Merge pull request #1057 from codeWorm2015/opencl

modified build.sh

src/framework/cl/cl_scope.h

@@ -54,7 +54,7 @@ class CLScope {
     }
 
     auto program =
-        CLEngine::Instance()->CreateProgramWith(context_.get(), file_name);
+        CLEngine::Instance()->CreateProgramWith(context_.get(), "./cl_kernel/" + file_name);
     programs_[file_name] = std::move(program);
 
     status_ = clBuildProgram(program.get(), 0, 0, 0, 0, 0);

src/operators/kernel/cl/cl_kernel/common.h → src/operators/kernel/cl/cl_kernel/cl_common.h

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #pragma once;
 
-/*
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
 
 inline hafl4 activation(half4 in
 #ifdef PRELU
@@ -32,4 +32,3 @@ inline hafl4 activation(half4 in
   return output;
 }
 
-*/

src/operators/kernel/cl/cl_kernel/conv_kernel.inc.cl

@@ -23,9 +23,7 @@ conv_add_bn_relu
 
 */
 
-/*
-
-#include "common.h"
+#include "cl_common.h"
 
 __kernel void conv_1x1(__private const int global_size_dim0,
                        __private const int global_size_dim1,
@@ -305,5 +303,3 @@ __kernel void depth_conv_3x3(__private const int global_size_dim0,
     int2 output_pos(out_c * global_size_dim1 + out_w, out_nh);
     write_imageh(output_image, output_pos, output);
 }
-
-*/

tools/build.sh

@@ -90,6 +90,8 @@ build_for_android() {
     fi
     cd "../build/release/${PLATFORM}"
     make -j 8
+    mkdir ./build/cl_kernel
+    cp ../../../src/operators/kernel/cl/cl_kernel/*  ./build/cl_kernel/
 }
 
 build_for_ios() {