Set -ffp-contract=off to aid cross-platform bit-reproducability

Currently when running VVenC on Arm machines, the md5sum of the output
video differs from x86 even when specifying `--SIMD=SCALAR`. This is due
to the default `-ffp-contract` setting on GCC and LLVM enabling fused
multiply-add instructions for floating-point arithmetic, which causes
small bit differences due to the lack of intermediate rounding in these
instructions.

Setting `-ffp-contract=off` disables this behaviour, ensuring the md5sum
output for the tested video now matches for `--SIMD=SCALAR`. There are
some remaining differences when other SIMD settings are used on Arm
which will be resolved in a separate PR.

Author: georges-arm

CMakeLists.txt

@@ -23,6 +23,13 @@ message( STATUS "CMAKE_MODULE_PATH: updating module path to: ${CMAKE_MODULE_PATH
 
 include( vvencCompilerSupport )
 
+# Avoid non-portable fp contractions leading to cross-platform bit-differences.
+if( CMAKE_CXX_COMPILER_ID STREQUAL "GNU" OR
+    CMAKE_CXX_COMPILER_ID STREQUAL "Clang" )
+  set( CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -ffp-contract=off" )
+  set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -ffp-contract=off" )
+endif()
+
 detect_target_architecture( VVENC_TARGET_ARCH )
 if( VVENC_TARGET_ARCH MATCHES "ARM\|AARCH64" )
   set( VVENC_ARM_SIMD_DEFAULT TRUE )