Merge branch 'main' into xtensa-codegen

halide · Jan 20, 2023 · 4023640 · 4023640
2 parents 2e9ae6a + c601e4e
commit 4023640
Show file tree

Hide file tree

Showing 45 changed files with 243 additions and 357 deletions.
diff --git a/.github/workflows/pip.yml b/.github/workflows/pip.yml
@@ -109,7 +109,7 @@ jobs:
           cmake -G Ninja -S llvm-src/llvm -B llvm-build
           -DCMAKE_BUILD_TYPE=Release
           "-DCMAKE_OSX_ARCHITECTURES=arm64;x86_64"
-          "-DLLVM_TARGETS_TO_BUILD=X86;ARM;NVPTX;AArch64;Mips;Hexagon;WebAssembly"
+          "-DLLVM_TARGETS_TO_BUILD=X86;ARM;NVPTX;AArch64;Hexagon;WebAssembly"
           "-DLLVM_ENABLE_PROJECTS=clang;lld"
           -DLLVM_ENABLE_ASSERTIONS=ON
           -DLLVM_ENABLE_RTTI=ON

diff --git a/Makefile b/Makefile
@@ -112,7 +112,6 @@ LLVM_CXX_FLAGS += -DLLVM_VERSION=$(LLVM_VERSION_TIMES_10)
 WITH_X86 ?= $(findstring x86, $(LLVM_COMPONENTS))
 WITH_ARM ?= $(findstring arm, $(LLVM_COMPONENTS))
 WITH_HEXAGON ?= $(findstring hexagon, $(LLVM_COMPONENTS))
-WITH_MIPS ?= $(findstring mips, $(LLVM_COMPONENTS))
 WITH_RISCV ?= $(findstring riscv, $(LLVM_COMPONENTS))
 WITH_AARCH64 ?= $(findstring aarch64, $(LLVM_COMPONENTS))
 WITH_POWERPC ?= $(findstring powerpc, $(LLVM_COMPONENTS))
@@ -139,9 +138,6 @@ X86_LLVM_CONFIG_LIB=$(if $(WITH_X86), x86, )
 ARM_CXX_FLAGS=$(if $(WITH_ARM), -DWITH_ARM, )
 ARM_LLVM_CONFIG_LIB=$(if $(WITH_ARM), arm, )
 
-MIPS_CXX_FLAGS=$(if $(WITH_MIPS), -DWITH_MIPS, )
-MIPS_LLVM_CONFIG_LIB=$(if $(WITH_MIPS), mips, )
-
 POWERPC_CXX_FLAGS=$(if $(WITH_POWERPC), -DWITH_POWERPC, )
 POWERPC_LLVM_CONFIG_LIB=$(if $(WITH_POWERPC), powerpc, )
 
@@ -209,7 +205,6 @@ CXX_FLAGS += $(OPENCL_CXX_FLAGS)
 CXX_FLAGS += $(METAL_CXX_FLAGS)
 CXX_FLAGS += $(OPENGLCOMPUTE_CXX_FLAGS)
 CXX_FLAGS += $(D3D12_CXX_FLAGS)
-CXX_FLAGS += $(MIPS_CXX_FLAGS)
 CXX_FLAGS += $(POWERPC_CXX_FLAGS)
 CXX_FLAGS += $(INTROSPECTION_CXX_FLAGS)
 CXX_FLAGS += $(EXCEPTIONS_CXX_FLAGS)
@@ -238,7 +233,6 @@ LLVM_STATIC_LIBFILES = \
 	$(METAL_LLVM_CONFIG_LIB) \
 	$(PTX_LLVM_CONFIG_LIB) \
 	$(AARCH64_LLVM_CONFIG_LIB) \
-	$(MIPS_LLVM_CONFIG_LIB) \
 	$(POWERPC_LLVM_CONFIG_LIB) \
 	$(HEXAGON_LLVM_CONFIG_LIB) \
 	$(AMDGPU_LLVM_CONFIG_LIB) \
@@ -433,7 +427,6 @@ SOURCE_FILES = \
   CodeGen_Internal.cpp \
   CodeGen_LLVM.cpp \
   CodeGen_Metal_Dev.cpp \
-  CodeGen_MIPS.cpp \
   CodeGen_OpenCL_Dev.cpp \
   CodeGen_OpenGLCompute_Dev.cpp \
   CodeGen_Posix.cpp \
@@ -738,7 +731,6 @@ HEADER_FILES = \
   StrictifyFloat.h \
   Substitute.h \
   Target.h \
-  ThreadPool.h \
   Tracing.h \
   TrimNoOps.h \
   Tuple.h \
@@ -796,7 +788,6 @@ RUNTIME_CPP_COMPONENTS = \
   metal \
   metal_objc_arm \
   metal_objc_x86 \
-  mips_cpu_features \
   module_aot_ref_count \
   module_jit_ref_count \
   msan \
@@ -857,7 +848,6 @@ RUNTIME_LL_COMPONENTS = \
   arm \
   arm_no_neon \
   hvx_128 \
-  mips \
   posix_math \
   powerpc \
   ptx_dev \
@@ -1624,7 +1614,7 @@ $(FILTERS_DIR)/autograd_grad.a: $(BIN_DIR)/autograd.generator $(BIN_MULLAPUDI201
 # all have the form nested_externs_*).
 $(FILTERS_DIR)/nested_externs_%.a: $(BIN_DIR)/nested_externs.generator
 	@mkdir -p $(@D)
-	$(CURDIR)/$< -g nested_externs_$* $(GEN_AOT_OUTPUTS) -o $(CURDIR)/$(FILTERS_DIR) target=$(TARGET)-no_runtime
+	$(CURDIR)/$< -g nested_externs_$* $(GEN_AOT_OUTPUTS) -o $(CURDIR)/$(FILTERS_DIR) target=$(TARGET)-no_runtime-user_context-c_plus_plus_name_mangling
 
 # Similarly, gpu_multi needs two different kernels to test compilation caching.
 # Also requies user-context.
@@ -2271,6 +2261,7 @@ install: $(LIB_DIR)/libHalide.a $(BIN_DIR)/libHalide.$(SHARED_EXT) $(INCLUDE_DIR
 	cp $(ROOT_DIR)/tools/halide_image_io.h $(PREFIX)/share/halide/tools
 	cp $(ROOT_DIR)/tools/halide_image_info.h $(PREFIX)/share/halide/tools
 	cp $(ROOT_DIR)/tools/halide_malloc_trace.h $(PREFIX)/share/halide/tools
+	cp $(ROOT_DIR)/tools/halide_thread_pool.h $(PREFIX)/share/halide/tools
 ifeq ($(UNAME), Darwin)
 	install_name_tool -id $(PREFIX)/lib/libHalide.$(SHARED_EXT) $(PREFIX)/lib/libHalide.$(SHARED_EXT)
 endif
@@ -2349,6 +2340,7 @@ $(DISTRIB_DIR)/lib/libHalide.$(SHARED_EXT): \
 	cp $(ROOT_DIR)/tools/halide_image_io.h $(DISTRIB_DIR)/tools
 	cp $(ROOT_DIR)/tools/halide_image_info.h $(DISTRIB_DIR)/tools
 	cp $(ROOT_DIR)/tools/halide_malloc_trace.h $(DISTRIB_DIR)/tools
+	cp $(ROOT_DIR)/tools/halide_thread_pool.h $(DISTRIB_DIR)/tools
 	cp $(ROOT_DIR)/tools/halide_trace_config.h $(DISTRIB_DIR)/tools
 	cp $(ROOT_DIR)/README*.md $(DISTRIB_DIR)
 	cp $(BUILD_DIR)/halide_config.* $(DISTRIB_DIR)

diff --git a/README.md b/README.md
@@ -4,7 +4,7 @@ Halide is a programming language designed to make it easier to write
 high-performance image and array processing code on modern machines. Halide
 currently targets:
 
-- CPU architectures: X86, ARM, MIPS, Hexagon, PowerPC, RISC-V
+- CPU architectures: X86, ARM, Hexagon, PowerPC, RISC-V
 - Operating systems: Linux, Windows, macOS, Android, iOS, Qualcomm QuRT
 - GPU Compute APIs: CUDA, OpenCL, OpenGL Compute Shaders, Apple Metal, Microsoft
   Direct X 12
@@ -138,7 +138,7 @@ Then build it like so:
 ```
 % cmake -DCMAKE_BUILD_TYPE=Release \
         -DLLVM_ENABLE_PROJECTS="clang;lld;clang-tools-extra" \
-        -DLLVM_TARGETS_TO_BUILD="X86;ARM;NVPTX;AArch64;Mips;Hexagon;WebAssembly" \
+        -DLLVM_TARGETS_TO_BUILD="X86;ARM;NVPTX;AArch64;Hexagon;WebAssembly" \
         -DLLVM_ENABLE_TERMINFO=OFF -DLLVM_ENABLE_ASSERTIONS=ON \
         -DLLVM_ENABLE_EH=ON -DLLVM_ENABLE_RTTI=ON -DLLVM_BUILD_32_BITS=OFF \
         -S llvm-project/llvm -B llvm-build
@@ -298,7 +298,7 @@ D:\> cmake -G Ninja ^
            -DCMAKE_BUILD_TYPE=Release ^
            -DLLVM_ENABLE_PROJECTS=clang;lld;clang-tools-extra ^
            -DLLVM_ENABLE_TERMINFO=OFF ^
-           -DLLVM_TARGETS_TO_BUILD=X86;ARM;NVPTX;AArch64;Mips;Hexagon ^
+           -DLLVM_TARGETS_TO_BUILD=X86;ARM;NVPTX;AArch64;Hexagon ^
            -DLLVM_ENABLE_ASSERTIONS=ON ^
            -DLLVM_ENABLE_EH=ON ^
            -DLLVM_ENABLE_RTTI=ON ^
@@ -313,7 +313,7 @@ D:\> cmake -G Ninja ^
            -DCMAKE_BUILD_TYPE=Release ^
            -DLLVM_ENABLE_PROJECTS=clang;lld;clang-tools-extra ^
            -DLLVM_ENABLE_TERMINFO=OFF ^
-           -DLLVM_TARGETS_TO_BUILD=X86;ARM;NVPTX;AArch64;Mips;Hexagon ^
+           -DLLVM_TARGETS_TO_BUILD=X86;ARM;NVPTX;AArch64;Hexagon ^
            -DLLVM_ENABLE_ASSERTIONS=ON ^
            -DLLVM_ENABLE_EH=ON ^
            -DLLVM_ENABLE_RTTI=ON ^

diff --git a/README_cmake.md b/README_cmake.md
@@ -409,7 +409,6 @@ LLVM component names):
 | `TARGET_AMDGPU`      | `ON`, _if available_ | Enable the AMD GPU backend          |
 | `TARGET_ARM`         | `ON`, _if available_ | Enable the ARM backend              |
 | `TARGET_HEXAGON`     | `ON`, _if available_ | Enable the Hexagon backend          |
-| `TARGET_MIPS`        | `ON`, _if available_ | Enable the MIPS backend             |
 | `TARGET_NVPTX`       | `ON`, _if available_ | Enable the NVidia PTX backend       |
 | `TARGET_POWERPC`     | `ON`, _if available_ | Enable the PowerPC backend          |
 | `TARGET_RISCV`       | `ON`, _if available_ | Enable the RISC V backend           |
@@ -770,14 +769,15 @@ Variables set by the package:
 
 Halide defines the following targets that are available to users:
 
-| Imported target      | Description                                                                                                                          |
-|----------------------|--------------------------------------------------------------------------------------------------------------------------------------|
-| `Halide::Halide`     | this is the JIT-mode library to use when using Halide from C++.                                                                      |
-| `Halide::Generator`  | this is the target to use when defining a generator executable. It supplies a `main()` function.                                     |
-| `Halide::Runtime`    | adds include paths to the Halide runtime headers                                                                                     |
-| `Halide::Tools`      | adds include paths to the Halide tools, including the benchmarking utility.                                                          |
-| `Halide::ImageIO`    | adds include paths to the Halide image IO utility and sets up dependencies to PNG / JPEG if they are available.                      |
-| `Halide::RunGenMain` | used with the `REGISTRATION` parameter of `add_halide_library` to create simple runners and benchmarking tools for Halide libraries. |
+| Imported target      | Description                                                                                                                                                                                    |
+|----------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| `Halide::Halide`     | this is the JIT-mode library to use when using Halide from C++.                                                                                                                                |
+| `Halide::Generator`  | this is the target to use when defining a generator executable. It supplies a `main()` function.                                                                                               |
+| `Halide::Runtime`    | adds include paths to the Halide runtime headers                                                                                                                                               |
+| `Halide::Tools`      | adds include paths to the Halide tools, including the benchmarking utility.                                                                                                                    |
+| `Halide::ImageIO`    | adds include paths to the Halide image IO utility. Depends on `PNG::PNG` and `JPEG::JPEG` if they exist or were loaded through the corresponding package components.                           |
+| `Halide::ThreadPool` | adds include paths to the Halide _simple_ thread pool utility library. This is not the same as the runtime's thread pool and is intended only for use by tests. Depends on `Threads::Threads`. |
+| `Halide::RunGenMain` | used with the `REGISTRATION` parameter of `add_halide_library` to create simple runners and benchmarking tools for Halide libraries.                                                           |
 
 The following targets are not guaranteed to be available:
 

diff --git a/README_webassembly.md b/README_webassembly.md
@@ -136,7 +136,7 @@ $ ctest -L "correctness|generator" -j
     (`"all"`) then it's already present, but otherwise, add it explicitly:
 
 ```
--DLLVM_TARGETS_TO_BUILD="X86;ARM;NVPTX;AArch64;Mips;PowerPC;Hexagon;WebAssembly
+-DLLVM_TARGETS_TO_BUILD="X86;ARM;NVPTX;AArch64;PowerPC;Hexagon;WebAssembly
 ```
 
 ## Enabling wasm JIT

diff --git a/apps/HelloAndroid/README.md b/apps/HelloAndroid/README.md
@@ -2,8 +2,8 @@ HelloHalide is a simple application which applies a tone curve and sharpening to
 a video preview from the camera on a phone or tablet.
 
 This application builds for multiple native ABIs. (At present armeabi,
-armeabi-v7a, arm64-v8a, mips, x86_64, and x86 are supported. mips64 is not
-presently working.) Halide code is generated for each architecture.
+armeabi-v7a, arm64-v8a, x86_64, and x86 are supported.) Halide code is
+generated for each architecture.
 
 This build is meant to use Android command line tools. (An IDE is not required.)
 In order to build, the following will be required:
@@ -47,10 +47,9 @@ application.
 
 Pay attention to the list of platforms supported by your Halide installation.
 They are listed in jni/Application.mk APP_ABI variable and in build.gradle archs
-map. For example, if your Halide installation was built without mips support or
-without arm64-v8a, remove them from APP_ABI and archs. Both list and map should
-match, otherwise you will be getting compilation errors complaining about a
-missing hello.h file:
+map. For example, if your Halide installation was built without arm64-v8a,
+remove it from APP_ABI and archs. Both list and map should match, otherwise
+you will be getting compilation errors complaining about a missing hello.h file:
 
 ```
 :compileDebugNdkClassic FAILED

diff --git a/apps/HelloAndroid/build.gradle b/apps/HelloAndroid/build.gradle
@@ -67,7 +67,6 @@ binaries.withType(NativeExecutableBinary) { binary ->
         "armeabi":      "arm-32-android",
         "armeabi-v7a":  "arm-32-android",
         "arm64-v8a":    "arm-64-android",
-        "mips":         "mips-32-android",
         "x86_64":       "x86-64-android-sse41",
         "x86":          "x86-32-android"
     ]

diff --git a/apps/HelloAndroid/build.sh b/apps/HelloAndroid/build.sh
@@ -8,11 +8,7 @@ c++ jni/hello_generator.cpp ../../tools/GenGen.cpp \
     -L ../../bin -lHalide -ldl -lpthread -lz \
     -o bin/hello_generator
 
-# 64-bit MIPS (mips-64-android,mips64) currently does not build since
-# llvm will not compile for the R6 version of the ISA without Nan2008
-# and the gcc toolchain used by the Android build setup requires those
-# two options together.
-for archs in arm-32-android,armeabi arm-32-android-armv7s,armeabi-v7a arm-64-android,arm64-v8a mips-32-android,mips x86-64-android-sse41,x86_64 x86-32-android,x86 ; do
+for archs in arm-32-android,armeabi arm-32-android-armv7s,armeabi-v7a arm-64-android,arm64-v8a x86-64-android-sse41,x86_64 x86-32-android,x86 ; do
     IFS=,
     set $archs
     HL_TARGET=$1

diff --git a/apps/HelloAndroid/jni/Application.mk b/apps/HelloAndroid/jni/Application.mk
@@ -1,8 +1,4 @@
-# Can't use "APP_ABI = all" as 64-bit MIPS currently does not build since
-# llvm will not compile for the R6 version of the ISA without Nan2008
-# and the gcc toolchain used by the Android build setup requires those
-# two options together.
-APP_ABI := armeabi armeabi-v7a arm64-v8a mips x86_64 x86
+APP_ABI := armeabi armeabi-v7a arm64-v8a x86_64 x86
 APP_PLATFORM := android-17
 APP_STL := gnustl_static
 APP_CPPFLAGS := -std=c++17
diff --git a/apps/HelloAndroidCamera2/README.md b/apps/HelloAndroidCamera2/README.md
@@ -21,8 +21,8 @@ E/halide_native( 6146): ANativeWindow buffer locked but its size was 1920 x
 1440, expected 1440 x 1080
 
 This application builds for multiple native ABIs. (At present armeabi,
-armeabi-v7a, arm64-v8a, mips, x86_64, and x86 are supported. mips64 is not
-presently working.) Halide code is generated for each architecture.
+armeabi-v7a, arm64-v8a, x86_64, and x86 are supported.) Halide code is
+generated for each architecture.
 
 This build is meant to use Android command line tools. (An IDE is not required.)
 In order to build, the following will be required:
@@ -65,10 +65,10 @@ application.
 
 Pay attention to the list of platforms supported by your Halide installation.
 They are listed in jni/Application.mk APP_ABI variable and in build.gradle archs
-map. For example, if your Halide installation was built without mips support or
-without arm64-v8a, remove them from APP_ABI and archs. Both list and map should
-match, otherwise you will be getting compilation errors complaining about a
-missing halide_generated.h file:
+map. For example, if your Halide installation was built without arm64-v8a,
+remove it from APP_ABI and archs. Both list and map should match, otherwise
+you will be getting compilation errors complaining about a missing
+halide_generated.h file:
 
 ```
 :ndkBuild FAILED

diff --git a/apps/HelloAndroidCamera2/build.gradle b/apps/HelloAndroidCamera2/build.gradle
@@ -72,7 +72,6 @@ binaries.withType(NativeExecutableBinary) { binary ->
         "armeabi":      "arm-32-android",
         "armeabi-v7a":  "arm-32-android",
         "arm64-v8a":    "arm-64-android",
-        "mips":         "mips-32-android",
         "x86_64":       "x86-64-android-sse41",
         "x86":          "x86-32-android"
     ]

diff --git a/apps/HelloAndroidCamera2/build.sh b/apps/HelloAndroidCamera2/build.sh
@@ -20,11 +20,7 @@ c++ jni/deinterleave_generator.cpp ../../tools/GenGen.cpp \
     -L ../../bin -lHalide -ldl -lpthread -lz \
     -o bin/deinterleave_generator
 
-# 64-bit MIPS (mips-64-android,mips64) currently does not build since
-# llvm will not compile for the R6 version of the ISA without Nan2008
-# and the gcc toolchain used by the Android build setup requires those
-# two options together.
-for archs in arm-32-android,armeabi arm-32-android-armv7s,armeabi-v7a arm-64-android,arm64-v8a mips-32-android,mips x86-64-android-sse41,x86_64 x86-32-android,x86 ; do
+for archs in arm-32-android,armeabi arm-32-android-armv7s,armeabi-v7a arm-64-android,arm64-v8a x86-64-android-sse41,x86_64 x86-32-android,x86 ; do
     IFS=,
     set $archs
     HL_TARGET=$1

diff --git a/apps/HelloAndroidCamera2/jni/Application.mk b/apps/HelloAndroidCamera2/jni/Application.mk
@@ -1,8 +1,4 @@
-# Can't use "APP_ABI = all" as 64-bit MIPS currently does not build since
-# llvm will not compile for the R6 version of the ISA without Nan2008
-# and the gcc toolchain used by the Android build setup requires those
-# two options together.
-APP_ABI := armeabi armeabi-v7a arm64-v8a mips x86_64 x86
+APP_ABI := armeabi armeabi-v7a arm64-v8a x86_64 x86
 APP_PLATFORM := android-21
 APP_STL := c++_static
 APP_CPPFLAGS := -std=c++17 -fno-rtti -fexceptions
diff --git a/apps/openglcompute/jni/Application.mk b/apps/openglcompute/jni/Application.mk
@@ -1,5 +1,5 @@
 # TODO(aam): Confirm that application builds and runs for all supported targets:
-# APP_ABI := armeabi armeabi-v7a arm64-v8a mips x86_64 x86
+# APP_ABI := armeabi armeabi-v7a arm64-v8a x86_64 x86
 APP_ABI := armeabi-v7a
 APP_PLATFORM := android-17
 

diff --git a/cmake/HalideGeneratorHelpers.cmake b/cmake/HalideGeneratorHelpers.cmake
@@ -50,7 +50,7 @@ function(add_halide_generator TARGET)
 
     if (NOT ARG_EXPORT_FILE)
         file(MAKE_DIRECTORY "${PROJECT_BINARY_DIR}/cmake")
-        set(ARG_EXPORT_FILE "${PROJECT_BINARY_DIR}/cmake/${ARG_PACKAGE_NAME}-config.cmake")
+        set(ARG_EXPORT_FILE "${PROJECT_BINARY_DIR}/cmake/${ARG_PACKAGE_NAME}Config.cmake")
     endif ()
 
     if (NOT ARG_SOURCES)
@@ -228,6 +228,16 @@ function(add_halide_library TARGET)
         message(FATAL_ERROR "Missing FROM argument specifying a Halide generator target")
     endif ()
 
+    if (NOT TARGET ${ARG_FROM})
+        # FROM is usually an unqualified name; if we are crosscompiling, we might need a
+        # fully-qualified name, so add the default package name and retry
+        set(FQ_ARG_FROM "${PROJECT_NAME}::halide_generators::${ARG_FROM}")
+        if (NOT TARGET ${FQ_ARG_FROM})
+            message(FATAL_ERROR "Unable to locate FROM as either ${ARG_FROM} or ${FQ_ARG_FROM}")
+        endif ()
+        set(ARG_FROM "${FQ_ARG_FROM}")
+    endif()
+
     get_property(py_src TARGET ${ARG_FROM} PROPERTY Halide_PYTHON_GENERATOR_SOURCE)
     if (py_src)
         # TODO: Python Generators need work to support crosscompiling (https://github.com/halide/Halide/issues/7014)
@@ -241,15 +251,6 @@ function(add_halide_library TARGET)
         set(GENERATOR_CMD ${CMAKE_COMMAND} -E env PYTHONPATH=${PYTHONPATH} ${Python3_EXECUTABLE} $<SHELL_PATH:${py_src}>)
         set(GENERATOR_CMD_DEPS ${ARG_FROM} Halide::Python ${py_src})
     else()
-        if (NOT TARGET ${ARG_FROM})
-            # FROM is usually an unqualified name; if we are crosscompiling, we might need a
-            # fully-qualified name, so add the default package name and retry
-            set(FQ_ARG_FROM "${PROJECT_NAME}::halide_generators::${ARG_FROM}")
-            if (NOT TARGET ${FQ_ARG_FROM})
-                message(FATAL_ERROR "Unable to locate FROM as either ${ARG_FROM} or ${FQ_ARG_FROM}")
-            endif ()
-            set(ARG_FROM "${FQ_ARG_FROM}")
-        endif()
         set(GENERATOR_CMD "${ARG_FROM}")
         set(GENERATOR_CMD_DEPS ${ARG_FROM})
         _Halide_place_dll(${ARG_FROM})
@@ -313,7 +314,7 @@ function(add_halide_library TARGET)
         set(ARG_USE_RUNTIME Halide::Runtime)
     elseif (NOT ARG_USE_RUNTIME)
         # If we're not using an existing runtime, create one.
-        add_halide_runtime("${TARGET}.runtime" TARGETS ${ARG_TARGETS})
+        add_halide_runtime("${TARGET}.runtime" TARGETS ${ARG_TARGETS} FROM ${ARG_FROM})
         set(ARG_USE_RUNTIME "${TARGET}.runtime")
     elseif (NOT TARGET ${ARG_USE_RUNTIME})
         message(FATAL_ERROR "Invalid runtime target ${ARG_USE_RUNTIME}")
@@ -555,7 +556,7 @@ endfunction()
 
 function(add_halide_runtime RT)
     set(options "")
-    set(oneValueArgs "")
+    set(oneValueArgs FROM)
     set(multiValueArgs TARGETS)
     cmake_parse_arguments(ARG "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
 
@@ -569,6 +570,17 @@ function(add_halide_runtime RT)
     # so that GCD calculation doesn't get confused.
     list(TRANSFORM ARG_TARGETS APPEND "-no_runtime")
 
+    if (ARG_FROM)
+        # Try to use generator which is available. This is essential for cross-compilation
+        # where we cannot use host compiler to build generator only for runtime.
+
+        # Need to check if the ones for python extension, which is not actually an executable
+        get_target_property(target_type ${ARG_FROM} TYPE)
+        get_target_property(aliased ${ARG_FROM} ALIASED_TARGET)
+        if (target_type STREQUAL "EXECUTABLE" AND NOT aliased)
+            add_executable(_Halide_gengen ALIAS ${ARG_FROM})
+        endif()
+    endif()
     # Ensure _Halide_gengen is defined
     _Halide_gengen_ensure()