[SYCL][Fusion] Interface with kernel fusion JIT

Signed-off-by: Lukas Sommer <[email protected]>

Author: sommerlukas

buildbot/configure.py

@@ -23,6 +23,11 @@ def do_configure(args):
     libclc_amd_target_names = ';amdgcn--;amdgcn--amdhsa'
     libclc_nvidia_target_names = ';nvptx64--;nvptx64--nvidiacl'
 
+    sycl_enable_fusion = "OFF"
+    if not args.disable_fusion:
+        llvm_external_projects += ";sycl-fusion"
+        sycl_enable_fusion = "ON"
+
     if args.llvm_external_projects:
         llvm_external_projects += ";" + args.llvm_external_projects.replace(",", ";")
 
@@ -32,6 +37,7 @@ def do_configure(args):
     xpti_dir = os.path.join(abs_src_dir, "xpti")
     xptifw_dir = os.path.join(abs_src_dir, "xptifw")
     libdevice_dir = os.path.join(abs_src_dir, "libdevice")
+    fusion_dir = os.path.join(abs_src_dir, "sycl-fusion")
     llvm_targets_to_build = args.host_target
     llvm_enable_projects = 'clang;' + llvm_external_projects
     libclc_targets_to_build = ''
@@ -144,6 +150,7 @@ def do_configure(args):
         "-DXPTI_SOURCE_DIR={}".format(xpti_dir),
         "-DLLVM_EXTERNAL_XPTIFW_SOURCE_DIR={}".format(xptifw_dir),
         "-DLLVM_EXTERNAL_LIBDEVICE_SOURCE_DIR={}".format(libdevice_dir),
+        "-DLLVM_EXTERNAL_SYCL_FUSION_SOURCE_DIR={}".format(fusion_dir),
         "-DLLVM_ENABLE_PROJECTS={}".format(llvm_enable_projects),
         "-DLIBCLC_TARGETS_TO_BUILD={}".format(libclc_targets_to_build),
         "-DLIBCLC_GENERATE_REMANGLED_VARIANTS={}".format(libclc_gen_remangled_variants),
@@ -159,7 +166,8 @@ def do_configure(args):
         "-DLLVM_ENABLE_LLD={}".format(llvm_enable_lld),
         "-DXPTI_ENABLE_WERROR={}".format(xpti_enable_werror),
         "-DSYCL_CLANG_EXTRA_FLAGS={}".format(sycl_clang_extra_flags),
-        "-DSYCL_ENABLE_PLUGINS={}".format(';'.join(set(sycl_enabled_plugins)))
+        "-DSYCL_ENABLE_PLUGINS={}".format(';'.join(set(sycl_enabled_plugins))),
+        "-DSYCL_ENABLE_KERNEL_FUSION={}".format(sycl_enable_fusion)
     ]
 
     if args.l0_headers and args.l0_loader:
@@ -238,6 +246,7 @@ def main():
     parser.add_argument("--llvm-external-projects", help="Add external projects to build. Add as comma seperated list.")
     parser.add_argument("--ci-defaults", action="store_true", help="Enable default CI parameters")
     parser.add_argument("--enable-plugin", action='append', help="Enable SYCL plugin")
+    parser.add_argument("--disable-fusion", action="store_true", help="Disable the kernel fusion JIT compiler")
     args = parser.parse_args()
 
     print("args:{}".format(args))

sycl-fusion/CMakeLists.txt

@@ -9,6 +9,10 @@ set(SYCL_JIT_BASE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
 # directories, similar to how clang/CMakeLists.txt does it.
 set(LLVM_SPIRV_INCLUDE_DIRS "${LLVM_MAIN_SRC_DIR}/../llvm-spirv/include")
 
-add_subdirectory(jit-compiler)
-add_subdirectory(passes)
-add_subdirectory(test)
+if(WIN32)
+  message(WARNING "Kernel fusion not yet supported on Windows")
+else(WIN32)
+  add_subdirectory(jit-compiler)
+  add_subdirectory(passes)
+  add_subdirectory(test)
+endif(WIN32)

sycl-fusion/jit-compiler/CMakeLists.txt

@@ -39,6 +39,15 @@ target_link_libraries(sycl-fusion
   ${CMAKE_THREAD_LIBS_INIT}
 )
 
+if(NOT MSVC AND NOT APPLE)
+  # Manage symbol visibility through the linker to make sure no LLVM symbols
+  # are exported and confuse the drivers.
+  set(linker_script "${CMAKE_CURRENT_SOURCE_DIR}/ld-version-script.txt")
+  target_link_libraries(
+    sycl-fusion PRIVATE "-Wl,--version-script=${linker_script}")
+  set_target_properties(sycl-fusion PROPERTIES LINK_DEPENDS ${linker_script})
+endif()
+
 install(TARGETS sycl-fusion
   LIBRARY DESTINATION "lib${LLVM_LIBDIR_SUFFIX}" COMPONENT sycl-fusion
   RUNTIME DESTINATION "bin" COMPONENT sycl-fusion)

sycl-fusion/jit-compiler/ld-version-script.txt

@@ -0,0 +1,8 @@
+{
+  global:
+    /* Export everything from jit_compiler namespace */
+    _ZN12jit_compiler*;
+
+  local:
+    *;
+};

sycl-fusion/jit-compiler/lib/fusion/FusionPipeline.cpp

@@ -78,7 +78,7 @@ FusionPipeline::runFusionPasses(Module &Mod, SYCLModuleInfo &InputInfo,
     FPM.addPass(createFunctionToLoopPassAdaptor(IndVarSimplifyPass{}));
     LoopUnrollOptions UnrollOptions;
     FPM.addPass(LoopUnrollPass{UnrollOptions});
-    FPM.addPass(SROAPass{});
+    FPM.addPass(SROAPass{SROAOptions::ModifyCFG});
     // Run the InferAddressSpace pass to remove as many address-space casts
     // to/from generic address-space as possible, because these hinder
     // internalization.
@@ -94,11 +94,11 @@ FusionPipeline::runFusionPasses(Module &Mod, SYCLModuleInfo &InputInfo,
   // Run additional optimization passes after completing fusion.
   {
     FunctionPassManager FPM;
-    FPM.addPass(SROAPass{});
+    FPM.addPass(SROAPass{SROAOptions::ModifyCFG});
     FPM.addPass(SCCPPass{});
     FPM.addPass(InstCombinePass{});
     FPM.addPass(SimplifyCFGPass{});
-    FPM.addPass(SROAPass{});
+    FPM.addPass(SROAPass{SROAOptions::ModifyCFG});
     FPM.addPass(InstCombinePass{});
     FPM.addPass(SimplifyCFGPass{});
     FPM.addPass(ADCEPass{});

sycl/CMakeLists.txt

@@ -145,7 +145,12 @@ install(DIRECTORY ${OpenCL_INCLUDE_DIR}/CL
         COMPONENT OpenCL-Headers)
 
 # Option to enable online kernel fusion via a JIT compiler
-option(SYCL_ENABLE_KERNEL_FUSION "Enable kernel fusion via JIT compiler" OFF)
+option(SYCL_ENABLE_KERNEL_FUSION "Enable kernel fusion via JIT compiler" ON)
+if(SYCL_ENABLE_KERNEL_FUSION AND WIN32)
+  message(WARNING "Kernel fusion not yet supported on Windows")
+  set(SYCL_ENABLE_KERNEL_FUSION OFF CACHE 
+      BOOL "Kernel fusion not yet supported on Windows" FORCE)
+endif()
 
 # Needed for feature_test.hpp
 if ("cuda" IN_LIST SYCL_ENABLE_PLUGINS)

sycl/cmake/modules/AddSYCLUnitTest.cmake

@@ -56,6 +56,11 @@ macro(add_sycl_unittest test_dirname link_variant)
       OpenCL-Headers
       ${SYCL_LINK_LIBS}
     )
+
+  if(SYCL_ENABLE_KERNEL_FUSION)
+    target_link_libraries(${test_dirname} PRIVATE sycl-fusion)
+  endif(SYCL_ENABLE_KERNEL_FUSION)
+  
   target_include_directories(${test_dirname}
     PRIVATE SYSTEM
       ${sycl_inc_dir}

sycl/doc/GetStartedGuide.md

@@ -12,6 +12,7 @@ and a wide range of compute accelerators such as GPU and FPGA.
     - [Build DPC++ toolchain with support for HIP AMD](#build-dpc-toolchain-with-support-for-hip-amd)
     - [Build DPC++ toolchain with support for HIP NVIDIA](#build-dpc-toolchain-with-support-for-hip-nvidia)
     - [Build DPC++ toolchain with support for ESIMD CPU Emulation](#build-dpc-toolchain-with-support-for-esimd-emulator)
+    - [Build DPC++ toolchain with support for runtime kernel fusion](#build-dpc-toolchain-with-support-for-runtime-kernel-fusion)
     - [Build Doxygen documentation](#build-doxygen-documentation)
     - [Deployment](#deployment)
   - [Use DPC++ toolchain](#use-dpc-toolchain)
@@ -298,6 +299,16 @@ Enabling this flag requires following packages installed.
 Currently, this feature was tested and verified on Ubuntu 20.04
 environment.
 
+### Build DPC++ toolchain with support for runtime kernel fusion
+
+Support for the experimental SYCL extension for user-driven kernel fusion 
+at runtime is enabled by default. 
+
+To disable support for this feature, follow the instructions for the 
+Linux DPC++ toolchain, but add the `--disable-fusion` flag.
+
+Kernel fusion is currently not yet supported on the Windows platform.
+
 ### Build Doxygen documentation
 
 Building Doxygen documentation is similar to building the product itself. First,

sycl/include/sycl/detail/cg.hpp

@@ -98,6 +98,10 @@ class CG {
 
   CGTYPE getType() { return MType; }
 
+  std::vector<std::vector<char>> &getArgsStorage() { return MArgsStorage; }
+
+  std::vector<detail::AccessorImplPtr> &getAccStorage() { return MAccStorage; }
+
   virtual ~CG() = default;
 
 private:

sycl/source/CMakeLists.txt

@@ -87,6 +87,13 @@ function(add_sycl_rt_library LIB_NAME LIB_OBJ_NAME)
     PRIVATE OpenCL-Headers
   )
 
+  if(SYCL_ENABLE_KERNEL_FUSION)
+    target_link_libraries(${LIB_NAME} PRIVATE sycl-fusion)
+    target_link_libraries(${LIB_OBJ_NAME} PRIVATE sycl-fusion)
+    set_property(GLOBAL APPEND PROPERTY SYCL_TOOLCHAIN_INSTALL_COMPONENTS
+      sycl-fusion)
+  endif(SYCL_ENABLE_KERNEL_FUSION)
+
   find_package(Threads REQUIRED)
 
   target_link_libraries(${LIB_NAME}
@@ -139,6 +146,8 @@ set(SYCL_SOURCES
     "detail/handler_proxy.cpp"
     "detail/image_accessor_util.cpp"
     "detail/image_impl.cpp"
+    "detail/jit_compiler.cpp"
+    "detail/jit_device_binaries.cpp"
     "detail/kernel_impl.cpp"
     "detail/kernel_program_cache.cpp"
     "detail/memory_manager.cpp"