ocl: avoid rescheduling of async kernels

modules/core/src/ocl.cpp

@@ -2755,7 +2755,7 @@ KernelArg KernelArg::Constant(const Mat& m)
 struct Kernel::Impl
 {
     Impl(const char* kname, const Program& prog) :
-        refcount(1), handle(NULL), isInProgress(false), nu(0)
+        refcount(1), handle(NULL), isInProgress(false), isAsyncRun(false), nu(0)
     {
         cl_program ph = (cl_program)prog.ptr();
         cl_int retval = 0;
@@ -2832,6 +2832,7 @@ struct Kernel::Impl
     enum { MAX_ARRS = 16 };
     UMatData* u[MAX_ARRS];
     bool isInProgress;
+    bool isAsyncRun;  // true if kernel was scheduled in async mode
     int nu;
     std::list<Image2D> images;
     bool haveTempDstUMats;
@@ -3111,13 +3112,45 @@ bool Kernel::run(int dims, size_t _globalsize[], size_t _localsize[],
 }
 
 
+static bool isRaiseErrorOnReuseAsyncKernel()
+{
+    static bool initialized = false;
+    static bool value = false;
+    if (!initialized)
+    {
+        value = cv::utils::getConfigurationParameterBool("OPENCV_OPENCL_RAISE_ERROR_REUSE_ASYNC_KERNEL", false);
+        initialized = true;
+    }
+    return value;
+}
+
 bool Kernel::Impl::run(int dims, size_t globalsize[], size_t localsize[],
         bool sync, int64* timeNS, const Queue& q)
 {
     CV_INSTRUMENT_REGION_OPENCL_RUN(name.c_str());
 
-    if (!handle || isInProgress)
+    if (!handle)
+    {
+        CV_LOG_ERROR(NULL, "OpenCL kernel has zero handle: " << name);
         return false;
+    }
+
+    if (isAsyncRun)
+    {
+        CV_LOG_ERROR(NULL, "OpenCL kernel can't be reused in async mode: " << name);
+        if (isRaiseErrorOnReuseAsyncKernel())
+            CV_Assert(0);
+        return false;  // OpenCV 5.0: raise error
+    }
+    isAsyncRun = !sync;
+
+    if (isInProgress)
+    {
+        CV_LOG_ERROR(NULL, "Previous OpenCL kernel launch is not finished: " << name);
+        if (isRaiseErrorOnReuseAsyncKernel())
+            CV_Assert(0);
+        return false;  // OpenCV 5.0: raise error
+    }
 
     cl_command_queue qq = getQueue(q);
     if (haveTempDstUMats)

modules/dnn/src/ocl4dnn/src/math_functions.cpp

@@ -46,6 +46,8 @@
 #include <vector>
 #include "opencl_kernels_dnn.hpp"
 
+#include "opencv2/core/utils/logger.hpp"
+
 namespace cv { namespace dnn { namespace ocl4dnn {
 
 enum gemm_data_type_t
@@ -238,10 +240,6 @@ static bool ocl4dnnFastImageGEMM(const CBLAS_TRANSPOSE TransA,
         kernel_name += "_float";
     }
 
-    ocl::Kernel oclk_gemm_float(kernel_name.c_str(), ocl::dnn::gemm_image_oclsrc, opts);
-    if (oclk_gemm_float.empty())
-        return false;
-
     while (C_start_y < M)
     {
         blockC_width = std::min(static_cast<int>(N) - C_start_x, blocksize);
@@ -348,6 +346,10 @@ static bool ocl4dnnFastImageGEMM(const CBLAS_TRANSPOSE TransA,
             }
             local[1] = 1;
 
+            ocl::Kernel oclk_gemm_float(kernel_name.c_str(), ocl::dnn::gemm_image_oclsrc, opts);
+            if (oclk_gemm_float.empty())
+                return false;
+
             cl_uint arg_idx = 0;
             if (is_image_a)
                 oclk_gemm_float.set(arg_idx++, ocl::KernelArg::PtrReadOnly(A));
@@ -378,7 +380,10 @@ static bool ocl4dnnFastImageGEMM(const CBLAS_TRANSPOSE TransA,
             oclk_gemm_float.set(arg_idx++, isFirstColBlock);
 
             if (!oclk_gemm_float.run(2, global, local, false))
+            {
+                CV_LOG_WARNING(NULL, "OpenCL kernel enqueue failed: " << kernel_name);
                 return false;
+            }
 
             if (TransA == CblasNoTrans)
                 A_start_x += blockA_width;