Merge branch 'develop' into develop

src/fpga/api.cpp

@@ -22,7 +22,7 @@ limitations under the License. */
 #include "fpga/filter.h"
 #include "fpga/image.h"
 #define FPGA_TEST_MODE
-#define PADDLE_MOBILE_OS_LINUX
+//#define PADDLE_MOBILE_OS_LINUX
 
 namespace paddle_mobile {
 namespace fpga {
@@ -125,6 +125,7 @@ float fp16_2_fp32(half fp16_num) {
 }
 
 int ComputeBasicConv(const struct ConvArgs &args) {
+#ifdef FPGA_TEST_MODE
   DLOG << "======Compute Basic Conv======";
   DLOG << "   relu_enabled:" << args.relu_enabled
        << "   sb_address:" << args.sb_address
@@ -144,7 +145,7 @@ int ComputeBasicConv(const struct ConvArgs &args) {
        << "   stride_w:" << args.kernel.stride_w;
   DLOG << "   out_address:" << args.output.address
        << "   out_scale_address:" << args.output.scale_address;
-
+#endif
   return do_ioctl(IOCTL_CONFIG_CONV, &args);
 }
 
@@ -192,8 +193,9 @@ int ComputeFpgaPool(const struct PoolingArgs &args) {
 int ComputeFpgaEWAdd(const struct EWAddArgs &args) {
 #ifdef FPGA_TEST_MODE
   DLOG << "=============ComputeFpgaEWAdd===========";
-  DLOG << "   relu_enabled:" << args.relu_enabled << "   const0:" << args.const0
-       << "   const1:" << args.const1;
+  DLOG << "   relu_enabled:" << args.relu_enabled
+       << "   const0:" << fp16_2_fp32(short(args.const0))
+       << "   const1:" << fp16_2_fp32(short(args.const1));
   DLOG << "   image0_address:" << args.image0.address
        << "   image0_scale_address:" << args.image0.scale_address
        << "   image0_channels:" << args.image0.channels
@@ -401,8 +403,8 @@ void fill_conv_arg(struct WrapperConvArgs *arg, framework::Tensor *input,
   arg->concat_arg.image_num = arg->split_num;
   arg->concat_arg.image_out = out_ptr;
   arg->concat_arg.scale_out = out->scale;
-  arg->concat_arg.height = (uint32_t)filter->dims()[2];
-  arg->concat_arg.width = (uint32_t)filter->dims()[3];
+  arg->concat_arg.height = (uint32_t)out->dims()[2];
+  arg->concat_arg.width = (uint32_t)out->dims()[3];
 
   int n = arg->split_num;
   arg->concat_arg.images_in =
@@ -411,7 +413,6 @@ void fill_conv_arg(struct WrapperConvArgs *arg, framework::Tensor *input,
       (float **)fpga_malloc(n * sizeof(float *));  // NOLINT
   arg->concat_arg.channel_num =
       (uint32_t *)fpga_malloc(n * sizeof(uint32_t));  // NOLINT
-  arg->concat_arg.image_out = out_ptr;
 
   auto channel = (int)out->dims()[1];  // NOLINT
   int filter_num_per_div = get_filter_num_per_div(filter, group_num);

src/fpga/bias_scale.cpp

@@ -27,6 +27,9 @@ void align_element(float **data_in, int num_per_div_before_alignment, int num) {
       (num + num_per_div_before_alignment - 1) / num_per_div_before_alignment;
   int num_per_div_after_alignment =
       align_to_x(num_per_div_before_alignment, BS_NUM_ALIGNMENT);
+  if (num_per_div_before_alignment == num_per_div_after_alignment) {
+    return;
+  }
   int num_element =
       2 * div_num * num_per_div_after_alignment;  // including bias & scale
   float *ptr_aligned =

src/fpga/filter.cpp

@@ -210,12 +210,12 @@ void format_filter(float **data_in, int num, int channel, int height, int width,
       align_to_x(num_per_div_before_alignment, FILTER_NUM_ALIGNMENT);
   int div_num =
       (num + num_per_div_before_alignment - 1) / num_per_div_before_alignment;
-  int num_after_alignment = num_per_div_after_alignment * div_num;
-
+  int residual = num % num_per_div_before_alignment;
+  int num_after_alignment = num_per_div_after_alignment *
+                                ((residual == 0) ? div_num : (div_num - 1)) +
+                            align_to_x(residual, FILTER_NUM_ALIGNMENT);
   quantize(data_in, data_size, max);
-
   char **quantize_data = (char **)data_in;  // NOLINT
-
   convert_to_hwc(quantize_data, num, channel, height, width);
   align_element(quantize_data, num, chw);
   align_num(quantize_data, num_per_div_before_alignment, num, chw);

src/operators/kernel/fpga/fc_relu_kernel.cpp

@@ -44,6 +44,7 @@ bool FusionFcReluKernel<FPGA, float>::Init(FusionFcReluParam<FPGA> *param) {
   int width = (uint32_t)input_x->dims()[3];
   int filter_channel = chw / height / width;
 
+  out->Resize(framework::make_ddim({1, channel, 1, 1}));
   filter->Resize(framework::make_ddim({num, filter_channel, height, width}));
   float max_value = fpga::filter_find_max(filter);
   fpga::format_fc_filter(filter, max_value);

src/operators/kernel/fpga/fusion_fc_kernel.cpp

@@ -45,6 +45,7 @@ bool FusionFcKernel<FPGA, float>::Init(FusionFcParam<FPGA> *param) {
   int width = (uint32_t)input_x->dims()[3];
   int filter_channel = chw / height / width;
 
+  out->Resize(framework::make_ddim({1, channel, 1, 1}));
   filter->Resize(framework::make_ddim({num, filter_channel, height, width}));
   float max_value = fpga::filter_find_max(filter);
   fpga::format_fc_filter(filter, max_value);

src/operators/kernel/fpga/mul_kernel.cpp

@@ -44,6 +44,7 @@ bool MulKernel<FPGA, float>::Init(MulParam<FPGA> *param) {
   int width = (uint32_t)input_x->dims()[3];
   int filter_channel = chw / height / width;
 
+  out->Resize(framework::make_ddim({1, channel, 1, 1}));
   filter->Resize(framework::make_ddim({num, filter_channel, height, width}));
   float max_value = fpga::filter_find_max(filter);
   fpga::format_fc_filter(filter, max_value);

src/operators/kernel/fpga/softmax_kernel.cpp

@@ -27,7 +27,7 @@ bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
   auto input = const_cast<Tensor *>(param->InputX());
   auto input_ptr = input->data<float>();
   auto float_input = new Tensor;
-  float_input->mutable_data<float>(input->dims());
+  float_input->mutable_data<float>({1, input->dims()[1]});
   fpga::format_fp32_ofm(float_input);
 
   fpga::BypassArgs args = {fpga::DATA_TYPE_FP16};
@@ -56,7 +56,6 @@ void SoftmaxKernel<FPGA, float>::Compute(
   fpga::fpga_invalidate(
       (void *)in_x->data<float>(),  // NOLINT
       fpga::get_align_image_cw(in_x->dims()[1]) * sizeof(float));
-
   math::SoftmaxFuntor<CPU, float>()(in_x, out);
   fpga::fpga_flush(out->data<float>(), out->memory_size());
 }