Merge pull request #12985 from wzw-intel:vkcom_refine

modules/dnn/src/dnn.cpp

@@ -1411,10 +1411,45 @@ struct Net::Impl
                 continue;
             }
 
+            if (ld.type == "Convolution")
+            {
+                std::vector<MatShape> in_shapes;
+                std::vector<MatShape> out_shapes;
+                CV_Assert(ld.inputBlobs.size() == ld.outputBlobs.size());
+
+                for (int i = 0; i < ld.inputBlobs.size(); i++)
+                {
+                    in_shapes.push_back(shape(*ld.inputBlobs[i]));
+                    out_shapes.push_back(shape(ld.outputBlobs[i]));
+                }
+                int64 flops = layer->getFLOPS(in_shapes, out_shapes);
+                // FIXME
+                //
+                // This is a workaround for GPU hang on heavy convolution workload ( > 10 GFLOPS).
+                // For the long time task, vkWaitForFences() return without error but next call on
+                // vkQueueSubmit() return -4, i.e. "VK_ERROR_DEVICE_LOST" and driver reports GPU hang.
+                //
+                // Need more investigation on root cause of GPU hang and need to optimize convolution shader
+                // to reduce process time.
+                if (flops > CV_BIG_INT(10) * 1000 * 1000 * 1000)
+                {
+                    continue;
+                }
+            }
+
             ld.skip = false;
+
+            try
+            {
                 ld.backendNodes[DNN_BACKEND_VKCOM] =
                     layer->initVkCom(ld.inputBlobsWrappers);
             }
+            catch (const cv::Exception& e)
+            {
+                CV_LOG_ERROR(NULL, "initVkCom failed, fallback to CPU implementation. " << e.what());
+                ld.backendNodes[DNN_BACKEND_VKCOM] = Ptr<BackendNode>();
+            }
+        }
 #endif
     }
 
@@ -2317,9 +2352,18 @@ struct Net::Impl
                     forwardInfEngine(node);
                 }
                 else if (preferableBackend == DNN_BACKEND_VKCOM)
+                {
+                    try
                     {
                         forwardVkCom(ld.outputBlobsWrappers, node);
                     }
+                    catch (const cv::Exception& e)
+                    {
+                        CV_LOG_ERROR(NULL, "forwardVkCom failed, fallback to CPU implementation. " << e.what());
+                        it->second = Ptr<BackendNode>();
+                        forwardLayer(ld);
+                    }
+                }
                 else
                 {
                     CV_Error(Error::StsNotImplemented, "Unknown backend identifier");

modules/dnn/src/vkcom/shader/softmax.comp

@@ -18,6 +18,7 @@ layout(push_constant) uniform pushBlock {
     int channel_size;
     int outer_size;
     int channels;
+    int logsoftmax;
 } p;
 layout(local_size_x = LOCAL_SZ_X, local_size_y = 1, local_size_z = 1) in;
 
@@ -68,9 +69,8 @@ void main()
         for (int i = 0; i < p.channel_size; ++i)
         {
             float v = output_buffer[index] / sum_buffer[reduced_buffer_off + i];
-#ifdef LOG_SOFTMAX
+            if (p.logsoftmax == 1)
                 v = log(v);
-#endif
             output_buffer[index] = v;
             index++;
         }

modules/dnn/src/vkcom/shader/spv_shader.hpp

@@ -16,7 +16,7 @@ extern const unsigned int permute_spv[765];
 extern const unsigned int lrn_spv[1845];
 extern const unsigned int concat_spv[541];
 extern const unsigned int avg_pool_spv[1538];
-extern const unsigned int softmax_spv[1440];
+extern const unsigned int softmax_spv[1496];
 extern const unsigned int prior_box_spv[1480];
 extern const unsigned int max_pool_spv[1449];
 extern const unsigned int relu_spv[502];

modules/dnn/src/vkcom/src/common.hpp

@@ -42,7 +42,8 @@ enum ShapeIdx
 { \
         if (f != VK_SUCCESS) \
         { \
-            CV_LOG_WARNING(NULL, "Vulkan check failed"); \
+            CV_LOG_ERROR(NULL, "Vulkan check failed, result = " << f); \
+            CV_Error(Error::StsError, "Vulkan check failed"); \
         } \
 }
 

modules/dnn/src/vkcom/src/op_softmax.cpp

@@ -22,6 +22,7 @@ struct SoftmaxParam {
     int channel_size;
     int outer_size;
     int channels;
+    int logsoftmax;
 };
 
 OpSoftmax::OpSoftmax(const int axis, const bool log_softmax)
@@ -90,7 +91,7 @@ bool OpSoftmax::forward(Tensor& in, Tensor& out)
     bindTensor(device_, *max_tensor_,  1, descriptor_set_);
     bindTensor(device_, *sum_tensor_,  2, descriptor_set_);
     bindTensor(device_, out, 3, descriptor_set_);
-    SoftmaxParam param = {channel_size_, outer_size_, channels_};
+    SoftmaxParam param = {channel_size_, outer_size_, channels_, log_softmax_ == true ? 1 : 0};
     recordCommandBuffer((void *)&param, sizeof(SoftmaxParam));
     runCommandBuffer();
     return true;

modules/dnn/src/vkcom/vulkan/vk_functions.cpp

@@ -17,7 +17,7 @@
 
 namespace cv { namespace dnn { namespace vkcom {
 
-#include "function_list.inl"
+#include "function_list.inl.hpp"
 
 }}} // namespace cv::dnn::vkcom
 #endif // HAVE_VULKAN

modules/dnn/src/vkcom/vulkan/vk_functions.hpp

@@ -20,7 +20,7 @@
 
 namespace cv { namespace dnn { namespace vkcom {
 
-#include "function_list.inl"
+#include "function_list.inl.hpp"
 
 }}} // namespace cv::dnn::vkcom
 #endif // HAVE_VULKAN

modules/dnn/src/vkcom/vulkan/vk_loader.cpp

@@ -57,7 +57,7 @@ bool loadVulkanFunctions(VkInstance& instance)
       return false; \
     }
 
-#include "function_list.inl"
+#include "function_list.inl.hpp"
     return true;
 }
 
@@ -74,7 +74,7 @@ bool loadVulkanGlobalFunctions()
       return false; \
     }
 
-#include "function_list.inl"
+#include "function_list.inl.hpp"
     return true;
 }