Cycles: Speed up oneAPI GPU binaries compilation by using parallel instances

This change speeds up the compilation at the cost of higher memory usage.
CMake implementation checks the amount of available memory to spawn a
reasonable number of parallel compiler jobs.

intern/cycles/kernel/CMakeLists.txt

@@ -727,6 +727,17 @@ if(WITH_CYCLES_DEVICE_ONEAPI)
     ${SRC_UTIL_HEADERS}
   )
 
+  set (ONEAPI_OFFLINE_COMPILER_PARALLEL_JOBS 1)
+  if (WITH_CYCLES_ONEAPI_BINARIES)
+    cmake_host_system_information(RESULT AVAILABLE_MEMORY_AMOUNT QUERY AVAILABLE_PHYSICAL_MEMORY)
+    # Conservative value of peak consumption here, just to be fully sure that other backend compilers will have enough memory as well
+    set(ONEAPI_GPU_COMPILER_MEMORY_AT_PEAK_MB 8150)
+    math(EXPR ONEAPI_OFFLINE_COMPILER_PARALLEL_JOBS "${AVAILABLE_MEMORY_AMOUNT} / ${ONEAPI_GPU_COMPILER_MEMORY_AT_PEAK_MB}")
+    if (ONEAPI_OFFLINE_COMPILER_PARALLEL_JOBS LESS 1)
+      set(ONEAPI_OFFLINE_COMPILER_PARALLEL_JOBS 1)
+    endif()
+    message(STATUS "${ONEAPI_OFFLINE_COMPILER_PARALLEL_JOBS} instance(s) of oneAPI offline compiler will be used.")
+  endif()
   # SYCL_CPP_FLAGS is a variable that the user can set to pass extra compiler options
   set(sycl_compiler_flags
       ${CMAKE_CURRENT_SOURCE_DIR}/${SRC_KERNEL_DEVICE_ONEAPI}
@@ -735,6 +746,8 @@ if(WITH_CYCLES_DEVICE_ONEAPI)
       -fdelayed-template-parsing
       -mllvm -inlinedefault-threshold=300
       -mllvm -inlinehint-threshold=400
+      -fsycl-device-code-split=per_kernel
+      -fsycl-max-parallel-link-jobs=${ONEAPI_OFFLINE_COMPILER_PARALLEL_JOBS}
       -shared
       -DWITH_ONEAPI
       -ffast-math