GPU RTC: Add deterministic mode

davidrohr · davidrohr · commit bccffa5eeccd · 2025-03-25T20:13:27.000+01:00
diff --git a/GPU/GPUTracking/Base/cuda/CMakeLists.txt b/GPU/GPUTracking/Base/cuda/CMakeLists.txt
@@ -85,22 +85,28 @@ add_custom_target(${MODULE}_CUDA_SRC_CHK ALL DEPENDS ${CMAKE_CURRENT_BINARY_DIR}
 add_custom_command(
     OUTPUT ${GPU_RTC_BIN}.command
     COMMAND echo -n "${CMAKE_CUDA_COMPILER} ${GPU_RTC_FLAGS_SEPARATED} ${GPU_RTC_DEFINES} -fatbin" > ${GPU_RTC_BIN}.command
-    COMMAND_EXPAND_LISTS
-    VERBATIM
+    COMMAND_EXPAND_LISTS VERBATIM
     COMMENT "Preparing CUDA RTC command file ${GPU_RTC_BIN}.command"
 )
 create_binary_resource(${GPU_RTC_BIN}.command ${GPU_RTC_BIN}.command.o)
 
 add_custom_command(
     OUTPUT ${GPU_RTC_BIN}.command.arch
     COMMAND echo -n "${GPU_RTC_FLAGS_ARCH}" > ${GPU_RTC_BIN}.command.arch
-    COMMAND_EXPAND_LISTS
-    VERBATIM
-    COMMENT "Preparing CUDA RTC ARCH file ${GPU_RTC_BIN}.command.arch"
+    COMMAND_EXPAND_LISTS VERBATIM
+    COMMENT "Preparing CUDA RTC ARCH command file ${GPU_RTC_BIN}.command.arch"
 )
 create_binary_resource(${GPU_RTC_BIN}.command.arch ${GPU_RTC_BIN}.command.arch.o)
 
-set(SRCS ${SRCS} ${GPU_RTC_BIN}.src.o ${GPU_RTC_BIN}.command.o ${GPU_RTC_BIN}.command.arch.o)
+add_custom_command(
+    OUTPUT ${GPU_RTC_BIN}.command.no_fast_math
+    COMMAND echo -n "${GPUCA_CUDA_NO_FAST_MATH_FLAGS}" > ${GPU_RTC_BIN}.command.no_fast_math
+    COMMAND_EXPAND_LISTS VERBATIM
+    COMMENT "Preparing CUDA RTC NO_FAST_MATH command file ${GPU_RTC_BIN}.command.arch"
+)
+create_binary_resource(${GPU_RTC_BIN}.command.no_fast_math ${GPU_RTC_BIN}.command.no_fast_math.o)
+
+set(SRCS ${SRCS} ${GPU_RTC_BIN}.src.o ${GPU_RTC_BIN}.command.o ${GPU_RTC_BIN}.command.arch.o ${GPU_RTC_BIN}.command.no_fast_math.o)
 # -------------------------------- End RTC -------------------------------------------------------
 
 if(ALIGPU_BUILD_TYPE STREQUAL "O2")
diff --git a/GPU/GPUTracking/Base/cuda/GPUReconstructionCUDAGenRTC.cxx b/GPU/GPUTracking/Base/cuda/GPUReconstructionCUDAGenRTC.cxx
@@ -29,11 +29,13 @@ using namespace o2::gpu;
 QGET_LD_BINARY_SYMBOLS(GPUReconstructionCUDArtc_src);
 QGET_LD_BINARY_SYMBOLS(GPUReconstructionCUDArtc_command);
 QGET_LD_BINARY_SYMBOLS(GPUReconstructionCUDArtc_command_arch);
+QGET_LD_BINARY_SYMBOLS(GPUReconstructionCUDArtc_command_no_fast_math);
 
 int32_t GPUReconstructionCUDA::genRTC(std::string& filename, uint32_t& nCompile)
 {
   std::string rtcparam = std::string("#define GPUCA_RTC_CODE\n") +
                          std::string(mProcessingSettings.rtc.optSpecialCode ? "#define GPUCA_RTC_SPECIAL_CODE(...) __VA_ARGS__\n" : "#define GPUCA_RTC_SPECIAL_CODE(...)\n") +
+                         std::string(mProcessingSettings.rtc.deterministic ? "#define GPUCA_DETERMINISTIC_CODE(det, indet) det\n" : "") +
                          GPUParamRTC::generateRTCCode(param(), mProcessingSettings.rtc.optConstexpr);
   if (filename == "") {
     filename = "/tmp/o2cagpu_rtc_";
@@ -52,6 +54,7 @@ int32_t GPUReconstructionCUDA::genRTC(std::string& filename, uint32_t& nCompile)
   std::string baseCommand = (mProcessingSettings.RTCprependCommand != "" ? (mProcessingSettings.RTCprependCommand + " ") : "");
   baseCommand += (getenv("O2_GPU_RTC_OVERRIDE_CMD") ? std::string(getenv("O2_GPU_RTC_OVERRIDE_CMD")) : std::string(_binary_GPUReconstructionCUDArtc_command_start, _binary_GPUReconstructionCUDArtc_command_len));
   baseCommand += std::string(" ") + (mProcessingSettings.RTCoverrideArchitecture != "" ? mProcessingSettings.RTCoverrideArchitecture : std::string(_binary_GPUReconstructionCUDArtc_command_arch_start, _binary_GPUReconstructionCUDArtc_command_arch_len));
+  baseCommand += mProcessingSettings.rtc.deterministic ? (std::string(" ") + std::string(_binary_GPUReconstructionCUDArtc_command_no_fast_math_start, _binary_GPUReconstructionCUDArtc_command_no_fast_math_len)) : std::string("");
 
   char shasource[21], shaparam[21], shacmd[21], shakernels[21];
   if (mProcessingSettings.rtc.cacheOutput) {
diff --git a/GPU/GPUTracking/Base/cuda/GPUReconstructionCUDArtc.cu b/GPU/GPUTracking/Base/cuda/GPUReconstructionCUDArtc.cu
@@ -15,6 +15,7 @@
 #define GPUCA_GPUCODE_GENRTC
 #define GPUCA_GPUCODE_COMPILEKERNELS
 #define GPUCA_RTC_SPECIAL_CODE(...) GPUCA_RTC_SPECIAL_CODE(__VA_ARGS__)
+#define GPUCA_DETERMINISTIC_CODE(...) GPUCA_DETERMINISTIC_CODE(__VA_ARGS__)
 #include "GPUReconstructionCUDADef.h"
 #include "GPUReconstructionIncludesDeviceAll.h"
 
diff --git a/GPU/GPUTracking/Base/hip/CMakeLists.txt b/GPU/GPUTracking/Base/hip/CMakeLists.txt
@@ -123,22 +123,28 @@ add_custom_target(${MODULE}_HIP_SRC_CHK ALL DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/
 add_custom_command(
     OUTPUT ${GPU_RTC_BIN}.command
     COMMAND echo -n "${CMAKE_HIP_COMPILER} ${GPU_RTC_FLAGS_SEPARATED} ${GPU_RTC_DEFINES} -x hip --cuda-device-only" > ${GPU_RTC_BIN}.command
-    COMMAND_EXPAND_LISTS
-    VERBATIM
+    COMMAND_EXPAND_LISTS VERBATIM
     COMMENT "Preparing HIP RTC command file ${GPU_RTC_BIN}.command"
 )
 create_binary_resource(${GPU_RTC_BIN}.command ${GPU_RTC_BIN}.command.o)
 
 add_custom_command(
     OUTPUT ${GPU_RTC_BIN}.command.arch
     COMMAND echo -n "${GPU_RTC_FLAGS_ARCH}" > ${GPU_RTC_BIN}.command.arch
-    COMMAND_EXPAND_LISTS
-    VERBATIM
-    COMMENT "Preparing HIP RTC ARCH file ${GPU_RTC_BIN}.command.arch"
+    COMMAND_EXPAND_LISTS VERBATIM
+    COMMENT "Preparing HIP RTC ARCH command file ${GPU_RTC_BIN}.command.arch"
 )
 create_binary_resource(${GPU_RTC_BIN}.command.arch ${GPU_RTC_BIN}.command.arch.o)
 
-set(SRCS ${SRCS} ${GPU_RTC_BIN}.src.o ${GPU_RTC_BIN}.command.o ${GPU_RTC_BIN}.command.arch.o)
+add_custom_command(
+    OUTPUT ${GPU_RTC_BIN}.command.no_fast_math
+    COMMAND echo -n "${GPUCA_CXX_NO_FAST_MATH_FLAGS}" > ${GPU_RTC_BIN}.command.no_fast_math
+    COMMAND_EXPAND_LISTS VERBATIM
+    COMMENT "Preparing HIP RTC NO_FAST_MATH command file ${GPU_RTC_BIN}.command.no_fast_math"
+)
+create_binary_resource(${GPU_RTC_BIN}.command.no_fast_math ${GPU_RTC_BIN}.command.no_fast_math.o)
+
+set(SRCS ${SRCS} ${GPU_RTC_BIN}.src.o ${GPU_RTC_BIN}.command.o ${GPU_RTC_BIN}.command.arch.o ${GPU_RTC_BIN}.command.no_fast_math.o)
 # -------------------------------- End RTC -------------------------------------------------------
 
 if(ALIGPU_BUILD_TYPE STREQUAL "O2")
diff --git a/GPU/GPUTracking/Definitions/GPUSettingsList.h b/GPU/GPUTracking/Definitions/GPUSettingsList.h
@@ -208,6 +208,7 @@ BeginSubConfig(GPUSettingsProcessingRTC, rtc, configStandalone.proc, "RTC", 0, "
 AddOption(cacheOutput, bool, false, "", 0, "Cache RTC compilation results")
 AddOption(optConstexpr, bool, true, "", 0, "Replace constant variables by static constexpr expressions")
 AddOption(optSpecialCode, int8_t, -1, "", 0, "Insert GPUCA_RTC_SPECIAL_CODE special code during RTC")
+AddOption(deterministic, bool, false, "", 0, "Compile RTC in deterministic mode, with NO_FAST_MATH flags and GPUCA_DETERMINISTIC_MODE define")
 AddOption(compilePerKernel, bool, true, "", 0, "Run one RTC compilation per kernel")
 AddOption(enable, bool, false, "", 0, "Use RTC to optimize GPU code")
 AddOption(runTest, int32_t, 0, "", 0, "Do not run the actual benchmark, but just test RTC compilation (1 full test, 2 test only compilation)")
