     case vkgraph::VkStorageType::TEXTURE_2D:
-      return api::StorageType::TEXTURE_2D;
+      return api::kTexture2D;
     default:
       break;
   }
-  return api::StorageType::UNKNOWN;
+  VK_THROW("Invalid storage type encountered!");
 }
 
 api::GPUMemoryLayout get_memory_layout(
     const vkgraph::VkMemoryLayout& vk_memory_layout) {
   switch (vk_memory_layout) {
     case vkgraph::VkMemoryLayout::TENSOR_WIDTH_PACKED:
-      return api::GPUMemoryLayout::TENSOR_WIDTH_PACKED;
+      return api::kWidthPacked;
     case vkgraph::VkMemoryLayout::TENSOR_HEIGHT_PACKED:
-      return api::GPUMemoryLayout::TENSOR_HEIGHT_PACKED;
+      return api::kHeightPacked;
     case vkgraph::VkMemoryLayout::TENSOR_CHANNELS_PACKED:
-      return api::GPUMemoryLayout::TENSOR_CHANNELS_PACKED;
+      return api::kChannelsPacked;
     default:
       break;
   }

backends/vulkan/runtime/api/Runtime.cpp

@@ -16,6 +16,22 @@
 namespace vkcompute {
 namespace api {
 
+#define PRINT_CASE(name)       \
+  case MemoryAccessType::name: \
+    out << #name;              \
+    break;
+
+std::ostream& operator<<(std::ostream& out, const MemoryAccessType& tag) {
+  switch (tag) {
+    PRINT_CASE(NONE)
+    PRINT_CASE(READ)
+    PRINT_CASE(WRITE)
+  }
+  return out;
+}
+
+#undef PRINT_CASE
+
 namespace {
 
 void find_requested_layers_and_extensions(

backends/vulkan/runtime/api/Shader.cpp

@@ -23,38 +23,19 @@ ShaderInfo::ShaderInfo()
           0u,
       } {}
 
-ShaderInfo::ShaderInfo(
-    std::string name,
-    const uint32_t* const spirv_bin,
-    const uint32_t size,
-    std::vector<VkDescriptorType>  layout)
-    : src_code{
-          spirv_bin,
-          size,
-      },
-      kernel_name{std::move(name)},
-      kernel_layout{std::move(layout)} {}
-
 ShaderInfo::ShaderInfo(
     std::string name,
     const uint32_t* const spirv_bin,
     const uint32_t size,
     std::vector<VkDescriptorType>  layout,
-    const std::vector<uint32_t>& tile_size,
-    const StorageType bias_storage_type,
-    const StorageType weight_storage_type)
+    const utils::uvec3 tile_size)
     : src_code{
           spirv_bin,
           size,
       },
       kernel_name{std::move(name)},
       kernel_layout{std::move(layout)},
-      tile_size(tile_size),
-      bias_storage_type(bias_storage_type),
-      weight_storage_type(weight_storage_type) {
-  for (uint64_t i = 0; i < tile_size.size(); ++i) {
-    out_tile_size.data[i] = tile_size[i];
-  }
+      out_tile_size(tile_size) {
 }
 
 bool operator==(const ShaderInfo& _1, const ShaderInfo& _2) {

backends/vulkan/runtime/api/Shader.h

@@ -62,25 +62,14 @@ struct ShaderInfo final {
   // Shader Metadata
   utils::uvec3 out_tile_size{1u, 1u, 1u};
 
-  std::vector<uint32_t> tile_size;
-  StorageType bias_storage_type{StorageType::UNKNOWN};
-  StorageType weight_storage_type{StorageType::UNKNOWN};
-
   explicit ShaderInfo();
-  explicit ShaderInfo(std::string, const char*);
-  explicit ShaderInfo(
-      std::string,
-      const uint32_t*,
-      const uint32_t,
-      std::vector<VkDescriptorType>);
+
   explicit ShaderInfo(
       std::string,
       const uint32_t*,
       const uint32_t,
       std::vector<VkDescriptorType>,
-      const std::vector<uint32_t>& tile_size,
-      const StorageType bias_storage_type,
-      const StorageType weight_storage_type);
+      const utils::uvec3 tile_size);
 };
 
 bool operator==(const ShaderInfo& _1, const ShaderInfo& _2);