Windows debug

.github/workflows/build-test.yml

@@ -44,10 +44,6 @@ on:
         type: boolean
         default: false
 
-  pull_request:
-    branches:
-      - main
-      - release/**
   push:
     branches:
       - main

CMakeLists.txt

@@ -18,16 +18,9 @@ include(CTest)
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 
 # Options
-if(WIN32)
-  set(DEFAULT_BUILD_PROTON OFF)
-else()
-  set(DEFAULT_BUILD_PROTON ON)
-endif()
-
-# Define the option with the determined default value
-option(TRITON_BUILD_PROTON "Build the Triton Proton profiler" ${DEFAULT_BUILD_PROTON})
 option(TRITON_BUILD_TUTORIALS "Build C++ Triton tutorials" ON)
 option(TRITON_BUILD_PYTHON_MODULE "Build Python Triton bindings" OFF)
+option(TRITON_BUILD_PROTON "Build the Triton Proton profiler" ON)
 option(TRITON_BUILD_UT "Build C++ Triton Unit Tests" ON)
 option(TRITON_BUILD_WITH_CCACHE "Build with ccache (if available)" ON)
 set(TRITON_CODEGEN_BACKENDS "" CACHE STRING "Enable different codegen backends")

third_party/amd/backend/include/hsa/hsa_ext_amd.h

@@ -69,17 +69,6 @@ extern "C" {
  *  @{
  */
 
-/**
- * @brief Macro to use to determine that a  flag is set when querying flags within uint8_t[8]
- * types
- */
-static __inline__ __attribute__((always_inline)) bool hsa_flag_isset64(uint8_t* value,
-                                                                       uint32_t bit) {
-  unsigned int index = bit / 8;
-  unsigned int subBit = bit % 8;
-  return ((uint8_t*)value)[index] & (1 << subBit);
-}
-
 /**
  * @brief A fixed-size type used to represent ::hsa_signal_condition_t constants.
  */

third_party/proton/csrc/include/Driver/Dispatch.h

@@ -1,7 +1,13 @@
 #ifndef PROTON_DRIVER_DISPATCH_H_
 #define PROTON_DRIVER_DISPATCH_H_
 
+#ifdef WIN32
+#define WIN32_LEAN_AND_MEAN
+#define NOMINMAX
+#include <windows.h>
+#else
 #include <dlfcn.h>
+#endif
 
 #include <stdexcept>
 #include <string>
@@ -57,6 +63,31 @@ template <typename ExternLib> class Dispatch {
 public:
   Dispatch() = delete;
 
+#ifdef WIN32
+  static void init(const char *name, void **lib) {
+    if (*lib == nullptr) {
+      // First reuse the existing handle
+      *lib = GetModuleHandle(name);
+    }
+    if (*lib == nullptr) {
+      // If not found, try to load it from LD_LIBRARY_PATH
+      *lib = LoadLibraryA(name);
+    }
+    if (*lib == nullptr) {
+      // If still not found, try to load it from the default path
+      auto dir = std::string(ExternLib::defaultDir);
+      if (dir.length() > 0) {
+        auto fullPath = dir + "/" + name;
+        *lib = LoadLibraryA(fullPath.c_str());
+      }
+    }
+    if (*lib == nullptr) {
+      throw std::runtime_error("Could not find `" + std::string(name) +
+                               "`. Make sure it is in your "
+                               "LD_LIBRARY_PATH.");
+    }
+  }
+#else
   static void init(const char *name, void **lib) {
     if (*lib == nullptr) {
       // First reuse the existing handle
@@ -80,6 +111,7 @@ template <typename ExternLib> class Dispatch {
                                "LD_LIBRARY_PATH.");
     }
   }
+#endif
 
   static void check(typename ExternLib::RetType ret, const char *functionName) {
     if (ret != ExternLib::success) {
@@ -94,7 +126,12 @@ template <typename ExternLib> class Dispatch {
   exec(FnT &handler, const char *functionName, Args... args) {
     init(ExternLib::name, &ExternLib::lib);
     if (handler == nullptr) {
+#ifdef WIN32
+      handler = reinterpret_cast<FnT>(
+          GetProcAddress((HMODULE)(ExternLib::lib), functionName));
+#else
       handler = reinterpret_cast<FnT>(dlsym(ExternLib::lib, functionName));
+#endif
       if (handler == nullptr) {
         throw std::runtime_error("Failed to load " +
                                  std::string(ExternLib::name));

third_party/proton/csrc/lib/Driver/GPU/CudaApi.cpp

@@ -10,7 +10,11 @@ struct ExternLibCuda : public ExternLibBase {
   // https://forums.developer.nvidia.com/t/wsl2-libcuda-so-and-libcuda-so-1-should-be-symlink/236301
   // On WSL, "libcuda.so" and "libcuda.so.1" may not be linked, so we use
   // "libcuda.so.1" instead.
+#ifdef WIN32
+  static constexpr const char *name = "nvcuda.dll";
+#else
   static constexpr const char *name = "libcuda.so.1";
+#endif
   static constexpr const char *defaultDir = "";
   static constexpr RetType success = CUDA_SUCCESS;
   static void *lib;

third_party/proton/csrc/lib/Driver/GPU/CuptiApi.cpp

@@ -10,7 +10,11 @@ namespace cupti {
 #define TOSTRING(x) STRINGIFY(x)
 struct ExternLibCupti : public ExternLibBase {
   using RetType = CUptiResult;
+#ifdef WIN32
+  static constexpr const char *name = "cupti.dll";
+#else
   static constexpr const char *name = "libcupti.so";
+#endif
 #ifdef CUPTI_LIB_DIR
   static constexpr const char *defaultDir = TOSTRING(CUPTI_LIB_DIR);
 #else

third_party/proton/csrc/lib/Driver/GPU/HipApi.cpp

@@ -9,7 +9,11 @@ namespace hip {
 
 struct ExternLibHip : public ExternLibBase {
   using RetType = hipError_t;
+#ifdef WIN32
+  static constexpr const char *name = "amdhip64.dll";
+#else
   static constexpr const char *name = "libamdhip64.so";
+#endif
   static constexpr const char *defaultDir = "";
   static constexpr RetType success = hipSuccess;
   static void *lib;
@@ -70,8 +74,13 @@ const char *getKernelNameRef(const hipFunction_t f) {
   static hipKernelNameRef_t func = nullptr;
   Dispatch<ExternLibHip>::init(ExternLibHip::name, &ExternLibHip::lib);
   if (func == nullptr)
+#ifdef WIN32
+    func = reinterpret_cast<hipKernelNameRef_t>(
+        GetProcAddress((HMODULE)ExternLibHip::lib, "hipKernelNameRef"));
+#else
     func = reinterpret_cast<hipKernelNameRef_t>(
         dlsym(ExternLibHip::lib, "hipKernelNameRef"));
+#endif
   return (func ? func(f) : NULL);
 }
 
@@ -81,8 +90,13 @@ const char *getKernelNameRefByPtr(const void *hostFunction,
   static hipKernelNameRefByPtr_t func = nullptr;
   Dispatch<ExternLibHip>::init(ExternLibHip::name, &ExternLibHip::lib);
   if (func == nullptr)
+#ifdef WIN32
+    func = reinterpret_cast<hipKernelNameRefByPtr_t>(
+        GetProcAddress((HMODULE)ExternLibHip::lib, "hipKernelNameRefByPtr"));
+#else
     func = reinterpret_cast<hipKernelNameRefByPtr_t>(
         dlsym(ExternLibHip::lib, "hipKernelNameRefByPtr"));
+#endif
   return (func ? func(hostFunction, stream) : NULL);
 }
 

third_party/proton/csrc/lib/Driver/GPU/HsaApi.cpp

@@ -7,7 +7,11 @@ namespace hsa {
 
 struct ExternLibHsa : public ExternLibBase {
   using RetType = hsa_status_t;
+#ifdef WIN32
+  static constexpr const char *name = "hsa-runtime64.dll";
+#else
   static constexpr const char *name = "libhsa-runtime64.so";
+#endif
   static constexpr const char *defaultDir = "";
   static constexpr RetType success = HSA_STATUS_SUCCESS;
   static void *lib;
@@ -26,8 +30,13 @@ hsa_status_t iterateAgents(hsa_status_t (*callback)(hsa_agent_t agent,
   static hsa_iterate_agents_t func = nullptr;
   Dispatch<ExternLibHsa>::init(ExternLibHsa::name, &ExternLibHsa::lib);
   if (func == nullptr)
+#ifdef WIN32
+    func = reinterpret_cast<hsa_iterate_agents_t>(
+        GetProcAddress((HMODULE)ExternLibHsa::lib, "hsa_iterate_agents"));
+#else
     func = reinterpret_cast<hsa_iterate_agents_t>(
         dlsym(ExternLibHsa::lib, "hsa_iterate_agents"));
+#endif
   return (func ? func(callback, data) : HSA_STATUS_ERROR_FATAL);
 }
 

third_party/proton/csrc/lib/Driver/GPU/RoctracerApi.cpp

@@ -7,7 +7,12 @@ namespace roctracer {
 
 struct ExternLibRoctracer : public ExternLibBase {
   using RetType = roctracer_status_t;
+#ifdef WIN32
+  // Looks like there is no support for Windows yet
+  static constexpr const char *name = "roctracer64.dll";
+#else
   static constexpr const char *name = "libroctracer64.so";
+#endif
   static constexpr const char *defaultDir = "";
   static constexpr RetType success = ROCTRACER_STATUS_SUCCESS;
   static void *lib;
@@ -27,8 +32,13 @@ void start() {
   Dispatch<ExternLibRoctracer>::init(ExternLibRoctracer::name,
                                      &ExternLibRoctracer::lib);
   if (func == nullptr)
+#ifdef WIN32
+    func = reinterpret_cast<roctracer_start_t>(
+        GetProcAddress((HMODULE)ExternLibRoctracer::lib, "roctracer_start"));
+#else
     func = reinterpret_cast<roctracer_start_t>(
         dlsym(ExternLibRoctracer::lib, "roctracer_start"));
+#endif
   if (func)
     func();
 }
@@ -39,8 +49,13 @@ void stop() {
   Dispatch<ExternLibRoctracer>::init(ExternLibRoctracer::name,
                                      &ExternLibRoctracer::lib);
   if (func == nullptr)
+#ifdef WIN32
+    func = reinterpret_cast<roctracer_stop_t>(
+        GetProcAddress((HMODULE)ExternLibRoctracer::lib, "roctracer_stop"));
+#else
     func = reinterpret_cast<roctracer_stop_t>(
         dlsym(ExternLibRoctracer::lib, "roctracer_stop"));
+#endif
   if (func)
     func();
 }
@@ -51,8 +66,13 @@ char *getOpString(uint32_t domain, uint32_t op, uint32_t kind) {
   Dispatch<ExternLibRoctracer>::init(ExternLibRoctracer::name,
                                      &ExternLibRoctracer::lib);
   if (func == nullptr)
+#ifdef WIN32
+    func = reinterpret_cast<roctracer_op_string_t>(GetProcAddress(
+        (HMODULE)ExternLibRoctracer::lib, "roctracer_op_string"));
+#else
     func = reinterpret_cast<roctracer_op_string_t>(
         dlsym(ExternLibRoctracer::lib, "roctracer_op_string"));
+#endif
   return (func ? func(domain, op, kind) : NULL);
 }
 

third_party/proton/csrc/lib/Profiler/Cupti/CuptiProfiler.cpp

@@ -12,6 +12,10 @@
 #include <memory>
 #include <stdexcept>
 
+#ifdef WIN32
+#define aligned_alloc _aligned_malloc
+#endif
+
 namespace proton {
 
 template <>
@@ -257,7 +261,11 @@ void CuptiProfiler::CuptiProfilerPimpl::completeBuffer(CUcontext ctx,
     }
   } while (true);
 
+#ifdef WIN32
+  _aligned_free(buffer);
+#else
   std::free(buffer);
+#endif
 
   profiler.correlation.complete(maxCorrelationId);
 }

third_party/proton/csrc/lib/Profiler/RocTracer/RoctracerProfiler.cpp

@@ -16,9 +16,6 @@
 #include <mutex>
 #include <tuple>
 
-#include <cxxabi.h>
-#include <unistd.h>
-
 namespace proton {
 
 template <>