From 05d7562a5df04979cbc61ef30217dfba6bdcd0a2 Mon Sep 17 00:00:00 2001
From: nihuini <nihuini@tencent.com>
Date: Mon, 2 Jul 2018 18:54:41 +0800
Subject: [PATCH] reorder kernel weight, pipeline friendly ;)

---
 src/layer/arm/convolution_1x1.h |  933 ++++++++++++++++------------
 src/layer/arm/convolution_3x3.h | 1001 +++++++++++++++++--------------
 2 files changed, 1089 insertions(+), 845 deletions(-)

diff --git a/src/layer/arm/convolution_1x1.h b/src/layer/arm/convolution_1x1.h
index a5b8cfbff..f0c974893 100644
--- a/src/layer/arm/convolution_1x1.h
+++ b/src/layer/arm/convolution_1x1.h
@@ -21,14 +21,14 @@ static void conv1x1s1_sgemm_transform_kernel_neon(const Mat& _kernel, Mat& kerne
     const float* kernel = _kernel;
 
     // interleave
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
     kernel_tm.create(4*8, inch/4 + inch%4, outch/8 + (outch%8)/4 + outch%4);
 #else
     kernel_tm.create(4*4, inch/4 + inch%4, outch/4 + outch%4);
-#endif // __aarch64__
+#endif // __ARM_NEON && __aarch64__
 
     int p = 0;
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
     for (; p+7<outch; p+=8)
     {
         const float* kernel0 = kernel + (p+0)*inch;
@@ -42,36 +42,7 @@ static void conv1x1s1_sgemm_transform_kernel_neon(const Mat& _kernel, Mat& kerne
 
         float* ktmp = kernel_tm.channel(p/8);
 
-        int q = 0;
-        for (; q+3<inch; q+=4)
-        {
-            vst1q_f32(ktmp, vld1q_f32(kernel0));
-            ktmp += 4;
-            vst1q_f32(ktmp, vld1q_f32(kernel1));
-            ktmp += 4;
-            vst1q_f32(ktmp, vld1q_f32(kernel2));
-            ktmp += 4;
-            vst1q_f32(ktmp, vld1q_f32(kernel3));
-            ktmp += 4;
-            vst1q_f32(ktmp, vld1q_f32(kernel4));
-            ktmp += 4;
-            vst1q_f32(ktmp, vld1q_f32(kernel5));
-            ktmp += 4;
-            vst1q_f32(ktmp, vld1q_f32(kernel6));
-            ktmp += 4;
-            vst1q_f32(ktmp, vld1q_f32(kernel7));
-            ktmp += 4;
-
-            kernel0 += 4;
-            kernel1 += 4;
-            kernel2 += 4;
-            kernel3 += 4;
-            kernel4 += 4;
-            kernel5 += 4;
-            kernel6 += 4;
-            kernel7 += 4;
-        }
-        for (; q<inch; q++)
+        for (int q=0; q<inch; q++)
         {
             // kernel0...7 0
             ktmp[0] = kernel0[0];
@@ -82,8 +53,8 @@ static void conv1x1s1_sgemm_transform_kernel_neon(const Mat& _kernel, Mat& kerne
             ktmp[5] = kernel5[0];
             ktmp[6] = kernel6[0];
             ktmp[7] = kernel7[0];
-            ktmp += 8;
 
+            ktmp += 8;
             kernel0 += 1;
             kernel1 += 1;
             kernel2 += 1;
@@ -94,7 +65,7 @@ static void conv1x1s1_sgemm_transform_kernel_neon(const Mat& _kernel, Mat& kerne
             kernel7 += 1;
         }
     }
-#endif // __aarch64__
+#endif // __ARM_NEON && __aarch64__
     for (; p+3<outch; p+=4)
     {
         const float* kernel0 = kernel + (p+0)*inch;
@@ -102,38 +73,21 @@ static void conv1x1s1_sgemm_transform_kernel_neon(const Mat& _kernel, Mat& kerne
         const float* kernel2 = kernel + (p+2)*inch;
         const float* kernel3 = kernel + (p+3)*inch;
 
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
         float* ktmp = kernel_tm.channel(p/8 + (p%8)/4);
 #else
         float* ktmp = kernel_tm.channel(p/4);
-#endif // __aarch64__
-
-        int q = 0;
-        for (; q+3<inch; q+=4)
-        {
-            vst1q_f32(ktmp, vld1q_f32(kernel0));
-            ktmp += 4;
-            vst1q_f32(ktmp, vld1q_f32(kernel1));
-            ktmp += 4;
-            vst1q_f32(ktmp, vld1q_f32(kernel2));
-            ktmp += 4;
-            vst1q_f32(ktmp, vld1q_f32(kernel3));
-            ktmp += 4;
+#endif // __ARM_NEON && __aarch64__
 
-            kernel0 += 4;
-            kernel1 += 4;
-            kernel2 += 4;
-            kernel3 += 4;
-        }
-        for (; q<inch; q++)
+        for (int q=0; q<inch; q++)
         {
-            // kernel0...7 0
+            // kernel0...3 0
             ktmp[0] = kernel0[0];
             ktmp[1] = kernel1[0];
             ktmp[2] = kernel2[0];
             ktmp[3] = kernel3[0];
-            ktmp += 4;
 
+            ktmp += 4;
             kernel0 += 1;
             kernel1 += 1;
             kernel2 += 1;
@@ -144,21 +98,13 @@ static void conv1x1s1_sgemm_transform_kernel_neon(const Mat& _kernel, Mat& kerne
     {
         const float* kernel0 = kernel + p*inch;
 
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
         float* ktmp = kernel_tm.channel(p/8 + (p%8)/4 + p%4);
 #else
         float* ktmp = kernel_tm.channel(p/4 + p%4);
-#endif // __aarch64__
-
-        int q = 0;
-        for (; q+3<inch; q+=4)
-        {
-            vst1q_f32(ktmp, vld1q_f32(kernel0));
-            ktmp += 4;
+#endif // __ARM_NEON && __aarch64__
 
-            kernel0 += 4;
-        }
-        for (; q<inch; q++)
+        for (int q=0; q<inch; q++)
         {
             ktmp[0] = kernel0[0];
             ktmp++;
@@ -193,51 +139,26 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
             int i = ii * 8;
 
             const float* img0 = bottom_blob.channel(0);
-            const float* img1 = bottom_blob.channel(1);
-            const float* img2 = bottom_blob.channel(2);
-            const float* img3 = bottom_blob.channel(3);
-
             img0 += i;
-            img1 += i;
-            img2 += i;
-            img3 += i;
-
-            const int imgstep = bottom_blob.cstep*4;
 
             float* tmpptr = tmp.channel(i/8);
 
-            int q = 0;
-
-            for (; q+3<inch; q+=4)
-            {
-                vst1q_f32(tmpptr, vld1q_f32(img0));
-                vst1q_f32(tmpptr+4, vld1q_f32(img0+4));
-                tmpptr += 8;
-
-                vst1q_f32(tmpptr, vld1q_f32(img1));
-                vst1q_f32(tmpptr+4, vld1q_f32(img1+4));
-                tmpptr += 8;
-
-                vst1q_f32(tmpptr, vld1q_f32(img2));
-                vst1q_f32(tmpptr+4, vld1q_f32(img2+4));
-                tmpptr += 8;
-
-                vst1q_f32(tmpptr, vld1q_f32(img3));
-                vst1q_f32(tmpptr+4, vld1q_f32(img3+4));
-                tmpptr += 8;
-
-                img0 += imgstep;
-                img1 += imgstep;
-                img2 += imgstep;
-                img3 += imgstep;
-            }
-
-            for (; q<inch; q++)
+            for (int q=0; q<inch; q++)
             {
+#if __ARM_NEON
                 vst1q_f32(tmpptr, vld1q_f32(img0));
                 vst1q_f32(tmpptr+4, vld1q_f32(img0+4));
+#else
+                tmpptr[0] = img0[0];
+                tmpptr[1] = img0[1];
+                tmpptr[2] = img0[2];
+                tmpptr[3] = img0[3];
+                tmpptr[4] = img0[4];
+                tmpptr[5] = img0[5];
+                tmpptr[6] = img0[6];
+                tmpptr[7] = img0[7];
+#endif // __ARM_NEON
                 tmpptr += 8;
-
                 img0 += bottom_blob.cstep;
             }
         }
@@ -250,46 +171,21 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
             int i = remain_size_start + ii * 4;
 
             const float* img0 = bottom_blob.channel(0);
-            const float* img1 = bottom_blob.channel(1);
-            const float* img2 = bottom_blob.channel(2);
-            const float* img3 = bottom_blob.channel(3);
-
             img0 += i;
-            img1 += i;
-            img2 += i;
-            img3 += i;
-
-            const int imgstep = bottom_blob.cstep*4;
 
             float* tmpptr = tmp.channel(i/8 + (i%8)/4);
 
-            int q = 0;
-
-            for (; q+3<inch; q+=4)
-            {
-                vst1q_f32(tmpptr, vld1q_f32(img0));
-                tmpptr += 4;
-
-                vst1q_f32(tmpptr, vld1q_f32(img1));
-                tmpptr += 4;
-
-                vst1q_f32(tmpptr, vld1q_f32(img2));
-                tmpptr += 4;
-
-                vst1q_f32(tmpptr, vld1q_f32(img3));
-                tmpptr += 4;
-
-                img0 += imgstep;
-                img1 += imgstep;
-                img2 += imgstep;
-                img3 += imgstep;
-            }
-
-            for (; q<inch; q++)
+            for (int q=0; q<inch; q++)
             {
+#if __ARM_NEON
                 vst1q_f32(tmpptr, vld1q_f32(img0));
+#else
+                tmpptr[0] = img0[0];
+                tmpptr[1] = img0[1];
+                tmpptr[2] = img0[2];
+                tmpptr[3] = img0[3];
+#endif // __ARM_NEON
                 tmpptr += 4;
-
                 img0 += bottom_blob.cstep;
             }
         }
@@ -300,40 +196,14 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
         for (int i=remain_size_start; i<size; i++)
         {
             const float* img0 = bottom_blob.channel(0);
-            const float* img1 = bottom_blob.channel(1);
-            const float* img2 = bottom_blob.channel(2);
-            const float* img3 = bottom_blob.channel(3);
-
             img0 += i;
-            img1 += i;
-            img2 += i;
-            img3 += i;
-
-            const int imgstep = bottom_blob.cstep*4;
 
             float* tmpptr = tmp.channel(i/8 + (i%8)/4 + i%4);
 
-            int q = 0;
-
-            for (; q+3<inch; q+=4)
-            {
-                tmpptr[0] = img0[0];
-                tmpptr[1] = img1[0];
-                tmpptr[2] = img2[0];
-                tmpptr[3] = img3[0];
-                tmpptr += 4;
-
-                img0 += imgstep;
-                img1 += imgstep;
-                img2 += imgstep;
-                img3 += imgstep;
-            }
-
-            for (; q<inch; q++)
+            for (int q=0; q<inch; q++)
             {
                 tmpptr[0] = img0[0];
                 tmpptr++;
-
                 img0 += bottom_blob.cstep;
             }
         }
@@ -342,7 +212,7 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
     int nn_outch = 0;
     int remain_outch_start = 0;
 
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
     nn_outch = outch >> 3;
     remain_outch_start = nn_outch << 3;
 
@@ -403,89 +273,89 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%9], #64     \n"
 
                 "fmla   v16.4s, v8.4s, v0.s[0]  \n"
-                "fmla   v18.4s, v8.4s, v1.s[0]  \n"
-                "fmla   v20.4s, v8.4s, v2.s[0]  \n"
-                "fmla   v22.4s, v8.4s, v3.s[0]  \n"
+                "fmla   v18.4s, v8.4s, v0.s[1]  \n"
+                "fmla   v20.4s, v8.4s, v0.s[2]  \n"
+                "fmla   v22.4s, v8.4s, v0.s[3]  \n"
 
                 "fmla   v17.4s, v9.4s, v0.s[0]  \n"
-                "fmla   v19.4s, v9.4s, v1.s[0]  \n"
-                "fmla   v21.4s, v9.4s, v2.s[0]  \n"
-                "fmla   v23.4s, v9.4s, v3.s[0]  \n"
+                "fmla   v19.4s, v9.4s, v0.s[1]  \n"
+                "fmla   v21.4s, v9.4s, v0.s[2]  \n"
+                "fmla   v23.4s, v9.4s, v0.s[3]  \n"
+
+                "fmla   v24.4s, v8.4s, v1.s[0]  \n"
+                "fmla   v26.4s, v8.4s, v1.s[1]  \n"
+                "fmla   v28.4s, v8.4s, v1.s[2]  \n"
+                "fmla   v30.4s, v8.4s, v1.s[3]  \n"
+
+                "fmla   v25.4s, v9.4s, v1.s[0]  \n"
+                "fmla   v27.4s, v9.4s, v1.s[1]  \n"
+                "fmla   v29.4s, v9.4s, v1.s[2]  \n"
+                "fmla   v31.4s, v9.4s, v1.s[3]  \n"
 
                 "prfm   pldl1keep, [%8, #512]   \n"
                 "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%8], #64 \n"
 
-                "fmla   v16.4s, v10.4s, v0.s[1] \n"
-                "fmla   v18.4s, v10.4s, v1.s[1] \n"
-                "fmla   v20.4s, v10.4s, v2.s[1] \n"
-                "fmla   v22.4s, v10.4s, v3.s[1] \n"
+                "fmla   v16.4s, v10.4s, v2.s[0] \n"
+                "fmla   v18.4s, v10.4s, v2.s[1] \n"
+                "fmla   v20.4s, v10.4s, v2.s[2] \n"
+                "fmla   v22.4s, v10.4s, v2.s[3] \n"
 
-                "fmla   v17.4s, v11.4s, v0.s[1] \n"
-                "fmla   v19.4s, v11.4s, v1.s[1] \n"
-                "fmla   v21.4s, v11.4s, v2.s[1] \n"
-                "fmla   v23.4s, v11.4s, v3.s[1] \n"
+                "fmla   v17.4s, v11.4s, v2.s[0] \n"
+                "fmla   v19.4s, v11.4s, v2.s[1] \n"
+                "fmla   v21.4s, v11.4s, v2.s[2] \n"
+                "fmla   v23.4s, v11.4s, v2.s[3] \n"
 
-                "fmla   v16.4s, v12.4s, v0.s[2] \n"
-                "fmla   v18.4s, v12.4s, v1.s[2] \n"
-                "fmla   v20.4s, v12.4s, v2.s[2] \n"
-                "fmla   v22.4s, v12.4s, v3.s[2] \n"
+                "fmla   v24.4s, v10.4s, v3.s[0] \n"
+                "fmla   v26.4s, v10.4s, v3.s[1] \n"
+                "fmla   v28.4s, v10.4s, v3.s[2] \n"
+                "fmla   v30.4s, v10.4s, v3.s[3] \n"
 
-                "fmla   v17.4s, v13.4s, v0.s[2] \n"
-                "fmla   v19.4s, v13.4s, v1.s[2] \n"
-                "fmla   v21.4s, v13.4s, v2.s[2] \n"
-                "fmla   v23.4s, v13.4s, v3.s[2] \n"
+                "fmla   v25.4s, v11.4s, v3.s[0] \n"
+                "fmla   v27.4s, v11.4s, v3.s[1] \n"
+                "fmla   v29.4s, v11.4s, v3.s[2] \n"
+                "fmla   v31.4s, v11.4s, v3.s[3] \n"
 
                 "prfm   pldl1keep, [%9, #512]   \n"
                 "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%9], #64     \n"
 
-                "fmla   v16.4s, v14.4s, v0.s[3] \n"
-                "fmla   v18.4s, v14.4s, v1.s[3] \n"
-                "fmla   v20.4s, v14.4s, v2.s[3] \n"
-                "fmla   v22.4s, v14.4s, v3.s[3] \n"
-
-                "fmla   v17.4s, v15.4s, v0.s[3] \n"
-                "fmla   v19.4s, v15.4s, v1.s[3] \n"
-                "fmla   v21.4s, v15.4s, v2.s[3] \n"
-                "fmla   v23.4s, v15.4s, v3.s[3] \n"
-
-                "fmla   v24.4s, v8.4s, v4.s[0]  \n"
-                "fmla   v26.4s, v8.4s, v5.s[0]  \n"
-                "fmla   v28.4s, v8.4s, v6.s[0]  \n"
-                "fmla   v30.4s, v8.4s, v7.s[0]  \n"
-
-                "fmla   v25.4s, v9.4s, v4.s[0]  \n"
-                "fmla   v27.4s, v9.4s, v5.s[0]  \n"
-                "fmla   v29.4s, v9.4s, v6.s[0]  \n"
-                "fmla   v31.4s, v9.4s, v7.s[0]  \n"
-
-                "fmla   v24.4s, v10.4s, v4.s[1] \n"
-                "fmla   v26.4s, v10.4s, v5.s[1] \n"
-                "fmla   v28.4s, v10.4s, v6.s[1] \n"
-                "fmla   v30.4s, v10.4s, v7.s[1] \n"
-
-                "fmla   v25.4s, v11.4s, v4.s[1] \n"
-                "fmla   v27.4s, v11.4s, v5.s[1] \n"
-                "fmla   v29.4s, v11.4s, v6.s[1] \n"
-                "fmla   v31.4s, v11.4s, v7.s[1] \n"
-
-                "fmla   v24.4s, v12.4s, v4.s[2] \n"
-                "fmla   v26.4s, v12.4s, v5.s[2] \n"
-                "fmla   v28.4s, v12.4s, v6.s[2] \n"
-                "fmla   v30.4s, v12.4s, v7.s[2] \n"
-
-                "fmla   v25.4s, v13.4s, v4.s[2] \n"
-                "fmla   v27.4s, v13.4s, v5.s[2] \n"
-                "fmla   v29.4s, v13.4s, v6.s[2] \n"
-                "fmla   v31.4s, v13.4s, v7.s[2] \n"
-
-                "fmla   v24.4s, v14.4s, v4.s[3] \n"
-                "fmla   v26.4s, v14.4s, v5.s[3] \n"
-                "fmla   v28.4s, v14.4s, v6.s[3] \n"
+                "fmla   v16.4s, v12.4s, v4.s[0] \n"
+                "fmla   v18.4s, v12.4s, v4.s[1] \n"
+                "fmla   v20.4s, v12.4s, v4.s[2] \n"
+                "fmla   v22.4s, v12.4s, v4.s[3] \n"
+
+                "fmla   v17.4s, v13.4s, v4.s[0] \n"
+                "fmla   v19.4s, v13.4s, v4.s[1] \n"
+                "fmla   v21.4s, v13.4s, v4.s[2] \n"
+                "fmla   v23.4s, v13.4s, v4.s[3] \n"
+
+                "fmla   v24.4s, v12.4s, v5.s[0] \n"
+                "fmla   v26.4s, v12.4s, v5.s[1] \n"
+                "fmla   v28.4s, v12.4s, v5.s[2] \n"
+                "fmla   v30.4s, v12.4s, v5.s[3] \n"
+
+                "fmla   v25.4s, v13.4s, v5.s[0] \n"
+                "fmla   v27.4s, v13.4s, v5.s[1] \n"
+                "fmla   v29.4s, v13.4s, v5.s[2] \n"
+                "fmla   v31.4s, v13.4s, v5.s[3] \n"
+
+                "fmla   v16.4s, v14.4s, v6.s[0] \n"
+                "fmla   v18.4s, v14.4s, v6.s[1] \n"
+                "fmla   v20.4s, v14.4s, v6.s[2] \n"
+                "fmla   v22.4s, v14.4s, v6.s[3] \n"
+
+                "fmla   v17.4s, v15.4s, v6.s[0] \n"
+                "fmla   v19.4s, v15.4s, v6.s[1] \n"
+                "fmla   v21.4s, v15.4s, v6.s[2] \n"
+                "fmla   v23.4s, v15.4s, v6.s[3] \n"
+
+                "fmla   v24.4s, v14.4s, v7.s[0] \n"
+                "fmla   v26.4s, v14.4s, v7.s[1] \n"
+                "fmla   v28.4s, v14.4s, v7.s[2] \n"
                 "fmla   v30.4s, v14.4s, v7.s[3] \n"
 
-                "fmla   v25.4s, v15.4s, v4.s[3] \n"
-                "fmla   v27.4s, v15.4s, v5.s[3] \n"
-                "fmla   v29.4s, v15.4s, v6.s[3] \n"
+                "fmla   v25.4s, v15.4s, v7.s[0] \n"
+                "fmla   v27.4s, v15.4s, v7.s[1] \n"
+                "fmla   v29.4s, v15.4s, v7.s[2] \n"
                 "fmla   v31.4s, v15.4s, v7.s[3] \n"
 
                 "subs   w4, w4, #1              \n"
@@ -599,39 +469,39 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%9], #64     \n"
 
                 "fmla   v16.4s, v8.4s, v0.s[0]  \n"
-                "fmla   v17.4s, v8.4s, v1.s[0]  \n"
-                "fmla   v18.4s, v8.4s, v2.s[0]  \n"
-                "fmla   v19.4s, v8.4s, v3.s[0]  \n"
-                "fmla   v20.4s, v8.4s, v4.s[0]  \n"
-                "fmla   v21.4s, v8.4s, v5.s[0]  \n"
-                "fmla   v22.4s, v8.4s, v6.s[0]  \n"
-                "fmla   v23.4s, v8.4s, v7.s[0]  \n"
-
-                "fmla   v16.4s, v9.4s, v0.s[1]  \n"
-                "fmla   v17.4s, v9.4s, v1.s[1]  \n"
-                "fmla   v18.4s, v9.4s, v2.s[1]  \n"
-                "fmla   v19.4s, v9.4s, v3.s[1]  \n"
-                "fmla   v20.4s, v9.4s, v4.s[1]  \n"
-                "fmla   v21.4s, v9.4s, v5.s[1]  \n"
-                "fmla   v22.4s, v9.4s, v6.s[1]  \n"
-                "fmla   v23.4s, v9.4s, v7.s[1]  \n"
-
-                "fmla   v16.4s, v10.4s, v0.s[2] \n"
-                "fmla   v17.4s, v10.4s, v1.s[2] \n"
-                "fmla   v18.4s, v10.4s, v2.s[2] \n"
-                "fmla   v19.4s, v10.4s, v3.s[2] \n"
-                "fmla   v20.4s, v10.4s, v4.s[2] \n"
-                "fmla   v21.4s, v10.4s, v5.s[2] \n"
-                "fmla   v22.4s, v10.4s, v6.s[2] \n"
-                "fmla   v23.4s, v10.4s, v7.s[2] \n"
-
-                "fmla   v16.4s, v11.4s, v0.s[3] \n"
-                "fmla   v17.4s, v11.4s, v1.s[3] \n"
-                "fmla   v18.4s, v11.4s, v2.s[3] \n"
-                "fmla   v19.4s, v11.4s, v3.s[3] \n"
-                "fmla   v20.4s, v11.4s, v4.s[3] \n"
-                "fmla   v21.4s, v11.4s, v5.s[3] \n"
-                "fmla   v22.4s, v11.4s, v6.s[3] \n"
+                "fmla   v17.4s, v8.4s, v0.s[1]  \n"
+                "fmla   v18.4s, v8.4s, v0.s[2]  \n"
+                "fmla   v19.4s, v8.4s, v0.s[3]  \n"
+                "fmla   v20.4s, v8.4s, v1.s[0]  \n"
+                "fmla   v21.4s, v8.4s, v1.s[1]  \n"
+                "fmla   v22.4s, v8.4s, v1.s[2]  \n"
+                "fmla   v23.4s, v8.4s, v1.s[3]  \n"
+
+                "fmla   v16.4s, v9.4s, v2.s[0]  \n"
+                "fmla   v17.4s, v9.4s, v2.s[1]  \n"
+                "fmla   v18.4s, v9.4s, v2.s[2]  \n"
+                "fmla   v19.4s, v9.4s, v2.s[3]  \n"
+                "fmla   v20.4s, v9.4s, v3.s[0]  \n"
+                "fmla   v21.4s, v9.4s, v3.s[1]  \n"
+                "fmla   v22.4s, v9.4s, v3.s[2]  \n"
+                "fmla   v23.4s, v9.4s, v3.s[3]  \n"
+
+                "fmla   v16.4s, v10.4s, v4.s[0] \n"
+                "fmla   v17.4s, v10.4s, v4.s[1] \n"
+                "fmla   v18.4s, v10.4s, v4.s[2] \n"
+                "fmla   v19.4s, v10.4s, v4.s[3] \n"
+                "fmla   v20.4s, v10.4s, v5.s[0] \n"
+                "fmla   v21.4s, v10.4s, v5.s[1] \n"
+                "fmla   v22.4s, v10.4s, v5.s[2] \n"
+                "fmla   v23.4s, v10.4s, v5.s[3] \n"
+
+                "fmla   v16.4s, v11.4s, v6.s[0] \n"
+                "fmla   v17.4s, v11.4s, v6.s[1] \n"
+                "fmla   v18.4s, v11.4s, v6.s[2] \n"
+                "fmla   v19.4s, v11.4s, v6.s[3] \n"
+                "fmla   v20.4s, v11.4s, v7.s[0] \n"
+                "fmla   v21.4s, v11.4s, v7.s[1] \n"
+                "fmla   v22.4s, v11.4s, v7.s[2] \n"
                 "fmla   v23.4s, v11.4s, v7.s[3] \n"
 
                 "subs   w4, w4, #1              \n"
@@ -734,25 +604,24 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 "prfm   pldl1keep, [%9, #512]   \n"
                 "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%9], #64     \n"
 
-                "fmla   v16.4s, v8.4s, v0.4s    \n"
-                "fmla   v17.4s, v8.4s, v1.4s    \n"
-                "fmla   v18.4s, v8.4s, v2.4s    \n"
-                "fmla   v19.4s, v8.4s, v3.4s    \n"
-                "fmla   v20.4s, v8.4s, v4.4s    \n"
-                "fmla   v21.4s, v8.4s, v5.4s    \n"
-                "fmla   v22.4s, v8.4s, v6.4s    \n"
-                "fmla   v23.4s, v8.4s, v7.4s    \n"
+                "fmla   v16.4s, v0.4s, v8.s[0]  \n"
+                "fmla   v17.4s, v1.4s, v8.s[0]  \n"
+                "fmla   v18.4s, v2.4s, v8.s[1]  \n"
+                "fmla   v19.4s, v3.4s, v8.s[1]  \n"
+                "fmla   v20.4s, v4.4s, v8.s[2]  \n"
+                "fmla   v21.4s, v5.4s, v8.s[2]  \n"
+                "fmla   v22.4s, v6.4s, v8.s[3]  \n"
+                "fmla   v23.4s, v7.4s, v8.s[3]  \n"
 
                 "subs   w4, w4, #1              \n"
                 "bne    0b                      \n"
 
-                "faddp  v0.4s, v16.4s, v17.4s   \n"
-                "faddp  v1.4s, v18.4s, v19.4s   \n"
-                "faddp  v2.4s, v20.4s, v21.4s   \n"
-                "faddp  v3.4s, v22.4s, v23.4s   \n"
-                "faddp  v16.4s, v0.4s, v1.4s    \n"
-                "faddp  v17.4s, v2.4s, v3.4s    \n"
-
+                "fadd   v16.4s, v16.4s, v18.4s  \n"
+                "fadd   v17.4s, v17.4s, v19.4s  \n"
+                "fadd   v20.4s, v20.4s, v22.4s  \n"
+                "fadd   v21.4s, v21.4s, v23.4s  \n"
+                "fadd   v16.4s, v16.4s, v20.4s  \n"
+                "fadd   v17.4s, v17.4s, v21.4s  \n"
                 "fadd   v24.4s, v24.4s, v16.4s  \n"
                 "fadd   v25.4s, v25.4s, v17.4s  \n"
 
@@ -814,7 +683,7 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
             );
         }
     }
-#endif // __aarch64__
+#endif // __ARM_NEON && __aarch64__
 
     nn_outch = (outch - remain_outch_start) >> 2;
 
@@ -836,12 +705,13 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
         for (; i+7<size; i+=8)
         {
             const float* tmpptr = tmp.channel(i/8);
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
             const float* kptr = kernel.channel(p/8 + (p%8)/4);
 #else
             const float* kptr = kernel.channel(p/4);
-#endif // __aarch64__
+#endif // __ARM_NEON && __aarch64__
 
+#if __ARM_NEON
 #if __aarch64__
             asm volatile(
                 "ld1    {v0.4s}, [%12]          \n"
@@ -868,46 +738,46 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%5], #64     \n"
 
                 "fmla   v8.4s, v4.4s, v0.s[0]   \n"
-                "fmla   v10.4s, v4.4s, v1.s[0]  \n"
-                "fmla   v12.4s, v4.4s, v2.s[0]  \n"
-                "fmla   v14.4s, v4.4s, v3.s[0]  \n"
+                "fmla   v10.4s, v4.4s, v0.s[1]  \n"
+                "fmla   v12.4s, v4.4s, v0.s[2]  \n"
+                "fmla   v14.4s, v4.4s, v0.s[3]  \n"
 
                 "fmla   v9.4s, v5.4s, v0.s[0]   \n"
-                "fmla   v11.4s, v5.4s, v1.s[0]  \n"
-                "fmla   v13.4s, v5.4s, v2.s[0]  \n"
-                "fmla   v15.4s, v5.4s, v3.s[0]  \n"
+                "fmla   v11.4s, v5.4s, v0.s[1]  \n"
+                "fmla   v13.4s, v5.4s, v0.s[2]  \n"
+                "fmla   v15.4s, v5.4s, v0.s[3]  \n"
 
                 "prfm   pldl1keep, [%4, #512]   \n"
                 "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
 
-                "fmla   v8.4s, v6.4s, v0.s[1]   \n"
+                "fmla   v8.4s, v6.4s, v1.s[0]   \n"
                 "fmla   v10.4s, v6.4s, v1.s[1]  \n"
-                "fmla   v12.4s, v6.4s, v2.s[1]  \n"
-                "fmla   v14.4s, v6.4s, v3.s[1]  \n"
+                "fmla   v12.4s, v6.4s, v1.s[2]  \n"
+                "fmla   v14.4s, v6.4s, v1.s[3]  \n"
 
-                "fmla   v9.4s, v7.4s, v0.s[1]   \n"
+                "fmla   v9.4s, v7.4s, v1.s[0]   \n"
                 "fmla   v11.4s, v7.4s, v1.s[1]  \n"
-                "fmla   v13.4s, v7.4s, v2.s[1]  \n"
-                "fmla   v15.4s, v7.4s, v3.s[1]  \n"
+                "fmla   v13.4s, v7.4s, v1.s[2]  \n"
+                "fmla   v15.4s, v7.4s, v1.s[3]  \n"
 
-                "fmla   v8.4s, v16.4s, v0.s[2]  \n"
-                "fmla   v10.4s, v16.4s, v1.s[2] \n"
+                "fmla   v8.4s, v16.4s, v2.s[0]  \n"
+                "fmla   v10.4s, v16.4s, v2.s[1] \n"
                 "fmla   v12.4s, v16.4s, v2.s[2] \n"
-                "fmla   v14.4s, v16.4s, v3.s[2] \n"
+                "fmla   v14.4s, v16.4s, v2.s[3] \n"
 
-                "fmla   v9.4s, v17.4s, v0.s[2]  \n"
-                "fmla   v11.4s, v17.4s, v1.s[2] \n"
+                "fmla   v9.4s, v17.4s, v2.s[0]  \n"
+                "fmla   v11.4s, v17.4s, v2.s[1] \n"
                 "fmla   v13.4s, v17.4s, v2.s[2] \n"
-                "fmla   v15.4s, v17.4s, v3.s[2] \n"
+                "fmla   v15.4s, v17.4s, v2.s[3] \n"
 
-                "fmla   v8.4s, v18.4s, v0.s[3]  \n"
-                "fmla   v10.4s, v18.4s, v1.s[3] \n"
-                "fmla   v12.4s, v18.4s, v2.s[3] \n"
+                "fmla   v8.4s, v18.4s, v3.s[0]  \n"
+                "fmla   v10.4s, v18.4s, v3.s[1] \n"
+                "fmla   v12.4s, v18.4s, v3.s[2] \n"
                 "fmla   v14.4s, v18.4s, v3.s[3] \n"
 
-                "fmla   v9.4s, v19.4s, v0.s[3]  \n"
-                "fmla   v11.4s, v19.4s, v1.s[3] \n"
-                "fmla   v13.4s, v19.4s, v2.s[3] \n"
+                "fmla   v9.4s, v19.4s, v3.s[0]  \n"
+                "fmla   v11.4s, v19.4s, v3.s[1] \n"
+                "fmla   v13.4s, v19.4s, v3.s[2] \n"
                 "fmla   v15.4s, v19.4s, v3.s[3] \n"
 
                 "subs   w4, w4, #1              \n"
@@ -994,48 +864,48 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 "vld1.f32   {d4-d7}, [%5 :128]! \n"
 
                 "vmla.f32   q8, q4, d0[0]       \n"
-                "vmla.f32   q10, q4, d2[0]      \n"
-                "vmla.f32   q12, q4, d4[0]      \n"
-                "vmla.f32   q14, q4, d6[0]      \n"
+                "vmla.f32   q10, q4, d0[1]      \n"
+                "vmla.f32   q12, q4, d1[0]      \n"
+                "vmla.f32   q14, q4, d1[1]      \n"
 
                 "vmla.f32   q9, q5, d0[0]       \n"
-                "vmla.f32   q11, q5, d2[0]      \n"
-                "vmla.f32   q13, q5, d4[0]      \n"
-                "vmla.f32   q15, q5, d6[0]      \n"
+                "vmla.f32   q11, q5, d0[1]      \n"
+                "vmla.f32   q13, q5, d1[0]      \n"
+                "vmla.f32   q15, q5, d1[1]      \n"
 
-                "vmla.f32   q8, q6, d0[1]       \n"
+                "vmla.f32   q8, q6, d2[0]       \n"
                 "vmla.f32   q10, q6, d2[1]      \n"
-                "vmla.f32   q12, q6, d4[1]      \n"
-                "vmla.f32   q14, q6, d6[1]      \n"
+                "vmla.f32   q12, q6, d3[0]      \n"
+                "vmla.f32   q14, q6, d3[1]      \n"
 
-                "vmla.f32   q9, q7, d0[1]       \n"
+                "vmla.f32   q9, q7, d2[0]       \n"
                 "vmla.f32   q11, q7, d2[1]      \n"
-                "vmla.f32   q13, q7, d4[1]      \n"
-                "vmla.f32   q15, q7, d6[1]      \n"
+                "vmla.f32   q13, q7, d3[0]      \n"
+                "vmla.f32   q15, q7, d3[1]      \n"
 
                 "pld        [%4, #512]          \n"
 //                 "vldm       %4!, {d8-d15}       \n"
                 "vld1.f32   {d8-d11}, [%4 :128]!    \n"
                 "vld1.f32   {d12-d15}, [%4 :128]!   \n"
 
-                "vmla.f32   q8, q4, d1[0]       \n"
-                "vmla.f32   q10, q4, d3[0]      \n"
+                "vmla.f32   q8, q4, d4[0]       \n"
+                "vmla.f32   q10, q4, d4[1]      \n"
                 "vmla.f32   q12, q4, d5[0]      \n"
-                "vmla.f32   q14, q4, d7[0]      \n"
+                "vmla.f32   q14, q4, d5[1]      \n"
 
-                "vmla.f32   q9, q5, d1[0]       \n"
-                "vmla.f32   q11, q5, d3[0]      \n"
+                "vmla.f32   q9, q5, d4[0]       \n"
+                "vmla.f32   q11, q5, d4[1]      \n"
                 "vmla.f32   q13, q5, d5[0]      \n"
-                "vmla.f32   q15, q5, d7[0]      \n"
+                "vmla.f32   q15, q5, d5[1]      \n"
 
-                "vmla.f32   q8, q6, d1[1]       \n"
-                "vmla.f32   q10, q6, d3[1]      \n"
-                "vmla.f32   q12, q6, d5[1]      \n"
+                "vmla.f32   q8, q6, d6[0]       \n"
+                "vmla.f32   q10, q6, d6[1]      \n"
+                "vmla.f32   q12, q6, d7[0]      \n"
                 "vmla.f32   q14, q6, d7[1]      \n"
 
-                "vmla.f32   q9, q7, d1[1]       \n"
-                "vmla.f32   q11, q7, d3[1]      \n"
-                "vmla.f32   q13, q7, d5[1]      \n"
+                "vmla.f32   q9, q7, d6[0]       \n"
+                "vmla.f32   q11, q7, d6[1]      \n"
+                "vmla.f32   q13, q7, d7[0]      \n"
                 "vmla.f32   q15, q7, d7[1]      \n"
 
                 "subs       r4, r4, #1          \n"
@@ -1093,17 +963,138 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
             );
 #endif // __aarch64__
+#else
+            float sum0_0 = biasptr[0];
+            float sum0_1 = biasptr[0];
+            float sum0_2 = biasptr[0];
+            float sum0_3 = biasptr[0];
+            float sum0_4 = biasptr[0];
+            float sum0_5 = biasptr[0];
+            float sum0_6 = biasptr[0];
+            float sum0_7 = biasptr[0];
+
+            float sum1_0 = biasptr[1];
+            float sum1_1 = biasptr[1];
+            float sum1_2 = biasptr[1];
+            float sum1_3 = biasptr[1];
+            float sum1_4 = biasptr[1];
+            float sum1_5 = biasptr[1];
+            float sum1_6 = biasptr[1];
+            float sum1_7 = biasptr[1];
+
+            float sum2_0 = biasptr[2];
+            float sum2_1 = biasptr[2];
+            float sum2_2 = biasptr[2];
+            float sum2_3 = biasptr[2];
+            float sum2_4 = biasptr[2];
+            float sum2_5 = biasptr[2];
+            float sum2_6 = biasptr[2];
+            float sum2_7 = biasptr[2];
+
+            float sum3_0 = biasptr[3];
+            float sum3_1 = biasptr[3];
+            float sum3_2 = biasptr[3];
+            float sum3_3 = biasptr[3];
+            float sum3_4 = biasptr[3];
+            float sum3_5 = biasptr[3];
+            float sum3_6 = biasptr[3];
+            float sum3_7 = biasptr[3];
+
+            for (int q=0; q<inch; q++)
+            {
+                sum0_0 += tmpptr[0] * kptr[0];
+                sum0_1 += tmpptr[1] * kptr[0];
+                sum0_2 += tmpptr[2] * kptr[0];
+                sum0_3 += tmpptr[3] * kptr[0];
+                sum0_4 += tmpptr[4] * kptr[0];
+                sum0_5 += tmpptr[5] * kptr[0];
+                sum0_6 += tmpptr[6] * kptr[0];
+                sum0_7 += tmpptr[7] * kptr[0];
+
+                sum1_0 += tmpptr[0] * kptr[1];
+                sum1_1 += tmpptr[1] * kptr[1];
+                sum1_2 += tmpptr[2] * kptr[1];
+                sum1_3 += tmpptr[3] * kptr[1];
+                sum1_4 += tmpptr[4] * kptr[1];
+                sum1_5 += tmpptr[5] * kptr[1];
+                sum1_6 += tmpptr[6] * kptr[1];
+                sum1_7 += tmpptr[7] * kptr[1];
+
+                sum2_0 += tmpptr[0] * kptr[2];
+                sum2_1 += tmpptr[1] * kptr[2];
+                sum2_2 += tmpptr[2] * kptr[2];
+                sum2_3 += tmpptr[3] * kptr[2];
+                sum2_4 += tmpptr[4] * kptr[2];
+                sum2_5 += tmpptr[5] * kptr[2];
+                sum2_6 += tmpptr[6] * kptr[2];
+                sum2_7 += tmpptr[7] * kptr[2];
+
+                sum3_0 += tmpptr[0] * kptr[3];
+                sum3_1 += tmpptr[1] * kptr[3];
+                sum3_2 += tmpptr[2] * kptr[3];
+                sum3_3 += tmpptr[3] * kptr[3];
+                sum3_4 += tmpptr[4] * kptr[3];
+                sum3_5 += tmpptr[5] * kptr[3];
+                sum3_6 += tmpptr[6] * kptr[3];
+                sum3_7 += tmpptr[7] * kptr[3];
+
+                tmpptr += 8;
+                kptr += 4;
+            }
+
+            outptr0[0] = sum0_0;
+            outptr0[1] = sum0_1;
+            outptr0[2] = sum0_2;
+            outptr0[3] = sum0_3;
+            outptr0[4] = sum0_4;
+            outptr0[5] = sum0_5;
+            outptr0[6] = sum0_6;
+            outptr0[7] = sum0_7;
+
+            outptr1[0] = sum1_0;
+            outptr1[1] = sum1_1;
+            outptr1[2] = sum1_2;
+            outptr1[3] = sum1_3;
+            outptr1[4] = sum1_4;
+            outptr1[5] = sum1_5;
+            outptr1[6] = sum1_6;
+            outptr1[7] = sum1_7;
+
+            outptr2[0] = sum2_0;
+            outptr2[1] = sum2_1;
+            outptr2[2] = sum2_2;
+            outptr2[3] = sum2_3;
+            outptr2[4] = sum2_4;
+            outptr2[5] = sum2_5;
+            outptr2[6] = sum2_6;
+            outptr2[7] = sum2_7;
+
+            outptr3[0] = sum3_0;
+            outptr3[1] = sum3_1;
+            outptr3[2] = sum3_2;
+            outptr3[3] = sum3_3;
+            outptr3[4] = sum3_4;
+            outptr3[5] = sum3_5;
+            outptr3[6] = sum3_6;
+            outptr3[7] = sum3_7;
+
+            outptr0 += 8;
+            outptr1 += 8;
+            outptr2 += 8;
+            outptr3 += 8;
+#endif // __ARM_NEON
         }
 
         for (; i+3<size; i+=4)
         {
             const float* tmpptr = tmp.channel(i/8 + (i%8)/4);
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
             const float* kptr = kernel.channel(p/8 + (p%8)/4);
 #else
             const float* kptr = kernel.channel(p/4);
-#endif // __aarch64__
+#endif // __ARM_NEON && __aarch64__
 
+#if __ARM_NEON
 #if __aarch64__
             asm volatile(
                 "ld1    {v0.4s}, [%12]          \n"
@@ -1126,23 +1117,23 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%5], #64     \n"
 
                 "fmla   v8.4s, v4.4s, v0.s[0]   \n"
-                "fmla   v9.4s, v4.4s, v1.s[0]   \n"
-                "fmla   v10.4s, v4.4s, v2.s[0]  \n"
-                "fmla   v11.4s, v4.4s, v3.s[0]  \n"
+                "fmla   v9.4s, v4.4s, v0.s[1]   \n"
+                "fmla   v10.4s, v4.4s, v0.s[2]  \n"
+                "fmla   v11.4s, v4.4s, v0.s[3]  \n"
 
-                "fmla   v8.4s, v5.4s, v0.s[1]   \n"
+                "fmla   v8.4s, v5.4s, v1.s[0]   \n"
                 "fmla   v9.4s, v5.4s, v1.s[1]   \n"
-                "fmla   v10.4s, v5.4s, v2.s[1]  \n"
-                "fmla   v11.4s, v5.4s, v3.s[1]  \n"
+                "fmla   v10.4s, v5.4s, v1.s[2]  \n"
+                "fmla   v11.4s, v5.4s, v1.s[3]  \n"
 
-                "fmla   v8.4s, v6.4s, v0.s[2]   \n"
-                "fmla   v9.4s, v6.4s, v1.s[2]   \n"
+                "fmla   v8.4s, v6.4s, v2.s[0]   \n"
+                "fmla   v9.4s, v6.4s, v2.s[1]   \n"
                 "fmla   v10.4s, v6.4s, v2.s[2]  \n"
-                "fmla   v11.4s, v6.4s, v3.s[2]  \n"
+                "fmla   v11.4s, v6.4s, v2.s[3]  \n"
 
-                "fmla   v8.4s, v7.4s, v0.s[3]   \n"
-                "fmla   v9.4s, v7.4s, v1.s[3]   \n"
-                "fmla   v10.4s, v7.4s, v2.s[3]  \n"
+                "fmla   v8.4s, v7.4s, v3.s[0]   \n"
+                "fmla   v9.4s, v7.4s, v3.s[1]   \n"
+                "fmla   v10.4s, v7.4s, v3.s[2]  \n"
                 "fmla   v11.4s, v7.4s, v3.s[3]  \n"
 
                 "subs   w4, w4, #1              \n"
@@ -1220,23 +1211,23 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 "vld1.f32   {d4-d7}, [%5 :128]! \n"
 
                 "vmla.f32   q8, q4, d0[0]       \n"
-                "vmla.f32   q9, q4, d2[0]       \n"
-                "vmla.f32   q10, q4, d4[0]      \n"
-                "vmla.f32   q11, q4, d6[0]      \n"
+                "vmla.f32   q9, q4, d0[1]       \n"
+                "vmla.f32   q10, q4, d1[0]      \n"
+                "vmla.f32   q11, q4, d1[1]      \n"
 
-                "vmla.f32   q8, q5, d0[1]       \n"
+                "vmla.f32   q8, q5, d2[0]       \n"
                 "vmla.f32   q9, q5, d2[1]       \n"
-                "vmla.f32   q10, q5, d4[1]      \n"
-                "vmla.f32   q11, q5, d6[1]      \n"
+                "vmla.f32   q10, q5, d3[0]      \n"
+                "vmla.f32   q11, q5, d3[1]      \n"
 
-                "vmla.f32   q8, q6, d1[0]       \n"
-                "vmla.f32   q9, q6, d3[0]       \n"
+                "vmla.f32   q8, q6, d4[0]       \n"
+                "vmla.f32   q9, q6, d4[1]       \n"
                 "vmla.f32   q10, q6, d5[0]      \n"
-                "vmla.f32   q11, q6, d7[0]      \n"
+                "vmla.f32   q11, q6, d5[1]      \n"
 
-                "vmla.f32   q8, q7, d1[1]       \n"
-                "vmla.f32   q9, q7, d3[1]       \n"
-                "vmla.f32   q10, q7, d5[1]      \n"
+                "vmla.f32   q8, q7, d6[0]       \n"
+                "vmla.f32   q9, q7, d6[1]       \n"
+                "vmla.f32   q10, q7, d7[0]      \n"
                 "vmla.f32   q11, q7, d7[1]      \n"
 
                 "subs       r4, r4, #1          \n"
@@ -1289,17 +1280,90 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11"
             );
 #endif // __aarch64__
+#else
+            float sum0_0 = biasptr[0];
+            float sum0_1 = biasptr[0];
+            float sum0_2 = biasptr[0];
+            float sum0_3 = biasptr[0];
+
+            float sum1_0 = biasptr[1];
+            float sum1_1 = biasptr[1];
+            float sum1_2 = biasptr[1];
+            float sum1_3 = biasptr[1];
+
+            float sum2_0 = biasptr[2];
+            float sum2_1 = biasptr[2];
+            float sum2_2 = biasptr[2];
+            float sum2_3 = biasptr[2];
+
+            float sum3_0 = biasptr[3];
+            float sum3_1 = biasptr[3];
+            float sum3_2 = biasptr[3];
+            float sum3_3 = biasptr[3];
+
+            for (int q=0; q<inch; q++)
+            {
+                sum0_0 += tmpptr[0] * kptr[0];
+                sum0_1 += tmpptr[1] * kptr[0];
+                sum0_2 += tmpptr[2] * kptr[0];
+                sum0_3 += tmpptr[3] * kptr[0];
+
+                sum1_0 += tmpptr[0] * kptr[1];
+                sum1_1 += tmpptr[1] * kptr[1];
+                sum1_2 += tmpptr[2] * kptr[1];
+                sum1_3 += tmpptr[3] * kptr[1];
+
+                sum2_0 += tmpptr[0] * kptr[2];
+                sum2_1 += tmpptr[1] * kptr[2];
+                sum2_2 += tmpptr[2] * kptr[2];
+                sum2_3 += tmpptr[3] * kptr[2];
+
+                sum3_0 += tmpptr[0] * kptr[3];
+                sum3_1 += tmpptr[1] * kptr[3];
+                sum3_2 += tmpptr[2] * kptr[3];
+                sum3_3 += tmpptr[3] * kptr[3];
+
+                tmpptr += 4;
+                kptr += 4;
+            }
+
+            outptr0[0] = sum0_0;
+            outptr0[1] = sum0_1;
+            outptr0[2] = sum0_2;
+            outptr0[3] = sum0_3;
+
+            outptr1[0] = sum1_0;
+            outptr1[1] = sum1_1;
+            outptr1[2] = sum1_2;
+            outptr1[3] = sum1_3;
+
+            outptr2[0] = sum2_0;
+            outptr2[1] = sum2_1;
+            outptr2[2] = sum2_2;
+            outptr2[3] = sum2_3;
+
+            outptr3[0] = sum3_0;
+            outptr3[1] = sum3_1;
+            outptr3[2] = sum3_2;
+            outptr3[3] = sum3_3;
+
+            outptr0 += 4;
+            outptr1 += 4;
+            outptr2 += 4;
+            outptr3 += 4;
+#endif // __ARM_NEON
         }
 
         for (; i<size; i++)
         {
             const float* tmpptr = tmp.channel(i/8 + (i%8)/4 + i%4);
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
             const float* kptr = kernel.channel(p/8 + (p%8)/4);
 #else
             const float* kptr = kernel.channel(p/4);
-#endif // __aarch64__
+#endif // __ARM_NEON && __aarch64__
 
+#if __ARM_NEON
 #if __aarch64__
             asm volatile(
                 "ld1    {v12.4s}, [%12]         \n"
@@ -1322,18 +1386,17 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 "prfm   pldl1keep, [%5, #512]   \n"
                 "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%5], #64     \n"
 
-                "fmla   v8.4s, v4.4s, v0.4s     \n"
-                "fmla   v9.4s, v4.4s, v1.4s     \n"
-                "fmla   v10.4s, v4.4s, v2.4s    \n"
-                "fmla   v11.4s, v4.4s, v3.4s    \n"
+                "fmla   v8.4s, v0.4s, v4.s[0]   \n"
+                "fmla   v9.4s, v1.4s, v4.s[1]   \n"
+                "fmla   v10.4s, v2.4s, v4.s[2]  \n"
+                "fmla   v11.4s, v3.4s, v4.s[3]  \n"
 
                 "subs   w4, w4, #1              \n"
                 "bne    0b                      \n"
 
-                "faddp  v0.4s, v8.4s, v9.4s     \n"
-                "faddp  v1.4s, v10.4s, v11.4s   \n"
-                "faddp  v8.4s, v0.4s, v1.4s     \n"
-
+                "fadd   v8.4s, v8.4s, v9.4s     \n"
+                "fadd   v10.4s, v10.4s, v11.4s  \n"
+                "fadd   v8.4s, v8.4s, v10.4s    \n"
                 "fadd   v12.4s, v12.4s, v8.4s   \n"
 
                 "1:                             \n"
@@ -1403,20 +1466,17 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 "vld1.f32   {d0-d3}, [%5 :128]! \n"
                 "vld1.f32   {d4-d7}, [%5 :128]! \n"
 
-                "vmla.f32   q8, q4, q0          \n"
-                "vmla.f32   q9, q4, q1          \n"
-                "vmla.f32   q10, q4, q2         \n"
-                "vmla.f32   q11, q4, q3         \n"
+                "vmla.f32   q8, q0, d8[0]       \n"
+                "vmla.f32   q9, q1, d8[1]       \n"
+                "vmla.f32   q10, q2, d9[0]      \n"
+                "vmla.f32   q11, q3, d9[1]      \n"
 
                 "subs       r4, r4, #1          \n"
                 "bne        0b                  \n"
 
-                "vadd.f32   d16, d16, d17       \n"
-                "vadd.f32   d20, d20, d21       \n"
-                "vadd.f32   d17, d18, d19       \n"
-                "vadd.f32   d21, d22, d23       \n"
-                "vpadd.f32  d16, d16, d17       \n"
-                "vpadd.f32  d17, d20, d21       \n"
+                "vadd.f32   q8, q8, q9          \n"
+                "vadd.f32   q10, q10, q11       \n"
+                "vadd.f32   q8, q8, q10         \n"
                 "vadd.f32   q12, q12, q8        \n"
 
                 "1:                             \n"
@@ -1463,6 +1523,33 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11", "q12"
             );
 #endif // __aarch64__
+#else
+            float sum0 = biasptr[0];
+            float sum1 = biasptr[1];
+            float sum2 = biasptr[2];
+            float sum3 = biasptr[3];
+
+            for (int q=0; q<inch; q++)
+            {
+                sum0 += tmpptr[0] * kptr[0];
+                sum1 += tmpptr[0] * kptr[1];
+                sum2 += tmpptr[0] * kptr[2];
+                sum3 += tmpptr[0] * kptr[3];
+
+                tmpptr++;
+                kptr += 4;
+            }
+
+            outptr0[0] = sum0;
+            outptr1[0] = sum1;
+            outptr2[0] = sum2;
+            outptr3[0] = sum3;
+
+            outptr0++;
+            outptr1++;
+            outptr2++;
+            outptr3++;
+#endif // __ARM_NEON
         }
     }
 
@@ -1482,12 +1569,13 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
         for (; i+7<size; i+=8)
         {
             const float* tmpptr = tmp.channel(i/8);
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
             const float* kptr = kernel.channel(p/8 + (p%8)/4 + p%4);
 #else
             const float* kptr = kernel.channel(p/4 + p%4);
-#endif // __aarch64__
+#endif // __ARM_NEON && __aarch64__
 
+#if __ARM_NEON
 #if __aarch64__
             asm volatile(
                 "dup    v8.4s, %w6              \n"
@@ -1635,17 +1723,54 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 : "cc", "memory", "r4", "q0", "q4", "q5", "q6", "q7", "q8", "q9", "q12", "q13", "q14", "q15"
             );
 #endif // __aarch64__
+#else
+            float sum0 = bias0;
+            float sum1 = bias0;
+            float sum2 = bias0;
+            float sum3 = bias0;
+            float sum4 = bias0;
+            float sum5 = bias0;
+            float sum6 = bias0;
+            float sum7 = bias0;
+
+            for (int q=0; q<inch; q++)
+            {
+                sum0 += tmpptr[0] * kptr[0];
+                sum1 += tmpptr[1] * kptr[0];
+                sum2 += tmpptr[2] * kptr[0];
+                sum3 += tmpptr[3] * kptr[0];
+                sum4 += tmpptr[4] * kptr[0];
+                sum5 += tmpptr[5] * kptr[0];
+                sum6 += tmpptr[6] * kptr[0];
+                sum7 += tmpptr[7] * kptr[0];
+
+                tmpptr += 8;
+                kptr++;
+            }
+
+            outptr0[0] = sum0;
+            outptr0[1] = sum1;
+            outptr0[2] = sum2;
+            outptr0[3] = sum3;
+            outptr0[4] = sum4;
+            outptr0[5] = sum5;
+            outptr0[6] = sum6;
+            outptr0[7] = sum7;
+
+            outptr0 += 8;
+#endif // __ARM_NEON
         }
 
         for (; i+3<size; i+=4)
         {
             const float* tmpptr = tmp.channel(i/8 + (i%8)/4);
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
             const float* kptr = kernel.channel(p/8 + (p%8)/4 + p%4);
 #else
             const float* kptr = kernel.channel(p/4 + p%4);
-#endif // __aarch64__
+#endif // __ARM_NEON && __aarch64__
 
+#if __ARM_NEON
 #if __aarch64__
             asm volatile(
                 "dup    v8.4s, %w6              \n"
@@ -1767,21 +1892,46 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 : "cc", "memory", "r4", "q0", "q4", "q5", "q6", "q7", "q8"
             );
 #endif // __aarch64__
+#else
+            float sum0 = bias0;
+            float sum1 = bias0;
+            float sum2 = bias0;
+            float sum3 = bias0;
+
+            for (int q=0; q<inch; q++)
+            {
+                sum0 += tmpptr[0] * kptr[0];
+                sum1 += tmpptr[1] * kptr[0];
+                sum2 += tmpptr[2] * kptr[0];
+                sum3 += tmpptr[3] * kptr[0];
+
+                tmpptr += 4;
+                kptr++;
+            }
+
+            outptr0[0] = sum0;
+            outptr0[1] = sum1;
+            outptr0[2] = sum2;
+            outptr0[3] = sum3;
+
+            outptr0 += 4;
+#endif // __ARM_NEON
         }
 
         for (; i<size; i++)
         {
             const float* tmpptr = tmp.channel(i/8 + (i%8)/4 + i%4);
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
             const float* kptr = kernel.channel(p/8 + (p%8)/4 + p%4);
 #else
             const float* kptr = kernel.channel(p/4 + p%4);
-#endif // __aarch64__
-
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
+#endif // __ARM_NEON && __aarch64__
 
             int q = 0;
 
+#if __ARM_NEON
+            float32x4_t _sum0 = vdupq_n_f32(0.f);
+
             for (; q+3<inch; q+=4)
             {
                 float32x4_t _p0 = vld1q_f32(tmpptr);
@@ -1798,11 +1948,14 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
             }
 
 #if __aarch64__
-            float sum0 = vaddvq_f32(_sum0);
+            float sum0 = bias0 + vaddvq_f32(_sum0);
 #else
             float32x2_t _ss = vadd_f32(vget_low_f32(_sum0), vget_high_f32(_sum0));
-            float sum0 = vget_lane_f32(vpadd_f32(_ss, _ss), 0);
+            float sum0 = bias0 + vget_lane_f32(vpadd_f32(_ss, _ss), 0);
 #endif
+#else
+            float sum0 = bias0;
+#endif // __ARM_NEON
 
             for (; q<inch; q++)
             {
@@ -1811,7 +1964,7 @@ static void conv1x1s1_sgemm_neon(const Mat& bottom_blob, Mat& top_blob, const Ma
                 kptr++;
             }
 
-            outptr0[0] = bias0 + sum0;
+            outptr0[0] = sum0;
 
             outptr0++;
         }
diff --git a/src/layer/arm/convolution_3x3.h b/src/layer/arm/convolution_3x3.h
index da0cbbaf8..83439b78c 100644
--- a/src/layer/arm/convolution_3x3.h
+++ b/src/layer/arm/convolution_3x3.h
@@ -1588,7 +1588,7 @@ static void conv3x3s1_winograd64_transform_kernel_neon5(const Mat& kernel, Mat&
     // interleave weights
 //     Mat kernel_tm2(8*8, inch, outch);
 //     Mat kernel_tm2(inch, 64, outch);
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
     Mat kernel_tm2(8*4*(inch/4) + 8*(inch%4), 64, outch/8 + (outch%8)/4 + outch%4);
 #else
     Mat kernel_tm2(4*4*(inch/4) + 4*(inch%4), 64, outch/4 + outch%4);
@@ -1613,106 +1613,7 @@ static void conv3x3s1_winograd64_transform_kernel_neon5(const Mat& kernel, Mat&
         {
             float* ktm2p = ktm2.row(r);
 
-            int q=0;
-            for (; q+3<inch; q+=4)
-            {
-                const float* ktm0_0 = kernel0_tm.row(q);
-                const float* ktm0_1 = kernel0_tm.row(q+1);
-                const float* ktm0_2 = kernel0_tm.row(q+2);
-                const float* ktm0_3 = kernel0_tm.row(q+3);
-
-                ktm2p[0] = ktm0_0[r];
-                ktm2p[1] = ktm0_1[r];
-                ktm2p[2] = ktm0_2[r];
-                ktm2p[3] = ktm0_3[r];
-
-                ktm2p += 4;
-
-                const float* ktm1_0 = kernel1_tm.row(q);
-                const float* ktm1_1 = kernel1_tm.row(q+1);
-                const float* ktm1_2 = kernel1_tm.row(q+2);
-                const float* ktm1_3 = kernel1_tm.row(q+3);
-
-                ktm2p[0] = ktm1_0[r];
-                ktm2p[1] = ktm1_1[r];
-                ktm2p[2] = ktm1_2[r];
-                ktm2p[3] = ktm1_3[r];
-
-                ktm2p += 4;
-
-                const float* ktm2_0 = kernel2_tm.row(q);
-                const float* ktm2_1 = kernel2_tm.row(q+1);
-                const float* ktm2_2 = kernel2_tm.row(q+2);
-                const float* ktm2_3 = kernel2_tm.row(q+3);
-
-                ktm2p[0] = ktm2_0[r];
-                ktm2p[1] = ktm2_1[r];
-                ktm2p[2] = ktm2_2[r];
-                ktm2p[3] = ktm2_3[r];
-
-                ktm2p += 4;
-
-                const float* ktm3_0 = kernel3_tm.row(q);
-                const float* ktm3_1 = kernel3_tm.row(q+1);
-                const float* ktm3_2 = kernel3_tm.row(q+2);
-                const float* ktm3_3 = kernel3_tm.row(q+3);
-
-                ktm2p[0] = ktm3_0[r];
-                ktm2p[1] = ktm3_1[r];
-                ktm2p[2] = ktm3_2[r];
-                ktm2p[3] = ktm3_3[r];
-
-                ktm2p += 4;
-
-                const float* ktm4_0 = kernel4_tm.row(q);
-                const float* ktm4_1 = kernel4_tm.row(q+1);
-                const float* ktm4_2 = kernel4_tm.row(q+2);
-                const float* ktm4_3 = kernel4_tm.row(q+3);
-
-                ktm2p[0] = ktm4_0[r];
-                ktm2p[1] = ktm4_1[r];
-                ktm2p[2] = ktm4_2[r];
-                ktm2p[3] = ktm4_3[r];
-
-                ktm2p += 4;
-
-                const float* ktm5_0 = kernel5_tm.row(q);
-                const float* ktm5_1 = kernel5_tm.row(q+1);
-                const float* ktm5_2 = kernel5_tm.row(q+2);
-                const float* ktm5_3 = kernel5_tm.row(q+3);
-
-                ktm2p[0] = ktm5_0[r];
-                ktm2p[1] = ktm5_1[r];
-                ktm2p[2] = ktm5_2[r];
-                ktm2p[3] = ktm5_3[r];
-
-                ktm2p += 4;
-
-                const float* ktm6_0 = kernel6_tm.row(q);
-                const float* ktm6_1 = kernel6_tm.row(q+1);
-                const float* ktm6_2 = kernel6_tm.row(q+2);
-                const float* ktm6_3 = kernel6_tm.row(q+3);
-
-                ktm2p[0] = ktm6_0[r];
-                ktm2p[1] = ktm6_1[r];
-                ktm2p[2] = ktm6_2[r];
-                ktm2p[3] = ktm6_3[r];
-
-                ktm2p += 4;
-
-                const float* ktm7_0 = kernel7_tm.row(q);
-                const float* ktm7_1 = kernel7_tm.row(q+1);
-                const float* ktm7_2 = kernel7_tm.row(q+2);
-                const float* ktm7_3 = kernel7_tm.row(q+3);
-
-                ktm2p[0] = ktm7_0[r];
-                ktm2p[1] = ktm7_1[r];
-                ktm2p[2] = ktm7_2[r];
-                ktm2p[3] = ktm7_3[r];
-
-                ktm2p += 4;
-            }
-            for (; q<inch; q++)
+            for (int q=0; q<inch; q++)
             {
                 const float* ktm0_0 = kernel0_tm.row(q);
                 const float* ktm1_0 = kernel1_tm.row(q);
@@ -1744,7 +1645,7 @@ static void conv3x3s1_winograd64_transform_kernel_neon5(const Mat& kernel, Mat&
         const Mat kernel2_tm = kernel_tm.channel(p+2);
         const Mat kernel3_tm = kernel_tm.channel(p+3);
 
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
         Mat ktm2 = kernel_tm2.channel(p/8+(p%8)/4);
 #else
         Mat ktm2 = kernel_tm2.channel(p/4);
@@ -1754,58 +1655,7 @@ static void conv3x3s1_winograd64_transform_kernel_neon5(const Mat& kernel, Mat&
         {
             float* ktm2p = ktm2.row(r);
 
-            int q=0;
-            for (; q+3<inch; q+=4)
-            {
-                const float* ktm0_0 = kernel0_tm.row(q);
-                const float* ktm0_1 = kernel0_tm.row(q+1);
-                const float* ktm0_2 = kernel0_tm.row(q+2);
-                const float* ktm0_3 = kernel0_tm.row(q+3);
-
-                ktm2p[0] = ktm0_0[r];
-                ktm2p[1] = ktm0_1[r];
-                ktm2p[2] = ktm0_2[r];
-                ktm2p[3] = ktm0_3[r];
-
-                ktm2p += 4;
-
-                const float* ktm1_0 = kernel1_tm.row(q);
-                const float* ktm1_1 = kernel1_tm.row(q+1);
-                const float* ktm1_2 = kernel1_tm.row(q+2);
-                const float* ktm1_3 = kernel1_tm.row(q+3);
-
-                ktm2p[0] = ktm1_0[r];
-                ktm2p[1] = ktm1_1[r];
-                ktm2p[2] = ktm1_2[r];
-                ktm2p[3] = ktm1_3[r];
-
-                ktm2p += 4;
-
-                const float* ktm2_0 = kernel2_tm.row(q);
-                const float* ktm2_1 = kernel2_tm.row(q+1);
-                const float* ktm2_2 = kernel2_tm.row(q+2);
-                const float* ktm2_3 = kernel2_tm.row(q+3);
-
-                ktm2p[0] = ktm2_0[r];
-                ktm2p[1] = ktm2_1[r];
-                ktm2p[2] = ktm2_2[r];
-                ktm2p[3] = ktm2_3[r];
-
-                ktm2p += 4;
-
-                const float* ktm3_0 = kernel3_tm.row(q);
-                const float* ktm3_1 = kernel3_tm.row(q+1);
-                const float* ktm3_2 = kernel3_tm.row(q+2);
-                const float* ktm3_3 = kernel3_tm.row(q+3);
-
-                ktm2p[0] = ktm3_0[r];
-                ktm2p[1] = ktm3_1[r];
-                ktm2p[2] = ktm3_2[r];
-                ktm2p[3] = ktm3_3[r];
-
-                ktm2p += 4;
-            }
-            for (; q<inch; q++)
+            for (int q=0; q<inch; q++)
             {
                 const float* ktm0_0 = kernel0_tm.row(q);
                 const float* ktm1_0 = kernel1_tm.row(q);
@@ -1825,7 +1675,7 @@ static void conv3x3s1_winograd64_transform_kernel_neon5(const Mat& kernel, Mat&
     {
         const Mat kernel0_tm = kernel_tm.channel(p);
 
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
         Mat ktm2 = kernel_tm2.channel(p/8+(p%8)/4+p%4);
 #else
         Mat ktm2 = kernel_tm2.channel(p/4+p%4);
@@ -1835,22 +1685,7 @@ static void conv3x3s1_winograd64_transform_kernel_neon5(const Mat& kernel, Mat&
         {
             float* ktm2p = ktm2.row(r);
 
-            int q=0;
-            for (; q+3<inch; q+=4)
-            {
-                const float* ktm0_0 = kernel0_tm.row(q);
-                const float* ktm0_1 = kernel0_tm.row(q+1);
-                const float* ktm0_2 = kernel0_tm.row(q+2);
-                const float* ktm0_3 = kernel0_tm.row(q+3);
-
-                ktm2p[0] = ktm0_0[r];
-                ktm2p[1] = ktm0_1[r];
-                ktm2p[2] = ktm0_2[r];
-                ktm2p[3] = ktm0_3[r];
-
-                ktm2p += 4;
-            }
-            for (; q<inch; q++)
+            for (int q=0; q<inch; q++)
             {
                 const float* ktm0_0 = kernel0_tm.row(q);
 
@@ -9236,23 +9071,23 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                 r0 += r*tiles + i;
 
-                int q=0;
-                for (; q<inch; q++)
+                for (int q=0; q<inch; q++)
                 {
-//                     asm volatile("prfm pldl1keep, [%0, #256] \n" : :"r"(r0) :);
+#if __ARM_NEON
                     float32x4_t _r0 = vld1q_f32(r0);
                     float32x4_t _r0n = vld1q_f32(r0+4);
                     vst1q_f32(tm2p, _r0);
                     vst1q_f32(tm2p+4, _r0n);
-
-//                     tm2p[0] = r0[0];
-//                     tm2p[1] = r0[1];
-//                     tm2p[2] = r0[2];
-//                     tm2p[3] = r0[3];
-//                     tm2p[4] = r0[4];
-//                     tm2p[5] = r0[5];
-//                     tm2p[6] = r0[6];
-//                     tm2p[7] = r0[7];
+#else
+                    tm2p[0] = r0[0];
+                    tm2p[1] = r0[1];
+                    tm2p[2] = r0[2];
+                    tm2p[3] = r0[3];
+                    tm2p[4] = r0[4];
+                    tm2p[5] = r0[5];
+                    tm2p[6] = r0[6];
+                    tm2p[7] = r0[7];
+#endif // __ARM_NEON
 
                     r0 += bottom_blob_tm.cstep;
                     tm2p += 8;
@@ -9266,17 +9101,17 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                 r0 += r*tiles + i;
 
-                int q=0;
-                for (; q<inch; q++)
+                for (int q=0; q<inch; q++)
                 {
-//                     asm volatile("prfm pldl1keep, [%0, #128] \n" : :"r"(r0) :);
+#if __ARM_NEON
                     float32x4_t _r0 = vld1q_f32(r0);
                     vst1q_f32(tm2p, _r0);
-
-//                     tm2p[0] = r0[0];
-//                     tm2p[1] = r0[1];
-//                     tm2p[2] = r0[2];
-//                     tm2p[3] = r0[3];
+#else
+                    tm2p[0] = r0[0];
+                    tm2p[1] = r0[1];
+                    tm2p[2] = r0[2];
+                    tm2p[3] = r0[3];
+#endif // __ARM_NEON
 
                     r0 += bottom_blob_tm.cstep;
                     tm2p += 4;
@@ -9290,8 +9125,7 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                 r0 += r*tiles + i;
 
-                int q=0;
-                for (; q<inch; q++)
+                for (int q=0; q<inch; q++)
                 {
                     tm2p[0] = r0[0];
 
@@ -9309,7 +9143,7 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
         int nn_outch = 0;
         int remain_outch_start = 0;
 
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
         nn_outch = outch >> 3;
         remain_outch_start = nn_outch << 3;
 
@@ -9383,82 +9217,81 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                         "fmla   v16.4s, v8.4s, v0.s[0]  \n"
                         "fmla   v17.4s, v9.4s, v0.s[0]  \n"
-                        "fmla   v18.4s, v8.4s, v1.s[0]  \n"
-                        "fmla   v19.4s, v9.4s, v1.s[0]  \n"
+                        "fmla   v18.4s, v8.4s, v0.s[1]  \n"
+                        "fmla   v19.4s, v9.4s, v0.s[1]  \n"
+                        "fmla   v20.4s, v8.4s, v0.s[2]  \n"
+                        "fmla   v21.4s, v9.4s, v0.s[2]  \n"
+                        "fmla   v22.4s, v8.4s, v0.s[3]  \n"
+                        "fmla   v23.4s, v9.4s, v0.s[3]  \n"
 
                         "prfm   pldl1keep, [%9, #512]   \n"
                         "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%9], #64   \n"
 
-                        "fmla   v20.4s, v8.4s, v2.s[0]  \n"
-                        "fmla   v21.4s, v9.4s, v2.s[0]  \n"
-                        "fmla   v22.4s, v8.4s, v3.s[0]  \n"
-                        "fmla   v23.4s, v9.4s, v3.s[0]  \n"
-
-                        "fmla   v24.4s, v8.4s, v4.s[0]  \n"
-                        "fmla   v25.4s, v9.4s, v4.s[0]  \n"
-                        "fmla   v26.4s, v8.4s, v5.s[0]  \n"
-                        "fmla   v27.4s, v9.4s, v5.s[0]  \n"
-                        "fmla   v28.4s, v8.4s, v6.s[0]  \n"
-                        "fmla   v29.4s, v9.4s, v6.s[0]  \n"
-                        "fmla   v30.4s, v8.4s, v7.s[0]  \n"
-                        "fmla   v31.4s, v9.4s, v7.s[0]  \n"
-
-                        "fmla   v16.4s, v10.4s, v0.s[1] \n"
-                        "fmla   v17.4s, v11.4s, v0.s[1] \n"
-                        "fmla   v18.4s, v10.4s, v1.s[1] \n"
-                        "fmla   v19.4s, v11.4s, v1.s[1] \n"
-                        "fmla   v20.4s, v10.4s, v2.s[1] \n"
-                        "fmla   v21.4s, v11.4s, v2.s[1] \n"
-                        "fmla   v22.4s, v10.4s, v3.s[1] \n"
-                        "fmla   v23.4s, v11.4s, v3.s[1] \n"
+                        "fmla   v24.4s, v8.4s, v1.s[0]  \n"
+                        "fmla   v25.4s, v9.4s, v1.s[0]  \n"
+                        "fmla   v26.4s, v8.4s, v1.s[1]  \n"
+                        "fmla   v27.4s, v9.4s, v1.s[1]  \n"
+                        "fmla   v28.4s, v8.4s, v1.s[2]  \n"
+                        "fmla   v29.4s, v9.4s, v1.s[2]  \n"
+                        "fmla   v30.4s, v8.4s, v1.s[3]  \n"
+                        "fmla   v31.4s, v9.4s, v1.s[3]  \n"
+
+                        "fmla   v16.4s, v10.4s, v2.s[0] \n"
+                        "fmla   v17.4s, v11.4s, v2.s[0] \n"
+                        "fmla   v18.4s, v10.4s, v2.s[1] \n"
+                        "fmla   v19.4s, v11.4s, v2.s[1] \n"
+                        "fmla   v20.4s, v10.4s, v2.s[2] \n"
+                        "fmla   v21.4s, v11.4s, v2.s[2] \n"
+                        "fmla   v22.4s, v10.4s, v2.s[3] \n"
+                        "fmla   v23.4s, v11.4s, v2.s[3] \n"
 
                         "prfm   pldl1keep, [%8, #512]   \n"
                         "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%8], #64 \n"
 
-                        "fmla   v24.4s, v10.4s, v4.s[1] \n"
-                        "fmla   v25.4s, v11.4s, v4.s[1] \n"
-                        "fmla   v26.4s, v10.4s, v5.s[1] \n"
-                        "fmla   v27.4s, v11.4s, v5.s[1] \n"
-                        "fmla   v28.4s, v10.4s, v6.s[1] \n"
-                        "fmla   v29.4s, v11.4s, v6.s[1] \n"
-                        "fmla   v30.4s, v10.4s, v7.s[1] \n"
-                        "fmla   v31.4s, v11.4s, v7.s[1] \n"
-
-                        "fmla   v16.4s, v12.4s, v0.s[2] \n"
-                        "fmla   v17.4s, v13.4s, v0.s[2] \n"
-                        "fmla   v18.4s, v12.4s, v1.s[2] \n"
-                        "fmla   v19.4s, v13.4s, v1.s[2] \n"
-                        "fmla   v20.4s, v12.4s, v2.s[2] \n"
-                        "fmla   v21.4s, v13.4s, v2.s[2] \n"
-                        "fmla   v22.4s, v12.4s, v3.s[2] \n"
-                        "fmla   v23.4s, v13.4s, v3.s[2] \n"
-
-                        "fmla   v24.4s, v12.4s, v4.s[2] \n"
-                        "fmla   v25.4s, v13.4s, v4.s[2] \n"
-                        "fmla   v26.4s, v12.4s, v5.s[2] \n"
-                        "fmla   v27.4s, v13.4s, v5.s[2] \n"
-                        "fmla   v28.4s, v12.4s, v6.s[2] \n"
-                        "fmla   v29.4s, v13.4s, v6.s[2] \n"
-                        "fmla   v30.4s, v12.4s, v7.s[2] \n"
-                        "fmla   v31.4s, v13.4s, v7.s[2] \n"
-
-                        "fmla   v16.4s, v14.4s, v0.s[3] \n"
-                        "fmla   v17.4s, v15.4s, v0.s[3] \n"
-                        "fmla   v18.4s, v14.4s, v1.s[3] \n"
-                        "fmla   v19.4s, v15.4s, v1.s[3] \n"
-                        "fmla   v20.4s, v14.4s, v2.s[3] \n"
-                        "fmla   v21.4s, v15.4s, v2.s[3] \n"
-                        "fmla   v22.4s, v14.4s, v3.s[3] \n"
-                        "fmla   v23.4s, v15.4s, v3.s[3] \n"
+                        "fmla   v24.4s, v10.4s, v3.s[0] \n"
+                        "fmla   v25.4s, v11.4s, v3.s[0] \n"
+                        "fmla   v26.4s, v10.4s, v3.s[1] \n"
+                        "fmla   v27.4s, v11.4s, v3.s[1] \n"
+                        "fmla   v28.4s, v10.4s, v3.s[2] \n"
+                        "fmla   v29.4s, v11.4s, v3.s[2] \n"
+                        "fmla   v30.4s, v10.4s, v3.s[3] \n"
+                        "fmla   v31.4s, v11.4s, v3.s[3] \n"
+
+                        "fmla   v16.4s, v12.4s, v4.s[0] \n"
+                        "fmla   v17.4s, v13.4s, v4.s[0] \n"
+                        "fmla   v18.4s, v12.4s, v4.s[1] \n"
+                        "fmla   v19.4s, v13.4s, v4.s[1] \n"
+                        "fmla   v20.4s, v12.4s, v4.s[2] \n"
+                        "fmla   v21.4s, v13.4s, v4.s[2] \n"
+                        "fmla   v22.4s, v12.4s, v4.s[3] \n"
+                        "fmla   v23.4s, v13.4s, v4.s[3] \n"
+
+                        "fmla   v24.4s, v12.4s, v5.s[0] \n"
+                        "fmla   v25.4s, v13.4s, v5.s[0] \n"
+                        "fmla   v26.4s, v12.4s, v5.s[1] \n"
+                        "fmla   v27.4s, v13.4s, v5.s[1] \n"
+                        "fmla   v28.4s, v12.4s, v5.s[2] \n"
+                        "fmla   v29.4s, v13.4s, v5.s[2] \n"
+                        "fmla   v30.4s, v12.4s, v5.s[3] \n"
+                        "fmla   v31.4s, v13.4s, v5.s[3] \n"
+
+                        "fmla   v16.4s, v14.4s, v6.s[0] \n"
+                        "fmla   v17.4s, v15.4s, v6.s[0] \n"
+                        "fmla   v18.4s, v14.4s, v6.s[1] \n"
+                        "fmla   v19.4s, v15.4s, v6.s[1] \n"
+                        "fmla   v20.4s, v14.4s, v6.s[2] \n"
+                        "fmla   v21.4s, v15.4s, v6.s[2] \n"
+                        "fmla   v22.4s, v14.4s, v6.s[3] \n"
+                        "fmla   v23.4s, v15.4s, v6.s[3] \n"
 
                         "subs   w4, w4, #1              \n"
 
-                        "fmla   v24.4s, v14.4s, v4.s[3] \n"
-                        "fmla   v25.4s, v15.4s, v4.s[3] \n"
-                        "fmla   v26.4s, v14.4s, v5.s[3] \n"
-                        "fmla   v27.4s, v15.4s, v5.s[3] \n"
-                        "fmla   v28.4s, v14.4s, v6.s[3] \n"
-                        "fmla   v29.4s, v15.4s, v6.s[3] \n"
+                        "fmla   v24.4s, v14.4s, v7.s[0] \n"
+                        "fmla   v25.4s, v15.4s, v7.s[0] \n"
+                        "fmla   v26.4s, v14.4s, v7.s[1] \n"
+                        "fmla   v27.4s, v15.4s, v7.s[1] \n"
+                        "fmla   v28.4s, v14.4s, v7.s[2] \n"
+                        "fmla   v29.4s, v15.4s, v7.s[2] \n"
                         "fmla   v30.4s, v14.4s, v7.s[3] \n"
                         "fmla   v31.4s, v15.4s, v7.s[3] \n"
 
@@ -9566,45 +9399,44 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                         "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%9], #64   \n"
 
                         "fmla   v16.4s, v8.4s, v0.s[0]  \n"
-                        "fmla   v17.4s, v8.4s, v1.s[0]  \n"
-                        "fmla   v18.4s, v8.4s, v2.s[0]  \n"
-                        "fmla   v19.4s, v8.4s, v3.s[0]  \n"
+                        "fmla   v17.4s, v8.4s, v0.s[1]  \n"
+                        "fmla   v18.4s, v8.4s, v0.s[2]  \n"
+                        "fmla   v19.4s, v8.4s, v0.s[3]  \n"
+                        "fmla   v20.4s, v8.4s, v1.s[0]  \n"
+                        "fmla   v21.4s, v8.4s, v1.s[1]  \n"
+                        "fmla   v22.4s, v8.4s, v1.s[2]  \n"
+                        "fmla   v23.4s, v8.4s, v1.s[3]  \n"
 
                         "prfm   pldl1keep, [%9, #512]   \n"
                         "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%9], #64   \n"
 
-                        "fmla   v20.4s, v8.4s, v4.s[0]  \n"
-                        "fmla   v21.4s, v8.4s, v5.s[0]  \n"
-                        "fmla   v22.4s, v8.4s, v6.s[0]  \n"
-                        "fmla   v23.4s, v8.4s, v7.s[0]  \n"
-
-                        "fmla   v16.4s, v9.4s, v0.s[1]  \n"
-                        "fmla   v17.4s, v9.4s, v1.s[1]  \n"
-                        "fmla   v18.4s, v9.4s, v2.s[1]  \n"
-                        "fmla   v19.4s, v9.4s, v3.s[1]  \n"
-                        "fmla   v20.4s, v9.4s, v4.s[1]  \n"
-                        "fmla   v21.4s, v9.4s, v5.s[1]  \n"
-                        "fmla   v22.4s, v9.4s, v6.s[1]  \n"
-                        "fmla   v23.4s, v9.4s, v7.s[1]  \n"
-
-                        "fmla   v16.4s, v10.4s, v0.s[2] \n"
-                        "fmla   v17.4s, v10.4s, v1.s[2] \n"
-                        "fmla   v18.4s, v10.4s, v2.s[2] \n"
-                        "fmla   v19.4s, v10.4s, v3.s[2] \n"
-                        "fmla   v20.4s, v10.4s, v4.s[2] \n"
-                        "fmla   v21.4s, v10.4s, v5.s[2] \n"
-                        "fmla   v22.4s, v10.4s, v6.s[2] \n"
-                        "fmla   v23.4s, v10.4s, v7.s[2] \n"
+                        "fmla   v16.4s, v9.4s, v2.s[0]  \n"
+                        "fmla   v17.4s, v9.4s, v2.s[1]  \n"
+                        "fmla   v18.4s, v9.4s, v2.s[2]  \n"
+                        "fmla   v19.4s, v9.4s, v2.s[3]  \n"
+                        "fmla   v20.4s, v9.4s, v3.s[0]  \n"
+                        "fmla   v21.4s, v9.4s, v3.s[1]  \n"
+                        "fmla   v22.4s, v9.4s, v3.s[2]  \n"
+                        "fmla   v23.4s, v9.4s, v3.s[3]  \n"
+
+                        "fmla   v16.4s, v10.4s, v4.s[0] \n"
+                        "fmla   v17.4s, v10.4s, v4.s[1] \n"
+                        "fmla   v18.4s, v10.4s, v4.s[2] \n"
+                        "fmla   v19.4s, v10.4s, v4.s[3] \n"
+                        "fmla   v20.4s, v10.4s, v5.s[0] \n"
+                        "fmla   v21.4s, v10.4s, v5.s[1] \n"
+                        "fmla   v22.4s, v10.4s, v5.s[2] \n"
+                        "fmla   v23.4s, v10.4s, v5.s[3] \n"
 
                         "subs   w4, w4, #1              \n"
 
-                        "fmla   v16.4s, v11.4s, v0.s[3] \n"
-                        "fmla   v17.4s, v11.4s, v1.s[3] \n"
-                        "fmla   v18.4s, v11.4s, v2.s[3] \n"
-                        "fmla   v19.4s, v11.4s, v3.s[3] \n"
-                        "fmla   v20.4s, v11.4s, v4.s[3] \n"
-                        "fmla   v21.4s, v11.4s, v5.s[3] \n"
-                        "fmla   v22.4s, v11.4s, v6.s[3] \n"
+                        "fmla   v16.4s, v11.4s, v6.s[0] \n"
+                        "fmla   v17.4s, v11.4s, v6.s[1] \n"
+                        "fmla   v18.4s, v11.4s, v6.s[2] \n"
+                        "fmla   v19.4s, v11.4s, v6.s[3] \n"
+                        "fmla   v20.4s, v11.4s, v7.s[0] \n"
+                        "fmla   v21.4s, v11.4s, v7.s[1] \n"
+                        "fmla   v22.4s, v11.4s, v7.s[2] \n"
                         "fmla   v23.4s, v11.4s, v7.s[3] \n"
 
                         "bne    0b                      \n"
@@ -9679,14 +9511,8 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                     const float* ktm0 = kernel_tm0.row(r);
 
-                    float32x4_t _sum0 = vdupq_n_f32(0.f);
-                    float32x4_t _sum1 = vdupq_n_f32(0.f);
-                    float32x4_t _sum2 = vdupq_n_f32(0.f);
-                    float32x4_t _sum3 = vdupq_n_f32(0.f);
-                    float32x4_t _sum4 = vdupq_n_f32(0.f);
-                    float32x4_t _sum5 = vdupq_n_f32(0.f);
-                    float32x4_t _sum6 = vdupq_n_f32(0.f);
-                    float32x4_t _sum7 = vdupq_n_f32(0.f);
+                    float32x4_t _sum0123 = vdupq_n_f32(0.f);
+                    float32x4_t _sum4567 = vdupq_n_f32(0.f);
 
                     int q=0;
                     for (; q+3<inch; q+=4)
@@ -9702,10 +9528,10 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                         float32x4_t _ktm3 = vld1q_f32(ktm0 + 12);
                         ktm0 += 16;
 
-                        _sum0 = vmlaq_f32(_sum0, _bb2p0, _ktm0);
-                        _sum1 = vmlaq_f32(_sum1, _bb2p0, _ktm1);
-                        _sum2 = vmlaq_f32(_sum2, _bb2p0, _ktm2);
-                        _sum3 = vmlaq_f32(_sum3, _bb2p0, _ktm3);
+                        _sum0123 = vmlaq_laneq_f32(_sum0123, _ktm0, _bb2p0, 0);
+                        _sum4567 = vmlaq_laneq_f32(_sum4567, _ktm1, _bb2p0, 0);
+                        _sum0123 = vmlaq_laneq_f32(_sum0123, _ktm2, _bb2p0, 1);
+                        _sum4567 = vmlaq_laneq_f32(_sum4567, _ktm3, _bb2p0, 1);
 
 //                         asm volatile("prfm pldl1keep, [%0, #512] \n" : :"r"(ktm0) :);
                         float32x4_t _ktm4 = vld1q_f32(ktm0 + 0);
@@ -9714,37 +9540,34 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                         float32x4_t _ktm7 = vld1q_f32(ktm0 + 12);
                         ktm0 += 16;
 
-                        _sum4 = vmlaq_f32(_sum4, _bb2p0, _ktm4);
-                        _sum5 = vmlaq_f32(_sum5, _bb2p0, _ktm5);
-                        _sum6 = vmlaq_f32(_sum6, _bb2p0, _ktm6);
-                        _sum7 = vmlaq_f32(_sum7, _bb2p0, _ktm7);
+                        _sum0123 = vmlaq_laneq_f32(_sum0123, _ktm4, _bb2p0, 2);
+                        _sum4567 = vmlaq_laneq_f32(_sum4567, _ktm5, _bb2p0, 2);
+                        _sum0123 = vmlaq_laneq_f32(_sum0123, _ktm6, _bb2p0, 3);
+                        _sum4567 = vmlaq_laneq_f32(_sum4567, _ktm7, _bb2p0, 3);
                     }
 
-                    // TODO transpose and acc
-                    float sum0 = vaddvq_f32(_sum0);
-                    float sum1 = vaddvq_f32(_sum1);
-                    float sum2 = vaddvq_f32(_sum2);
-                    float sum3 = vaddvq_f32(_sum3);
-                    float sum4 = vaddvq_f32(_sum4);
-                    float sum5 = vaddvq_f32(_sum5);
-                    float sum6 = vaddvq_f32(_sum6);
-                    float sum7 = vaddvq_f32(_sum7);
-
                     for (; q<inch; q++)
                     {
-                        sum0 += bb2p0[0] * ktm0[0];
-                        sum1 += bb2p0[0] * ktm0[1];
-                        sum2 += bb2p0[0] * ktm0[2];
-                        sum3 += bb2p0[0] * ktm0[3];
-                        sum4 += bb2p0[0] * ktm0[4];
-                        sum5 += bb2p0[0] * ktm0[5];
-                        sum6 += bb2p0[0] * ktm0[6];
-                        sum7 += bb2p0[0] * ktm0[7];
+                        float32x4_t _bb2p0 = vld1q_dup_f32(bb2p0);
+                        float32x4_t _ktm0123 = vld1q_f32(ktm0 + 0);
+                        float32x4_t _ktm4567 = vld1q_f32(ktm0 + 4);
+
+                        _sum0123 = vmlaq_f32(_sum0123, _bb2p0, _ktm0123);
+                        _sum4567 = vmlaq_f32(_sum4567, _bb2p0, _ktm4567);
 
                         bb2p0 += 1;
                         ktm0 += 8;
                     }
 
+                    float sum0 = vgetq_lane_f32(_sum0123, 0);
+                    float sum1 = vgetq_lane_f32(_sum0123, 1);
+                    float sum2 = vgetq_lane_f32(_sum0123, 2);
+                    float sum3 = vgetq_lane_f32(_sum0123, 3);
+                    float sum4 = vgetq_lane_f32(_sum4567, 0);
+                    float sum5 = vgetq_lane_f32(_sum4567, 1);
+                    float sum6 = vgetq_lane_f32(_sum4567, 2);
+                    float sum7 = vgetq_lane_f32(_sum4567, 3);
+
                     output0_tm[0] = sum0;
                     output1_tm[0] = sum1;
                     output2_tm[0] = sum2;
@@ -9774,7 +9597,7 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
         {
             int p = remain_outch_start + pp * 4;
 
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
             const Mat kernel_tm0 = kernel_tm.channel(p/8+(p%8)/4);
 #else
             const Mat kernel_tm0 = kernel_tm.channel(p/4);
@@ -9801,6 +9624,7 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                     const float* bb2p0 = bb2.row(i/8);
 
                     const float* ktm0 = kernel_tm0.row(r);
+#if __ARM_NEON
 #if __aarch64__
                     asm volatile(
                         "eor    v8.16b, v8.16b, v8.16b     \n"
@@ -9827,40 +9651,40 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                         "fmla   v8.4s, v4.4s, v0.s[0]   \n"
                         "fmla   v9.4s, v5.4s, v0.s[0]   \n"
-                        "fmla   v10.4s, v4.4s, v1.s[0]  \n"
-                        "fmla   v11.4s, v5.4s, v1.s[0]  \n"
-                        "fmla   v12.4s, v4.4s, v2.s[0]  \n"
-                        "fmla   v13.4s, v5.4s, v2.s[0]  \n"
-                        "fmla   v14.4s, v4.4s, v3.s[0]  \n"
-                        "fmla   v15.4s, v5.4s, v3.s[0]  \n"
-
-                        "fmla   v8.4s, v6.4s, v0.s[1]   \n"
-                        "fmla   v9.4s, v7.4s, v0.s[1]   \n"
-                        "fmla   v10.4s, v6.4s, v1.s[1]  \n"
-                        "fmla   v11.4s, v7.4s, v1.s[1]  \n"
-                        "fmla   v12.4s, v6.4s, v2.s[1]  \n"
-                        "fmla   v13.4s, v7.4s, v2.s[1]  \n"
-                        "fmla   v14.4s, v6.4s, v3.s[1]  \n"
-                        "fmla   v15.4s, v7.4s, v3.s[1]  \n"
+                        "fmla   v10.4s, v4.4s, v0.s[1]  \n"
+                        "fmla   v11.4s, v5.4s, v0.s[1]  \n"
+                        "fmla   v12.4s, v4.4s, v0.s[2]  \n"
+                        "fmla   v13.4s, v5.4s, v0.s[2]  \n"
+                        "fmla   v14.4s, v4.4s, v0.s[3]  \n"
+                        "fmla   v15.4s, v5.4s, v0.s[3]  \n"
 
                         "prfm   pldl1keep, [%4, #512]   \n"
                         "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
 
-                        "fmla   v8.4s, v16.4s, v0.s[2]  \n"
-                        "fmla   v9.4s, v17.4s, v0.s[2]  \n"
-                        "fmla   v10.4s, v16.4s, v1.s[2] \n"
-                        "fmla   v11.4s, v17.4s, v1.s[2] \n"
+                        "fmla   v8.4s, v6.4s, v1.s[0]   \n"
+                        "fmla   v9.4s, v7.4s, v1.s[0]   \n"
+                        "fmla   v10.4s, v6.4s, v1.s[1]  \n"
+                        "fmla   v11.4s, v7.4s, v1.s[1]  \n"
+                        "fmla   v12.4s, v6.4s, v1.s[2]  \n"
+                        "fmla   v13.4s, v7.4s, v1.s[2]  \n"
+                        "fmla   v14.4s, v6.4s, v1.s[3]  \n"
+                        "fmla   v15.4s, v7.4s, v1.s[3]  \n"
+
+                        "fmla   v8.4s, v16.4s, v2.s[0]  \n"
+                        "fmla   v9.4s, v17.4s, v2.s[0]  \n"
+                        "fmla   v10.4s, v16.4s, v2.s[1] \n"
+                        "fmla   v11.4s, v17.4s, v2.s[1] \n"
                         "fmla   v12.4s, v16.4s, v2.s[2] \n"
                         "fmla   v13.4s, v17.4s, v2.s[2] \n"
-                        "fmla   v14.4s, v16.4s, v3.s[2] \n"
-                        "fmla   v15.4s, v17.4s, v3.s[2] \n"
-
-                        "fmla   v8.4s, v18.4s, v0.s[3]  \n"
-                        "fmla   v9.4s, v19.4s, v0.s[3]  \n"
-                        "fmla   v10.4s, v18.4s, v1.s[3] \n"
-                        "fmla   v11.4s, v19.4s, v1.s[3] \n"
-                        "fmla   v12.4s, v18.4s, v2.s[3] \n"
-                        "fmla   v13.4s, v19.4s, v2.s[3] \n"
+                        "fmla   v14.4s, v16.4s, v2.s[3] \n"
+                        "fmla   v15.4s, v17.4s, v2.s[3] \n"
+
+                        "fmla   v8.4s, v18.4s, v3.s[0]  \n"
+                        "fmla   v9.4s, v19.4s, v3.s[0]  \n"
+                        "fmla   v10.4s, v18.4s, v3.s[1] \n"
+                        "fmla   v11.4s, v19.4s, v3.s[1] \n"
+                        "fmla   v12.4s, v18.4s, v3.s[2] \n"
+                        "fmla   v13.4s, v19.4s, v3.s[2] \n"
                         "fmla   v14.4s, v18.4s, v3.s[3] \n"
                         "fmla   v15.4s, v19.4s, v3.s[3] \n"
 
@@ -9946,42 +9770,42 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                         "vmla.f32   q8, q4, d0[0]   \n"
                         "vmla.f32   q9, q5, d0[0]   \n"
-                        "vmla.f32   q10, q4, d2[0]  \n"
-                        "vmla.f32   q11, q5, d2[0]  \n"
-                        "vmla.f32   q12, q4, d4[0]  \n"
-                        "vmla.f32   q13, q5, d4[0]  \n"
-                        "vmla.f32   q14, q4, d6[0]  \n"
-                        "vmla.f32   q15, q5, d6[0]  \n"
-
-                        "vmla.f32   q8, q6, d0[1]   \n"
-                        "vmla.f32   q9, q7, d0[1]   \n"
+                        "vmla.f32   q10, q4, d0[1]  \n"
+                        "vmla.f32   q11, q5, d0[1]  \n"
+                        "vmla.f32   q12, q4, d1[0]  \n"
+                        "vmla.f32   q13, q5, d1[0]  \n"
+                        "vmla.f32   q14, q4, d1[1]  \n"
+                        "vmla.f32   q15, q5, d1[1]  \n"
+
+                        "vmla.f32   q8, q6, d2[0]   \n"
+                        "vmla.f32   q9, q7, d2[0]   \n"
                         "vmla.f32   q10, q6, d2[1]  \n"
                         "vmla.f32   q11, q7, d2[1]  \n"
-                        "vmla.f32   q12, q6, d4[1]  \n"
-                        "vmla.f32   q13, q7, d4[1]  \n"
-                        "vmla.f32   q14, q6, d6[1]  \n"
-                        "vmla.f32   q15, q7, d6[1]  \n"
+                        "vmla.f32   q12, q6, d3[0]  \n"
+                        "vmla.f32   q13, q7, d3[0]  \n"
+                        "vmla.f32   q14, q6, d3[1]  \n"
+                        "vmla.f32   q15, q7, d3[1]  \n"
 
                         "pld        [%4, #512]      \n"
 //                         "vldm       %4!, {d8-d15}   \n"
                         "vld1.f32   {d8-d11}, [%4 :128]! \n"
                         "vld1.f32   {d12-d15}, [%4 :128]! \n"
 
-                        "vmla.f32   q8, q4, d1[0]   \n"
-                        "vmla.f32   q9, q5, d1[0]   \n"
-                        "vmla.f32   q10, q4, d3[0]  \n"
-                        "vmla.f32   q11, q5, d3[0]  \n"
+                        "vmla.f32   q8, q4, d4[0]   \n"
+                        "vmla.f32   q9, q5, d4[0]   \n"
+                        "vmla.f32   q10, q4, d4[1]  \n"
+                        "vmla.f32   q11, q5, d4[1]  \n"
                         "vmla.f32   q12, q4, d5[0]  \n"
                         "vmla.f32   q13, q5, d5[0]  \n"
-                        "vmla.f32   q14, q4, d7[0]  \n"
-                        "vmla.f32   q15, q5, d7[0]  \n"
-
-                        "vmla.f32   q8, q6, d1[1]   \n"
-                        "vmla.f32   q9, q7, d1[1]   \n"
-                        "vmla.f32   q10, q6, d3[1]  \n"
-                        "vmla.f32   q11, q7, d3[1]  \n"
-                        "vmla.f32   q12, q6, d5[1]  \n"
-                        "vmla.f32   q13, q7, d5[1]  \n"
+                        "vmla.f32   q14, q4, d5[1]  \n"
+                        "vmla.f32   q15, q5, d5[1]  \n"
+
+                        "vmla.f32   q8, q6, d6[0]   \n"
+                        "vmla.f32   q9, q7, d6[0]   \n"
+                        "vmla.f32   q10, q6, d6[1]  \n"
+                        "vmla.f32   q11, q7, d6[1]  \n"
+                        "vmla.f32   q12, q6, d7[0]  \n"
+                        "vmla.f32   q13, q7, d7[0]  \n"
                         "vmla.f32   q14, q6, d7[1]  \n"
                         "vmla.f32   q15, q7, d7[1]  \n"
 
@@ -10038,12 +9862,133 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                         : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
                     );
 #endif // __aarch64__
+#else
+                    float sum0_0 = 0.f;
+                    float sum0_1 = 0.f;
+                    float sum0_2 = 0.f;
+                    float sum0_3 = 0.f;
+                    float sum0_4 = 0.f;
+                    float sum0_5 = 0.f;
+                    float sum0_6 = 0.f;
+                    float sum0_7 = 0.f;
+
+                    float sum1_0 = 0.f;
+                    float sum1_1 = 0.f;
+                    float sum1_2 = 0.f;
+                    float sum1_3 = 0.f;
+                    float sum1_4 = 0.f;
+                    float sum1_5 = 0.f;
+                    float sum1_6 = 0.f;
+                    float sum1_7 = 0.f;
+
+                    float sum2_0 = 0.f;
+                    float sum2_1 = 0.f;
+                    float sum2_2 = 0.f;
+                    float sum2_3 = 0.f;
+                    float sum2_4 = 0.f;
+                    float sum2_5 = 0.f;
+                    float sum2_6 = 0.f;
+                    float sum2_7 = 0.f;
+
+                    float sum3_0 = 0.f;
+                    float sum3_1 = 0.f;
+                    float sum3_2 = 0.f;
+                    float sum3_3 = 0.f;
+                    float sum3_4 = 0.f;
+                    float sum3_5 = 0.f;
+                    float sum3_6 = 0.f;
+                    float sum3_7 = 0.f;
+
+                    for (int q=0; q<inch; q++)
+                    {
+                        sum0_0 += bb2p0[0] * ktm0[0];
+                        sum0_1 += bb2p0[1] * ktm0[0];
+                        sum0_2 += bb2p0[2] * ktm0[0];
+                        sum0_3 += bb2p0[3] * ktm0[0];
+                        sum0_4 += bb2p0[4] * ktm0[0];
+                        sum0_5 += bb2p0[5] * ktm0[0];
+                        sum0_6 += bb2p0[6] * ktm0[0];
+                        sum0_7 += bb2p0[7] * ktm0[0];
+
+                        sum1_0 += bb2p0[0] * ktm0[1];
+                        sum1_1 += bb2p0[1] * ktm0[1];
+                        sum1_2 += bb2p0[2] * ktm0[1];
+                        sum1_3 += bb2p0[3] * ktm0[1];
+                        sum1_4 += bb2p0[4] * ktm0[1];
+                        sum1_5 += bb2p0[5] * ktm0[1];
+                        sum1_6 += bb2p0[6] * ktm0[1];
+                        sum1_7 += bb2p0[7] * ktm0[1];
+
+                        sum2_0 += bb2p0[0] * ktm0[2];
+                        sum2_1 += bb2p0[1] * ktm0[2];
+                        sum2_2 += bb2p0[2] * ktm0[2];
+                        sum2_3 += bb2p0[3] * ktm0[2];
+                        sum2_4 += bb2p0[4] * ktm0[2];
+                        sum2_5 += bb2p0[5] * ktm0[2];
+                        sum2_6 += bb2p0[6] * ktm0[2];
+                        sum2_7 += bb2p0[7] * ktm0[2];
+
+                        sum3_0 += bb2p0[0] * ktm0[3];
+                        sum3_1 += bb2p0[1] * ktm0[3];
+                        sum3_2 += bb2p0[2] * ktm0[3];
+                        sum3_3 += bb2p0[3] * ktm0[3];
+                        sum3_4 += bb2p0[4] * ktm0[3];
+                        sum3_5 += bb2p0[5] * ktm0[3];
+                        sum3_6 += bb2p0[6] * ktm0[3];
+                        sum3_7 += bb2p0[7] * ktm0[3];
+
+                        bb2p0 += 8;
+                        ktm0 += 4;
+                    }
+
+                    output0_tm[0] = sum0_0;
+                    output0_tm[1] = sum0_1;
+                    output0_tm[2] = sum0_2;
+                    output0_tm[3] = sum0_3;
+                    output0_tm[4] = sum0_4;
+                    output0_tm[5] = sum0_5;
+                    output0_tm[6] = sum0_6;
+                    output0_tm[7] = sum0_7;
+
+                    output1_tm[0] = sum1_0;
+                    output1_tm[1] = sum1_1;
+                    output1_tm[2] = sum1_2;
+                    output1_tm[3] = sum1_3;
+                    output1_tm[4] = sum1_4;
+                    output1_tm[5] = sum1_5;
+                    output1_tm[6] = sum1_6;
+                    output1_tm[7] = sum1_7;
+
+                    output2_tm[0] = sum2_0;
+                    output2_tm[1] = sum2_1;
+                    output2_tm[2] = sum2_2;
+                    output2_tm[3] = sum2_3;
+                    output2_tm[4] = sum2_4;
+                    output2_tm[5] = sum2_5;
+                    output2_tm[6] = sum2_6;
+                    output2_tm[7] = sum2_7;
+
+                    output3_tm[0] = sum3_0;
+                    output3_tm[1] = sum3_1;
+                    output3_tm[2] = sum3_2;
+                    output3_tm[3] = sum3_3;
+                    output3_tm[4] = sum3_4;
+                    output3_tm[5] = sum3_5;
+                    output3_tm[6] = sum3_6;
+                    output3_tm[7] = sum3_7;
+
+                    output0_tm += 8;
+                    output1_tm += 8;
+                    output2_tm += 8;
+                    output3_tm += 8;
+#endif // __ARM_NEON
                 }
                 for (; i+3<tiles; i+=4)
                 {
                     const float* bb2p0 = bb2.row(i/8+(i%8)/4);
 
                     const float* ktm0 = kernel_tm0.row(r);
+#if __ARM_NEON
 #if __aarch64__
                     asm volatile(
                         "eor    v8.16b, v8.16b, v8.16b     \n"
@@ -10065,23 +10010,23 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                         "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%5], #64     \n"
 
                         "fmla   v8.4s, v4.4s, v0.s[0]   \n"
-                        "fmla   v9.4s, v4.4s, v1.s[0]   \n"
-                        "fmla   v10.4s, v4.4s, v2.s[0]  \n"
-                        "fmla   v11.4s, v4.4s, v3.s[0]  \n"
+                        "fmla   v9.4s, v4.4s, v0.s[1]   \n"
+                        "fmla   v10.4s, v4.4s, v0.s[2]  \n"
+                        "fmla   v11.4s, v4.4s, v0.s[3]  \n"
 
-                        "fmla   v8.4s, v5.4s, v0.s[1]   \n"
+                        "fmla   v8.4s, v5.4s, v1.s[0]   \n"
                         "fmla   v9.4s, v5.4s, v1.s[1]   \n"
-                        "fmla   v10.4s, v5.4s, v2.s[1]  \n"
-                        "fmla   v11.4s, v5.4s, v3.s[1]  \n"
+                        "fmla   v10.4s, v5.4s, v1.s[2]  \n"
+                        "fmla   v11.4s, v5.4s, v1.s[3]  \n"
 
-                        "fmla   v8.4s, v6.4s, v0.s[2]   \n"
-                        "fmla   v9.4s, v6.4s, v1.s[2]   \n"
+                        "fmla   v8.4s, v6.4s, v2.s[0]   \n"
+                        "fmla   v9.4s, v6.4s, v2.s[1]   \n"
                         "fmla   v10.4s, v6.4s, v2.s[2]  \n"
-                        "fmla   v11.4s, v6.4s, v3.s[2]  \n"
+                        "fmla   v11.4s, v6.4s, v2.s[3]  \n"
 
-                        "fmla   v8.4s, v7.4s, v0.s[3]   \n"
-                        "fmla   v9.4s, v7.4s, v1.s[3]   \n"
-                        "fmla   v10.4s, v7.4s, v2.s[3]  \n"
+                        "fmla   v8.4s, v7.4s, v3.s[0]   \n"
+                        "fmla   v9.4s, v7.4s, v3.s[1]   \n"
+                        "fmla   v10.4s, v7.4s, v3.s[2]  \n"
                         "fmla   v11.4s, v7.4s, v3.s[3]  \n"
 
                         "subs   w4, w4, #1              \n"
@@ -10157,23 +10102,23 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                         "vld1.f32   {d4-d7}, [%5 :128]!  \n"
 
                         "vmla.f32   q8, q4, d0[0]   \n"
-                        "vmla.f32   q9, q4, d2[0]   \n"
-                        "vmla.f32   q10, q4, d4[0]  \n"
-                        "vmla.f32   q11, q4, d6[0]  \n"
+                        "vmla.f32   q9, q4, d0[1]   \n"
+                        "vmla.f32   q10, q4, d1[0]  \n"
+                        "vmla.f32   q11, q4, d1[1]  \n"
 
-                        "vmla.f32   q8, q5, d0[1]   \n"
+                        "vmla.f32   q8, q5, d2[0]   \n"
                         "vmla.f32   q9, q5, d2[1]   \n"
-                        "vmla.f32   q10, q5, d4[1]  \n"
-                        "vmla.f32   q11, q5, d6[1]  \n"
+                        "vmla.f32   q10, q5, d3[0]  \n"
+                        "vmla.f32   q11, q5, d3[1]  \n"
 
-                        "vmla.f32   q8, q6, d1[0]   \n"
-                        "vmla.f32   q9, q6, d3[0]   \n"
+                        "vmla.f32   q8, q6, d4[0]   \n"
+                        "vmla.f32   q9, q6, d4[1]   \n"
                         "vmla.f32   q10, q6, d5[0]  \n"
-                        "vmla.f32   q11, q6, d7[0]  \n"
+                        "vmla.f32   q11, q6, d5[1]  \n"
 
-                        "vmla.f32   q8, q7, d1[1]   \n"
-                        "vmla.f32   q9, q7, d3[1]   \n"
-                        "vmla.f32   q10, q7, d5[1]  \n"
+                        "vmla.f32   q8, q7, d6[0]   \n"
+                        "vmla.f32   q9, q7, d6[1]   \n"
+                        "vmla.f32   q10, q7, d7[0]  \n"
                         "vmla.f32   q11, q7, d7[1]  \n"
 
                         "subs       r4, r4, #1      \n"
@@ -10225,6 +10170,78 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                         : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11"
                     );
 #endif // __aarch64__
+#else
+                    float sum0_0 = 0.f;
+                    float sum0_1 = 0.f;
+                    float sum0_2 = 0.f;
+                    float sum0_3 = 0.f;
+
+                    float sum1_0 = 0.f;
+                    float sum1_1 = 0.f;
+                    float sum1_2 = 0.f;
+                    float sum1_3 = 0.f;
+
+                    float sum2_0 = 0.f;
+                    float sum2_1 = 0.f;
+                    float sum2_2 = 0.f;
+                    float sum2_3 = 0.f;
+
+                    float sum3_0 = 0.f;
+                    float sum3_1 = 0.f;
+                    float sum3_2 = 0.f;
+                    float sum3_3 = 0.f;
+
+                    for (int q=0; q<inch; q++)
+                    {
+                        sum0_0 += bb2p0[0] * ktm0[0];
+                        sum0_1 += bb2p0[1] * ktm0[0];
+                        sum0_2 += bb2p0[2] * ktm0[0];
+                        sum0_3 += bb2p0[3] * ktm0[0];
+
+                        sum1_0 += bb2p0[0] * ktm0[1];
+                        sum1_1 += bb2p0[1] * ktm0[1];
+                        sum1_2 += bb2p0[2] * ktm0[1];
+                        sum1_3 += bb2p0[3] * ktm0[1];
+
+                        sum2_0 += bb2p0[0] * ktm0[2];
+                        sum2_1 += bb2p0[1] * ktm0[2];
+                        sum2_2 += bb2p0[2] * ktm0[2];
+                        sum2_3 += bb2p0[3] * ktm0[2];
+
+                        sum3_0 += bb2p0[0] * ktm0[3];
+                        sum3_1 += bb2p0[1] * ktm0[3];
+                        sum3_2 += bb2p0[2] * ktm0[3];
+                        sum3_3 += bb2p0[3] * ktm0[3];
+
+                        bb2p0 += 4;
+                        ktm0 += 4;
+                    }
+
+                    output0_tm[0] = sum0_0;
+                    output0_tm[1] = sum0_1;
+                    output0_tm[2] = sum0_2;
+                    output0_tm[3] = sum0_3;
+
+                    output1_tm[0] = sum1_0;
+                    output1_tm[1] = sum1_1;
+                    output1_tm[2] = sum1_2;
+                    output1_tm[3] = sum1_3;
+
+                    output2_tm[0] = sum2_0;
+                    output2_tm[1] = sum2_1;
+                    output2_tm[2] = sum2_2;
+                    output2_tm[3] = sum2_3;
+
+                    output3_tm[0] = sum3_0;
+                    output3_tm[1] = sum3_1;
+                    output3_tm[2] = sum3_2;
+                    output3_tm[3] = sum3_3;
+
+                    output0_tm += 4;
+                    output1_tm += 4;
+                    output2_tm += 4;
+                    output3_tm += 4;
+#endif // __ARM_NEON
                 }
                 for (; i<tiles; i++)
                 {
@@ -10232,10 +10249,8 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                     const float* ktm0 = kernel_tm0.row(r);
 
-                    float32x4_t _sum0 = vdupq_n_f32(0.f);
-                    float32x4_t _sum1 = vdupq_n_f32(0.f);
-                    float32x4_t _sum2 = vdupq_n_f32(0.f);
-                    float32x4_t _sum3 = vdupq_n_f32(0.f);
+#if __ARM_NEON
+                    float32x4_t _sum0123 = vdupq_n_f32(0.f);
 
                     int q=0;
                     for (; q+3<inch; q+=4)
@@ -10251,30 +10266,41 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                         float32x4_t _ktm3 = vld1q_f32(ktm0 + 12);
                         ktm0 += 16;
 
-                        _sum0 = vmlaq_f32(_sum0, _bb2p0, _ktm0);
-                        _sum1 = vmlaq_f32(_sum1, _bb2p0, _ktm1);
-                        _sum2 = vmlaq_f32(_sum2, _bb2p0, _ktm2);
-                        _sum3 = vmlaq_f32(_sum3, _bb2p0, _ktm3);
-                    }
-
-                    // TODO transpose and acc
 #if __aarch64__
-                    float sum0 = vaddvq_f32(_sum0);
-                    float sum1 = vaddvq_f32(_sum1);
-                    float sum2 = vaddvq_f32(_sum2);
-                    float sum3 = vaddvq_f32(_sum3);
+                        _sum0123 = vmlaq_laneq_f32(_sum0123, _ktm0, _bb2p0, 0);
+                        _sum0123 = vmlaq_laneq_f32(_sum0123, _ktm1, _bb2p0, 1);
+                        _sum0123 = vmlaq_laneq_f32(_sum0123, _ktm2, _bb2p0, 2);
+                        _sum0123 = vmlaq_laneq_f32(_sum0123, _ktm3, _bb2p0, 3);
 #else
-                    float32x2_t _ss0 = vadd_f32(vget_low_f32(_sum0), vget_high_f32(_sum0));
-                    float32x2_t _ss1 = vadd_f32(vget_low_f32(_sum1), vget_high_f32(_sum1));
-                    float32x2_t _ss2 = vadd_f32(vget_low_f32(_sum2), vget_high_f32(_sum2));
-                    float32x2_t _ss3 = vadd_f32(vget_low_f32(_sum3), vget_high_f32(_sum3));
-                    float sum0 = vget_lane_f32(vpadd_f32(_ss0, _ss0), 0);
-                    float sum1 = vget_lane_f32(vpadd_f32(_ss1, _ss1), 0);
-                    float sum2 = vget_lane_f32(vpadd_f32(_ss2, _ss2), 0);
-                    float sum3 = vget_lane_f32(vpadd_f32(_ss3, _ss3), 0);
+                        _sum0123 = vmlaq_lane_f32(_sum0123, _ktm0, vget_low_f32(_bb2p0), 0);
+                        _sum0123 = vmlaq_lane_f32(_sum0123, _ktm1, vget_low_f32(_bb2p0), 1);
+                        _sum0123 = vmlaq_lane_f32(_sum0123, _ktm2, vget_high_f32(_bb2p0), 0);
+                        _sum0123 = vmlaq_lane_f32(_sum0123, _ktm3, vget_high_f32(_bb2p0), 1);
 #endif // __aarch64__
+                    }
 
                     for (; q<inch; q++)
+                    {
+                        float32x4_t _bb2p0 = vld1q_dup_f32(bb2p0);
+                        float32x4_t _ktm0 = vld1q_f32(ktm0);
+
+                        _sum0123 = vmlaq_f32(_sum0123, _bb2p0, _ktm0);
+
+                        bb2p0 += 1;
+                        ktm0 += 4;
+                    }
+
+                    float sum0 = vgetq_lane_f32(_sum0123, 0);
+                    float sum1 = vgetq_lane_f32(_sum0123, 1);
+                    float sum2 = vgetq_lane_f32(_sum0123, 2);
+                    float sum3 = vgetq_lane_f32(_sum0123, 3);
+#else
+                    float sum0 = 0.f;
+                    float sum1 = 0.f;
+                    float sum2 = 0.f;
+                    float sum3 = 0.f;
+
+                    for (int q=0; q<inch; q++)
                     {
                         sum0 += bb2p0[0] * ktm0[0];
                         sum1 += bb2p0[0] * ktm0[1];
@@ -10284,6 +10310,7 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                         bb2p0 += 1;
                         ktm0 += 4;
                     }
+#endif // __ARM_NEON
 
                     output0_tm[0] = sum0;
                     output1_tm[0] = sum1;
@@ -10302,7 +10329,7 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
         for (int p=remain_outch_start; p<outch; p++)
         {
-#if __aarch64__
+#if __ARM_NEON && __aarch64__
             const Mat kernel_tm0 = kernel_tm.channel(p/8+(p%8)/4+p%4);
 #else
             const Mat kernel_tm0 = kernel_tm.channel(p/4+p%4);
@@ -10323,6 +10350,7 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                     const float* bb2p0 = bb2.row(i/8);
 
                     const float* ktm0 = kernel_tm0.row(r);
+#if __ARM_NEON
 #if __aarch64__
                     asm volatile(
                         "eor    v8.16b, v8.16b, v8.16b     \n"
@@ -10335,19 +10363,19 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                         "0:                             \n"
 
-                        "prfm   pldl1keep, [%4, #512]   \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%4], #64     \n"
+                        "prfm   pldl1keep, [%1, #512]   \n"
+                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64     \n"
 
-                        "prfm   pldl1keep, [%5, #128]   \n"
-                        "ld1    {v0.4s}, [%5], #16      \n"
+                        "prfm   pldl1keep, [%2, #128]   \n"
+                        "ld1    {v0.4s}, [%2], #16      \n"
 
                         "fmla   v8.4s, v4.4s, v0.s[0]   \n"
                         "fmla   v9.4s, v5.4s, v0.s[0]   \n"
                         "fmla   v8.4s, v6.4s, v0.s[1]   \n"
                         "fmla   v9.4s, v7.4s, v0.s[1]   \n"
 
-                        "prfm   pldl1keep, [%4, #512]   \n"
-                        "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%4], #64 \n"
+                        "prfm   pldl1keep, [%1, #512]   \n"
+                        "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%1], #64 \n"
 
                         "fmla   v8.4s, v12.4s, v0.s[2]  \n"
                         "fmla   v9.4s, v13.4s, v0.s[2]  \n"
@@ -10366,11 +10394,11 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                         "2:                             \n"
 
-                        "prfm   pldl1keep, [%4, #256]   \n"
-                        "ld1    {v4.4s, v5.4s}, [%4], #32      \n"
+                        "prfm   pldl1keep, [%1, #256]   \n"
+                        "ld1    {v4.4s, v5.4s}, [%1], #32      \n"
 
-                        "prfm   pldl1keep, [%5, #32]    \n"
-                        "ld1r   {v0.4s}, [%5], #4       \n"
+                        "prfm   pldl1keep, [%2, #32]    \n"
+                        "ld1r   {v0.4s}, [%2], #4       \n"
 
                         "fmla   v8.4s, v4.4s, v0.4s     \n"
                         "fmla   v9.4s, v5.4s, v0.4s     \n"
@@ -10403,23 +10431,23 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                         "0:                             \n"
 
-                        "pld        [%4, #512]          \n"
-//                         "vldm       %4!, {d8-d15}       \n"
-                        "vld1.f32   {d8-d11}, [%4 :128]! \n"
-                        "vld1.f32   {d12-d15}, [%4 :128]! \n"
+                        "pld        [%1, #512]          \n"
+//                         "vldm       %1!, {d8-d15}       \n"
+                        "vld1.f32   {d8-d11}, [%1 :128]! \n"
+                        "vld1.f32   {d12-d15}, [%1 :128]! \n"
 
-                        "pld        [%5, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%5 :128]! \n"
+                        "pld        [%2, #128]          \n"
+                        "vld1.f32   {d0-d1}, [%2 :128]! \n"
 
                         "vmla.f32   q8, q4, d0[0]       \n"
                         "vmla.f32   q9, q5, d0[0]       \n"
                         "vmla.f32   q8, q6, d0[1]       \n"
                         "vmla.f32   q9, q7, d0[1]       \n"
 
-                        "pld        [%4, #512]          \n"
-//                         "vldm       %4!, {d24-d31}      \n"
-                        "vld1.f32   {d24-d27}, [%4 :128]! \n"
-                        "vld1.f32   {d28-d31}, [%4 :128]! \n"
+                        "pld        [%1, #512]          \n"
+//                         "vldm       %1!, {d24-d31}      \n"
+                        "vld1.f32   {d24-d27}, [%1 :128]! \n"
+                        "vld1.f32   {d28-d31}, [%1 :128]! \n"
 
                         "vmla.f32   q8, q12, d1[0]      \n"
                         "vmla.f32   q9, q13, d1[0]      \n"
@@ -10438,11 +10466,11 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                         "2:                             \n"
 
-                        "pld        [%4, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%4 :128]! \n"
+                        "pld        [%1, #256]          \n"
+                        "vld1.f32   {d8-d11}, [%1 :128]! \n"
 
-                        "pld        [%5, #32]           \n"
-                        "vld1.f32   {d0[],d1[]}, [%5]!  \n"
+                        "pld        [%2, #32]           \n"
+                        "vld1.f32   {d0[],d1[]}, [%2]!  \n"
 
                         "vmla.f32   q8, q4, q0          \n"
                         "vmla.f32   q9, q5, q0          \n"
@@ -10464,12 +10492,49 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                         : "cc", "memory", "r4", "q0", "q4", "q5", "q6", "q7", "q8", "q9", "q12", "q13", "q14", "q15"
                     );
 #endif // __aarch64__
+#else
+                    float sum0 = 0.f;
+                    float sum1 = 0.f;
+                    float sum2 = 0.f;
+                    float sum3 = 0.f;
+                    float sum4 = 0.f;
+                    float sum5 = 0.f;
+                    float sum6 = 0.f;
+                    float sum7 = 0.f;
+
+                    for (int q=0; q<inch; q++)
+                    {
+                        sum0 += bb2p0[0] * ktm0[0];
+                        sum1 += bb2p0[1] * ktm0[0];
+                        sum2 += bb2p0[2] * ktm0[0];
+                        sum3 += bb2p0[3] * ktm0[0];
+                        sum4 += bb2p0[4] * ktm0[0];
+                        sum5 += bb2p0[5] * ktm0[0];
+                        sum6 += bb2p0[6] * ktm0[0];
+                        sum7 += bb2p0[7] * ktm0[0];
+
+                        bb2p0 += 8;
+                        ktm0 += 1;
+                    }
+
+                    output0_tm[0] = sum0;
+                    output0_tm[1] = sum1;
+                    output0_tm[2] = sum2;
+                    output0_tm[3] = sum3;
+                    output0_tm[4] = sum4;
+                    output0_tm[5] = sum5;
+                    output0_tm[6] = sum6;
+                    output0_tm[7] = sum7;
+
+                    output0_tm += 8;
+#endif // __ARM_NEON
                 }
                 for (; i+3<tiles; i+=4)
                 {
                     const float* bb2p0 = bb2.row(i/8+(i%8)/4);
 
                     const float* ktm0 = kernel_tm0.row(r);
+#if __ARM_NEON
 #if __aarch64__
                     asm volatile(
                         "eor    v8.16b, v8.16b, v8.16b     \n"
@@ -10589,6 +10654,30 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                         : "cc", "memory", "r4", "q0", "q4", "q5", "q6", "q7", "q8"
                     );
 #endif // __aarch64__
+#else
+                    float sum0 = 0.f;
+                    float sum1 = 0.f;
+                    float sum2 = 0.f;
+                    float sum3 = 0.f;
+
+                    for (int q=0; q<inch; q++)
+                    {
+                        sum0 += bb2p0[0] * ktm0[0];
+                        sum1 += bb2p0[1] * ktm0[0];
+                        sum2 += bb2p0[2] * ktm0[0];
+                        sum3 += bb2p0[3] * ktm0[0];
+
+                        bb2p0 += 4;
+                        ktm0 += 1;
+                    }
+
+                    output0_tm[0] = sum0;
+                    output0_tm[1] = sum1;
+                    output0_tm[2] = sum2;
+                    output0_tm[3] = sum3;
+
+                    output0_tm += 4;
+#endif // __ARM_NEON
                 }
                 for (; i<tiles; i++)
                 {
@@ -10596,9 +10685,9 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
 
                     const float* ktm0 = kernel_tm0.row(r);
 
-                    float32x4_t _sum0 = vdupq_n_f32(0.f);
-
                     int q=0;
+#if __ARM_NEON
+                    float32x4_t _sum0 = vdupq_n_f32(0.f);
                     for (; q+3<inch; q+=4)
                     {
 //                         asm volatile("prfm pldl1keep, [%0, #128] \n" : :"r"(bb2p0) :);
@@ -10617,7 +10706,9 @@ static void conv3x3s1_winograd64_neon5(const Mat& bottom_blob, Mat& top_blob, co
                     float32x2_t _ss0 = vadd_f32(vget_low_f32(_sum0), vget_high_f32(_sum0));
                     float sum0 = vget_lane_f32(vpadd_f32(_ss0, _ss0), 0);
 #endif // __aarch64__
-
+#else
+                    float sum0 = 0.f;
+#endif
                     for (; q<inch; q++)
                     {
                         sum0 += bb2p0[0] * ktm0[0];