Revert to non-posic_memalign version of STREAM

Also request medium mcmodel to support large arrays with STREAM.

This reverts commit da10cd8.

examples/stream/CMakeLists.txt

@@ -4,7 +4,7 @@
 cmake_minimum_required (VERSION 3.18.3)
 
 find_package(OpenMP REQUIRED)
-set(STREAM_ARRAY_SIZE 400000000 CACHE STRING "STREAM benchmark array size")
+set(STREAM_ARRAY_SIZE 80000000 CACHE STRING "STREAM benchmark array size")
 
 project(stream C)
 
@@ -14,7 +14,10 @@ target_link_libraries(${PROJECT_NAME} PRIVATE OpenMP::OpenMP_CXX)
 if (MSVC)
 	target_compile_options(${PROJECT_NAME} PRIVATE /arch:AVX2)
 else()
-	target_compile_options(${PROJECT_NAME} PRIVATE -march=native)
+	target_compile_options(${PROJECT_NAME} PRIVATE -march=native -mcmodel=medium) # mcmodel needed for large global vars
+	if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
+		target_link_options(${PROJECT_NAME} PRIVATE -shared-intel)
+	endif()
 endif()
 
 project(llama-stream CXX)
@@ -30,5 +33,8 @@ target_link_libraries(${PROJECT_NAME} PRIVATE llama::llama OpenMP::OpenMP_CXX)
 if (MSVC)
 	target_compile_options(${PROJECT_NAME} PRIVATE /arch:AVX2)
 else()
-	target_compile_options(${PROJECT_NAME} PRIVATE -march=native)
+	target_compile_options(${PROJECT_NAME} PRIVATE -march=native -mcmodel=medium) # mcmodel needed for large global vars
+	if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
+		target_link_options(${PROJECT_NAME} PRIVATE -shared-intel)
+	endif()
 endif()

examples/stream/README.md

@@ -1,6 +1,7 @@
 This is a compliant version of the [STREAM benchmark](https://www.cs.virginia.edu/stream/).
 There are a few official such versions, which you can download [here](https://www.cs.virginia.edu/stream/FTP/Code/).
-The [original version](https://www.cs.virginia.edu/stream/FTP/Code/stream.c) places the input and ouput arrays into the static program segment.
-This causes the linker to complain because it breaks limits of ELF/PE (Linux/Windows executable file format) for big array sizes.
-We therefore took the [version using `posix_memalign`](https://www.cs.virginia.edu/stream/FTP/Code/Versions/stream_5-10_posix_memalign.c) for allocation.
-Some fixes were applied to make it compile with MSVC.
+This example is based on the [original version](https://www.cs.virginia.edu/stream/FTP/Code/stream.c) which places the input and ouput arrays into the static program segment.
+This requires to set corresponding command line flags, notable `-mcmodel=medium` for gcc, clang and icpx.
+Unfortunately, there is no equivalent for MSVC und you are therefore limited in the values of `STREAM_ARRAY_SIZE`.
+For MSVC, you can checkout the STREAM [version using `posix_memalign`](https://www.cs.virginia.edu/stream/FTP/Code/Versions/stream_5-10_posix_memalign.c) for allocation.
+Some fixes were applied to this version regardless to make it compile with MSVC.