[SYCL][Fusion] Interface with kernel fusion JIT

CONTRIBUTING.md

@@ -81,6 +81,15 @@ approved by an Intel representative.
   - A new approval is needed if the PR was updated (e.g. during code review).
 - Once the PR is approved and all checks have passed, the pull request is
 ready for merge.
+- Sometimes unrelated fails can be observed in the PR. It's author
+responsibility to find/guess the reason of these fails and post a comment in
+the PR with:
+  - possible reason of fails, ideally with a link to the PR, which caused fails
+  - link to other PR(s), which expected to fix fails
+  - person who is currently looking into fails
+  - link to existing open issue
+  - if author cannot identify any of these, the minimal action expected is to
+    open a new [issue](/../../issues)
 
 ### Merge
 
@@ -89,7 +98,8 @@ and merge] and using PR description as the commit message, replacing all
 individual comments made per commit. Authors of the change must ensure PR
 description is up to date at the merge stage, as sometimes comments addressed
 during code reviews can invalidate original PR description. Feel free to ping
-@intel/llvm-gatekeepers if your PR is green and can be merged.
+@intel/llvm-gatekeepers if your PR is green and can be merged. Note that
+gatekeepers will require explanation for any failures observed.
 
 Pulldown from LLVM upstream is done through merge commits to preserve hashes of
 the original commits pulled from the LLVM community repository.

README.md

@@ -6,6 +6,8 @@ Intel LLVM-based projects:
 - [oneAPI DPC++ compiler](#oneapi-dpc-compiler)
 - [Late-outline OpenMP and OpenMP Offload](#late-outline-openmp-and-openmp-offload)
 
+For general contribution process see [CONTRIBUTING.md](./CONTRIBUTING.md)
+
 ## oneAPI DPC++ compiler
 
 [![](https://spec.oneapi.io/oneapi-logo-white-scaled.jpg)](https://www.oneapi.io/)

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/CMakeLists.txt

@@ -145,7 +145,7 @@ 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)
 
 # 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 +299,14 @@ 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.
+
 ### Build Doxygen documentation
 
 Building Doxygen documentation is similar to building the product itself. First,

sycl/doc/developer/ContributeToDPCPP.md

@@ -1,5 +1,9 @@
 # Contributing to DPC++
 
+## General guidelines
+
+Read [CONTRIBUTING.md](/CONTRIBUTING.md) first.
+
 ## Maintaining stable ABI/API
 
 All changes made to the DPC++ compiler and runtime library should generally

sycl/include/CL/__spirv/spirv_types.hpp

@@ -108,13 +108,22 @@ enum class GroupOperation : uint32_t {
   ExclusiveScan = 2
 };
 
+#if (SYCL_EXT_ONEAPI_MATRIX_VERSION > 1)
+enum class MatrixLayout : uint32_t {
+  RowMajor = 0,
+  ColumnMajor = 1,
+  Packed = 2,
+  Dynamic = 3
+};
+#else
 enum class MatrixLayout : uint32_t {
   RowMajor = 0,
   ColumnMajor = 1,
   PackedA = 2,
   PackedB = 3,
   Unused = 4
 };
+#endif
 
 enum class MatrixUse : uint32_t { MatrixA = 0, MatrixB = 1, Accumulator = 2 };
 

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/include/sycl/detail/defines.hpp

@@ -21,7 +21,7 @@
 #endif
 #endif
 
-#if __has_attribute(sycl_special_class)
+#if __has_attribute(sycl_special_class) && (defined __SYCL_DEVICE_ONLY__)
 #define __SYCL_SPECIAL_CLASS __attribute__((sycl_special_class))
 #else
 #define __SYCL_SPECIAL_CLASS

sycl/include/sycl/ext/oneapi/device_global/device_global.hpp

@@ -15,6 +15,7 @@
 #include <sycl/exception.hpp>
 #include <sycl/ext/oneapi/device_global/properties.hpp>
 #include <sycl/ext/oneapi/properties/properties.hpp>
+#include <sycl/pointers.hpp>
 
 #ifdef __SYCL_DEVICE_ONLY__
 #define __SYCL_HOST_NOT_SUPPORTED(Op)
@@ -42,9 +43,27 @@ struct HasArrowOperator<
 template <typename T, typename PropertyListT, typename = void>
 class device_global_base {
 protected:
-  T *usmptr;
-  T *get_ptr() noexcept { return usmptr; }
-  const T *get_ptr() const noexcept { return usmptr; }
+  using pointer_t = typename decorated_global_ptr<T>::pointer;
+  pointer_t usmptr;
+  pointer_t get_ptr() noexcept { return usmptr; }
+  const pointer_t get_ptr() const noexcept { return usmptr; }
+
+public:
+  template <access::decorated IsDecorated>
+  multi_ptr<T, access::address_space::global_space, IsDecorated>
+  get_multi_ptr() noexcept {
+    __SYCL_HOST_NOT_SUPPORTED("get_multi_ptr()")
+    return multi_ptr<T, access::address_space::global_space, IsDecorated>{
+        get_ptr()};
+  }
+
+  template <access::decorated IsDecorated>
+  multi_ptr<const T, access::address_space::global_space, IsDecorated>
+  get_multi_ptr() const noexcept {
+    __SYCL_HOST_NOT_SUPPORTED("get_multi_ptr()")
+    return multi_ptr<const T, access::address_space::global_space, IsDecorated>{
+        get_ptr()};
+  }
 };
 
 // Specialization of device_global base class for when device_image_scope is in
@@ -58,6 +77,23 @@ class device_global_base<
   T val{};
   T *get_ptr() noexcept { return &val; }
   const T *get_ptr() const noexcept { return &val; }
+
+public:
+  template <access::decorated IsDecorated>
+  multi_ptr<T, access::address_space::global_space, IsDecorated>
+  get_multi_ptr() noexcept {
+    __SYCL_HOST_NOT_SUPPORTED("get_multi_ptr()")
+    return address_space_cast<access::address_space::global_space, IsDecorated,
+                              T>(this->get_ptr());
+  }
+
+  template <access::decorated IsDecorated>
+  multi_ptr<const T, access::address_space::global_space, IsDecorated>
+  get_multi_ptr() const noexcept {
+    __SYCL_HOST_NOT_SUPPORTED("get_multi_ptr()")
+    return address_space_cast<access::address_space::global_space, IsDecorated,
+                              const T>(this->get_ptr());
+  }
 };
 } // namespace detail
 
@@ -113,22 +149,6 @@ class
   device_global &operator=(const device_global &) = delete;
   device_global &operator=(const device_global &&) = delete;
 
-  template <access::decorated IsDecorated>
-  multi_ptr<T, access::address_space::global_space, IsDecorated>
-  get_multi_ptr() noexcept {
-    __SYCL_HOST_NOT_SUPPORTED("get_multi_ptr()")
-    return address_space_cast<access::address_space::global_space, IsDecorated>(
-        this->get_ptr());
-  }
-
-  template <access::decorated IsDecorated>
-  multi_ptr<const T, access::address_space::global_space, IsDecorated>
-  get_multi_ptr() const noexcept {
-    __SYCL_HOST_NOT_SUPPORTED("get_multi_ptr()")
-    return address_space_cast<access::address_space::global_space, IsDecorated,
-                              const T>(this->get_ptr());
-  }
-
   T &get() noexcept {
     __SYCL_HOST_NOT_SUPPORTED("get()")
     return *this->get_ptr();