update FPGA v2 pe cpp file and ew kernel files, test=develop , closes #2432 (#2433)

* update proposal and psroipool kernel file in FPGA V2 track

* update, test=develop

* update FPGA v2 pe cpp file and ew kernel files, test=develop

mobile/src/fpga/V2/api.cpp

@@ -623,7 +623,7 @@ void fill_split_arg(struct SplitConvArgs *arg, framework::Tensor *input,
 
     arg->concat_arg.images_in[i] =
         (int8_t *)arg->conv_arg[i].output.address;  // NOLINT
-    arg->concat_arg.scales_in[i] = arg->conv_arg[i].output.scale_address;
+    arg->concat_arg.scales_in[i] = out->scale;
     arg->concat_arg.channel_num[i] = arg->conv_arg[i].filter_num;
 
     expand_conv_arg(&arg->conv_arg[i]);

mobile/src/operators/kernel/fpga/V2/anchor_generator_kernel.cpp

@@ -37,7 +37,7 @@ bool AnchorGeneratorKernel<FPGA, float>::Init(
   int anchors_offset[] = {-2,  -2,   18,   18,  -10, -9,   26,   25,   -23,
                           -20, 39,   36,   -43, -34, 59,   49,   -63,  -54,
                           79,  69,   -96,  -77, 112, 93,   -137, -118, 153,
-                          134, -204, -188, 220, 204, -281, -395, 296,  441};
+                          134, -204, -188, 220, 204, -281, -395, 296,  411};
 
   int anchors_offset2[] = {0, 0, 51, 77, 0, 0, 30, 35, 0, 0, 81, 103,
                            0, 0, 20, 21, 0, 0, 36, 44, 0, 0, 43, 58,

mobile/src/operators/kernel/fpga/V2/elementwise_add_kernel.cpp

@@ -12,12 +12,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 #ifdef ELEMENTWISEADD_OP
-
+#include <math.h>
 #include "operators/kernel/elementwise_add_kernel.h"
 
-#include <string>
-#include "fpga/V2/api.h"
-
 namespace paddle_mobile {
 namespace operators {
 
@@ -60,10 +57,36 @@ bool ElementwiseAddKernel<FPGA, float>::Init(ElementwiseAddParam<FPGA> *param) {
   return true;
 }
 
+void ComputeCPUEWAdd(fpga::EWAddArgs ewaddArgs) {
+  int inputc = ewaddArgs.image0.channels;
+  int inputh = ewaddArgs.image0.height;
+  int inputw = ewaddArgs.image0.width;
+  float inScale0 =
+          (reinterpret_cast<float*>(ewaddArgs.image0.scale_address))[0];
+  float inScale1 =
+          (reinterpret_cast<float*>(ewaddArgs.image1.scale_address))[0];
+  float outScale =
+          (reinterpret_cast<float*>(ewaddArgs.output.scale_address))[0];
+  int8_t* inPtr0 = reinterpret_cast<int8_t*>(ewaddArgs.image0.address);
+  int8_t* inPtr1 = reinterpret_cast<int8_t*>(ewaddArgs.image1.address);
+  int8_t* outPtr = reinterpret_cast<int8_t*>(ewaddArgs.output.address);
+  int datasize = inputc * inputh * inputw;
+  float const0 = inScale0 / outScale;
+  float const1 = inScale1 / outScale;
+  fpga::fpga_invalidate(inPtr0, datasize * sizeof(int8_t));
+  fpga::fpga_invalidate(inPtr1, datasize * sizeof(int8_t));
+  for (int i = 0; i < datasize; i++) {
+    float tmpF = inPtr0[i] * const0 + inPtr1[i] * const1;
+    int tmpI = static_cast<int>(round(tmpF));
+    outPtr[i] = (int8_t)((tmpI > 127 ? 127 : (tmpI < -127 ? -127 : tmpI)));
+  }
+  fpga::fpga_flush(outPtr, datasize * sizeof(int8_t));
+}
 template <>
 void ElementwiseAddKernel<FPGA, float>::Compute(
     const ElementwiseAddParam<FPGA> &param) {
-  fpga::ComputeFpgaEWAdd(param.FpgaArgs());
+  // fpga::ComputeFpgaEWAdd(param.FpgaArgs());
+  ComputeCPUEWAdd(param.FpgaArgs());
 }
 }  // namespace operators
 }  // namespace paddle_mobile

mobile/src/operators/kernel/fpga/V2/elementwise_add_relu_kernel.cpp

@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 #ifdef FUSION_ELEMENTWISEADDRELU_OP
-
+#include <math.h>
 #include "operators/kernel/elementwise_add_relu_kernel.h"
 
 namespace paddle_mobile {
@@ -58,10 +58,37 @@ bool ElementwiseAddReluKernel<FPGA, float>::Init(
   return true;
 }
 
+void ComputeCPUEWAddRelu(fpga::EWAddArgs ewaddArgs) {
+  int inputc = ewaddArgs.image0.channels;
+  int inputh = ewaddArgs.image0.height;
+  int inputw = ewaddArgs.image0.width;
+  float inScale0 =
+          (reinterpret_cast<float*>(ewaddArgs.image0.scale_address))[0];
+  float inScale1 =
+          (reinterpret_cast<float*>(ewaddArgs.image1.scale_address))[0];
+  float outScale =
+          (reinterpret_cast<float*>(ewaddArgs.output.scale_address))[0];
+  int8_t* inPtr0 = reinterpret_cast<int8_t*>(ewaddArgs.image0.address);
+  int8_t* inPtr1 = reinterpret_cast<int8_t*>(ewaddArgs.image1.address);
+  int8_t* outPtr = reinterpret_cast<int8_t*>(ewaddArgs.output.address);
+  int datasize = inputc * inputh * inputw;
+  float const0 = inScale0 / outScale;
+  float const1 = inScale1 / outScale;
+  fpga::fpga_invalidate(inPtr0, datasize * sizeof(int8_t));
+  fpga::fpga_invalidate(inPtr1, datasize * sizeof(int8_t));
+  for (int i = 0; i < datasize; i++) {
+    float tmpF = inPtr0[i] * const0 + inPtr1[i] * const1;
+    int tmpI = static_cast<int>(round(tmpF));
+    outPtr[i] = (int8_t)((tmpI > 127 ? 127 : (tmpI < 0 ? 0 : tmpI)));
+  }
+  fpga::fpga_flush(outPtr, datasize * sizeof(int8_t));
+}
+
 template <>
 void ElementwiseAddReluKernel<FPGA, float>::Compute(
     const ElementwiseAddReluParam<FPGA> &param) {
-  fpga::ComputeFpgaEWAdd(param.FpgaArgs());
+  // fpga::ComputeFpgaEWAdd(param.FpgaArgs());
+  ComputeCPUEWAddRelu(param.FpgaArgs());
 }
 }  // namespace operators
 }  // namespace paddle_mobile