From 992f29cfb983e7c93735f6a26322a5fd258ec56e Mon Sep 17 00:00:00 2001
From: yongqiang <mayongqiang01@baidu.com>
Date: Mon, 27 Jul 2020 07:36:52 +0000
Subject: [PATCH] change pool3x3s2p1_max to intrinsic. test=develop

---
 lite/backends/arm/math/pooling.cc | 76 +++++++++++++++++++++++++++++++
 1 file changed, 76 insertions(+)

diff --git a/lite/backends/arm/math/pooling.cc b/lite/backends/arm/math/pooling.cc
index 8303851ece..f9bf52d00a 100644
--- a/lite/backends/arm/math/pooling.cc
+++ b/lite/backends/arm/math/pooling.cc
@@ -2235,6 +2235,7 @@ void pooling3x3s2p1_max(const float* din,
         int cnt_num = w_unroll_size;
         if (w_unroll_size > 0) {
 #ifdef __aarch64__
+#if 0
           asm volatile(
               /* preocess left */
               P3x3S2_INIT P3x3S2P1_MAX P3x3S2P0_MAX "2: \n" /* end */
@@ -2259,6 +2260,81 @@ void pooling3x3s2p1_max(const float* din,
                 "v10",
                 "v11",
                 "v31");
+#else 
+
+          float32x4_t vr0_1234 = vld1q_f32(dr0);
+          float32x4_t vr0_5678 = vld1q_f32(dr0 += 4);
+          float32x4_t vr0_9101112 = vld1q_f32(dr0 += 4);
+          float32x4_t vr1_1234 = vld1q_f32(dr1);
+          float32x4_t vr1_5678 = vld1q_f32(dr1 += 4);
+          float32x4_t vr1_9101112 = vld1q_f32(dr1 += 4);
+          float32x4_t vr2_1234 = vld1q_f32(dr2);
+          float32x4_t vr2_5678 = vld1q_f32(dr2 += 4);
+          float32x4_t vr2_9101112 = vld1q_f32(dr2 += 4);
+
+          float32x4_t vmax_1234 = vmaxq_f32(vr0_1234, vr1_1234);
+          vmax_1234 = vmaxq_f32(vmax_1234, vr2_1234);
+          float32x4_t vmax_5678 = vmaxq_f32(vr0_5678, vr1_5678);
+          vmax_5678 = vmaxq_f32(vmax_5678, vr2_5678);
+          float32x4_t vmax_9101112 = vmaxq_f32(vr0_9101112, vr1_9101112);
+          vmax_9101112 = vmaxq_f32(vmax_9101112, vr2_9101112);
+          float32x4_t vmax_0123 = vextq_f32(vmin, vmax_1234, 1);
+          float32x4_t vmax_4567 = vextq_f32(vmax_1234, vmax_5678, 1);
+          float32x2_t vmax_12_34 =
+              vpmax_f32(vget_low_f32(vmax_1234), vget_high_f32(vmax_1234));
+          float32x2_t vmax_01_23 =
+              vpmax_f32(vget_low_f32(vmax_0123), vget_high_f32(vmax_0123));
+          float32x2_t vmax_56_78 =
+              vpmax_f32(vget_low_f32(vmax_5678), vget_high_f32(vmax_5678));
+          float32x2_t vmax_45_67 =
+              vpmax_f32(vget_low_f32(vmax_4567), vget_high_f32(vmax_4567));
+          float32x2_t vmax_012_234 = vmax_f32(vmax_12_34, vmax_01_23);
+          float32x2_t vmax_456_678 = vmax_f32(vmax_56_78, vmax_45_67);
+          vst1_f32(dr_out, vmax_012_234);
+          dr_out += 2;
+          vst1_f32(dr_out, vmax_456_678);
+          dr_out += 2;
+          cnt_num--;
+          dr0--;
+          dr1--;
+          dr2--;
+          while (cnt_num--) {
+            float32x4_t vr0_1234 = vld1q_f32(dr0);
+            float32x4_t vr0_5678 = vld1q_f32(dr0 += 4);
+            float32x4_t vr0_9101112 = vld1q_f32(dr0 += 4);
+            float32x4_t vr1_1234 = vld1q_f32(dr1);
+            float32x4_t vr1_5678 = vld1q_f32(dr1 += 4);
+            float32x4_t vr1_9101112 = vld1q_f32(dr1 += 4);
+            float32x4_t vr2_1234 = vld1q_f32(dr2);
+            float32x4_t vr2_5678 = vld1q_f32(dr2 += 4);
+            float32x4_t vr2_9101112 = vld1q_f32(dr2 += 4);
+            float32x4_t vmax_1234 = vmaxq_f32(vr0_1234, vr1_1234);
+            vmax_1234 = vmaxq_f32(vmax_1234, vr2_1234);
+            float32x4_t vmax_5678 = vmaxq_f32(vr0_5678, vr1_5678);
+            vmax_5678 = vmaxq_f32(vmax_5678, vr2_5678);
+            float32x4_t vmax_9101112 = vmaxq_f32(vr0_9101112, vr1_9101112);
+            vmax_9101112 = vmaxq_f32(vmax_9101112, vr2_9101112);
+            float32x4_t vmax_2345 = vextq_f32(vmax_1234, vmax_5678, 1);
+            float32x4_t vmax_6789 = vextq_f32(vmax_5678, vmax_9101112, 1);
+            float32x2_t vmax_12_34 =
+                vpmax_f32(vget_low_f32(vmax_1234), vget_high_f32(vmax_1234));
+            float32x2_t vmax_23_45 =
+                vpmax_f32(vget_low_f32(vmax_2345), vget_high_f32(vmax_2345));
+            float32x2_t vmax_56_78 =
+                vpmax_f32(vget_low_f32(vmax_5678), vget_high_f32(vmax_5678));
+            float32x2_t vmax_67_89 =
+                vpmax_f32(vget_low_f32(vmax_6789), vget_high_f32(vmax_6789));
+            float32x2_t vmax_123_345 = vmax_f32(vmax_12_34, vmax_23_45);
+            float32x2_t vmax_567_789 = vmax_f32(vmax_56_78, vmax_67_89);
+            vst1_f32(dr_out, vmax_123_345);
+            dr_out += 2;
+            vst1_f32(dr_out, vmax_567_789);
+            dr_out += 2;
+          }
+          dr0 += 8;
+          dr1 += 8;
+          dr2 += 8;
+#endif
 #else
           asm volatile(
               /* preocess left */
-- 
GitLab