From 3ac6335ba30da9af5ef8232286a59cdbe3bd1bfd Mon Sep 17 00:00:00 2001
From: nihui <shuizhuyuanluo@126.com>
Date: Sat, 5 Oct 2019 16:11:15 +0800
Subject: [PATCH] hardsigmoid and hardswish pack4

---
 src/layer/arm/hardsigmoid_arm.cpp | 40 ++++++++++++++++++++++++++++++
 src/layer/arm/hardsigmoid_arm.h   |  2 ++
 src/layer/arm/hardswish_arm.cpp   | 41 +++++++++++++++++++++++++++++++
 src/layer/arm/hardswish_arm.h     |  2 ++
 4 files changed, 85 insertions(+)

diff --git a/src/layer/arm/hardsigmoid_arm.cpp b/src/layer/arm/hardsigmoid_arm.cpp
index 54a9e9843..6c1489b22 100644
--- a/src/layer/arm/hardsigmoid_arm.cpp
+++ b/src/layer/arm/hardsigmoid_arm.cpp
@@ -22,12 +22,52 @@ namespace ncnn {
 
 DEFINE_LAYER_CREATOR(HardSigmoid_arm)
 
+HardSigmoid_arm::HardSigmoid_arm()
+{
+#if __ARM_NEON
+    support_packing = true;
+#endif // __ARM_NEON
+}
+
 int HardSigmoid_arm::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
 {
     int w = bottom_top_blob.w;
     int h = bottom_top_blob.h;
     int channels = bottom_top_blob.c;
     int size = w * h;
+    int elempack = bottom_top_blob.elempack;
+
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+
+    if (elempack == 4)
+    {
+        #pragma omp parallel for num_threads(opt.num_threads)
+        for (int q=0; q<channels; q++)
+        {
+            float* ptr = bottom_top_blob.channel(q);
+
+            float32x4_t _zero = vdupq_n_f32(0.f);
+            float32x4_t _one = vdupq_n_f32(1.f);
+            for (int i=0; i<size; i++)
+            {
+                float32x4_t _p = vld1q_f32(ptr);
+                float32x4_t _ans = vdupq_n_f32(beta);
+                _ans = vmlaq_n_f32(_ans, _p, alpha);
+                _ans = vmaxq_f32(_ans, _zero);
+                _ans = vminq_f32(_ans, _one);
+                vst1q_f32(ptr, _ans);
+
+                ptr += 4;
+            }
+        }
+
+        return 0;
+    }
+
+    } // opt.use_packing_layout
+#endif // __ARM_NEON
 
     #pragma omp parallel for num_threads(opt.num_threads)
     for (int q=0; q<channels; q++)
diff --git a/src/layer/arm/hardsigmoid_arm.h b/src/layer/arm/hardsigmoid_arm.h
index 3d73211c9..6dbbc43fc 100644
--- a/src/layer/arm/hardsigmoid_arm.h
+++ b/src/layer/arm/hardsigmoid_arm.h
@@ -22,6 +22,8 @@ namespace ncnn {
 class HardSigmoid_arm : virtual public HardSigmoid
 {
 public:
+    HardSigmoid_arm();
+
     virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
 };
 
diff --git a/src/layer/arm/hardswish_arm.cpp b/src/layer/arm/hardswish_arm.cpp
index 1f6b78f7a..9f119d474 100644
--- a/src/layer/arm/hardswish_arm.cpp
+++ b/src/layer/arm/hardswish_arm.cpp
@@ -22,12 +22,53 @@ namespace ncnn {
 
 DEFINE_LAYER_CREATOR(HardSwish_arm)
 
+HardSwish_arm::HardSwish_arm()
+{
+#if __ARM_NEON
+    support_packing = true;
+#endif // __ARM_NEON
+}
+
 int HardSwish_arm::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
 {
     int w = bottom_top_blob.w;
     int h = bottom_top_blob.h;
     int channels = bottom_top_blob.c;
     int size = w * h;
+    int elempack = bottom_top_blob.elempack;
+
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+
+    if (elempack == 4)
+    {
+        #pragma omp parallel for num_threads(opt.num_threads)
+        for (int q=0; q<channels; q++)
+        {
+            float* ptr = bottom_top_blob.channel(q);
+
+            float32x4_t _zero = vdupq_n_f32(0.f);
+            float32x4_t _one = vdupq_n_f32(1.f);
+            for (int i=0; i<size; i++)
+            {
+                float32x4_t _p = vld1q_f32(ptr);
+                float32x4_t _ans = vdupq_n_f32(beta);
+                _ans = vmlaq_n_f32(_ans, _p, alpha);
+                _ans = vmaxq_f32(_ans, _zero);
+                _ans = vminq_f32(_ans, _one);
+                _ans = vmulq_f32(_ans, _p);
+                vst1q_f32(ptr, _ans);
+
+                ptr += 4;
+            }
+        }
+
+        return 0;
+    }
+
+    } // opt.use_packing_layout
+#endif // __ARM_NEON
 
     #pragma omp parallel for num_threads(opt.num_threads)
     for (int q=0; q<channels; q++)
diff --git a/src/layer/arm/hardswish_arm.h b/src/layer/arm/hardswish_arm.h
index 1743d0952..c1eea15c4 100644
--- a/src/layer/arm/hardswish_arm.h
+++ b/src/layer/arm/hardswish_arm.h
@@ -22,6 +22,8 @@ namespace ncnn {
 class HardSwish_arm : virtual public HardSwish
 {
 public:
+    HardSwish_arm();
+
     virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
 };