Merge remote-tracking branch 'origin/fpga_from_lite' into merge

X# especially if it merges an updated upstream into a topic branch.

lite/api/resnet50_test_fpga.cc

@@ -31,11 +31,7 @@ TEST(ResNet50, test) {
   std::vector<Place> valid_places(
       {Place{TARGET(kFPGA), PRECISION(kFP16), DATALAYOUT(kNHWC)}});
 
-  predictor.Build(FLAGS_model_dir,
-                  "",
-                  "",
-                  Place{TARGET(kFPGA), PRECISION(kFP16), DATALAYOUT(kNHWC)},
-                  valid_places);
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
 
   auto* input_tensor = predictor.GetInput(0);
   input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));

lite/gen_code/paddle_infer.h

@@ -46,7 +46,7 @@ class Tensor {
  */
 class PaddlePredictor {
  public:
-  void Init();
+  void Init() {}
 
   std::unique_ptr<Tensor> GetTensor(const std::string &id) const;
   std::unique_ptr<Tensor> GetMutableTensor(const std::string &id);

lite/kernels/fpga/conv_compute.cc

@@ -13,9 +13,9 @@
 // limitations under the License.
 
 #include "lite/kernels/fpga/conv_compute.h"
+#include <vector>
 #include "lite/core/op_registry.h"
 #include "lite/core/type_system.h"
-
 namespace paddle {
 namespace lite {
 namespace kernels {
@@ -26,9 +26,11 @@ using float16 = zynqmp::float16;
 void ConvCompute::PrepareForRun() {
   auto& param = this->Param<param_t>();
   param.output->mutable_data<float16>();
+  int pad_h = (*param.paddings)[0];
+  int pad_w = (*param.paddings)[2];
   // ====================================================
-  if (param.x->ZynqTensor()->shape().channel() != 1 && 
-    param.groups == param.x->ZynqTensor()->shape().channel()) {
+  if (param.x->ZynqTensor()->shape().channel() != 1 &&
+      param.groups == param.x->ZynqTensor()->shape().channel()) {
     zynqmp::DepthwiseConvParam& conv_param = dw_conv_pe_.param();
 
     conv_param.input = param.x->ZynqTensor();
@@ -37,8 +39,8 @@ void ConvCompute::PrepareForRun() {
     conv_param.filter->setDataType(zynqmp::FP32);
     conv_param.groups = param.groups;
     conv_param.strides = param.strides;
-    conv_param.paddings = param.paddings;
-    conv_param.dilations = param.dilations;
+    conv_param.paddings = std::vector<int>({pad_h, pad_w});
+    conv_param.dilations = *param.dilations;
     fill_scale_bias_const(&conv_param);
     conv_param.bias()->copyFrom(param.bias->ZynqTensor());
     conv_param.relu.enabled = param.fuse_relu;
@@ -53,14 +55,14 @@ void ConvCompute::PrepareForRun() {
     conv_param.filter->setDataType(zynqmp::FP32);
     conv_param.groups = param.groups;
     conv_param.strides = param.strides;
-    conv_param.paddings = param.paddings;
-    conv_param.dilations = param.dilations;
+    conv_param.paddings = std::vector<int>({pad_h, pad_w});
+    conv_param.dilations = *param.dilations;
     fill_scale_bias_const(&conv_param);
     if (param.bias != nullptr) {
       conv_param.bias()->copyFrom(param.bias->ZynqTensor());
       std::cout << "copy bias \n";
     }
-    
+
     conv_param.relu.enabled = param.fuse_relu;
 
     // conv_param.filter->saveToFile("filter", true);
@@ -74,8 +76,8 @@ void ConvCompute::PrepareForRun() {
 void ConvCompute::Run() {
   auto& param = this->Param<param_t>();
   // std::cout << "in:" << param.x->ZynqTensor()->data<void>() << std::endl;
-  if (param.x->ZynqTensor()->shape().channel() != 1 && 
-    param.groups == param.x->ZynqTensor()->shape().channel()) {
+  if (param.x->ZynqTensor()->shape().channel() != 1 &&
+      param.groups == param.x->ZynqTensor()->shape().channel()) {
     dw_conv_pe_.dispatch();
     // param.output->ZynqTensor()->saveToFile("dw", true);
   } else {

lite/kernels/fpga/conv_compute_test.cc

@@ -143,11 +143,11 @@ void conv_compute_ref(const operators::ConvParam& param) {
   int kernel_h = param.filter->dims()[3];
   int stride_w = param.strides[0];
   int stride_h = param.strides[1];
-  int dila_w = param.dilations[0];
-  int dila_h = param.dilations[1];
+  int dila_w = (*param.dilations)[0];
+  int dila_h = (*param.dilations)[1];
 
-  int pad_w = param.paddings[0];
-  int pad_h = param.paddings[1];
+  int pad_w = (*param.paddings)[2];
+  int pad_h = (*param.paddings)[0];
   bool flag_bias = (param.bias != nullptr);
   bool flag_relu = param.fuse_relu;
 
@@ -277,9 +277,10 @@ TEST(conv_fpga, compute) {
                             param.bias = &bias;
                           }
                           param.fuse_relu = flag_relu;
-                          param.paddings = std::vector<int>({padding, padding});
+                          *param.paddings = std::vector<int>(
+                              {padding, padding, padding, padding});
                           param.strides = std::vector<int>({stride, stride});
-                          param.dilations =
+                          *param.dilations =
                               std::vector<int>({dilation, dilation});
                           param.groups = group;
                           conv.SetParam(param);

lite/kernels/fpga/pooling_compute.cc

@@ -38,7 +38,9 @@ void PoolCompute::PrepareForRun() {
   pool_param.globalPooling = param.global_pooling;
   pool_param.kernelSize = param.ksize;
   pool_param.strides = param.strides;
-  pool_param.paddings = param.paddings;
+  int pad_h = (*param.paddings)[0];
+  int pad_w = (*param.paddings)[2];
+  pool_param.paddings = std::vector<int>({pad_h, pad_w});
   pe_.init();
   pe_.apply();
 }

lite/kernels/fpga/pooling_compute_test.cc

@@ -46,7 +46,7 @@ std::vector<int64_t> compute_output_shape(operators::PoolParam* param_) {
   if (param_->global_pooling) {
     ksize.resize(static_cast<size_t>(x_dims.size()) - 2);
     for (size_t i = 0; i < ksize.size(); ++i) {
-      param_->paddings[i] = 0;
+      (*param_->paddings)[i] = 0;
       ksize[i] = static_cast<int>(x_dims[i + 2]);
     }
   }
@@ -59,7 +59,7 @@ std::vector<int64_t> compute_output_shape(operators::PoolParam* param_) {
     for (size_t i = 0; i < param_->ksize.size(); ++i) {
       output_shape.push_back(PoolOutputSize(x_dims[i + 2],
                                             param_->ksize[i],
-                                            param_->paddings[i],
+                                            (*param_->paddings)[i],
                                             param_->strides[i],
                                             param_->ceil_mode));
     }
@@ -76,7 +76,7 @@ void pool_compute_ref(const operators::PoolParam& param) {
 
   std::vector<int> ksize = param.ksize;
   std::vector<int> strides = param.strides;
-  std::vector<int> paddings = param.paddings;
+  std::vector<int> paddings = *param.paddings;
 
   std::string pooling_type = param.pooling_type;
   bool global_pooling = param.global_pooling;
@@ -103,7 +103,7 @@ void pool_compute_ref(const operators::PoolParam& param) {
   int stride_h = strides[0];
   int stride_w = strides[1];
   int pad_h = paddings[0];
-  int pad_w = paddings[1];
+  int pad_w = paddings[2];
 
   if (global_pooling == true) {
     for (int n = 0; n < in_n; ++n) {
@@ -230,7 +230,7 @@ TEST(pool_fpga, compute) {
                         }
                         param.global_pooling = global_pooling;
                         param.strides = {stride, stride};
-                        param.paddings = {pad, pad};
+                        *param.paddings = {pad, pad, pad, pad};
                         param.exclusive = exclusive;
                         param.ceil_mode = ceil_mode;
                         param.adaptive = false;

lite/tools/build_fpga.sh

@@ -2,12 +2,16 @@
 
 build_dir=build_fpga
 mkdir -p ${build_dir}
-cd ${build_dir}
 
-GEN_CODE_PATH_PREFIX=lite/gen_code
-mkdir -p ./${GEN_CODE_PATH_PREFIX}
-touch ./${GEN_CODE_PATH_PREFIX}/__generated_code__.cc
+root_dir=$(pwd)
+build_dir=${build_dir}
+# in build directory
+# 1. Prepare gen_code file
+GEN_CODE_PATH_PREFIX=${build_dir}/lite/gen_code
+mkdir -p ${GEN_CODE_PATH_PREFIX}
+touch ${GEN_CODE_PATH_PREFIX}/__generated_code__.cc
 
+cd ${build_dir}
 cmake .. \
         -DWITH_GPU=OFF \
         -DWITH_MKL=OFF \
@@ -23,6 +27,5 @@ cmake .. \
         -DLITE_BUILD_EXTRA=ON \
         -DLITE_WITH_PROFILE=ON
 
-make -j32
-
+make -j4
 cd -