GPU: Add GPUCA_DETERMINISTIC_NO_FTC

Author: davidrohr

GPU/GPUTracking/Base/GPUReconstruction.cxx

@@ -263,17 +263,17 @@ int32_t GPUReconstruction::InitPhaseBeforeDevice()
   if (GetProcessingSettings().debugLevel > 0) {
     mProcessingSettings->recoTaskTiming = true;
   }
-  if (GetProcessingSettings().deterministicGPUReconstruction == -1) {
+  bool detMode = false;
 #ifdef GPUCA_DETERMINISTIC_MODE
-    mProcessingSettings->deterministicGPUReconstruction = 1;
-#else
-    mProcessingSettings->deterministicGPUReconstruction = GetProcessingSettings().debugLevel >= 6;
+  detMode = true;
 #endif
+  if (GetProcessingSettings().deterministicGPUReconstruction == -1) {
+    mProcessingSettings->deterministicGPUReconstruction = detMode ? 1 : (GetProcessingSettings().debugLevel >= 6);
   }
   if (GetProcessingSettings().deterministicGPUReconstruction) {
-#ifndef GPUCA_DETERMINISTIC_MODE
-    GPUError("WARNING, deterministicGPUReconstruction needs GPUCA_DETERMINISTIC_MODE for being fully deterministic, without only most indeterminism by concurrency is removed, but floating point effects remain!");
-#endif
+    if (!detMode) {
+      GPUError("WARNING, deterministicGPUReconstruction needs GPUCA_DETERMINISTIC_MODE for being fully deterministic, without only most indeterminism by concurrency is removed, but floating point effects remain!");
+    }
     if (mProcessingSettings->debugLevel >= 6 && ((mProcessingSettings->debugMask + 1) & mProcessingSettings->debugMask)) {
       GPUError("WARNING: debugMask %d - debug output might not be deterministic with intermediate steps missing", mProcessingSettings->debugMask);
     }
@@ -283,9 +283,9 @@ int32_t GPUReconstruction::InitPhaseBeforeDevice()
     }
     mProcessingSettings->rtc.deterministic = 1;
   } else {
-#ifdef GPUCA_DETERMINISTIC_MODE
-    GPUError("WARNING, compiled with GPUCA_DETERMINISTIC_MODE but deterministicGPUReconstruction not set, only compile-time determinism and deterministic math enforced, not fully deterministic!");
-#endif
+    if (detMode) {
+      GPUError("WARNING, compiled with GPUCA_DETERMINISTIC_MODE but deterministicGPUReconstruction not set, only compile-time determinism and deterministic math enforced, not fully deterministic!");
+    }
   }
   if (GetProcessingSettings().deterministicGPUReconstruction && GetProcessingSettings().debugLevel >= 6) {
     mProcessingSettings->nTPCClustererLanes = 1;

GPU/GPUTracking/DataCompression/GPUTPCCompressionKernels.cxx

@@ -271,7 +271,7 @@ GPUdii() void GPUTPCCompressionKernels::Thread<GPUTPCCompressionKernels::step1un
 #ifdef GPUCA_GPUCODE
         static_assert(GPUCA_GET_THREAD_COUNT(GPUCA_LB_GPUTPCCompressionKernels_step1unattached) * 2 <= constants::TPC_COMP_CHUNK_SIZE);
 #endif
-#ifdef GPUCA_DETERMINISTIC_MODE // Not using GPUCA_DETERMINISTIC_CODE, which is enforced in TPC compression
+#ifdef GPUCA_DETERMINISTIC_MODE
         CAAlgo::sortInBlock(sortBuffer, sortBuffer + count, GPUTPCCompressionKernels_Compare<GPUSettings::SortZPadTime>(clusters->clusters[iSector][iRow]));
 #else  // GPUCA_DETERMINISTIC_MODE
         if (param.rec.tpc.compressionSortOrder == GPUSettings::SortZPadTime) {

GPU/GPUTracking/Definitions/GPUSettingsList.h

@@ -613,7 +613,7 @@ AddOption(inputcontrolmem, uint64_t, 0, "inputMemory", 0, "Use predefined input
 AddOption(cpuAffinity, int32_t, -1, "", 0, "Pin CPU affinity to this CPU core", min(-1))
 AddOption(fifoScheduler, bool, false, "", 0, "Use FIFO realtime scheduler", message("Setting FIFO scheduler: %s"))
 AddOption(fpe, int8_t, -1, "", 0, "Trap on floating point exceptions (-1 = if no ffast-math)")
-AddOption(flushDenormals, bool, true, "", 0, "Enable FTZ and DAZ (Flush all denormals to zero)")
+AddOption(flushDenormals, int8_t, -1, "", 0, "Enable FTZ and DAZ (Flush all denormals to zero), -1 = enable automatically if not prevented by deterministic mode")
 AddOption(solenoidBzNominalGPU, float, -1e6f, "", 0, "Field strength of solenoid Bz in kGaus")
 AddOption(constBz, bool, false, "", 0, "Force constand Bz")
 AddOption(overrideMaxTimebin, bool, false, "", 0, "Override max time bin setting for continuous data with max time bin in time frame")

GPU/GPUTracking/Standalone/Benchmark/CMakeLists.txt

@@ -35,6 +35,10 @@ install(DIRECTORY ../tools DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/GPU)
 install(DIRECTORY ../../Definitions/Parameters/ DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/GPU/tools FILES_MATCHING REGEX "\\.(python|sh|cmake)")
 target_compile_definitions(${targetName} PRIVATE $<TARGET_PROPERTY:O2::GPUTracking,COMPILE_DEFINITIONS>)
 
+if(GPUCA_DETERMINISTIC_MODE GREATER_EQUAL ${GPUCA_DETERMINISTIC_MODE_MAP_NO_FAST_MATH} AND GPUCA_DETERMINISTIC_NO_FTZ)
+target_compile_definitions(${targetName} PRIVATE GPUCA_DETERMINISTIC_NO_FTZ)
+endif()
+
 if(ROOT_FOUND)
   target_sources(${targetName} PRIVATE ../../qa/genEvents.cxx)
 endif()

GPU/GPUTracking/Standalone/Benchmark/standalone.cxx

@@ -141,7 +141,14 @@ int32_t ReadConfiguration(int argc, char** argv)
 #endif
     feenableexcept(FE_INVALID | FE_DIVBYZERO | FE_OVERFLOW);
   }
-  if (configStandalone.flushDenormals) {
+  bool detMode = false, noFTZMode = false;
+#ifdef GPUCA_DETERMINISTIC_MODE
+  detMode = true;
+#endif
+#ifdef GPUCA_DETERMINISTIC_NO_FTZ
+  noFTZMode = true;
+#endif
+  if (configStandalone.flushDenormals >= 1 || (configStandalone.flushDenormals == -1 && (configStandalone.proc.deterministicGPUReconstruction >= 1 || (configStandalone.proc.deterministicGPUReconstruction == -1 && detMode)) && !noFTZMode)) {
     disable_denormals();
   }
 

GPU/GPUTracking/Standalone/cmake/config.cmake

@@ -19,26 +19,27 @@ set(GPUCA_CONFIG_VC 1)
 set(GPUCA_CONFIG_FMT 1)
 set(GPUCA_CONFIG_ROOT 1)
 set(GPUCA_CONFIG_ONNX 0)
-set(GPUCA_BUILD_EVENT_DISPLAY 1)
-set(GPUCA_BUILD_EVENT_DISPLAY_FREETYPE 1)
-set(GPUCA_BUILD_EVENT_DISPLAY_VULKAN 1)
-set(GPUCA_BUILD_EVENT_DISPLAY_WAYLAND 1)
-set(GPUCA_BUILD_EVENT_DISPLAY_QT 1)
-set(GPUCA_CONFIG_GL3W 0)
-set(GPUCA_CONFIG_O2 1)
-set(GPUCA_BUILD_DEBUG 0)
-set(GPUCA_BUILD_DEBUG_SANITIZE 0)
-set(GPUCA_BUILD_DEBUG_HOSTONLY 0)
-set(GPUCA_DETERMINISTIC_MODE 0)             # OFF / NO_FAST_MATH / OPTO2 / GPU / WHOLEO2
-#set(GPUCA_CUDA_GCCBIN c++-14)
-#set(GPUCA_OPENCL_CLANGBIN clang-20)
-set(HIP_AMDGPUTARGET "default")             # "gfx906;gfx908;gfx90a"
-set(CUDA_COMPUTETARGET "default")           # 86 89
-#set(GPUCA_CUDA_COMPILE_MODE perkernel)     # onefile / perkernel / rtc
-#set(GPUCA_HIP_COMPILE_MODE perkernel)
-#set(GPUCA_RTC_NO_COMPILED_KERNELS 1)
-#set(GPUCA_KERNEL_RESOURCE_USAGE_VERBOSE 1)
-#set(GPUCA_CONFIG_COMPILER gcc)             # gcc / clang
-#set(GPUCA_CONFIG_WERROR 1)
-#add_definitions(-DGPUCA_GPU_DEBUG_PRINT)
-#set(GPUCA_OVERRIDE_PARAMETER_FILE "foo.csv")
+set(GPUCA_BUILD_EVENT_DISPLAY 1)              # Enable compilation of event display
+set(GPUCA_BUILD_EVENT_DISPLAY_FREETYPE 1)     # Use FreeType library to render fonts for event display
+set(GPUCA_BUILD_EVENT_DISPLAY_VULKAN 1)       # Enable Vulkan backend for event display (otherwise only OpenGL / Win32)
+set(GPUCA_BUILD_EVENT_DISPLAY_WAYLAND 1)      # Enable native wayland frontend for event display
+set(GPUCA_BUILD_EVENT_DISPLAY_QT 1)           # Use QT for Event Display GUI
+set(GPUCA_CONFIG_GL3W 0)                      # Use GL3W instead of glew
+set(GPUCA_CONFIG_O2 1)                        # Compile for O2 data, 0 for Run 2 data
+set(GPUCA_BUILD_DEBUG 0)                      # Enable debug mode (-O0, -ggdb, enable asserts)
+set(GPUCA_BUILD_DEBUG_SANITIZE 0)             # Enable undefined behavior and address sanitizers
+set(GPUCA_BUILD_DEBUG_HOSTONLY 0)             # Only compile host code in debug mode, GPU code compiled normally
+set(GPUCA_DETERMINISTIC_MODE 0)               # OFF / NO_FAST_MATH / OPTO2 / GPU / WHOLEO2
+set(GPUCA_DETERMINISTIC_NO_FTZ 0)             # If 1 and deterministic mode active, do not apply flush denormals to zero
+#set(GPUCA_CUDA_GCCBIN c++-14)                # Override which GCC to use for CUDA
+#set(GPUCA_OPENCL_CLANGBIN clang-20)          # Override which clang to use for OpenCL
+set(HIP_AMDGPUTARGET "default")               # Set AMD GPU tragets to compile for: e.g. "gfx906;gfx908;gfx90a"
+set(CUDA_COMPUTETARGET "default")             # Set NVIDIA GPU targets to compile for: e.g. "89;120"
+#set(GPUCA_CUDA_COMPILE_MODE perkernel)       # Mode to compile kernels for CUDA: onefile / perkernel / rtc
+#set(GPUCA_HIP_COMPILE_MODE perkernel)        # Mode to compile kernels for HIP:  onefile / perkernel / rtc
+#set(GPUCA_RTC_NO_COMPILED_KERNELS 1)         # Do not compile "perkernel" kernels at compile time, support only RTC
+#set(GPUCA_KERNEL_RESOURCE_USAGE_VERBOSE 1)   # Verbose resource usage output during kernel compilation
+#set(GPUCA_CONFIG_COMPILER gcc)               # Compiler to use for standalone compilation: gcc / clang
+#set(GPUCA_CONFIG_WERROR 1)                   # Enforce Werror
+#add_definitions(-DGPUCA_GPU_DEBUG_PRINT)     # Enable LOG(...) macros and GPUInfo(...) etc. in GPU code
+#set(GPUCA_OVERRIDE_PARAMETER_FILE "foo.csv") # Override the CSV or JSON file that contains GPU parameters

GPU/GPUTracking/utils/qmaths_helpers.h

@@ -26,7 +26,7 @@
 
 static void disable_denormals()
 {
-#if !(defined(__ARM_NEON) || defined(__aarch64__)) && __has_include(<xmmintrin.h>) // clang-format off
+#if !(defined(__ARM_NEON) || defined(__aarch64__)) && __has_include(<xmmintrin.h>)
   #if defined(_MM_FLUSH_ZERO_OFF) && defined(_MM_DENORMALS_ZERO_ON)
     _MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
     _MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);

GPU/documentation/deterministic-mode.md

@@ -30,3 +30,9 @@ Beyond comparing only the number of clusters and number of tracks, it is also po
 It will create a dump container all (most) intermediate results in text form, which can be compared. The output files is called `CPU.out` if using the CPU backend, and `GPU.out` for the GPU backend.
 Note that the dump files will be huge and the processing will be slow and consume much more memory than normal with `--debug 6 . It has been tested with datasets containing up to 50 Pb-Pb collisions, and might fail for larger data.
 The dump files (if the deterministic mode is used with both compile- and runtime-activation), the files should be 100% identical and can just be compared with `diff`.
+
+By default, the deterministic mode will apply flush-to-zero and denormals-are-zero to denormal floats.
+This can be disabled bia `-DDGPUCA_DETERMINISTIC_MODE`.
+Note that some GPUs cannot do precise float computation with denormals flushed to zero, while other GPUs do not support denormals at all.
+Thus, comparison between CPU and GPU deterministic results might require that this setting is either set or not set.
+CPU results for the 2 cases will always differ, since the floating point math will be slightly different.

dependencies/FindO2GPU.cmake

@@ -10,7 +10,7 @@
 # or submit itself to any jurisdiction.
 
 # NOTE!!!! - Whenever this file is changed, move it over to alidist/resources
-# FindO2GPU.cmake Version 14
+# FindO2GPU.cmake Version 15
 
 set(CUDA_COMPUTETARGET_DEFAULT_FULL 80-real 86-real 89-real 120-real 75-virtual)
 set(HIP_AMDGPUTARGET_DEFAULT_FULL gfx906;gfx908)
@@ -137,16 +137,23 @@ elseif(NOT GPUCA_DETERMINISTIC_MODE MATCHES "^[0-9]+$")
   endif()
   set(GPUCA_DETERMINISTIC_MODE ${GPUCA_DETERMINISTIC_MODE_MAP_${GPUCA_DETERMINISTIC_MODE}})
 endif()
-if (CMAKE_SYSTEM_NAME MATCHES Darwin OR NOT CMAKE_SYSTEM_PROCESSOR MATCHES "(x86)|(X86)|(amd64)|(AMD64)")
+if(GPUCA_DETERMINISTIC_NO_FTZ)
   set(GPUCA_CXX_DENORMALS_FLAGS "")
+  set(GPUCA_CUDA_DENORMALS_FLAGS "--ftz=false")
+  set(GPUCA_OCL_DENORMALS_FLAGS "")
+  set(GPUCA_HIP_DENORMALS_FLAGS "-fno-gpu-flush-denormals-to-zero")
 else()
-  set(GPUCA_CXX_DENORMALS_FLAGS "-mdaz-ftz")
+  if (CMAKE_SYSTEM_NAME MATCHES Darwin OR NOT CMAKE_SYSTEM_PROCESSOR MATCHES "(x86)|(X86)|(amd64)|(AMD64)")
+    set(GPUCA_CXX_DENORMALS_FLAGS "")
+  else()
+    set(GPUCA_CXX_DENORMALS_FLAGS "-mdaz-ftz")
+  endif()
+  set(GPUCA_CUDA_DENORMALS_FLAGS "--ftz=true")
+  set(GPUCA_OCL_DENORMALS_FLAGS "-cl-denorms-are-zero")
+  set(GPUCA_HIP_DENORMALS_FLAGS "-fgpu-flush-denormals-to-zero")
 endif()
-set(GPUCA_CUDA_DENORMALS_FLAGS "--ftz=true")
-set(GPUCA_OCL_DENORMALS_FLAGS "-cl-denorms-are-zero")
-set(GPUCA_HIP_DENORMALS_FLAGS "-fgpu-flush-denormals-to-zero")
 set(GPUCA_CXX_NO_FAST_MATH_FLAGS "-fno-fast-math -ffp-contract=off")
-set(GPUCA_CUDA_NO_FAST_MATH_FLAGS "--prec-div=true --prec-sqrt=true --fmad false")
+set(GPUCA_CUDA_NO_FAST_MATH_FLAGS "--prec-div=true --prec-sqrt=true --fmad false -Xcompiler -fno-fast-math -Xcompiler -ffp-contract=off")
 set(GPUCA_OCL_NO_FAST_MATH_FLAGS -cl-fp32-correctly-rounded-divide-sqrt )
 if(GPUCA_DETERMINISTIC_MODE GREATER_EQUAL ${GPUCA_DETERMINISTIC_MODE_MAP_WHOLEO2})
   add_definitions(-DGPUCA_DETERMINISTIC_MODE)