Merge pull request #967 from zhangyang0701/develop

modify Softmax for FPGA track close #966

src/fpga/api.cpp

@@ -181,10 +181,12 @@ int ComputeFPGAConcat(const struct ConcatArgs &args) {
   return 0;
 }
 
+int get_align_image_cw(int cw) { return align_to_x(cw, IMAGE_ALIGNMENT); }
+
 void format_image(framework::Tensor *image_tensor) {
   auto dims = image_tensor->dims();
   auto channel = dims[1], height = dims[2], width = dims[3];
-  auto data_ptr = image_tensor->mutable_data<float>();
+  auto data_ptr = image_tensor->data<float>();
   size_t memory_size = channel * height * width * sizeof(float);
   float *new_data = (float *)fpga_malloc(memory_size);
   fpga_copy(new_data, data_ptr, memory_size);
@@ -192,7 +194,7 @@ void format_image(framework::Tensor *image_tensor) {
   image_tensor->reset_data_ptr(new_data);
 }
 
-void format_ofm(framework::Tensor *ofm_tensor) {
+void format_fp16_ofm(framework::Tensor *ofm_tensor) {
   auto dims = ofm_tensor->dims();
   size_t memory_size = 0;
   if (dims.size() == 4) {
@@ -209,6 +211,23 @@ void format_ofm(framework::Tensor *ofm_tensor) {
   ofm_tensor->reset_data_ptr(p);
 }
 
+void format_fp32_ofm(framework::Tensor *ofm_tensor) {
+  auto dims = ofm_tensor->dims();
+  size_t memory_size = 0;
+  if (dims.size() == 4) {
+    auto channel = dims[1], height = dims[2], width = dims[3];
+    memory_size =
+        height * align_to_x(channel * width, IMAGE_ALIGNMENT) * sizeof(float);
+  } else if (dims.size() == 2) {
+    memory_size = align_to_x(dims[1], IMAGE_ALIGNMENT) * sizeof(float);
+  } else {
+    DLOG << "Wrong ofm dimension";
+  }
+  auto p = fpga_malloc(memory_size);
+  memset(p, 0, memory_size);
+  ofm_tensor->reset_data_ptr(p);
+}
+
 float filter_find_max(framework::Tensor *filter_tensor) {
   auto filter_ptr = filter_tensor->data<float>();
   return filter::find_max(filter_ptr, filter_tensor->numel());
@@ -242,7 +261,7 @@ void format_filter(framework::Tensor *filter_tensor, float max_value,
                    int group_num) {
   auto dims = filter_tensor->dims();
   auto num = dims[0], channel = dims[1], height = dims[2], width = dims[3];
-  auto data_ptr = filter_tensor->mutable_data<float>();
+  auto data_ptr = filter_tensor->data<float>();
   size_t memory_size = num * channel * height * width * sizeof(float);
   auto new_data = (float *)fpga_malloc(memory_size);
   fpga_copy(new_data, data_ptr, memory_size);
@@ -277,7 +296,7 @@ void fill_conv_arg(struct WrapperConvArgs *arg, framework::Tensor *input,
                    int padding_h, int padding_w, float *bs_ptr) {
   auto input_ptr = input->data<float>();
   auto filter_ptr = filter->data<float>();
-  auto out_ptr = out->mutable_data<float>();
+  auto out_ptr = out->data<float>();
 
   arg->group_num = (uint32_t)group_num;
   arg->split_num = (uint32_t)fpga::get_plit_num(filter);
@@ -300,7 +319,7 @@ void fill_conv_arg(struct WrapperConvArgs *arg, framework::Tensor *input,
       (uint32_t *)fpga::fpga_malloc(n * sizeof(uint32_t));
   arg->concat_arg.image_out = out_ptr;
 
-  const int channel = (int)out->dims()[1];
+  auto channel = (int)out->dims()[1];
   int filter_num_per_div = fpga::get_filter_num_per_div(filter, group_num);
   int element_num = fpga::get_aligned_filter_element_num(
       filter->dims()[1] * filter->dims()[2] * filter->dims()[3]);

src/fpga/api.h

@@ -206,8 +206,10 @@ int ComputeFPGAConcat(const struct ConcatArgs& args);
 
 static inline int align_to_x(int num, int x) { return (num + x - 1) / x * x; }
 
+int get_align_image_cw(int cw);
 void format_image(framework::Tensor* image_tensor);
-void format_ofm(framework::Tensor* ofm_tensor);  // only allocate memory
+void format_fp16_ofm(framework::Tensor* ofm_tensor);  // only allocate memory
+void format_fp32_ofm(framework::Tensor* ofm_tensor);
 
 float filter_find_max(framework::Tensor* filter_tensor);
 int get_filter_num_per_div(framework::Tensor* filter_tensor, int group_num);

src/operators/feed_op.h

@@ -45,7 +45,7 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
 
   void Init() {
     Tensor *output = param_.Out();
-    fpga::format_ofm(output);
+    fpga::format_fp16_ofm(output);
   }
 
   void RunImpl() const {
@@ -53,7 +53,7 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
     auto input_ptr = input->data<float>();
     fpga::format_image(input);
     Tensor *output = param_.Out();
-    auto output_ptr = output->mutable_data<half>();
+    auto output_ptr = output->data<half>();
 
     fpga::BypassArgs args;
 
@@ -62,9 +62,9 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
     args.input_layout_type = fpga::LAYOUT_CHW;
     args.output_layout_type = fpga::LAYOUT_HWC;
     args.image.address = (void *)input_ptr;
-    args.image.channels = input->dims()[1];
-    args.image.height = input->dims()[2];
-    args.image.width = input->dims()[3];
+    args.image.channels = (uint32_t)input->dims()[1];
+    args.image.height = (uint32_t)input->dims()[2];
+    args.image.width = (uint32_t)input->dims()[3];
     args.image.pad_height = 0;
     args.image.pad_width = 0;
     args.output.address = output_ptr;

src/operators/kernel/fpga/conv_add_bn_kernel.cpp

@@ -64,7 +64,7 @@ bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam<FPGA> *param) {
   int element_num_per_div =
       fpga::get_filter_num_per_div(filter, param->Groups());
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
-  fpga::format_ofm(out);
+  fpga::format_fp16_ofm(out);
 
   fpga::WrapperConvArgs conv_arg;
   fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,

src/operators/kernel/fpga/conv_add_bn_relu_kernel.cpp

@@ -62,7 +62,7 @@ bool ConvAddBNReluKernel<FPGA, float>::Init(
       fpga::get_filter_num_per_div(filter, param->Groups());
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
 
-  fpga::format_ofm(out);
+  fpga::format_fp16_ofm(out);
 
   fpga::WrapperConvArgs conv_arg;
   fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,

src/operators/kernel/fpga/conv_add_relu_kernel.cpp

@@ -44,7 +44,7 @@ bool ConvAddReluKernel<FPGA, float>::Init(FusionConvAddReluParam<FPGA> *param) {
       fpga::get_filter_num_per_div(filter, param->Groups());
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
 
-  fpga::format_ofm(out);
+  fpga::format_fp16_ofm(out);
 
   fpga::WrapperConvArgs conv_arg;
   fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,

src/operators/kernel/fpga/conv_bn_kernel.cpp

@@ -15,7 +15,6 @@ limitations under the License. */
 #ifdef FUSION_CONVBN_OP
 
 #include "operators/kernel/conv_bn_kernel.h"
-#include "fpga/api.h"
 
 namespace paddle_mobile {
 namespace operators {
@@ -33,10 +32,8 @@ bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam<FPGA> *param) {
   const float epsilon = param->Epsilon();
   PADDLE_MOBILE_ENFORCE(out->dims()[1] == param->InputBias()->dims()[0],
                         "Output channel should be equal to bias number");
-
   const int channel = out->dims()[1];
-  auto bs_ptr =
-      reinterpret_cast<float *>(fpga::fpga_malloc(2 * channel * sizeof(float)));
+  auto bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
   auto new_scale = new Tensor();
   auto new_bias = new Tensor();
   auto new_scale_ptr = new_scale->mutable_data<float>({channel});
@@ -59,7 +56,7 @@ bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam<FPGA> *param) {
       fpga::get_filter_num_per_div(filter, param->Groups());
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
 
-  fpga::format_ofm(out);
+  fpga::format_fp16_ofm(out);
 
   fpga::WrapperConvArgs conv_arg;
   fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,

src/operators/kernel/fpga/conv_bn_relu_kernel.cpp

@@ -56,7 +56,7 @@ bool ConvBNReluKernel<FPGA, float>::Init(FusionConvBNReluParam<FPGA> *param) {
       fpga::get_filter_num_per_div(filter, param->Groups());
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
 
-  fpga::format_ofm(out);
+  fpga::format_fp16_ofm(out);
 
   fpga::WrapperConvArgs conv_arg;
   fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,

src/operators/kernel/fpga/elementwise_add_relu_kernel.cpp

@@ -27,7 +27,7 @@ bool ElementwiseAddReluKernel<FPGA, float>::Init(
   auto *out = param->Out();
   auto input_x_ptr = input_x->data<float>();
   auto input_y_ptr = input_y->data<float>();
-  fpga::format_ofm(out);
+  fpga::format_fp16_ofm(out);
   auto out_ptr = out->mutable_data<float>();
 
   fpga::EWAddArgs ewaddArgs;

src/operators/kernel/fpga/fc_relu_kernel.cpp

@@ -49,7 +49,7 @@ bool FusionFcReluKernel<FPGA, float>::Init(FusionFcReluParam<FPGA> *param) {
 
   int element_num_per_div = fpga::get_filter_num_per_div(filter, 1);
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
-  fpga::format_ofm(out);
+  fpga::format_fp16_ofm(out);
 
   fpga::WrapperConvArgs conv_arg;
   fpga::fill_conv_arg(&conv_arg, input_x, out, filter, relu_enabled, 1, 1, 1, 0,

src/operators/kernel/fpga/fusion_fc_kernel.cpp

@@ -50,7 +50,7 @@ bool FusionFcKernel<FPGA, float>::Init(FusionFcParam<FPGA> *param) {
 
   int element_num_per_div = fpga::get_filter_num_per_div(filter, 1);
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
-  fpga::format_ofm(out);
+  fpga::format_fp16_ofm(out);
 
   fpga::WrapperConvArgs conv_arg;
   fpga::fill_conv_arg(&conv_arg, input_x, out, filter, relu_enabled, 1, 1, 1, 0,

src/operators/kernel/fpga/pool_kernel.cpp

@@ -24,7 +24,7 @@ bool PoolKernel<FPGA, float>::Init(PoolParam<FPGA> *param) {
   auto *input = const_cast<Tensor *>(param->Input());
   auto input_ptr = input->data<float>();
   Tensor *output = param->Output();
-  fpga::format_ofm(output);
+  fpga::format_fp16_ofm(output);
   auto output_ptr = output->mutable_data<float>();
   vector<int> ksize = param->Ksize();
   vector<int> strides = param->Strides();

src/operators/kernel/fpga/softmax_kernel.cpp

@@ -24,22 +24,23 @@ namespace operators {
 
 template <>
 bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
-  const Tensor *input = param->InputX();
+  auto input = const_cast<Tensor *>(param->InputX());
   auto input_ptr = input->data<float>();
-  auto output_ptr = param->Out();
-  Tensor *floatInput = new Tensor(*input);
+  auto float_input = new Tensor(*input);
+  fpga::format_fp32_ofm(float_input);
+
   fpga::BypassArgs args;
   args.input_layout_type = fpga::LAYOUT_HWC;
   args.output_layout_type = fpga::LAYOUT_CHW;
   args.input_data_type = fpga::DATA_TYPE_FP16;
   args.output_data_type = fpga::DATA_TYPE_FP32;
-  args.image.address = (void *)(input_ptr);
-  args.image.height = (uint32_t)input->dims()[0];
-  args.image.width = (uint32_t)input->dims()[1];
-  args.image.channels = 1;
-  args.output.address = (void *)floatInput->mutable_data<float>();
+  args.image.address = input_ptr;
+  args.image.height = 1;
+  args.image.width = 1;
+  args.image.channels = (uint32_t)input->dims()[1];
+  args.output.address = float_input->mutable_data<float>();
 
-  param->SetFloatInput(floatInput);
+  param->SetFloatInput(float_input);
   param->SetFpgaArgs(args);
   return true;
 }
@@ -47,17 +48,16 @@ bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
 template <>
 void SoftmaxKernel<FPGA, float>::Compute(
     const SoftmaxParam<FPGA> &param) const {
-  DLOG << "======================================= FPGA SoftMAX "
-          "===============================================";
-  const Tensor *in_x = param.FloatInput();
+  Tensor *in_x = param.FloatInput();
   Tensor *out = param.Out();
-  fpga::fpga_flush((void *)in_x->data<float>(), in_x->memory_size());
+
   fpga::PerformBypass(param.FpgaArgs());
-  fpga::fpga_invalidate(out->data<float>(), out->memory_size());
+  fpga::fpga_invalidate(
+      (void *)in_x->data<float>(),
+      (size_t)fpga::get_align_image_cw((int)in_x->dims()[1]) * sizeof(float));
 
-  auto x_dims = in_x->dims();
-  out->Resize(x_dims);
   math::SoftmaxFuntor<CPU, float>()(in_x, out);
+  fpga::fpga_flush(out->data<float>(), out->memory_size());
 }
 
 }  // namespace operators

test/fpga/test_format_data.cpp

@@ -71,7 +71,7 @@ void test_fill_conv_arg() {
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, 1001);
 
   DLOG << "format ofm";
-  fpga::format_ofm(&out);
+  fpga::format_fp16_ofm(&out);
   DLOG << "Build arg";
 
   fpga::WrapperConvArgs arg;