From 4af3ec223b4b477d72f3d6cf5402c985409ca3ee Mon Sep 17 00:00:00 2001
From: nihuini <nihuini@tencent.com>
Date: Tue, 16 Oct 2018 11:29:08 +0800
Subject: [PATCH] fix clip arm

---
 src/layer/arm/clip_arm.cpp | 95 +++++++++++++++++---------------------
 1 file changed, 43 insertions(+), 52 deletions(-)

diff --git a/src/layer/arm/clip_arm.cpp b/src/layer/arm/clip_arm.cpp
index f2aa2c45a..e42379ba7 100644
--- a/src/layer/arm/clip_arm.cpp
+++ b/src/layer/arm/clip_arm.cpp
@@ -29,77 +29,68 @@ int Clip_arm::forward_inplace(Mat &bottom_top_blob, const Option &opt) const
     int channels = bottom_top_blob.c;
     int size = w * h;
 
-#if __ARM_NEON
-    int nn = size >> 2;
-    int remain = size - (nn << 2);
-#else
-    int remian = size;
-#endif
+    #pragma omp parallel for num_threads(opt.num_threads)
+    for (int q=0; q<channels; q++)
+    {
+        float* ptr = bottom_top_blob.channel(q);
 
 #if __ARM_NEON
-    float32x4_t maxf32 = vmovq_n_f32(max);
-    float32x4_t minf32 = vmovq_n_f32(min);
+        int nn = size >> 2;
+        int remain = size & 3;
+#else
+        int remian = size;
 #endif
 
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int i = 0; i < channels; ++i)
-    {
-        float *channel_ptr = bottom_top_blob.channel(i);
 #if __ARM_NEON
+        float32x4_t _max = vdupq_n_f32(max);
+        float32x4_t _min = vdupq_n_f32(min);
 #if __aarch64__
-        for (; nn > 0; --nn)
+        for (; nn>0; nn--)
         {
-            float32x4_t clip_f32 = vld1q_f32(channel_ptr);
-            float32x4_t clip_min_f32 = vmaxq_f32(minf32, clip_f32);
-            float32x4_t clip_max_f32 = vminq_f32(maxf32, clip_min_f32);
-            vst1q_f32(channel_ptr, clip_max_f32);
-            channel_ptr += 4;
+            float32x4_t _ptr = vld1q_f32(ptr);
+            _ptr = vmaxq_f32(_ptr, _min);
+            _ptr = vminq_f32(_ptr, _max);
+            vst1q_f32(ptr, _ptr);
+            ptr += 4;
         }
 #else
         if (nn > 0)
         {
-            asm volatile(
-            "0:"
-            "pld        [%1,    #128]           \n"
-            "vld1.f32   {d0-d1},    [%1:128]    \n"
-
-            "vmax.f32   q1, %q4,    q0          \n"
-            "vmin.f32   q2, %q5,    q1          \n"
-
-            "subs       %0,          #1          \n"
-            "vst1.f32   {d4-d5},    [%1:128]!   \n"
-
-            "bne        0b                      \n"
-
-            :"=r"(nn),              //%0
-            "=r"(channel_ptr)       //%1
-            :"0"(nn),
-            "1"(channel_ptr),
-            "w"(minf32),            //%q4
-            "w"(maxf32)             //%q5
-            :"cc", "memory", "q0", "q1", "q2"
-            );
-
+        asm volatile(
+            "0:                             \n"
+            "pld        [%1, #128]          \n"
+            "vld1.f32   {d0-d1}, [%1: 128]  \n"
+
+            "vmax.f32   q0, q0, %q4         \n"
+            "vmin.f32   q0, q0, %q5         \n"
+
+            "subs       %0, #1              \n"
+            "vst1.f32   {d0-d1}, [%1: 128]! \n"
+
+            "bne        0b                  \n"
+
+            : "=r"(nn),     // %0
+              "=r"(ptr)     // %1
+            : "0"(nn),
+              "1"(ptr),
+              "w"(_min),    // %q4
+              "w"(_max)     // %q5
+            : "cc", "memory", "q0"
+        );
         }
 #endif // __aarch64__
 #endif // __ARM_NEON
 
-        for (; remain > 0; --remain)
+        for (; remain>0; remain--)
         {
+            if (*ptr < min)
+                *ptr = min;
 
-            if (*channel_ptr < min)
-            {
-                *channel_ptr = min;
-            }
-
-            if (*channel_ptr > max)
-            {
-                *channel_ptr = max;
-            }
+            if (*ptr > max)
+                *ptr = max;
 
-            ++channel_ptr;
+            ptr++;
         }
-
     }
 
     return 0;