GPU Compile Flags: Set denormals to zero unconditionally

Before I kept them in NO_FAST_MATH mode, but this yields warning by nvcc,
which for some stupid reason cannot be disabled. And in principle,
with denormals globally disabled, it is also deterministic again,
and disabling it everywhere is as simple as enabling them everywhere.

Author: davidrohr

GPU/GPUTracking/Base/opencl/CMakeLists.txt

@@ -27,7 +27,7 @@ set(OCL_FLAGS -Dcl_clang_storage_class_specifiers -cl-std=CLC++2021)
 if(NOT GPUCA_DETERMINISTIC_MODE GREATER_EQUAL ${GPUCA_DETERMINISTIC_MODE_MAP_NO_FAST_MATH})
   set(OCL_FLAGS ${OCL_FLAGS} -cl-denorms-are-zero -cl-mad-enable -cl-no-signed-zeros -cl-fast-relaxed-math)
 else()
-  set(OCL_FLAGS ${OCL_FLAGS} -cl-fp32-correctly-rounded-divide-sqrt)
+  set(OCL_FLAGS ${OCL_FLAGS} ${GPUCA_OCL_NO_FAST_MATH_FLAGS})
 endif()
 set(OCL_DEFINECL "-D$<JOIN:$<TARGET_PROPERTY:O2::GPUTracking,COMPILE_DEFINITIONS>,$<SEMICOLON>-D>"
             "-I$<JOIN:$<FILTER:$<TARGET_PROPERTY:O2::GPUTracking,INCLUDE_DIRECTORIES>,EXCLUDE,^/usr/include/?>,$<SEMICOLON>-I>"

GPU/GPUTracking/CMakeLists.txt

@@ -17,10 +17,10 @@ set(MODULE GPUTracking)
 if(GPUCA_DETERMINISTIC_MODE GREATER_EQUAL ${GPUCA_DETERMINISTIC_MODE_MAP_NO_FAST_MATH})
   set(CMAKE_CXX_FLAGS_${CMAKE_BUILD_TYPE_UPPER} "${CMAKE_CXX_FLAGS_${CMAKE_BUILD_TYPE_UPPER}} ${GPUCA_CXX_NO_FAST_MATH_FLAGS}")
   if(GPUCA_DETERMINISTIC_MODE GREATER_EQUAL ${GPUCA_DETERMINISTIC_MODE_MAP_OPTO2})
-    set(CMAKE_CXX_FLAGS_${CMAKE_BUILD_TYPE_UPPER} "${CMAKE_CXX_FLAGS_${CMAKE_BUILD_TYPE_UPPER}} -O2")
+    set(CMAKE_CXX_FLAGS_${CMAKE_BUILD_TYPE_UPPER} "${CMAKE_CXX_FLAGS_${CMAKE_BUILD_TYPE_UPPER}} -O2 ${GPUCA_CLANG_FTZ}")
   endif()
 elseif(NOT CMAKE_BUILD_TYPE_UPPER STREQUAL "DEBUG")
-  set(CMAKE_CXX_FLAGS_${CMAKE_BUILD_TYPE_UPPER} "${CMAKE_CXX_FLAGS_${CMAKE_BUILD_TYPE_UPPER}} -O3 -ffast-math")
+  set(CMAKE_CXX_FLAGS_${CMAKE_BUILD_TYPE_UPPER} "${CMAKE_CXX_FLAGS_${CMAKE_BUILD_TYPE_UPPER}} -O3 -ffast-math ${GPUCA_CLANG_FTZ}")
 endif()
 
 include(cmake/helpers.cmake)

GPU/GPUTracking/Definitions/GPUDefGPUParameters.h

@@ -603,11 +603,11 @@
 // #define GPUCA_KERNEL_DEBUGGER_OUTPUT
 
 // Some assertions to make sure out parameters are not invalid
-  static_assert(GPUCA_MAXN >= GPUCA_NEIGHBOURS_FINDER_MAX_NNEIGHUP, "Invalid GPUCA_NEIGHBOURS_FINDER_MAX_NNEIGHUP");
-  static_assert(GPUCA_ROW_COUNT >= GPUCA_TRACKLET_SELECTOR_HITS_REG_SIZE, "Invalid GPUCA_TRACKLET_SELECTOR_HITS_REG_SIZE");
-  #ifdef GPUCA_GPUCODE
-    static_assert(GPUCA_M_FIRST(GPUCA_LB_GPUTPCCompressionKernels_step1unattached) * 2 <= GPUCA_TPC_COMP_CHUNK_SIZE, "Invalid GPUCA_TPC_COMP_CHUNK_SIZE");
-  #endif
+static_assert(GPUCA_MAXN >= GPUCA_NEIGHBOURS_FINDER_MAX_NNEIGHUP, "Invalid GPUCA_NEIGHBOURS_FINDER_MAX_NNEIGHUP");
+static_assert(GPUCA_ROW_COUNT >= GPUCA_TRACKLET_SELECTOR_HITS_REG_SIZE, "Invalid GPUCA_TRACKLET_SELECTOR_HITS_REG_SIZE");
+#ifdef GPUCA_GPUCODE
+  static_assert(GPUCA_M_FIRST(GPUCA_LB_GPUTPCCompressionKernels_step1unattached) * 2 <= GPUCA_TPC_COMP_CHUNK_SIZE, "Invalid GPUCA_TPC_COMP_CHUNK_SIZE");
+#endif
 
 // Derived parameters
 #ifdef GPUCA_USE_TEXTURES
@@ -621,5 +621,5 @@
 #define GPUCA_NEW_ALIGNMENT (std::align_val_t{GPUCA_BUFFER_ALIGNMENT})
 #define GPUCA_OPERATOR_NEW_ALIGNMENT ,GPUCA_NEW_ALIGNMENT
 
-    // clang-format on
+// clang-format on
 #endif

dependencies/FindO2GPU.cmake

@@ -84,8 +84,14 @@ elseif(NOT GPUCA_DETERMINISTIC_MODE MATCHES "^[0-9]+$")
   set(GPUCA_DETERMINISTIC_MODE ${GPUCA_DETERMINISTIC_MODE_MAP_${GPUCA_DETERMINISTIC_MODE}})
   message(STATUS "Set to ${GPUCA_DETERMINISTIC_MODE}")
 endif()
-set(GPUCA_CXX_NO_FAST_MATH_FLAGS "-fno-fast-math -ffp-contract=off")
-set(GPUCA_CUDA_NO_FAST_MATH_FLAGS "--ftz=false --prec-div=true --prec-sqrt=true --fmad false")
+if (CMAKE_SYSTEM_NAME MATCHES Darwin OR NOT CMAKE_SYSTEM_PROCESSOR MATCHES "(x86)|(X86)|(amd64)|(AMD64)")
+  set(GPUCA_CLANG_FTZ "")
+else()
+  set(GPUCA_CLANG_FTZ "-mdaz-ftz")
+endif()
+set(GPUCA_CXX_NO_FAST_MATH_FLAGS "-fno-fast-math -ffp-contract=off ${GPUCA_CLANG_FTZ}")
+set(GPUCA_CUDA_NO_FAST_MATH_FLAGS "--ftz=true --prec-div=true --prec-sqrt=true --fmad false")
+set(GPUCA_OCL_NO_FAST_MATH_FLAGS -cl-fp32-correctly-rounded-divide-sqrt -cl-denorms-are-zero)
 if(GPUCA_DETERMINISTIC_MODE GREATER_EQUAL ${GPUCA_DETERMINISTIC_MODE_MAP_WHOLEO2})
   add_definitions(-DGPUCA_DETERMINISTIC_MODE)
   set(CMAKE_CXX_FLAGS_${CMAKE_BUILD_TYPE_UPPER} "${CMAKE_CXX_FLAGS_${CMAKE_BUILD_TYPE_UPPER}} ${GPUCA_CXX_NO_FAST_MATH_FLAGS}")
@@ -172,7 +178,7 @@ if(ENABLE_CUDA)
   endif()
 endif()
 
-# ---------------------------------- HIP ----------------------------------
+# ---------------------------------- OpenCL ----------------------------------
 if(ENABLE_OPENCL)
   find_package(OpenCL)
   if(ENABLE_OPENCL AND NOT ENABLE_OPENCL STREQUAL "AUTO")