cleanup && add pti.txt

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

Author: anmyachev

.github/pins/pti.txt

@@ -0,0 +1 @@
+15a201d25e5659692613b98ee33513263b689101

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

@@ -288,9 +288,10 @@ jobs:
       - name: Build PTI && Run Proton tests
         if: matrix.suite == 'rest' && inputs.driver_version == 'rolling' && inputs.device == 'max1100'
         run: |
+          PTI_COMMIT_ID="$(<.github/pins/pti.txt)"
           git clone https://github.com/intel/pti-gpu.git
           cd pti-gpu
-          git checkout 15a201d25e5659692613b98ee33513263b689101
+          git checkout $PTI_COMMIT_ID
           cd sdk
           cmake --preset linux-icpx-release
           BUILD_TESTING=1 PTI_BUILD_SAMPLES=1 cmake --build --preset linux-icpx-release

.github/workflows/triton-benchmarks.yml

@@ -117,11 +117,11 @@ jobs:
           pip install .
 
       - name: Build PTI from source
-        id: build-pti
         run: |
+          PTI_COMMIT_ID="$(<.github/pins/pti.txt)"
           git clone https://github.com/intel/pti-gpu.git
           cd pti-gpu
-          git checkout 15a201d25e5659692613b98ee33513263b689101
+          git checkout $PTI_COMMIT_ID
           cd sdk
           cmake --preset linux-icpx-release
           BUILD_TESTING=1 PTI_BUILD_SAMPLES=1 cmake --build --preset linux-icpx-release

python/tutorials/09-persistent-matmul.py

@@ -673,7 +673,7 @@ def bench_fn(label, reps, warmup_reps, fn, *args):
     for _ in range(warmup_reps):
         fn(*args)
     #FIXME: Enable for XPU once proton support works.
-    if True or is_cuda():
+    if is_cuda():
         with proton_context():
             for _ in range(reps):
                 fn(*args)
@@ -783,11 +783,11 @@ def show_profile(precision, profile_name):
 
         validate(32, 32, 32, dtype)
         validate(8192, 8192, args.K_range[0], dtype)
-        if True or is_cuda():
+        if is_cuda():
             proton.start("matmul", hook="triton")
             proton.deactivate()
         for K in range(args.K_range[0], args.K_range[1] + 1, args.K_step):
             bench(K, dtype)
-        if True or is_cuda():
+        if is_cuda():
             proton.finalize()
             show_profile(args.prec, "matmul")

third_party/proton/csrc/lib/Profiler/Xpupti/XpuptiProfiler.cpp

@@ -264,8 +264,6 @@ struct XpuptiProfiler::XpuptiProfilerPimpl
 
   static void allocBuffer(uint8_t **buffer, size_t *bufferSize);
   static void completeBuffer(uint8_t *buffer, size_t size, size_t validSize);
-  // static void callbackFn(void *userData, CUpti_CallbackDomain domain,
-  //                        CUpti_CallbackId cbId, const void *cbData);
   static void callbackFn(pti_callback_domain domain,
                          pti_api_group_id driver_api_group_id,
                          uint32_t driver_api_id,
@@ -354,32 +352,6 @@ void XpuptiProfiler::XpuptiProfilerPimpl::callbackFn(
   } else {
     throw std::runtime_error("[PROTON] callbackFn failed");
   }
-
-  /*                        switch (domain) {
-    case PTI_DOMAIN_DRIVER_GPU_OPERATION_APPEND:
-      CallbackGPUOperationAppend(domain, driver_api_group_id, driver_api_id,
-                                backend_context, cb_data, user_data);
-      break;
-    case PTI_DOMAIN_GPU_OPERATION_COMPLETED:
-      CallbackGPUOperationCompletion(domain, driver_api_group_id, driver_api_id,
-                                   backend_context, cb_data, user_data);
-      break;
-    default: {
-      std::cout << "In " << __func__ << ", domain: " << domain
-                << ", driver_group_id: " << driver_api_group_id
-                << ", driver_api_id: " << driver_api_id << std::endl;
-
-      const char* name_ptr = nullptr;
-
-      if (PTI_SUCCESS == ptiViewGetApiIdName(driver_api_group_id, driver_api_id,
-  &name_ptr) ) { std::cout << ", API name: " << name_ptr << std::endl; } else {
-        std::cout << ", Unknown API name" << std::endl;
-      }
-      break;
-    }
-  }
-  std::cout << std::endl;
-  */
 }
 
 void CallbackCommon(pti_callback_domain domain,
@@ -458,7 +430,6 @@ int callWaitOnSyclQueue(const std::string &utils_cache_path, void *syclQueue) {
 }
 
 void XpuptiProfiler::XpuptiProfilerPimpl::doStart() {
-  std::cout << "doStart\n" << std::flush;
   // should be call to shared lib
   XpuptiProfiler &profiler = threadState.profiler;
   if (profiler.utils_cache_path != "") {
@@ -501,7 +472,6 @@ void XpuptiProfiler::XpuptiProfilerPimpl::doStart() {
   xpupti::viewEnable<true>(PTI_VIEW_DEVICE_GPU_KERNEL);
   xpupti::viewEnable<true>(PTI_VIEW_DEVICE_GPU_MEM_FILL);
   xpupti::viewEnable<true>(PTI_VIEW_DEVICE_GPU_MEM_COPY);
-  std::cout << "doStart2\n" << std::flush;
   xpupti::subscribe<true>(&subscriber, callbackFn, &subscriber);
   // xpupti::viewEnable<true>(PTI_VIEW_DEVICE_GPU_MEM_COPY);
   // xpupti::viewEnable<true>(PTI_VIEW_DEVICE_GPU_MEM_FILL);
@@ -511,10 +481,8 @@ void XpuptiProfiler::XpuptiProfilerPimpl::doStart() {
   // xpupti::viewEnable<true>(PTI_VIEW_LEVEL_ZERO_CALLS);
   // setGraphCallbacks(subscriber, /*enable=*/true);
   // setRuntimeCallbacks(subscriber, /*enable=*/true);
-  std::cout << "doStart::enableDomain\n" << std::flush;
   xpupti::enableDomain<true>(subscriber,
                              PTI_CB_DOMAIN_DRIVER_GPU_OPERATION_APPENDED, 1, 1);
-  std::cout << "doStart::enableDomain after\n" << std::flush;
   // setDriverCallbacks(subscriber, /*enable=*/true);
 }
 
@@ -551,10 +519,8 @@ void XpuptiProfiler::XpuptiProfilerPimpl::doStop() {
   // setGraphCallbacks(subscriber, /*enable=*/false);
   // setRuntimeCallbacks(subscriber, /*enable=*/false);
   // setDriverCallbacks(subscriber, /*enable=*/false);
-  std::cout << "doStop::disableDomain\n" << std::flush;
   xpupti::disableDomain<true>(subscriber,
                               PTI_CB_DOMAIN_DRIVER_GPU_OPERATION_APPENDED);
-  std::cout << "doStop::disableDomain after\n" << std::flush;
   xpupti::unsubscribe<true>(subscriber);
   // cupti::finalize<true>();
 }