[LITE][OPENCL] Add local work size, global work size for opencl profiler; fix...

[LITE][OPENCL] Add local work size, global work size for opencl profiler; fix bug for variable inference pass. test=develop (#3741)

lite/core/mir/variable_place_inference_pass.h

@@ -174,7 +174,8 @@ class VariablePlaceInferencePass : public DebugPass {
           // If is quantization, infer the Int8 type.
           if (type->precision() == PRECISION(kInt8)) {
             x_out->AsArg().type = type;
-          } else if (type->precision() == PRECISION(kFP16)) {
+          } else if (type->precision() == PRECISION(kFP16) &&
+                     type->target() != TARGET(kOpenCL)) {
             x_out->AsArg().type = type;
           } else {
             PrecisionType tmp_ptype = x_out->AsArg().type->precision();

lite/core/profile/profiler.cc

@@ -164,6 +164,10 @@ std::string Profiler::Summary(Type type, bool concise, size_t w) {
      << " " << setw(9) << left << "clMin(ms)"
      << " " << setw(9) << left << "clMax(ms)"
      << " " << setw(9) << left << "clAvg(%)";
+  if (!concise) {
+    ss << " " << setw(12) << left << "GlobalWorkSize"
+       << " " << setw(12) << left << "LocalWorkSize";
+  }
 #endif
   ss << std::endl;
 
@@ -301,7 +305,10 @@ std::string Profiler::Summary(Type type, bool concise, size_t w) {
          << cl_times.Avg(w) << " " << setw(9) << left << fixed
          << setprecision(3) << cl_times.Min(w) << " " << setw(9) << left
          << fixed << setprecision(3) << cl_times.Max(w) << " " << left
-         << setprecision(2) << cl_percent << "%   ";
+         << setprecision(2) << cl_percent << "%   "
+         << " " << setw(12) << left << fixed
+         << unit.Character().global_work_size << " " << setw(12) << left
+         << fixed << unit.Character().local_work_size;
 #endif
       ss << std::endl;
     }

lite/core/profile/profiler.h

@@ -19,6 +19,9 @@
 #include <vector>
 #include "lite/core/profile/timer.h"
 #include "lite/core/tensor.h"
+#ifdef LITE_WITH_OPENCL
+#include "lite/backends/opencl/cl_include.h"
+#endif
 
 namespace paddle {
 namespace lite {
@@ -63,6 +66,21 @@ struct OpCharacter {
 
 #ifdef LITE_WITH_OPENCL
   cl::Event cl_event{};
+  std::string global_work_size{"N/A"};
+  std::string local_work_size{"N/A"};
+
+  std::string NDRangeToStr(const cl::NDRange& range) {
+    std::string range_str{""};
+    const size_t range_size = 3;
+    for (size_t i = 0; i < range_size /*range.size()*/; ++i) {
+      LOG(INFO) << "range[" << i << "]:" << std::to_string(range[i]);
+      range_str += std::to_string(range[i]);
+      if (i != range_size - 1) {
+        range_str += ",";
+      }
+    }
+    return range_str;
+  }
 #else
   void* cl_event{nullptr};
 #endif

lite/kernels/opencl/conv_image_compute.h

@@ -48,6 +48,8 @@ class ConvImageCompute : public KernelLite<TARGET(kOpenCL),
 #ifdef LITE_WITH_PROFILE
   void SetProfileRuntimeKernelInfo(paddle::lite::profile::OpCharacter* ch) {
     ch->kernel_func_name = kernel_func_names_[0];
+    ch->global_work_size = ch->NDRangeToStr(global_work_size_);
+    ch->local_work_size = ch->NDRangeToStr(local_work_size_);
     ch->cl_event =
         event_;  // `event_` defined in `kernel.h`, valid after kernel::Run
   }