add relu asm implementation

CMakeLists.txt

@@ -4,7 +4,8 @@ project(paddle-mobile)
 option(DEBUGING "enable debug mode" ON)
 option(USE_OPENMP "openmp support" OFF)
 option(USE_EXCEPTION "use std exception" ON)
-
+option(LOG_PROFILE "log profile" ON)
+# select the platform to build
 option(CPU "cpu" ON)
 option(MALI_GPU "mali gpu" OFF)
 option(FPGA "fpga" OFF)
@@ -45,6 +46,10 @@ else()
     add_definitions(-fno-exceptions)
 endif ()
 
+if (LOG_PROFILE)
+    add_definitions(-DPADDLE_MOBILE_PROFILE)
+endif()
+
 if(IS_MAC)
     add_definitions(-DX86)
 elseif(IS_IOS)

src/operators/kernel/arm/relu_kernel.cpp

@@ -37,13 +37,70 @@ void ReluKernel<CPU, float>::Compute(const ReluParam &param) const {
   auto *out = param.Out();
   auto *out_ptr = out->mutable_data<float>();
 
-  ReluFunctor<float> func_;
-  math::Transform trans;
-  trans(input_x_ptr, input_x_ptr + input_x->numel(), out_ptr, func_);
+  int numel = input_x->numel();
+  if (numel > 32) {
+    asm volatile(
+    "pld        [%[input_x_ptr], #0]        \n\t"
+    "vmov.f32   q8,    #0.0                 \n\t"
+    "subs %[num], %[num], #32                \n\t"
+    "blt        end_num_%=                  \n\t"
+    "loop_num_%=:                           \n\t"
+    "pld        [%[input_x_ptr], #1024]      \n\t"
 
-  //  for (int i = 0; i < input_x->numel(); i++) {
-  //    out_ptr[i] = input_x_ptr[i] > 0 ? input_x_ptr[i] : 0;
-  //  }
+    "vld1.32 {q0, q1}, [%[input_x_ptr]]!    \n\t"
+    "vld1.32 {q2, q3}, [%[input_x_ptr]]!    \n\t"
+    "vld1.32 {q4, q5}, [%[input_x_ptr]]!    \n\t"
+    "vld1.32 {q6, q7}, [%[input_x_ptr]]!    \n\t"
+
+    "vmax.f32 q0, q0, q8                   \n\t"
+    "vmax.f32 q1, q1, q8                    \n\t"
+    "vmax.f32 q2, q2, q8                   \n\t"
+    "vmax.f32 q3, q3, q8                   \n\t"
+    "vmax.f32 q4, q4, q8                   \n\t"
+    "vmax.f32 q5, q5, q8                   \n\t"
+    "vmax.f32 q6, q6, q8                   \n\t"
+    "vmax.f32 q7, q7, q8                   \n\t"
+
+    "vst1.32 {q0, q1}, [%[out_ptr]]!        \n\t"
+    "vst1.32 {q2, q3}, [%[out_ptr]]!       \n\t"
+    "vst1.32 {q4, q5}, [%[out_ptr]]!       \n\t"
+    "vst1.32 {q6, q7}, [%[out_ptr]]!       \n\t"
+
+    "subs %[num], %[num], #32              \n\t"
+    "bge        loop_num_%=                \n\t"
+    "end_num_%=:                           \n\t"
+    "cmp %[num], #0                         \n\t"
+    "bge   end_%=                          \n\t"
+    "mov r6, #4                             \n\t"
+    "mul r5, %[num], r6                     \n\t"
+    "add %[input_x_ptr], %[input_x_ptr], r5     \n\t"
+    "vld1.32 {q0, q1}, [%[input_x_ptr]]!    \n\t"
+    "vld1.32 {q2, q3}, [%[input_x_ptr]]!    \n\t"
+    "vld1.32 {q4, q5}, [%[input_x_ptr]]!    \n\t"
+    "vld1.32 {q6, q7}, [%[input_x_ptr]]!    \n\t"
+    "vmax.f32 q0, q0, q8                   \n\t"
+    "vmax.f32 q1, q1, q8                    \n\t"
+    "vmax.f32 q2, q2, q8                   \n\t"
+    "vmax.f32 q3, q3, q8                   \n\t"
+    "vmax.f32 q4, q4, q8                   \n\t"
+    "vmax.f32 q5, q5, q8                   \n\t"
+    "vmax.f32 q6, q6, q8                   \n\t"
+    "vmax.f32 q7, q7, q8                   \n\t"
+    "add %[out_ptr], %[out_ptr], r5       \n\t"
+    "vst1.32 {q0, q1}, [%[out_ptr]]!        \n\t"
+    "vst1.32 {q2, q3}, [%[out_ptr]]!       \n\t"
+    "vst1.32 {q4, q5}, [%[out_ptr]]!       \n\t"
+    "vst1.32 {q6, q7}, [%[out_ptr]]!       \n\t"
+    "end_%=:                                \n\t"
+    :
+    :[out_ptr]"r"(out_ptr), [input_x_ptr]"r"(input_x_ptr), [num]"r"(numel)
+    :"memory", "q0", "q1", "q2", "q3", "q4","q5","q6", "q7", "q8", "r5", "r6"
+    );
+  } else {
+    ReluFunctor<float> func_;
+    math::Transform trans;
+    trans(input_x_ptr, input_x_ptr + numel, out_ptr, func_);
+  }
 }
 }  // namespace operators
 }  // namespace paddle_mobile

test/net/test_googlenet.cpp

@@ -20,9 +20,9 @@ int main() {
   paddle_mobile::Loader<paddle_mobile::CPU> loader;
   bool optimize = false;
   auto time1 = time();
-  //  auto program = loader.Load(g_googlenet, optimize);
-  auto program = loader.Load(g_googlenet_combine + "/model",
-                             g_googlenet_combine + "/params", optimize);
+  auto program = loader.Load(g_googlenet, optimize);
+  //  auto program = loader.Load(g_googlenet_combine + "/model",
+  //                             g_googlenet_combine + "/params", optimize);
   auto time2 = time();
   DLOG << "load cost :" << time_diff(time1, time2) << "ms\n";
   paddle_mobile::Executor<paddle_mobile::CPU> executor(program, 1, optimize);