Expand Kineto platform support (ci-all) (#56323)

Summary:
Expanding support to all builds

Pull Request resolved: https://github.com/pytorch/pytorch/pull/56323

Test Plan: CI

Reviewed By: malfet

Differential Revision: D28171478

Pulled By: ilia-cher

fbshipit-source-id: 16bc752d1be3cbaeda5316f5d8a687ae05a83d22

cmake/Dependencies.cmake

@@ -1824,14 +1824,6 @@ if(USE_KINETO AND INTERN_BUILD_MOBILE)
   message(STATUS "Not using libkineto in a mobile build.")
   set(USE_KINETO OFF)
 endif()
-if(USE_KINETO AND MSVC)
-  message(STATUS "Not using libkineto in a Windows build.")
-  set(USE_KINETO OFF)
-endif()
-if(USE_KINETO AND APPLE)
-  message(STATUS "Not using libkineto in a Mac build.")
-  set(USE_KINETO OFF)
-endif()
 if(USE_KINETO AND (NOT USE_CUDA))
   set(LIBKINETO_NOCUPTI ON CACHE STRING "")
   message(STATUS "Using CPU-only Kineto in the non-CUDA build.")

cmake/TorchConfig.cmake.in

@@ -123,6 +123,10 @@ else()
   append_torchlib_if_found(sleef asmjit)
 endif()
 
+if(@USE_KINETO@)
+  append_torchlib_if_found(kineto)
+endif()
+
 if(@USE_CUDA@)
   if(MSVC)
     if(NOT NVTOOLEXT_HOME)

scripts/xcode_build.rb

@@ -51,7 +51,7 @@ end
 
 # link static libraries
 target.frameworks_build_phases.clear
-libs = ['libc10.a', 'libclog.a', 'libpthreadpool.a', 'libXNNPACK.a', 'libeigen_blas.a', 'libcpuinfo.a', 'libpytorch_qnnpack.a', 'libtorch_cpu.a', 'libtorch.a']
+libs = ['libc10.a', 'libclog.a', 'libpthreadpool.a', 'libXNNPACK.a', 'libeigen_blas.a', 'libcpuinfo.a', 'libpytorch_qnnpack.a', 'libtorch_cpu.a', 'libtorch.a', 'libkineto.a']
 for lib in libs do
     path = "#{install_path}/lib/#{lib}"
     if File.exist?(path)

test/test_profiler.py

@@ -482,6 +482,7 @@ class TestProfiler(TestCase):
                 assert is_int, "Invalid stacks record"
 
     @unittest.skipIf(not kineto_available(), "Kineto is required")
+    @unittest.skipIf(IS_WINDOWS, "Test is flaky on Windows")
     def test_tensorboard_trace_handler(self):
         use_cuda = torch.cuda.is_available()
         with _profile(use_cuda=use_cuda, use_kineto=True):

third_party/kineto

@@ -1 +1 @@
-Subproject commit 5bc9386b6d60c3b34b77961ea2900947103304b9
+Subproject commit 07344e6d29fb17283fbc909d7631b5256b98537d

torch/csrc/autograd/profiler_kineto.cpp

@@ -9,12 +9,6 @@
 #ifdef USE_KINETO
 #include <libkineto.h>
 
-#ifndef USE_KINETO_UPDATED
-#include <pthread.h>
-#include <unistd.h>
-#include <sys/syscall.h>
-#endif
-
 #ifndef _MSC_VER
 // TODO: TO be removed, once this properly works from libkineto
 // Literal copy-n-paste from third_party/kineto/libkineto/src/WeakSymbols.cpp
@@ -41,14 +35,6 @@ inline int64_t getTimeUs() {
   return duration_cast<microseconds>(high_resolution_clock::now().time_since_epoch()).count();
 }
 
-#ifndef USE_KINETO_UPDATED
-// Getting the linux tid is expensive, so cache it.
-// Caching linux pids and tids is not advisable in the general case,
-// but this is only for profiling purposes and we don't need to handle
-// special cases during fork, clone etc.
-static thread_local pid_t cachedTid;
-#endif
-
 std::string shapesToStr(const std::vector<std::vector<int64_t>>& shapes);
 std::string stacksToStr(const std::vector<std::string>& stacks);
 std::string dtypesToStr(const std::vector<std::string>& types);
@@ -64,14 +50,11 @@ struct KinetoThreadLocalState : public ProfilerThreadLocalState {
     if (!ctx) {
       return;
     }
-#ifdef USE_KINETO_UPDATED
+
     libkineto::GenericTraceActivity op;
     op.activityType = libkineto::ActivityType::CPU_OP;
     op.activityName = std::string(fn.name().str());
-#else
+
-    libkineto::ClientTraceActivity op;
-    op.opType = std::string(fn.name().str());
-#endif
     op.startTime = ctx->startUs;
     op.endTime = getTimeUs();
     op.device = 0;
@@ -82,16 +65,8 @@ struct KinetoThreadLocalState : public ProfilerThreadLocalState {
     //   op.inputDims = shapesToStr(*ctx->shapes);
     // }
 
-#ifdef USE_KINETO_UPDATED
     libkineto::api().activityProfiler().recordThreadInfo();
     op.sysThreadId = libkineto::systemThreadId();
-#else
-    if (!cachedTid) {
-      cachedTid = (pid_t)syscall(SYS_gettid);
-      libkineto::api().activityProfiler().recordThreadInfo(cachedTid, pthread_self());
-    }
-    op.sysThreadId = cachedTid;
-#endif
 
     {
       std::lock_guard<std::mutex> guard(state_mutex_);

torch/csrc/autograd/profiler_kineto.h

@@ -2,9 +2,9 @@
 
 #include <torch/csrc/autograd/profiler_legacy.h>
 
-// Kineto is currently available on Linux server-side
 #ifdef USE_KINETO
-#if !defined(__linux__) || defined(_WIN32) || defined(C10_MOBILE) || defined(__APPLE__) || defined(DISABLE_KINETO)
+// skip Kineto dependency on mobile
+#ifdef C10_MOBILE
 #undef USE_KINETO
 #endif
 #endif