eltwise layer SUM op update

[platform/upstream/opencv.git] / modules / dnn / src / layers / eltwise_layer.cpp

diff --git a/modules/dnn/src/layers/eltwise_layer.cpp b/modules/dnn/src/layers/eltwise_layer.cpp
index 4037573..27bd7ee 100644 (file)
--- a/modules/dnn/src/layers/eltwise_layer.cpp
+++ b/modules/dnn/src/layers/eltwise_layer.cpp
@@ -43,6 +43,7 @@
 #include "../precomp.hpp"
 #include "layers_common.hpp"
 #include "op_halide.hpp"
+#include "opencl_kernels_dnn.hpp"
 
 namespace cv
 {
@@ -271,22 +272,47 @@ public:
         switch (op)
         {
             case SUM:
-                if (coeffs.empty())
-                {
-                    add(inputs[0], inputs[1], outputs[0]);
-                    for (int i = 2; i < inputs.size(); ++i)
-                        add(outputs[0], inputs[i], outputs[0]);
-                }
-                else
                 {
-                    UMat mul0, mul1;
-                    multiply(coeffs[0], inputs[0], mul0);
-                    multiply(coeffs[1], inputs[1], mul1);
-                    add(mul0, mul1, outputs[0]);
-                    for (int i = 2; i < inputs.size(); ++i)
+                    int channels = total(shape(outputs[0]), 0, 2);
+                    int plane_size = total(shape(outputs[0]), 2);
+                    if (channels % 4 == 0 && plane_size % 4 == 0)
+                    {
+                        size_t localsize[] = { 128 };
+                        size_t globalsize[] = { (size_t)channels / 4 * localsize[0] };
+
+                        for (int i = 0; i < (inputs.size() - 1); ++i)
+                        {
+                            String buildopt = format("-DLOOP=%d", i);
+                            ocl::Kernel kernel("op_sum4", ocl::dnn::eltwise_oclsrc, buildopt);
+                            int idx = 0;
+                            UMat inpMat = (i == 0) ? inputs[0] : UMat();
+                            float coeff1 = (coeffs.empty() || i > 0) ? 1.0f : coeffs[i];
+                            float coeff2 = coeffs.empty() ? 1.0f : coeffs[i + 1];
+                            kernel.set(idx++, ocl::KernelArg::PtrReadOnly(inputs[0]));
+                            kernel.set(idx++, ocl::KernelArg::PtrReadOnly(inputs[1]));
+                            kernel.set(idx++, (int)plane_size);
+                            kernel.set(idx++, (float)coeff1);
+                            kernel.set(idx++, (float)coeff2);
+                            kernel.set(idx++, ocl::KernelArg::PtrReadWrite(outputs[0]));
+                            bool ret = kernel.run(1, globalsize, localsize, false);
+                            if (!ret)
+                                return false;
+                        }
+                    }
+                    else
                     {
-                        multiply(coeffs[i], inputs[i], mul0);
-                        add(mul0, outputs[0], outputs[0]);
+                        float coeff1 = coeffs.empty() ? 1.f : coeffs[0];
+                        float coeff2 = coeffs.empty() ? 1.f : coeffs[1];
+                        UMat mul0, mul1;
+                        multiply(coeff1, inputs[0], mul0);
+                        multiply(coeff2, inputs[1], mul1);
+                        add(mul0, mul1, outputs[0]);
+                        for (int i = 2; i < inputs.size(); ++i)
+                        {
+                            float coeff = coeffs.empty() ? 1.f : coeffs[i];
+                            multiply(coeff, inputs[i], mul0);
+                            add(mul0, outputs[0], outputs[0]);
+                        }
                     }
                 }
                 break;