Merge pull request #2583 from alibaba/feature/bugfix

Feature/bugfix
alibaba · Sep 7, 2023 · 32f72f4 · 32f72f4
2 parents 9e3cc72 + 8514c07
commit 32f72f4
Show file tree

Hide file tree

Showing 46 changed files with 125 additions and 167 deletions.
diff --git a/source/backend/opencl/core/ImageBufferConvertor.cpp b/source/backend/opencl/core/ImageBufferConvertor.cpp
@@ -59,8 +59,7 @@ bool convertNCHWBufferToImage(const Tensor *input, Tensor *output, cl::Kernel &b
     }
 
     #ifdef ENABLE_OPENCL_TIME_PROFILER
-        int costTime = (int)runtime->getCostTime(&event);
-        MNN_PRINT("kernel cost:%d    us inputFormatTransform\n",costTime);
+        runtime->pushEvent({"inputFormatTransform", event});
     #endif
     return true;
 }
@@ -111,8 +110,7 @@ bool convertNHWCBufferToImage(const Tensor *input, Tensor *output, cl::Kernel &b
     }
 
     #ifdef ENABLE_OPENCL_TIME_PROFILER
-        int costTime = (int)runtime->getCostTime(&event);
-        MNN_PRINT("kernel cost:%d    us inputFormatTransform\n",costTime);
+        runtime->pushEvent({"inputFormatTransform", event});
     #endif
     return true;
 }
@@ -167,8 +165,7 @@ bool convertImageToNCHWBuffer(const Tensor *input, Tensor *output, cl::Kernel &i
     }
 
     #ifdef ENABLE_OPENCL_TIME_PROFILER
-        int costTime = (int)runtime->getCostTime(&event);
-        MNN_PRINT("kernel cost:%d    us outputFormatTransform\n",costTime);
+        runtime->pushEvent({"outputFormatTransform", event});
     #endif
     return true;
 }
@@ -220,8 +217,7 @@ bool convertNC4HW4BufferToImage(const Tensor *input, Tensor *output, cl::Kernel
     }
 
     #ifdef ENABLE_OPENCL_TIME_PROFILER
-        int costTime = (int)runtime->getCostTime(&event);
-        MNN_PRINT("kernel cost:%d    us inputFormatTransform\n",costTime);
+        runtime->pushEvent({"inputFormatTransform", event});
     #endif
     return true;
 }
@@ -285,8 +281,7 @@ bool convertImageToNC4HW4Buffer(const Tensor *input, Tensor *output, cl::Kernel
     }
 
     #ifdef ENABLE_OPENCL_TIME_PROFILER
-        int costTime = (int)runtime->getCostTime(&event);
-        MNN_PRINT("kernel cost:%d    us outputFormatTransform\n",costTime);
+        runtime->pushEvent({"outputFormatTransform", event});
     #endif
     return true;
 }
@@ -341,8 +336,7 @@ bool convertImageToNHWCBuffer(const Tensor *input, Tensor *output, cl::Kernel &i
     }
 
     #ifdef ENABLE_OPENCL_TIME_PROFILER
-        int costTime = (int)runtime->getCostTime(&event);
-        MNN_PRINT("kernel cost:%d    us outputFormatTransform\n",costTime);
+        runtime->pushEvent({"outputFormatTransform", event});
     #endif
 
     return true;

diff --git a/source/backend/opencl/core/OpenCLBackend.cpp b/source/backend/opencl/core/OpenCLBackend.cpp
@@ -514,14 +514,15 @@ void OpenCLBackend::onResizeEnd() {
 
 void OpenCLBackend::onExecuteBegin() const {
     mOpenCLRuntime->mQueueCount = 0;
-    mOpenCLRuntime->mKernelTime = 0;
     mOpenCLRuntime->clearRecord();
+    mOpenCLRuntime->clearEvent();   
 }
 
 void OpenCLBackend::onExecuteEnd() const {
     mOpenCLRuntime->mQueueCount = 0;
     mOpenCLRuntime->clearRecord();
     mOpenCLRuntime->enqeueRecord();
+    mOpenCLRuntime->printEventTime();
 }
 
 
@@ -698,7 +699,7 @@ void OpenCLBackend::copyFromDevice(const Tensor* srcTensor, const Tensor* dstTen
     mOpenCLRuntime->clearRecord();
     //Convert format
     mCLRuntime->convertFromDevice(srcTensor, (const Tensor*)&interTensor, data_format, false);
-
+    mOpenCLRuntime->printEventTime();
 
 #ifdef ENABLE_OPENCL_TIME_PROFILER
     mOpenCLRuntime->commandQueue().finish();
@@ -743,10 +744,6 @@ void OpenCLBackend::copyFromDevice(const Tensor* srcTensor, const Tensor* dstTen
             hostPtr = nullptr;
         }
     }
-
-#ifdef ENABLE_OPENCL_TIME_PROFILER
-    MNN_PRINT("total kernel time:%d us\n", (int)mOpenCLRuntime->mKernelTime);
-#endif
 }
 
 

diff --git a/source/backend/opencl/core/runtime/OpenCLRuntime.cpp b/source/backend/opencl/core/runtime/OpenCLRuntime.cpp
@@ -364,6 +364,7 @@ OpenCLRuntime::~OpenCLRuntime() {
 #ifdef LOG_VERBOSE
     MNN_PRINT("start ~OpenCLRuntime !\n");
 #endif
+    clearEvent();
     releaseRecord();
     mBuildProgramMap.clear();
     mRecordings.clear();
@@ -779,4 +780,24 @@ void OpenCLRuntime::releaseRecord(){
     }
 #endif
 }
+
+void OpenCLRuntime::printEventTime(){
+#ifdef ENABLE_OPENCL_TIME_PROFILER
+    if(mEvents.empty()){
+        return;
+    }
+    for(int i = 0; i < mEvents.size(); ++i){
+        auto event = &mEvents[i].second;
+        cl_int res = event->wait();
+        MNN_CHECK_CL_SUCCESS(res, "clEvent");
+        auto StartNanos = event->getProfilingInfo<CL_PROFILING_COMMAND_START>();
+        auto StopNanos = event->getProfilingInfo<CL_PROFILING_COMMAND_END>();
+        auto kernel_time = (unsigned int)((StopNanos - StartNanos) / 1000.0);
+        mKernelTime += kernel_time;
+        MNN_PRINT("kernel time = %d    us %s\n", kernel_time, mEvents[i].first.c_str());
+    }
+    mEvents.clear();
+    MNN_PRINT("total kernel time = %d  us\n", mKernelTime);
+#endif
+}
 } // namespace MNN
diff --git a/source/backend/opencl/core/runtime/OpenCLRuntime.hpp b/source/backend/opencl/core/runtime/OpenCLRuntime.hpp
@@ -113,6 +113,14 @@ class OpenCLRuntime {
     std::string getDeviceName() {
         return mDeviceName;
     }
+    void pushEvent(std::pair<std::string, cl::Event> data) {
+        return mEvents.push_back(data);
+    }
+    void printEventTime();
+    void clearEvent(){
+        mKernelTime = 0;
+        mEvents.clear();
+    }
     uint64_t maxAllocSize() const;
     void setCommandQueueProfileEnable();
     void setCommandQueueProfileDisable();
@@ -181,6 +189,7 @@ class OpenCLRuntime {
     GpuType mGpuType;
     MaliAr mMaliAr;
     float mCLVersion = 1.0f;
+    std::vector<std::pair<std::string, cl::Event>> mEvents;
 
 #ifdef MNN_OPENCL_SVM_ENABLE
     cl_device_svm_capabilities mSvmCapabilities;

diff --git a/source/backend/opencl/execution/buffer/ArgMaxBufExecution.cpp b/source/backend/opencl/execution/buffer/ArgMaxBufExecution.cpp
@@ -107,8 +107,7 @@ ErrorCode ArgMaxBufExecution::onExecute(const std::vector<Tensor*>& inputs, cons
     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalSize,
                        mOpenCLBackend->getOpenCLRuntime(), &event);
 
-    int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-    MNN_PRINT("kernel cost:%d    us ArgMax\n",costTime);
+    mOpenCLBackend->getOpenCLRuntime()->pushEvent({"ArgMax", event});
 #else
     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalSize,
                        mOpenCLBackend->getOpenCLRuntime());

diff --git a/source/backend/opencl/execution/buffer/CastBufExecution.cpp b/source/backend/opencl/execution/buffer/CastBufExecution.cpp
@@ -23,11 +23,6 @@ ErrorCode CastBufExecution::onResize(const std::vector<Tensor*>& inputs, const s
     Tensor* output     = outputs[0];
     auto openCLBackend = static_cast<OpenCLBackend*>(backend());
     auto runtime       = openCLBackend->getOpenCLRuntime();
-#ifdef MNN_SUPPORT_INTEL_SUBGROUP
-    if (runtime->isSupportedIntelSubgroup()) {
-        return SubgrouponResize(inputs, outputs);
-    }
-#endif /* MNN_SUPPORT_INTEL_SUBGROUP */
     mKernel = runtime->buildKernel("cast_buf", "cast_buf", mBuildOptions);
     mMaxWorkGroupSize = static_cast<uint32_t>(runtime->getMaxWorkGroupSize(mKernel));
 
@@ -75,8 +70,7 @@ ErrorCode CastBufExecution::onExecute(const std::vector<Tensor*>& inputs, const
     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalSize,
                        mOpenCLBackend->getOpenCLRuntime(), &event);
 
-    int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-    MNN_PRINT("kernel cost:%d    us Cast\n",costTime);
+    mOpenCLBackend->getOpenCLRuntime()->pushEvent({"Cast", event});
 #else
     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalSize,
                        mOpenCLBackend->getOpenCLRuntime());

diff --git a/source/backend/opencl/execution/buffer/ConvBufExecution.cpp b/source/backend/opencl/execution/buffer/ConvBufExecution.cpp
@@ -674,8 +674,7 @@ ErrorCode ConvBufExecution::onExecute(const std::vector<Tensor *> &inputs, const
 #ifdef ENABLE_OPENCL_TIME_PROFILER
     cl::Event event;
     runKernel2D(mKernel, mGlobalWorkSize, mLocalWorkSize, mOpenCLBackend->getOpenCLRuntime(), &event);
-    int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-    MNN_PRINT("kernel cost:%d    us ConvBuf2D\n",costTime); 
+    mOpenCLBackend->getOpenCLRuntime()->pushEvent({"ConvBuf2D", event});
 #else
     runKernel2D(mKernel, mGlobalWorkSize, mLocalWorkSize, mOpenCLBackend->getOpenCLRuntime());
 #endif

diff --git a/source/backend/opencl/execution/buffer/ConvBufWinograd.cpp b/source/backend/opencl/execution/buffer/ConvBufWinograd.cpp
@@ -570,9 +570,6 @@ ErrorCode ConvBufWinograd::onExecute(const std::vector<Tensor*>& inputs, const s
     auto input  = inputs[0];
     auto output = outputs[0];
 
-    #ifdef ENABLE_OPENCL_TIME_PROFILER
-    int costTime = 0;
-    #endif
     for (int b = 0; b < input->batch(); ++b) {
         int index = b;
         /*Source Transform*/
@@ -581,10 +578,7 @@ ErrorCode ConvBufWinograd::onExecute(const std::vector<Tensor*>& inputs, const s
             cl::Event event;
             runKernel2D(mSourceTransform[index], mGWS_S[index], mLWS_S[index],
                         mOpenCLBackend->getOpenCLRuntime(), &event);
-
-            int costTime0 = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-            costTime += costTime0;
-            MNN_PRINT("kernel cost:%d    us ConvWino0\n",costTime0);
+            mOpenCLBackend->getOpenCLRuntime()->pushEvent({"ConvWino0", event});
         #else
             runKernel2D(mSourceTransform[index], mGWS_S[index], mLWS_S[index],
                         mOpenCLBackend->getOpenCLRuntime());
@@ -600,10 +594,7 @@ ErrorCode ConvBufWinograd::onExecute(const std::vector<Tensor*>& inputs, const s
             } else {
                 runKernel2D(mMatMul[index], mGWS_M[index], mLWS_M[index], mOpenCLBackend->getOpenCLRuntime(), &event);
             }
-
-            int costTime1 = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-            costTime += costTime1;
-            MNN_PRINT("kernel cost:%d    us ConvWino1\n",costTime1);
+            mOpenCLBackend->getOpenCLRuntime()->pushEvent({"ConvWino1", event});
         #else
             if (mUseSubgroup) {
                 run3DKernelDefault(mMatMul[index], mGWS_M[index], mLWS_M[index], mOpenCLBackend->getOpenCLRuntime());
@@ -619,19 +610,13 @@ ErrorCode ConvBufWinograd::onExecute(const std::vector<Tensor*>& inputs, const s
             cl::Event event;
             runKernel2D(mDestTransform[index], mGWS_D[index], mLWS_D[index],
                         mOpenCLBackend->getOpenCLRuntime(), &event);
-
-            int costTime2 = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-            costTime += costTime2;
-            MNN_PRINT("kernel cost:%d    us ConvWino2\n",costTime2);
+            mOpenCLBackend->getOpenCLRuntime()->pushEvent({"ConvWino2", event});
         #else
             runKernel2D(mDestTransform[index], mGWS_D[index], mLWS_D[index],
                         mOpenCLBackend->getOpenCLRuntime());
         #endif
         }
     }
-    #ifdef ENABLE_OPENCL_TIME_PROFILER
-    MNN_PRINT("kernel cost:%d    us ConvWino total\n",costTime);
-    #endif
 
     return NO_ERROR;
 }

diff --git a/source/backend/opencl/execution/buffer/ConvSubgroupBufExecution.cpp b/source/backend/opencl/execution/buffer/ConvSubgroupBufExecution.cpp
@@ -397,8 +397,7 @@ ErrorCode ConvSubgroupBuf::onExecute(const std::vector<Tensor *> &inputs, const
 
          cl::Event event;
          run3DKernelDefault(mTranseKernel, mTranseGlobalWorkSize, mTranseLocalWorkSize, mOpenCLBackend->getOpenCLRuntime(), &event);
-         int costTime0 = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-         MNN_PRINT("kernel cost:%d    us ConvSubgroup transe\n", costTime0);
+         mOpenCLBackend->getOpenCLRuntime()->pushEvent({"ConvSubgroup", event});
 #else
          run3DKernelDefault(mTranseKernel, mTranseGlobalWorkSize, mTranseLocalWorkSize, mOpenCLBackend->getOpenCLRuntime());
 #endif
@@ -407,8 +406,7 @@ ErrorCode ConvSubgroupBuf::onExecute(const std::vector<Tensor *> &inputs, const
 #ifdef ENABLE_OPENCL_TIME_PROFILER
          cl::Event event;
          run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalWorkSize, mOpenCLBackend->getOpenCLRuntime(), &event);
-         int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-         MNN_PRINT("kernel cost:%d    us ConvSubgroupBuf2D\n", costTime);
+         mOpenCLBackend->getOpenCLRuntime()->pushEvent({"ConvSubgroupBuf2D", event});
 #else
          run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalWorkSize, mOpenCLBackend->getOpenCLRuntime());
 #endif

diff --git a/source/backend/opencl/execution/buffer/DeconvBufExecution.cpp b/source/backend/opencl/execution/buffer/DeconvBufExecution.cpp
@@ -177,8 +177,8 @@ ErrorCode DeconvBufExecution::onExecute(const std::vector<Tensor *> &inputs, con
                        mOpenCLBackend->getOpenCLRuntime(),
                        &event);
 
-    int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-    MNN_PRINT("kernel cost:%d    us DeconvBuf\n",costTime);
+
+    mOpenCLBackend->getOpenCLRuntime()->pushEvent({"DeconvBuf", event});
 #else
     run3DKernelDefault(mKernel, mGWS, mLWS,
                        mOpenCLBackend->getOpenCLRuntime());

diff --git a/source/backend/opencl/execution/buffer/DepthwiseConvBufExecution.cpp b/source/backend/opencl/execution/buffer/DepthwiseConvBufExecution.cpp
@@ -307,9 +307,7 @@ ErrorCode DepthwiseConvBufExecution::onExecute(const std::vector<Tensor *> &inpu
     runKernel2D(mKernel, mGlobalWorkSize, mLocalWorkSize,
                 mOpenCLBackend->getOpenCLRuntime(),
                 &event);
-
-    int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-    MNN_PRINT("kernel cost:%d    us DepthwiseConvBuf\n",costTime);
+    mOpenCLBackend->getOpenCLRuntime()->pushEvent({"DepthwiseConvBuf", event});
 #else
     runKernel2D(mKernel, mGlobalWorkSize, mLocalWorkSize,
                 mOpenCLBackend->getOpenCLRuntime());

diff --git a/source/backend/opencl/execution/buffer/DepthwiseConvSubgroupBufExecution.cpp b/source/backend/opencl/execution/buffer/DepthwiseConvSubgroupBufExecution.cpp
@@ -260,8 +260,8 @@ ErrorCode DepthwiseConvSubgroupBufExecution::onExecute(const std::vector<Tensor
 
         run3DKernelDefault(mTranseKernel, mTranseGlobalWorkSize, mTranseLocalWorkSize,
                            mOpenCLBackend->getOpenCLRuntime(), &event);
-        int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-        MNN_PRINT("kernel cost:%d    us DepthwiseConvSubgroup transe\n", costTime);
+
+        mOpenCLBackend->getOpenCLRuntime()->pushEvent({"DepthwiseConvSubgroup transe", event});
 #else
         run3DKernelDefault(mTranseKernel, mTranseGlobalWorkSize, mTranseLocalWorkSize,
                            mOpenCLBackend->getOpenCLRuntime());
@@ -274,8 +274,7 @@ ErrorCode DepthwiseConvSubgroupBufExecution::onExecute(const std::vector<Tensor
                 mOpenCLBackend->getOpenCLRuntime(),
                 &event);
 
-    int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-    MNN_PRINT("kernel cost:%d    us DepthwiseConvSubgroupBuf\n",costTime);
+    mOpenCLBackend->getOpenCLRuntime()->pushEvent({"DepthwiseConvSubgroupBuf", event});
 #else
     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalWorkSize,
                         mOpenCLBackend->getOpenCLRuntime());

diff --git a/source/backend/opencl/execution/buffer/GridSampleBufExecution.cpp b/source/backend/opencl/execution/buffer/GridSampleBufExecution.cpp
@@ -90,9 +90,8 @@ ErrorCode GridSampleBufExecution::onExecute(const std::vector<Tensor *> &inputs,
     cl::Event event;
     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalWorkSize,
         mOpenCLBackend->getOpenCLRuntime(), &event);
-
-    int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-    MNN_PRINT("kernel cost:%d    us GridSample\n", costTime);
+
+    mOpenCLBackend->getOpenCLRuntime()->pushEvent({"GridSample", event});
 #else
     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalWorkSize, mOpenCLBackend->getOpenCLRuntime());
 #endif

diff --git a/source/backend/opencl/execution/buffer/Interp3DBufExecution.cpp b/source/backend/opencl/execution/buffer/Interp3DBufExecution.cpp
@@ -102,8 +102,7 @@ ErrorCode Interp3DBufExecution::onExecute(const std::vector<Tensor *> &inputs, c
     run3DKernelDefault(mKernel, mGWS, mLWS,
                        mOpenCLBackend->getOpenCLRuntime(), &event);
 
-    int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-    MNN_PRINT("kernel cost:%d    us Interp\n",costTime);
+    mOpenCLBackend->getOpenCLRuntime()->pushEvent({"Interp", event});
 #else
     run3DKernelDefault(mKernel, mGWS, mLWS, mOpenCLBackend->getOpenCLRuntime());
 #endif

diff --git a/source/backend/opencl/execution/buffer/InterpBufExecution.cpp b/source/backend/opencl/execution/buffer/InterpBufExecution.cpp
@@ -96,8 +96,7 @@ ErrorCode InterpBufExecution::onExecute(const std::vector<Tensor *> &inputs, con
     run3DKernelDefault(mKernel, mGWS, mLWS,
                        mOpenCLBackend->getOpenCLRuntime(), &event);
 
-    int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-    MNN_PRINT("kernel cost:%d    us Interp\n",costTime);
+    mOpenCLBackend->getOpenCLRuntime()->pushEvent({"Interp", event});
 #else
     run3DKernelDefault(mKernel, mGWS, mLWS, mOpenCLBackend->getOpenCLRuntime());
 #endif

diff --git a/source/backend/opencl/execution/buffer/LayerNormBufExecution.cpp b/source/backend/opencl/execution/buffer/LayerNormBufExecution.cpp
@@ -175,8 +175,7 @@ ErrorCode LayerNormBufExecution::onExecute(const std::vector<Tensor *> &inputs,
     run3DKernelDefault(mKernel, mGWS, mLWS,
                        mOpenCLBackend->getOpenCLRuntime(), &event);
 
-    int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-    MNN_PRINT("kernel cost:%d    us LayerNormBuf\n",costTime);
+    mOpenCLBackend->getOpenCLRuntime()->pushEvent({"LayerNormBuf", event});
 #else
     run3DKernelDefault(mKernel, mGWS, mLWS, mOpenCLBackend->getOpenCLRuntime());
 #endif

diff --git a/source/backend/opencl/execution/buffer/MatmulBufExecution.cpp b/source/backend/opencl/execution/buffer/MatmulBufExecution.cpp
@@ -111,9 +111,8 @@ ErrorCode MatMulBufExecution::onExecute(const std::vector<Tensor *> &inputs, con
     #ifdef ENABLE_OPENCL_TIME_PROFILER
         cl::Event event;
         runKernel2D(mKernel, mGlobalWorkSize, mLocalWorkSize, runtime, &event);
-
-        int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-        MNN_PRINT("kernel cost:%d    us MatmulBuf\n",costTime);
+
+        mOpenCLBackend->getOpenCLRuntime()->pushEvent({"MatmulBuf", event});
     #else
     runKernel2D(mKernel, mGlobalWorkSize, mLocalWorkSize, runtime, nullptr);
     #endif

diff --git a/source/backend/opencl/execution/buffer/PoolBufExecution.cpp b/source/backend/opencl/execution/buffer/PoolBufExecution.cpp
@@ -254,8 +254,7 @@ ErrorCode PoolBufExecution::onExecute(const std::vector<Tensor *> &inputs, const
     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalWorkSize,
                        mOpenCLBackend->getOpenCLRuntime(), &event);
 
-    int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-    MNN_PRINT("kernel cost:%d    us Pooling\n",costTime);
+    mOpenCLBackend->getOpenCLRuntime()->pushEvent({"Pooling", event});
 #else
     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalWorkSize,
                        mOpenCLBackend->getOpenCLRuntime());

diff --git a/source/backend/opencl/execution/buffer/RangeBufExecution.cpp b/source/backend/opencl/execution/buffer/RangeBufExecution.cpp
@@ -69,8 +69,7 @@ ErrorCode RangeBufExecution::onExecute(const std::vector<Tensor*>& inputs, const
     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalSize,
                        mOpenCLBackend->getOpenCLRuntime(), &event);
 
-    int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-    MNN_PRINT("kernel cost:%d    us Range\n",costTime);
+    mOpenCLBackend->getOpenCLRuntime()->pushEvent({"Range", event});
 #else
     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalSize,
                        mOpenCLBackend->getOpenCLRuntime());

diff --git a/source/backend/opencl/execution/buffer/ReductionBufExecution.cpp b/source/backend/opencl/execution/buffer/ReductionBufExecution.cpp
@@ -209,8 +209,8 @@ ErrorCode ReductionBufExecution::onExecute(const std::vector<Tensor *> &inputs,
         cl::Event event;
         run3DKernelDefault(mReduct1DKernel, mGlobalWorkSize, mLocalWorkSize,
                                mOpenCLBackend->getOpenCLRuntime(), &event);
-        int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
-        MNN_PRINT("kernel cost:%d    us Reduct1D\n",costTime);
+
+        mOpenCLBackend->getOpenCLRuntime()->pushEvent({"Reduct1D", event});
     #else
         run3DKernelDefault(mReduct1DKernel, mGlobalWorkSize, mLocalWorkSize,
                            mOpenCLBackend->getOpenCLRuntime());