[LITE][OPENCL][Image] fix issue in concat and nearest_interp thx for… (#3011)

* [LITE][OPENCL][Image] fix issue in concat and nearest_interp  thx for chenj and ys,test=develop

* [LITE][OPENCL][Image] fix issue in concat and nearest_interp  thx for chenj and ys,test=develop

lite/backends/opencl/cl_kernel/image/concat_kernel.cl

 /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-
 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
 You may obtain a copy of the License at
-
     http://www.apache.org/licenses/LICENSE-2.0
-
 Unless required by applicable law or agreed to in writing, software
 distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
@@ -52,7 +49,7 @@ __kernel void concat2(__read_only image2d_t input0,
         input_data = READ_IMG_TYPE(CL_DTYPE_CHAR, input1, sampler, input_pos);
       }
       int value_offset = c_in % 4;
-        float value;
+      CL_DTYPE value;
       if (value_offset == 0) {
         value = input_data.x;
       } else if (value_offset == 1) {
@@ -132,7 +129,7 @@ __kernel void concat_mul(__read_only image2d_t input,
       int2 output_pos;
       output_pos.x = (c_out / 4) * in_W + in_w;
       output_pos.y = in_nh;
-        float val;
+      CL_DTYPE val;
       if (i == 0) {
         val = input_data.x;
       } else if (i == 1) {

lite/kernels/opencl/concat_image_compute.cc

@@ -96,7 +96,7 @@ class ConcatComputeImage : public KernelLite<TARGET(kOpenCL),
     auto& param = *param_.get_mutable<param_t>();
     const auto& x_dims = param.output->dims();
     auto image_shape = InitImageDimInfoWith(x_dims);
-    auto* out_buf = param.output->mutable_data<float, cl::Image2D>(
+    auto* out_buf = param.output->mutable_data<half_t, cl::Image2D>(
         image_shape["width"], image_shape["height"]);
     const auto& y_dims = param.output->dims();  // useless: check dim only
 
@@ -107,21 +107,41 @@ class ConcatComputeImage : public KernelLite<TARGET(kOpenCL),
 
     auto inputs = param.x;
     int arg_idx = 0;
-    int width = inputs[0]->dims()[-1];
-    auto global_work_size = cl::NDRange{
-        static_cast<cl::size_type>(x_dims[-1]),
-        static_cast<cl::size_type>(image_shape["width"] / x_dims[-1]),
+    int width = inputs[0]->dims()[inputs[0]->dims().size() - 1];
+
+    LOG(INFO) << "concat 输入尺寸:  ";
+    for (size_t i = 0; i < inputs.size(); i++) {
+      LOG(INFO) << "inputs [" << i << "]"
+                << "[" << inputs[i]->dims().size() << "D]:"
+                << "   dims:" << inputs[i]->dims()[0] << " "
+                << inputs[i]->dims()[1] << " " << inputs[i]->dims()[2] << " "
+                << inputs[i]->dims()[3];
+    }
+    LOG(INFO) << "concat 输出尺寸:  ";
+    LOG(INFO) << " out  dims:  "
+              << "[" << x_dims.size() << "D]:" << x_dims[0] << " " << x_dims[1]
+              << " " << x_dims[2] << " " << x_dims[3];
+    LOG(INFO) << "axis_: " << axis_;
+    LOG(INFO) << "flag_: " << flag_;
+    auto global_work_size =
+        cl::NDRange{static_cast<cl::size_type>(x_dims[x_dims.size() - 1]),
+                    static_cast<cl::size_type>(image_shape["width"] /
+                                               x_dims[x_dims.size() - 1]),
                     static_cast<cl::size_type>(image_shape["height"])};
     VLOG(4) << TargetToStr(param.output->target());
     VLOG(4) << "image_shape(w,h):" << image_shape["width"] << " "
             << image_shape["height"];
     VLOG(4) << "x_dims[" << x_dims.size() << "D]:" << x_dims[0] << " "
-            << x_dims[1] << " " << x_dims[2] << " " << x_dims[3];
+            << x_dims[1] << " " << x_dims[2] << " " << x_dims[3]
+            << "x_dims[x_dims.size() - 1]" << x_dims[x_dims.size() - 1];
     VLOG(4) << "y_dims[" << y_dims.size() << "D]:" << y_dims[0] << " "
             << y_dims[1] << " " << y_dims[2] << " " << y_dims[3];
-    VLOG(4) << "width_: " << width_ << ", flag_: " << flag_;
+    LOG(INFO) << "width_: " << width_ << ", flag_: " << flag_;
+    VLOG(4) << "global_work_size: " << x_dims[x_dims.size() - 1] << "  "
+            << (image_shape["width"] / x_dims[x_dims.size() - 1]) << "  "
+            << (image_shape["height"]);
     auto kernel = context.cl_context()->GetKernel(kernel_key.str());
-    int out_w = x_dims[-1];
+    int out_w = x_dims[x_dims.size() - 1];
     int out_c = x_dims[1];
     if (inputs.size() == 2) {
       auto* x_buf0 = inputs[0]->data<float, cl::Image2D>();
@@ -159,12 +179,13 @@ class ConcatComputeImage : public KernelLite<TARGET(kOpenCL),
         auto in_dims = inputs[i]->dims();
         image_shape = InitImageDimInfoWith(in_dims);
         auto* x_buf = inputs[i]->data<float, cl::Image2D>();
-        auto in_w = in_dims[-1];
+        int in_w = in_dims[in_dims.size() - 1];
         VLOG(4) << "image_shape(w,h):" << image_shape["width"] << " "
                 << image_shape["height"];
-        global_work_size = cl::NDRange{
-            static_cast<cl::size_type>(in_dims[-1]),
-            static_cast<cl::size_type>(image_shape["width"] / in_dims[-1]),
+        global_work_size =
+            cl::NDRange{static_cast<cl::size_type>(in_dims[in_dims.size() - 1]),
+                        static_cast<cl::size_type>(image_shape["width"] /
+                                                   in_dims[in_dims.size() - 1]),
                         static_cast<cl::size_type>(image_shape["height"])};
         cl_int status = kernel.setArg(arg_idx, *x_buf);
         CL_CHECK_FATAL(status);
@@ -205,7 +226,7 @@ class ConcatComputeImage : public KernelLite<TARGET(kOpenCL),
   int width_ = 1;
   param_t* concat_param_{nullptr};
   std::string kernel_func_name_{};
-  std::string build_options_{"-DCL_DTYPE_half"};
+  std::string build_options_{" -DCL_DTYPE_half"};
   std::shared_ptr<cl::Event> event_{new cl::Event};
 };
 

lite/kernels/opencl/nearest_interp_image_compute.cc

@@ -45,15 +45,16 @@ class NearestInterpComputeImageDefault
   void Run() override {
     auto& param = *param_.get_mutable<param_t>();
     const auto& x_dims = param.X->dims();
+    const auto& y_dims = param.Out->dims();
     auto* x_buf =
         param.X->data<half_t,
                       cl::Image2D>();  // use half_t represents half float
-    auto image_shape = InitImageDimInfoWith(x_dims);
+    auto out_image_shape = InitImageDimInfoWith(y_dims);
     auto* out_buf = param.Out->mutable_data<half_t, cl::Image2D>(  // use half_t
         // represents half float
-        image_shape["width"],
-        image_shape["height"]);
-    const auto& y_dims = param.Out->dims();  // useless: check dim only
+        out_image_shape["width"],
+        out_image_shape["height"]);
+
     float scale_h = y_dims[2] / x_dims[2];
     float scale_w = y_dims[3] / x_dims[3];
     int in_dims_h = x_dims[2];
@@ -87,16 +88,22 @@ class NearestInterpComputeImageDefault
 
     VLOG(4) << TargetToStr(param.X->target());
     VLOG(4) << TargetToStr(param.Out->target());
-    VLOG(4) << "image_shape(w,h):" << image_shape["width"] << " "
-            << image_shape["height"];
+    VLOG(4) << "out_image_shape(w,h):" << out_image_shape["width"] << " "
+            << out_image_shape["height"];
     VLOG(4) << "x_dims[" << x_dims.size() << "D]:" << x_dims[0] << " "
             << x_dims[1] << " " << x_dims[2] << " " << x_dims[3];
     VLOG(4) << "y_dims[" << y_dims.size() << "D]:" << y_dims[0] << " "
             << y_dims[1] << " " << y_dims[2] << " " << y_dims[3];
 
+    const std::vector<size_t>& default_work_size =
+        DefaultWorkSize(y_dims,
+                        DDim(std::vector<DDim::value_type>{
+                            static_cast<int64_t>(out_image_shape["width"]),
+                            static_cast<int64_t>(out_image_shape["height"])}));
     auto global_work_size =
-        cl::NDRange{static_cast<cl::size_type>(image_shape["width"]),
-                    static_cast<cl::size_type>(image_shape["height"])};
+        cl::NDRange{static_cast<cl::size_type>(default_work_size.data()[0]),
+                    static_cast<cl::size_type>(default_work_size.data()[1]),
+                    static_cast<cl::size_type>(default_work_size.data()[2])};
     status = context.cl_context()->GetCommandQueue().enqueueNDRangeKernel(
         kernel,
         cl::NullRange,
@@ -112,7 +119,7 @@ class NearestInterpComputeImageDefault
 
  private:
   std::string kernel_func_name_{"nearest_interp"};
-  std::string build_options_{"-DCL_DTYPE_half"};
+  std::string build_options_{" -DCL_DTYPE_half"};
   std::shared_ptr<cl::Event> event_{new cl::Event};
 };