From ad0b1a8a0811a5cb8cd29aa28312b5b6e4008c07 Mon Sep 17 00:00:00 2001
From: Bin Li <libin11@xiaomi.com>
Date: Fri, 27 Apr 2018 11:20:50 +0800
Subject: [PATCH] optimize depthwise conv3x3 s1 s2 armv7 neon

---
 mace/kernels/arm/depthwise_conv2d_neon_3x3.cc | 43 +++++++++++++++++--
 1 file changed, 39 insertions(+), 4 deletions(-)

diff --git a/mace/kernels/arm/depthwise_conv2d_neon_3x3.cc b/mace/kernels/arm/depthwise_conv2d_neon_3x3.cc
index d0ba9ce2..489a3ce8 100644
--- a/mace/kernels/arm/depthwise_conv2d_neon_3x3.cc
+++ b/mace/kernels/arm/depthwise_conv2d_neon_3x3.cc
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#if defined(MACE_ENABLE_NEON) && defined(__aarch64__)
+#if defined(MACE_ENABLE_NEON)
 #include <arm_neon.h>
 #endif
 
@@ -99,7 +99,7 @@ void DepthwiseConv2dNeonK3x3S1(const float *input,
         }
       }
 
-#if defined(MACE_ENABLE_NEON) && defined(__aarch64__)
+#if defined(MACE_ENABLE_NEON)
       // load filter (1 outch x 3 height x 3 width): vf_outch_height
       float32x4_t vf00, vf01, vf02;
       vf00 = vld1q_f32(filter_ptr);
@@ -172,6 +172,7 @@ void DepthwiseConv2dNeonK3x3S1(const float *input,
           vo00 = vld1q_f32(out_base + out_offset);
           vo01 = vld1q_f32(out_base + out_offset + out_width);
 
+#if defined(__aarch64__)
           // outch 0, height 0
           vo00 = vfmaq_laneq_f32(vo00, vi00, vf00, 0);
           vo00 = vfmaq_laneq_f32(vo00, vi01, vf00, 1);
@@ -193,7 +194,29 @@ void DepthwiseConv2dNeonK3x3S1(const float *input,
           vo01 = vfmaq_laneq_f32(vo01, vi30, vf02, 0);
           vo01 = vfmaq_laneq_f32(vo01, vi31, vf02, 1);
           vo01 = vfmaq_laneq_f32(vo01, vi32, vf02, 2);
+#else
+          // outch 0, height 0
+          vo00 = vmlaq_lane_f32(vo00, vi00, vget_low_f32(vf00), 0);
+          vo00 = vmlaq_lane_f32(vo00, vi01, vget_low_f32(vf00), 1);
+          vo00 = vmlaq_lane_f32(vo00, vi02, vget_high_f32(vf00), 0);
+          vo00 = vmlaq_lane_f32(vo00, vi10, vget_low_f32(vf01), 0);
+          vo00 = vmlaq_lane_f32(vo00, vi11, vget_low_f32(vf01), 1);
+          vo00 = vmlaq_lane_f32(vo00, vi12, vget_high_f32(vf01), 0);
+          vo00 = vmlaq_lane_f32(vo00, vi20, vget_low_f32(vf02), 0);
+          vo00 = vmlaq_lane_f32(vo00, vi21, vget_low_f32(vf02), 1);
+          vo00 = vmlaq_lane_f32(vo00, vi22, vget_high_f32(vf02), 0);
 
+          // outch 0, height 1
+          vo01 = vmlaq_lane_f32(vo01, vi10, vget_low_f32(vf00), 0);
+          vo01 = vmlaq_lane_f32(vo01, vi11, vget_low_f32(vf00), 1);
+          vo01 = vmlaq_lane_f32(vo01, vi12, vget_high_f32(vf00), 0);
+          vo01 = vmlaq_lane_f32(vo01, vi20, vget_low_f32(vf01), 0);
+          vo01 = vmlaq_lane_f32(vo01, vi21, vget_low_f32(vf01), 1);
+          vo01 = vmlaq_lane_f32(vo01, vi22, vget_high_f32(vf01), 0);
+          vo01 = vmlaq_lane_f32(vo01, vi30, vget_low_f32(vf02), 0);
+          vo01 = vmlaq_lane_f32(vo01, vi31, vget_low_f32(vf02), 1);
+          vo01 = vmlaq_lane_f32(vo01, vi32, vget_high_f32(vf02), 0);
+#endif
           vst1q_f32(out_base + out_offset, vo00);
           vst1q_f32(out_base + out_offset + out_width, vo01);
         }  // w
@@ -316,7 +339,7 @@ void DepthwiseConv2dNeonK3x3S2(const float *input,
         }
       }
 
-#if defined(MACE_ENABLE_NEON) && defined(__aarch64__)
+#if defined(MACE_ENABLE_NEON)
       // load filter (1 outch x 3 height x 3 width): vf_outch_height
       float32x4_t vf00, vf01, vf02;
       vf00 = vld1q_f32(filter_ptr);
@@ -378,6 +401,7 @@ void DepthwiseConv2dNeonK3x3S2(const float *input,
           vi21 = vi2.val[1];
           vi22 = vextq_f32(vi20, vi2n, 1);
 
+#if defined(__aarch64__)
           // outch 0, height 0
           vo = vfmaq_laneq_f32(vo, vi00, vf00, 0);
           vo = vfmaq_laneq_f32(vo, vi01, vf00, 1);
@@ -388,7 +412,18 @@ void DepthwiseConv2dNeonK3x3S2(const float *input,
           vo = vfmaq_laneq_f32(vo, vi20, vf02, 0);
           vo = vfmaq_laneq_f32(vo, vi21, vf02, 1);
           vo = vfmaq_laneq_f32(vo, vi22, vf02, 2);
-
+#else
+          // outch 0, height 0
+          vo = vmlaq_lane_f32(vo, vi00, vget_low_f32(vf00), 0);
+          vo = vmlaq_lane_f32(vo, vi01, vget_low_f32(vf00), 1);
+          vo = vmlaq_lane_f32(vo, vi02, vget_high_f32(vf00), 0);
+          vo = vmlaq_lane_f32(vo, vi10, vget_low_f32(vf01), 0);
+          vo = vmlaq_lane_f32(vo, vi11, vget_low_f32(vf01), 1);
+          vo = vmlaq_lane_f32(vo, vi12, vget_high_f32(vf01), 0);
+          vo = vmlaq_lane_f32(vo, vi20, vget_low_f32(vf02), 0);
+          vo = vmlaq_lane_f32(vo, vi21, vget_low_f32(vf02), 1);
+          vo = vmlaq_lane_f32(vo, vi22, vget_high_f32(vf02), 0);
+#endif
           vst1q_f32(out_base + out_offset, vo);
         }  // w
 
-- 
GitLab