[PROTON][XPU] Always build XpuPti profiler

Signed-off-by: Anatoly Myachev <[email protected]>

Author: anmyachev

.github/workflows/build-test-reusable.yml

@@ -118,10 +118,6 @@ jobs:
           ref: ${{ inputs.pytorch_ref }}
           mode: ${{ inputs.pytorch_mode }}
 
-      - name: Build Proton with XPU support
-        run: |
-          echo TRITON_BUILD_PROTON_XPU=1 | tee -a $GITHUB_ENV
-
       - name: Build Triton
         uses: ./.github/actions/setup-triton
         with:

setup.py

@@ -519,7 +519,6 @@ def build_extension(self, ext):
         # environment variables we will pass through to cmake
         passthrough_args = [
             "TRITON_BUILD_PROTON",
-            "TRITON_BUILD_PROTON_XPU",
             "TRITON_BUILD_WITH_CCACHE",
             "TRITON_PARALLEL_LINK_JOBS",
         ]

third_party/proton/CMakeLists.txt

@@ -3,11 +3,6 @@ project(Proton LANGUAGES CXX)
 set(PROTON_SRC_DIR "${CMAKE_CURRENT_SOURCE_DIR}/csrc")
 set(PROTON_COMMON_DIR "${CMAKE_CURRENT_SOURCE_DIR}/common")
 
-option(TRITON_BUILD_PROTON_XPU "Build Proton with XPU support" OFF)
-if(TRITON_BUILD_PROTON_XPU)
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DTRITON_BUILD_PROTON_XPU")
-endif()
-
 # ============ Check for includes =============
 if(NOT CUPTI_INCLUDE_DIR)
   message(FATAL_ERROR "CUPTI include directory not defined")

third_party/proton/csrc/lib/Driver/CMakeLists.txt

@@ -1,16 +1,11 @@
-set(SOURCE_FILES
+add_proton_library(ProtonDriver
 	Device.cpp
 	GPU/CudaApi.cpp
 	GPU/CuptiApi.cpp
 	GPU/HipApi.cpp
 	GPU/HsaApi.cpp
 	GPU/RoctracerApi.cpp
 	GPU/NvtxApi.cpp
+	GPU/XpuApi.cpp
+	GPU/XpuptiApi.cpp
 )
-
-if(TRITON_BUILD_PROTON_XPU)
-    list(APPEND SOURCE_FILES GPU/XpuApi.cpp)
-	list(APPEND SOURCE_FILES GPU/XpuptiApi.cpp)
-endif()
-
-add_proton_library(ProtonDriver ${SOURCE_FILES})

third_party/proton/csrc/lib/Driver/Device.cpp

@@ -1,9 +1,7 @@
 #include "Device.h"
 #include "Driver/GPU/CudaApi.h"
 #include "Driver/GPU/HipApi.h"
-#ifdef TRITON_BUILD_PROTON_XPU
 #include "Driver/GPU/XpuApi.h"
-#endif
 
 #include "Utility/Errors.h"
 
@@ -16,11 +14,9 @@ Device getDevice(DeviceType type, uint64_t index) {
   if (type == DeviceType::HIP) {
     return hip::getDevice(index);
   }
-#ifdef TRITON_BUILD_PROTON_XPU
   if (type == DeviceType::XPU) {
     return xpu::getDevice(index);
   }
-#endif
   throw std::runtime_error("DeviceType not supported");
 }
 

third_party/proton/csrc/lib/Profiler/CMakeLists.txt

@@ -1,15 +1,10 @@
-set(SOURCE_FILES
-    Cupti/CuptiPCSampling.cpp
-    Cupti/CuptiProfiler.cpp
-    RocTracer/RoctracerProfiler.cpp
-	  Instrumentation/InstrumentationProfiler.cpp
-	  Instrumentation/CudaRuntime.cpp
-	  Instrumentation/HipRuntime.cpp
-	  Instrumentation/Metadata.cpp
+add_proton_library(ProtonProfiler
+	Cupti/CuptiPCSampling.cpp
+	Cupti/CuptiProfiler.cpp
+	RocTracer/RoctracerProfiler.cpp
+	Xpupti/XpuptiProfiler.cpp
+	Instrumentation/InstrumentationProfiler.cpp
+	Instrumentation/CudaRuntime.cpp
+	Instrumentation/HipRuntime.cpp
+	Instrumentation/Metadata.cpp
 )
-
-if(TRITON_BUILD_PROTON_XPU)
-    list(APPEND SOURCE_FILES Xpupti/XpuptiProfiler.cpp)
-endif()
-
-add_proton_library(ProtonProfiler ${SOURCE_FILES})

third_party/proton/csrc/lib/Session/Session.cpp

@@ -3,13 +3,11 @@
 #include "Context/Shadow.h"
 #include "Data/TraceData.h"
 #include "Data/TreeData.h"
+#include "Driver/GPU/XpuApi.h"
 #include "Profiler/Cupti/CuptiProfiler.h"
 #include "Profiler/Instrumentation/InstrumentationProfiler.h"
 #include "Profiler/Roctracer/RoctracerProfiler.h"
-#ifdef TRITON_BUILD_PROTON_XPU
-#include "Driver/GPU/XpuApi.h"
 #include "Profiler/Xpupti/XpuptiProfiler.h"
-#endif
 #include "Utility/String.h"
 
 namespace proton {
@@ -24,13 +22,10 @@ Profiler *makeProfiler(const std::string &name, void *sycl_queue = nullptr,
     return &RoctracerProfiler::instance();
   } else if (proton::toLower(name) == "instrumentation") {
     return &InstrumentationProfiler::instance();
-  }
-#ifdef TRITON_BUILD_PROTON_XPU
-  if (proton::toLower(name) == "xpupti") {
+  } else if (proton::toLower(name) == "xpupti") {
     xpu::PROTON_UTILS = utils_cache_path;
     return &XpuptiProfiler::instance().setSyclQueue(sycl_queue);
   }
-#endif
   throw std::runtime_error("Unknown profiler: " + name);
 }