[lite] fix fc bias and enable armv7 fc (#17695)

cmake/cross_compiling/android.cmake

@@ -45,7 +45,8 @@ if (NOT ANDROID_ARCH_ABI IN_LIST ANDROID_ARCH_ABI_LIST)
 endif()
 
 if(ANDROID_ARCH_ABI STREQUAL "armeabi-v7a")
-    message(STATUS "NEON is enabled on arm-v7a")
+    set(CMAKE_ANDROID_ARM_NEON ON)
+    message(STATUS "NEON is enabled on arm-v7a with softfp")
 endif()
 
 set(ANDROID_STL_TYPE_LITS "gnustl_static" "c++_static")

paddle/fluid/lite/arm/math/funcs.cc

@@ -21,14 +21,14 @@ namespace arm {
 namespace math {
 
 template <>
-void fill_bias_fc<float>(float *tensor, const float *bias, const int num,
+void fill_bias_fc<float>(float *out, const float *bias, const int num,
                          const int channel) {
   int cnt = channel >> 4;
   int remain = channel & 15;
 
   for (int j = 0; j < num; ++j) {
     const float *ptr_bias = bias;
-    float *ptr_out = tensor + j * channel;
+    float *ptr_out = out + j * channel;
 
     float32x4_t vout1;
     float32x4_t vout2;
@@ -61,7 +61,6 @@ void fill_bias_fc<float>(float *tensor, const float *bias, const int num,
       ptr_out += 16;
       ptr_bias += 16;
     }
-
 #if 0
         if (cnt > 0) {
             asm(
@@ -79,21 +78,21 @@ void fill_bias_fc<float>(float *tensor, const float *bias, const int num,
             );
         }
 #endif
-    for (; remain > 0; remain--) {
+    for (int i = 0; i < remain; ++i) {
       *(ptr_out++) += *(ptr_bias++);
     }
   }
 }
 
 template <>
-void fill_bias_fc<int>(int *tensor, const int *bias, const int num,
+void fill_bias_fc<int>(int *out, const int *bias, const int num,
                        const int channel) {
   int cnt = channel >> 4;
   int remain = channel & 15;
 
   for (int j = 0; j < num; ++j) {
     const int *ptr_bias = bias;
-    int *ptr_out = tensor + j * channel;
+    int *ptr_out = out + j * channel;
 
     int32x4_t vout1;
     int32x4_t vout2;
@@ -144,7 +143,7 @@ void fill_bias_fc<int>(int *tensor, const int *bias, const int num,
         );
     }
 #endif
-    for (; remain > 0; remain--) {
+    for (int i = 0; i < remain; ++i) {
       *(ptr_out++) += *(ptr_bias++);
     }
   }

paddle/fluid/lite/arm/math/packed_sgemm.cc

@@ -2751,7 +2751,7 @@ void sgemm_conv_4x8(const float* A_packed, const float* B, const float* bias,
                     bool transB, ARMContext* ctx) {
   size_t l2_cache =
       ctx->l2_cache_size() > 0 ? ctx->l2_cache_size() : 512 * 1024;
-  void* workspace = ctx->get_work_space();
+  auto* workspace = ctx->workspace_data<float>();
   int threads = ctx->threads();
   //! MBLOCK * x (result) + MBLOCK * k (A) + x * k (B) = l2
   int x_block =
@@ -2785,7 +2785,7 @@ void sgemm_conv_4x8(const float* A_packed, const float* B, const float* bias,
       flag_p_remain = true;
     }
     //! load bpanel
-    float* b_pannel = static_cast<float*>(workspace);
+    float* b_pannel = workspace;
     if (transB) {
       loadb_trans(b_pannel, B, K, 0, K, x0, xmax);
     } else {

paddle/fluid/lite/core/cpu_info.cc

@@ -20,7 +20,7 @@ namespace lite {
 
 #ifdef LITE_WITH_ARM
 
-void DeviceInfo::get_info(DeviceInfo* dev) {
+void DeviceInfo::InitInternal(DeviceInfo* dev) {
   set_default_cache(dev);
   dev->compute_core_num_ = arm_get_cpucount();
   dev->max_memory_ = arm_get_meminfo();

paddle/fluid/lite/core/cpu_info.h

@@ -85,14 +85,14 @@ class DeviceInfo {
     return *x;
   }
 
-  static void init_info() {
+  static void Init() {
     auto& info = Global();
-    get_info(&info);
+    InitInternal(&info);
   }
 
  private:
   DeviceInfo() = default;
-  static void get_info(DeviceInfo* dev);
+  static void InitInternal(DeviceInfo* dev);
 };
 
 size_t arm_get_meminfo();

paddle/fluid/lite/core/mir/runtime_context_assign_pass.cc

@@ -64,7 +64,7 @@ class RuntimeContextAssignPass : public StmtPass {
 
 #ifdef LITE_WITH_ARM
   std::unique_ptr<KernelContext> NewARMContext() {
-    DeviceInfo::init_info();
+    DeviceInfo::Init();
     std::unique_ptr<KernelContext> ctx(new KernelContext);
     ctx->As<ARMContext>();
     return ctx;

paddle/fluid/lite/kernels/arm/CMakeLists.txt

@@ -9,7 +9,7 @@ cc_library(relu_compute_arm SRCS relu_compute.cc DEPS ${lite_kernel_deps})
 cc_library(mul_compute_arm SRCS mul_compute.cc DEPS ${lite_kernel_deps} eigen3)
 cc_library(scale_compute_arm SRCS scale_compute.cc DEPS ${lite_kernel_deps} eigen3)
 
-lite_cc_test(test_fc_compute_arm SRCS fc_compute_test.cc DEPS fc_compute_arm eigen3)
+lite_cc_test(test_fc_compute_arm SRCS fc_compute_test.cc DEPS fc_compute_arm math_arm)
 
 set(arm_kernels
     fc_compute_arm

paddle/fluid/lite/kernels/arm/fc_compute_test.cc

@@ -61,8 +61,6 @@ TEST(fc_arm, compare_test) {
     b_data[i] = static_cast<float>(i);
   }
 
-  // TODO(TJ): enable bias soon
-  b_data = nullptr;
   lite::arm::math::fc_compute_eigen(x_data, batch_size, 3,  //
                                     w_data, 3, 4,           //
                                     b_data, ref_data);
@@ -74,11 +72,11 @@ TEST(fc_arm, compare_test) {
   param.in_num_col_dims = 1;
   param.input = &x;
   param.w = &w;
-  param.bias = nullptr;
+  param.bias = &b;
   param.output = &out;
   param.in_mat_dims = x.dims();
 
-  DeviceInfo::init_info();
+  DeviceInfo::Init();
   std::unique_ptr<KernelContext> ctx(new KernelContext);
   ctx->As<ARMContext>();
   fc.SetParam(param);
@@ -136,7 +134,7 @@ TEST(fc_arm, num_col_dims) {
 
   std::unique_ptr<KernelContext> ctx(new KernelContext);
   ctx->As<ARMContext>();
-  DeviceInfo::init_info();
+  DeviceInfo::Init();
 
   fc.SetParam(param);
   fc.SetContext(std::move(ctx));