[SYCL] Change kernel/program cache data structures and get rid of boost (#18599)

I'm working to ease Linux distros in packaging this repo, and our use of
boost was hindering that.
The only place we use boost today is the data structures for the kernel
and program cache.

Change it to be an equally fast data structure that was already use for
`xptifw` (albeit with a manually implemented wrapper for
`unordered_multimap` that the upstream repo doesn't provide) and delete
boost.

I've done various benchmarks to make sure there is no performance impact
using the cache benchmark test sent to me by Slawomir.

Here the the performance results. Each entry is an average of 10 runs,
each run using 3000 kernels with 10000 iterations. Let me know if you'd
like the full data. All times are in milliseconds and lower is better.

Linux:

Fast kernel cache:
| Boost  | This PR | Percent Diff
| ------------- | ------------- | ------------ |
| 44.95478  | 45.04912 | 0.21% |


Regular kernel cache/program cache:
| Boost  | This PR | Percent Diff
| ------------- | ------------- | ------------ |
| 21541.69   | 21150.15 |  -1.83%



Windows:

Fast kernel cache:
| Boost  | This PR | Percent Diff
| ------------- | ------------- | ------------ |
| 71.19915  | 69.38018 | -2.59% |


Regular kernel cache/program cache:
| Boost  | This PR | Percent Diff
| ------------- | ------------- | ------------ |
| 7295.955   | 7298.057 | 0.03%

My interpretation of the results is that there is basically no
performance impact, either positive or negative.

Different test kinds may have been run on different hardware, though of
course tests of the same os/test kind were done on the same machine.

And just for fun:

`configure.py` time:
| Boost  | This PR | Percent Diff
| ------------- | ------------- | ------------ |
| 29.812 | 21.1625 | -33.94% |

Signed-off-by: Sarnie, Nick <[email protected]>

Author: sarnex

sycl/CMakeLists.txt

@@ -82,6 +82,7 @@ if(MSVC)
   endif()
 endif()
 
+include(FetchEmhash)
 include(FetchUnifiedRuntime)
 
 # The change in SYCL_MAJOR_VERSION must be accompanied with the same update in
@@ -213,9 +214,6 @@ install(FILES
   DESTINATION "${SYCL_INCLUDE_DIR}/sycl"
   COMPONENT sycl-headers)
 
-include(AddBoostMp11Headers)
-include(FetchBoostUnorderedHeaders)
-
 # This is workaround to detect changes (add or modify) in subtree which
 # are not detected by copy_directory command.
 # TODO: detect and process remove header/directory case
@@ -250,7 +248,7 @@ add_custom_target(sycl-headers
           ${OUT_HEADERS_IN_SYCLCOMPAT_DIR}
           ${OUT_UR_HEADERS}
           sycl-device-aspect-macros-header
-          boost_mp11-headers)
+          )
 
 add_custom_command(
   OUTPUT  ${OUT_HEADERS_IN_SYCL_DIR}
@@ -275,7 +273,6 @@ add_custom_command(
 install(DIRECTORY "${sycl_inc_dir}/sycl" DESTINATION ${SYCL_INCLUDE_DIR} COMPONENT sycl-headers)
 install(DIRECTORY "${sycl_inc_dir}/CL" DESTINATION ${SYCL_INCLUDE_DIR}/ COMPONENT sycl-headers)
 install(DIRECTORY "${sycl_inc_dir}/std" DESTINATION ${SYCL_INCLUDE_DIR} COMPONENT sycl-headers)
-install(DIRECTORY ${BOOST_MP11_DESTINATION_DIR} DESTINATION ${SYCL_INCLUDE_DIR}/sycl/detail COMPONENT boost_mp11-headers)
 install(DIRECTORY "${sycl_inc_dir}/syclcompat" DESTINATION ${SYCL_INCLUDE_DIR} COMPONENT sycl-headers)
 install(FILES "${sycl_inc_dir}/syclcompat.hpp" DESTINATION ${SYCL_INCLUDE_DIR} COMPONENT sycl-headers)
 install(FILES "${UNIFIED_RUNTIME_INCLUDE_DIR}/ur_api.h" DESTINATION ${SYCL_INCLUDE_DIR}
@@ -458,7 +455,6 @@ get_property(SYCL_TOOLCHAIN_DEPS GLOBAL PROPERTY SYCL_TOOLCHAIN_INSTALL_COMPONEN
 # Listed here are component names contributing the package
 set( SYCL_TOOLCHAIN_DEPLOY_COMPONENTS
      append-file
-     boost_mp11-headers
      clang
      clang-offload-wrapper
      clang-offload-bundler

sycl/cmake/modules/AddBoostMp11Headers.cmake

@@ -0,0 +1,21 @@
+# Finds or fetches emhash.
+find_file(EMHASH_SYS_LOC "hash_table8.hpp" PATH_SUFFIXES "emhash")
+if(NOT EMHASH_SYS_LOC)
+  set(EMHASH_REPO https://github.com/ktprime/emhash)
+  message(STATUS "Will fetch emhash from ${EMHASH_REPO}")
+  FetchContent_Declare(emhash
+    GIT_REPOSITORY    ${EMHASH_REPO}
+    GIT_TAG           3ba9abdfdc2e0430fcc2fd8993cad31945b6a02b
+    SOURCE_SUBDIR  emhash
+  )
+  FetchContent_MakeAvailable(emhash)
+
+  # FetchContent downloads the files into a directory with
+  # '-src' as the suffix and emhash has the headers in the
+  # top level directory in the repo, so copy the headers to a directory
+  # named `emhash` so source files can include with <emhash/header.hpp>
+  file(MAKE_DIRECTORY ${CMAKE_BINARY_DIR}/include/emhash)
+  file(GLOB HEADERS "${emhash_SOURCE_DIR}/*.h*")
+  file(COPY ${HEADERS} DESTINATION ${CMAKE_BINARY_DIR}/include/emhash)
+  set(SYCL_EMHASH_DIR ${CMAKE_BINARY_DIR}/include/ CACHE INTERNAL "")
+endif()

sycl/cmake/modules/FetchBoostUnorderedHeaders.cmake

@@ -358,6 +358,7 @@ if(CMAKE_SYSTEM_NAME STREQUAL Windows)
       -DUMF_BUILD_SHARED_LIBRARY:BOOL=${UMF_BUILD_SHARED_LIBRARY}
       -DUMF_LINK_HWLOC_STATICALLY:BOOL=${UMF_LINK_HWLOC_STATICALLY}
       -DUMF_DISABLE_HWLOC:BOOL=${UMF_DISABLE_HWLOC}
+      -DSYCL_EMHASH_DIR:STRING=${SYCL_EMHASH_DIR}
       # Enable d suffix in UMF
       -DUMF_USE_DEBUG_POSTFIX:BOOL=ON
   )

sycl/cmake/modules/FetchEmhash.cmake

@@ -149,6 +149,15 @@ function(add_sycl_rt_library LIB_NAME LIB_OBJ_NAME)
     target_include_directories(${LIB_OBJ_NAME} PRIVATE ${zstd_INCLUDE_DIR})
   endif()
 
+  if (EMHASH_SYS_LOC)
+    # If we are using system emhash, include the system directory
+    # containing the emhash folder and note this location so
+    # other modules using emhash can find it.
+    cmake_path(GET EMHASH_SYS_LOC PARENT_PATH EMHASH_SYS_LOC)
+    set(SYCL_EMHASH_DIR ${EMHASH_SYS_LOC} CACHE INTERNAL "")
+    target_include_directories(${LIB_OBJ_NAME} PRIVATE ${EMHASH_SYS_LOC})
+  endif()
+
   # ur_win_proxy_loader
   if (WIN32)
     include_directories(${LLVM_EXTERNAL_SYCL_SOURCE_DIR}/ur_win_proxy_loader)

sycl/cmake/modules/FetchUnifiedRuntime.cmake

@@ -520,7 +520,7 @@ std::optional<ur_program_handle_t> context_impl::getProgramForDevImgs(
 
       auto KeyMappingsIt = KeyMap.find(OuterKey);
       assert(KeyMappingsIt != KeyMap.end());
-      auto CachedProgIt = Cache.find(KeyMappingsIt->second);
+      auto CachedProgIt = Cache.find((*KeyMappingsIt).second);
       assert(CachedProgIt != Cache.end());
       BuildRes = CachedProgIt->second;
     }