diff --git a/build/android/app/src/main/jni/src/armv8/arm64.c b/build/android/app/src/main/jni/src/armv8/arm64.c
index de0cffb..9b422e1 100644
--- a/build/android/app/src/main/jni/src/armv8/arm64.c
+++ b/build/android/app/src/main/jni/src/armv8/arm64.c
@@ -166,13 +166,12 @@ void uavs3e_funs_init_arm64()
     uavs3e_funs_handle.cost_ssd[5] = uavs3e_get_ssd_128_arm64;
 
     uavs3e_funs_handle.cost_sad_x3[0] = uavs3e_get_sad_x3_4_arm64;
-    //uavs3e_funs_handle.cost_sad_x3[1] = uavs3e_get_sad_x3_8_arm64;
-    //uavs3e_funs_handle.cost_sad_x3[2] = uavs3e_get_sad_x3_16_arm64;
-    //uavs3e_funs_handle.cost_sad_x3[3] = uavs3e_get_sad_x3_32_arm64;
-    //uavs3e_funs_handle.cost_sad_x3[4] = uavs3e_get_sad_x3_64_arm64;
-    //uavs3e_funs_handle.cost_sad_x3[5] = uavs3e_get_sad_x3_128_arm64;
+    uavs3e_funs_handle.cost_sad_x3[1] = uavs3e_get_sad_x3_8_arm64;
+    uavs3e_funs_handle.cost_sad_x3[2] = uavs3e_get_sad_x3_16_arm64;
+    uavs3e_funs_handle.cost_sad_x3[3] = uavs3e_get_sad_x3_32_arm64;
+    uavs3e_funs_handle.cost_sad_x3[4] = uavs3e_get_sad_x3_64_arm64;
+    uavs3e_funs_handle.cost_sad_x3[5] = uavs3e_get_sad_x3_128_arm64;
 
-    /*
     uavs3e_funs_handle.cost_sad_x4[0] = uavs3e_get_sad_x4_4_arm64;
     uavs3e_funs_handle.cost_sad_x4[1] = uavs3e_get_sad_x4_8_arm64;
     uavs3e_funs_handle.cost_sad_x4[2] = uavs3e_get_sad_x4_16_arm64;
@@ -180,13 +179,14 @@ void uavs3e_funs_init_arm64()
     uavs3e_funs_handle.cost_sad_x4[4] = uavs3e_get_sad_x4_64_arm64;
     uavs3e_funs_handle.cost_sad_x4[5] = uavs3e_get_sad_x4_128_arm64;
 
-    uavs3e_funs_handle.cost_satd[0][0] = uavs3e_had_4x4_arm64;
-    uavs3e_funs_handle.cost_satd[1][0] = uavs3e_had_8x4_arm64;
-    uavs3e_funs_handle.cost_satd[0][1] = uavs3e_had_4x8_arm64;
-    uavs3e_funs_handle.cost_satd[1][1] = uavs3e_had_8x8_arm64;
-    uavs3e_funs_handle.cost_satd[2][1] = uavs3e_had_16x8_arm64;
-    uavs3e_funs_handle.cost_satd[1][2] = uavs3e_had_8x16_arm64;
+    //uavs3e_funs_handle.cost_satd[0][0] = uavs3e_had_4x4_arm64;
+    //uavs3e_funs_handle.cost_satd[1][0] = uavs3e_had_8x4_arm64;
+    //uavs3e_funs_handle.cost_satd[0][1] = uavs3e_had_4x8_arm64;
+    //uavs3e_funs_handle.cost_satd[1][1] = uavs3e_had_8x8_arm64;
+    //uavs3e_funs_handle.cost_satd[2][1] = uavs3e_had_16x8_arm64;
+    //uavs3e_funs_handle.cost_satd[1][2] = uavs3e_had_8x16_arm64;
 
+    /*
     uavs3e_funs_handle.cost_var[0] = uavs3e_get_var_4_arm64;
     uavs3e_funs_handle.cost_var[1] = uavs3e_get_var_8_arm64;
     uavs3e_funs_handle.cost_var[2] = uavs3e_get_var_16_arm64;
diff --git a/build/android/app/src/main/jni/src/armv8/cost_arm64.S b/build/android/app/src/main/jni/src/armv8/cost_arm64.S
index 6d75694..7b0157e 100644
--- a/build/android/app/src/main/jni/src/armv8/cost_arm64.S
+++ b/build/android/app/src/main/jni/src/armv8/cost_arm64.S
@@ -1981,5 +1981,1354 @@ get_sad_x3_128_y:
 
 ret
 
+//void uavs3e_get_sad_x4_4_arm64(pel *p_org, int i_org, pel *pred0, pel *pred1, pel *pred2, pel, *pred3, int i_pred, u32 sad[3], int height)
+//*p_org->x0, i_org->x1, *pred0->x2, *pred1->x3, *pred2->x4, *pred3->x5, i_pred->x6, sad[4]->x7, height->[x8]
+function uavs3e_get_sad_x4_4_arm64
+    ldr x8, [sp]
+    lsl x1, x1, #1
+    lsl x6, x6, #1
+    movi v18.16b, #0
+    movi v19.16b, #0
+    movi v20.16b, #0
+    movi v21.16b, #0
+
+get_sad_x4_4_y:
+    //load p_org
+    ld1 {v0.d}[0], [x0], x1
+    ld1 {v0.d}[1], [x0], x1
+    ld1 {v1.d}[0], [x0], x1
+    ld1 {v1.d}[1], [x0], x1
+    //load pred
+    ld1 {v2.d}[0], [x2], x6
+    ld1 {v2.d}[1], [x2], x6
+    ld1 {v3.d}[0], [x2], x6
+    ld1 {v3.d}[1], [x2], x6
+    ld1 {v4.d}[0], [x3], x6
+    ld1 {v4.d}[1], [x3], x6
+    ld1 {v5.d}[0], [x3], x6
+    ld1 {v5.d}[1], [x3], x6
+    ld1 {v6.d}[0], [x4], x6
+    ld1 {v6.d}[1], [x4], x6
+    ld1 {v7.d}[0], [x4], x6
+    ld1 {v7.d}[1], [x4], x6
+    ld1 {v16.d}[0], [x5], x6
+    ld1 {v16.d}[1], [x5], x6
+    ld1 {v17.d}[0], [x5], x6
+    ld1 {v17.d}[1], [x5], x6
+    //abs
+    uabd v2.8h, v0.8h, v2.8h
+    uabd v3.8h, v1.8h, v3.8h
+    uabd v4.8h, v0.8h, v4.8h
+    uabd v5.8h, v1.8h, v5.8h
+    uabd v6.8h, v0.8h, v6.8h
+    uabd v7.8h, v1.8h, v7.8h
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+
+    uaddl v0.4s, v2.4h, v3.4h
+    uaddl2 v1.4s, v2.8h, v3.8h
+    add v18.4s, v18.4s, v0.4s
+    add v18.4s, v18.4s, v1.4s
+    uaddl v0.4s, v4.4h, v5.4h
+    uaddl2 v1.4s, v4.8h, v5.8h
+    add v19.4s, v19.4s, v0.4s
+    add v19.4s, v19.4s, v1.4s
+    uaddl v0.4s, v6.4h, v7.4h
+    uaddl2 v1.4s, v6.8h, v7.8h
+    add v20.4s, v20.4s, v0.4s
+    add v20.4s, v20.4s, v1.4s
+    uaddl v0.4s, v16.4h, v17.4h
+    uaddl2 v1.4s, v16.8h, v17.8h
+    add v21.4s, v21.4s, v0.4s
+    add v21.4s, v21.4s, v1.4s
+    subs w8, w8, #4
+    bgt get_sad_x4_4_y
+
+    addp v18.4s, v18.4s, v18.4s
+    addp v18.4s, v18.4s, v18.4s
+    addp v19.4s, v19.4s, v19.4s
+    addp v19.4s, v19.4s, v19.4s
+    addp v20.4s, v20.4s, v20.4s
+    addp v20.4s, v20.4s, v20.4s
+    addp v21.4s, v21.4s, v21.4s
+    addp v21.4s, v21.4s, v21.4s
+
+    st1 {v18.s}[0], [x7], #4
+    st1 {v19.s}[0], [x7], #4
+    st1 {v20.s}[0], [x7], #4
+    st1 {v21.s}[0], [x7], #4
+
+    ret
+
+//void uavs3e_get_sad_x4_8_arm64(pel *p_org, int i_org, pel *pred0, pel *pred1, pel *pred2, pel, *pred3, int i_pred, u32 sad[3], int height)
+//*p_org->x0, i_org->x1, *pred0->x2, *pred1->x3, *pred2->x4, *pred3->x5, i_pred->x6, sad[4]->x7, height->[x8]
+function uavs3e_get_sad_x4_8_arm64
+    ldr x8, [sp]
+    lsl x1, x1, #1
+    lsl x6, x6, #1
+    movi v28.16b, #0
+    movi v29.16b, #0
+    movi v30.16b, #0
+    movi v31.16b, #0
+
+get_sad_x4_8_y:
+    //load p_org
+    ld1 {v0.8h}, [x0], x1
+    ld1 {v1.8h}, [x0], x1
+    ld1 {v2.8h}, [x0], x1
+    ld1 {v3.8h}, [x0], x1
+    //load pred
+    ld1 {v4.8h}, [x2], x6
+    ld1 {v5.8h}, [x2], x6
+    ld1 {v6.8h}, [x2], x6
+    ld1 {v7.8h}, [x2], x6
+    ld1 {v16.8h}, [x3], x6
+    ld1 {v17.8h}, [x3], x6
+    ld1 {v18.8h}, [x3], x6
+    ld1 {v19.8h}, [x3], x6
+    ld1 {v20.8h}, [x4], x6
+    ld1 {v21.8h}, [x4], x6
+    ld1 {v22.8h}, [x4], x6
+    ld1 {v23.8h}, [x4], x6
+    ld1 {v24.8h}, [x5], x6
+    ld1 {v25.8h}, [x5], x6
+    ld1 {v26.8h}, [x5], x6
+    ld1 {v27.8h}, [x5], x6
+
+    //abs
+    uabd v4.8h, v0.8h, v4.8h
+    uabd v5.8h, v1.8h, v5.8h
+    uabd v6.8h, v2.8h, v6.8h
+    uabd v7.8h, v3.8h, v7.8h
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v0.8h, v20.8h
+    uabd v21.8h, v1.8h, v21.8h
+    uabd v22.8h, v2.8h, v22.8h
+    uabd v23.8h, v3.8h, v23.8h
+    uabd v24.8h, v0.8h, v24.8h
+    uabd v25.8h, v1.8h, v25.8h
+    uabd v26.8h, v2.8h, v26.8h
+    uabd v27.8h, v3.8h, v27.8h
+
+    uaddl v0.4s, v4.4h, v5.4h
+    uaddl2 v1.4s, v4.8h, v5.8h
+    uaddl v2.4s, v6.4h, v7.4h
+    uaddl2 v3.4s, v6.8h, v7.8h
+    add v28.4s, v28.4s, v0.4s
+    add v28.4s, v28.4s, v1.4s
+    add v28.4s, v28.4s, v2.4s
+    add v28.4s, v28.4s, v3.4s
+    uaddl v0.4s, v16.4h, v17.4h
+    uaddl2 v1.4s, v16.8h, v17.8h
+    uaddl v2.4s, v18.4h, v19.4h
+    uaddl2 v3.4s, v18.8h, v19.8h
+    add v29.4s, v29.4s, v0.4s
+    add v29.4s, v29.4s, v1.4s
+    add v29.4s, v29.4s, v2.4s
+    add v29.4s, v29.4s, v3.4s
+    uaddl v0.4s, v20.4h, v21.4h
+    uaddl2 v1.4s, v20.8h, v21.8h
+    uaddl v2.4s, v22.4h, v23.4h
+    uaddl2 v3.4s, v22.8h, v23.8h
+    add v30.4s, v30.4s, v0.4s
+    add v30.4s, v30.4s, v1.4s
+    add v30.4s, v30.4s, v2.4s
+    add v30.4s, v30.4s, v3.4s
+    uaddl v0.4s, v24.4h, v25.4h
+    uaddl2 v1.4s, v24.8h, v25.8h
+    uaddl v2.4s, v26.4h, v27.4h
+    uaddl2 v3.4s, v26.8h, v27.8h
+    add v31.4s, v31.4s, v0.4s
+    add v31.4s, v31.4s, v1.4s
+    add v31.4s, v31.4s, v2.4s
+    add v31.4s, v31.4s, v3.4s
+    subs w8, w8, #4
+    bgt get_sad_x4_8_y
+
+    addp v28.4s, v28.4s, v28.4s
+    addp v28.4s, v28.4s, v28.4s
+    addp v29.4s, v29.4s, v29.4s
+    addp v29.4s, v29.4s, v29.4s
+    addp v30.4s, v30.4s, v30.4s
+    addp v30.4s, v30.4s, v30.4s
+    addp v31.4s, v31.4s, v31.4s
+    addp v31.4s, v31.4s, v31.4s
+
+    st1 {v28.s}[0], [x7], #4
+    st1 {v29.s}[0], [x7], #4
+    st1 {v30.s}[0], [x7], #4
+    st1 {v31.s}[0], [x7], #4
+
+ret
+
+//void uavs3e_get_sad_x4_16_arm64(pel *p_org, int i_org, pel *pred0, pel *pred1, pel *pred2, pel, *pred3, int i_pred, u32 sad[3], int height)
+//*p_org->x0, i_org->x1, *pred0->x2, *pred1->x3, *pred2->x4, *pred3->x5, i_pred->x6, sad[4]->x7, height->[x8]
+function uavs3e_get_sad_x4_16_arm64
+    ldr x8, [sp]
+    lsl x1, x1, #1
+    lsl x6, x6, #1
+    movi v26.16b, #0
+    movi v27.16b, #0
+    movi v28.16b, #0
+    movi v29.16b, #0
+
+get_sad_x4_16_y:
+    //load p_org
+    ld1 {v0.8h, v1.8h}, [x0], x1
+    ld1 {v2.8h, v3.8h}, [x0], x1
+    ld1 {v4.8h, v5.8h}, [x0], x1
+    ld1 {v6.8h, v7.8h}, [x0], x1
+
+    //load pred0
+    ld1 {v16.8h, v17.8h}, [x2], x6
+    ld1 {v18.8h, v19.8h}, [x2], x6
+    ld1 {v20.8h, v21.8h}, [x2], x6
+    ld1 {v22.8h, v23.8h}, [x2], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+
+    //load pred1
+    ld1 {v16.8h, v17.8h}, [x3], x6
+    ld1 {v18.8h, v19.8h}, [x3], x6
+    ld1 {v20.8h, v21.8h}, [x3], x6
+    ld1 {v22.8h, v23.8h}, [x3], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+
+    //load pred2
+    ld1 {v16.8h, v17.8h}, [x4], x6
+    ld1 {v18.8h, v19.8h}, [x4], x6
+    ld1 {v20.8h, v21.8h}, [x4], x6
+    ld1 {v22.8h, v23.8h}, [x4], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+
+    //load pred3
+    ld1 {v16.8h, v17.8h}, [x5], x6
+    ld1 {v18.8h, v19.8h}, [x5], x6
+    ld1 {v20.8h, v21.8h}, [x5], x6
+    ld1 {v22.8h, v23.8h}, [x5], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+
+    subs w8, w8, #4
+    bgt get_sad_x4_16_y
+
+    addp v26.4s, v26.4s, v26.4s
+    addp v26.4s, v26.4s, v26.4s
+    addp v27.4s, v27.4s, v27.4s
+    addp v27.4s, v27.4s, v27.4s
+    addp v28.4s, v28.4s, v28.4s
+    addp v28.4s, v28.4s, v28.4s
+    addp v29.4s, v29.4s, v29.4s
+    addp v29.4s, v29.4s, v29.4s
+
+    st1 {v26.s}[0], [x7], #4
+    st1 {v27.s}[0], [x7], #4
+    st1 {v28.s}[0], [x7], #4
+    st1 {v29.s}[0], [x7], #4
+ret
+
+//void uavs3e_get_sad_x4_32_arm64(pel *p_org, int i_org, pel *pred0, pel *pred1, pel *pred2, pel, *pred3, int i_pred, u32 sad[3], int height)
+//*p_org->x0, i_org->x1, *pred0->x2, *pred1->x3, *pred2->x4, *pred3->x5, i_pred->x6, sad[3]->x7, height->[x8]
+function uavs3e_get_sad_x4_32_arm64
+    ldr x8, [sp]
+    lsl x1, x1, #1
+    lsl x6, x6, #1
+    movi v26.16b, #0
+    movi v27.16b, #0
+    movi v28.16b, #0
+    movi v29.16b, #0
+
+get_sad_x4_32_y:
+    //load p_org
+    ld1 {v0.8h, v1.8h, v2.8h, v3.8h}, [x0], x1
+    ld1 {v4.8h, v5.8h, v6.8h, v7.8h}, [x0], x1
+    //load pred0
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x2], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x2], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+
+    //load pred1
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x3], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x3], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+
+    //load pred2
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x4], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x4], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+
+    //load pred3
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x5], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x5], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+
+    subs w8, w8, #2
+    bgt get_sad_x4_32_y
+
+    addp v26.4s, v26.4s, v26.4s
+    addp v26.4s, v26.4s, v26.4s
+    addp v27.4s, v27.4s, v27.4s
+    addp v27.4s, v27.4s, v27.4s
+    addp v28.4s, v28.4s, v28.4s
+    addp v28.4s, v28.4s, v28.4s
+    addp v29.4s, v29.4s, v29.4s
+    addp v29.4s, v29.4s, v29.4s
+
+    st1 {v26.s}[0], [x7], #4
+    st1 {v27.s}[0], [x7], #4
+    st1 {v28.s}[0], [x7], #4
+    st1 {v29.s}[0], [x7], #4
+
+ret
+
+
+//void uavs3e_get_sad_x4_64_arm64(pel *p_org, int i_org, pel *pred0, pel *pred1, pel *pred2, pel, *pred3, int i_pred, u32 sad[4], int height)
+//*p_org->x0, i_org->x1, *pred0->x2, *pred1->x3, *pred2->x4, *pred3->x5, i_pred->x6, sad[4]->x7, height->[x8]
+function uavs3e_get_sad_x4_64_arm64
+    ldr x15, [sp]
+    lsl x1, x1, #1
+    lsl x6, x6, #1
+    lsl x13, x1, #1     //2 * i_org
+    lsl x14, x6, #1     //2 * i_pred
+    movi v26.16b, #0
+    movi v27.16b, #0
+    movi v28.16b, #0
+    movi v29.16b, #0
+
+get_sad_x4_64_y:
+    mov x8, x0
+    mov x9, x2
+    mov x10, x3
+    mov x11, x4
+    mov x12, x5
+    //load p_org
+    ld1 {v0.8h, v1.8h, v2.8h, v3.8h}, [x8], x1
+    ld1 {v4.8h, v5.8h, v6.8h, v7.8h}, [x8], x1
+    //load pred0
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x9], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x9], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+
+    //load pred1
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x10], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x10], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+
+    //load pred2
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x11], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x11], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+
+    //load pred3
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x12], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x12], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+
+    add x8, x0, #64
+    add x9, x2, #64
+    add x10, x3, #64
+    add x11, x4, #64
+    add x12, x5, #64
+    //load p_org
+    ld1 {v0.8h, v1.8h, v2.8h, v3.8h}, [x8], x1
+    ld1 {v4.8h, v5.8h, v6.8h, v7.8h}, [x8], x1
+    //load pred0
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x9], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x9], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+
+    //load pred1
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x10], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x10], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+
+    //load pred2
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x11], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x11], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+
+    //load pred3
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x12], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x12], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+
+    add x0, x0, x13
+    add x2, x2, x14
+    add x3, x3, x14
+    add x4, x4, x14
+    add x5, x5, x14
+
+    subs w15, w15, #2
+    bgt get_sad_x4_64_y
+
+    addp v26.4s, v26.4s, v26.4s
+    addp v26.4s, v26.4s, v26.4s
+    addp v27.4s, v27.4s, v27.4s
+    addp v27.4s, v27.4s, v27.4s
+    addp v28.4s, v28.4s, v28.4s
+    addp v28.4s, v28.4s, v28.4s
+    addp v29.4s, v29.4s, v29.4s
+    addp v29.4s, v29.4s, v29.4s
+
+    st1 {v26.s}[0], [x7], #4
+    st1 {v27.s}[0], [x7], #4
+    st1 {v28.s}[0], [x7], #4
+    st1 {v29.s}[0], [x7], #4
+
+ret
+
+//void uavs3e_get_sad_x4_128_arm64(pel *p_org, int i_org, pel *pred0, pel *pred1, pel *pred2, pel, *pred3, int i_pred, u32 sad[4], int height)
+//*p_org->x0, i_org->x1, *pred0->x2, *pred1->x3, *pred2->x4, *pred3->x5, i_pred->x6, sad[4]->x7, height->[x8]
+function uavs3e_get_sad_x4_128_arm64
+    ldr x15, [sp]
+    lsl x1, x1, #1
+    lsl x6, x6, #1
+    lsl x13, x1, #1     //2 * i_org
+    lsl x14, x6, #1     //2 * i_pred
+    movi v26.16b, #0
+    movi v27.16b, #0
+    movi v28.16b, #0
+    movi v29.16b, #0
+
+get_sad_x4_128_y:
+    mov x8, x0
+    mov x9, x2
+    mov x10, x3
+    mov x11, x4
+    mov x12, x5
+    //load p_org
+    ld1 {v0.8h, v1.8h, v2.8h, v3.8h}, [x8], x1
+    ld1 {v4.8h, v5.8h, v6.8h, v7.8h}, [x8], x1
+    //load pred0
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x9], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x9], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+
+    //load pred1
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x10], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x10], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+
+    //load pred2
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x11], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x11], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+
+    //load pred3
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x12], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x12], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+
+    add x8, x0, #64
+    add x9, x2, #64
+    add x10, x3, #64
+    add x11, x4, #64
+    add x12, x4, #64
+    //load p_org
+    ld1 {v0.8h, v1.8h, v2.8h, v3.8h}, [x8], x1
+    ld1 {v4.8h, v5.8h, v6.8h, v7.8h}, [x8], x1
+    //load pred0
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x9], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x9], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+
+    //load pred1
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x10], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x10], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+
+    //load pred2
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x11], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x11], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+
+    //load pred3
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x12], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x12], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+
+    add x8, x0, #128
+    add x9, x2, #128
+    add x10, x3, #128
+    add x11, x4, #128
+    add x12, x4, #128
+    //load p_org
+    ld1 {v0.8h, v1.8h, v2.8h, v3.8h}, [x8], x1
+    ld1 {v4.8h, v5.8h, v6.8h, v7.8h}, [x8], x1
+    //load pred0
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x9], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x9], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+
+    //load pred1
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x10], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x10], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+
+    //load pred2
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x11], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x11], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+
+    //load pred3
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x12], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x12], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+
+    add x8, x0, #192
+    add x9, x2, #192
+    add x10, x3, #192
+    add x11, x4, #192
+    add x12, x5, #192
+    //load p_org
+    ld1 {v0.8h, v1.8h, v2.8h, v3.8h}, [x8], x1
+    ld1 {v4.8h, v5.8h, v6.8h, v7.8h}, [x8], x1
+    //load pred0
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x9], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x9], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v26.4s, v26.4s, v24.4s
+    add v26.4s, v26.4s, v25.4s
+
+    //load pred1
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x10], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x10], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v27.4s, v27.4s, v24.4s
+    add v27.4s, v27.4s, v25.4s
+
+    //load pred2
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x11], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x11], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v28.4s, v28.4s, v24.4s
+    add v28.4s, v28.4s, v25.4s
+
+    //load pred3
+    ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x12], x6
+    ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x12], x6
+
+    uabd v16.8h, v0.8h, v16.8h
+    uabd v17.8h, v1.8h, v17.8h
+    uabd v18.8h, v2.8h, v18.8h
+    uabd v19.8h, v3.8h, v19.8h
+    uabd v20.8h, v4.8h, v20.8h
+    uabd v21.8h, v5.8h, v21.8h
+    uabd v22.8h, v6.8h, v22.8h
+    uabd v23.8h, v7.8h, v23.8h
+
+    uaddl v24.4s, v16.4h, v17.4h
+    uaddl2 v25.4s, v16.8h, v17.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v18.4h, v19.4h
+    uaddl2 v25.4s, v18.8h, v19.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v20.4h, v21.4h
+    uaddl2 v25.4s, v20.8h, v21.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+    uaddl v24.4s, v22.4h, v23.4h
+    uaddl2 v25.4s, v22.8h, v23.8h
+    add v29.4s, v29.4s, v24.4s
+    add v29.4s, v29.4s, v25.4s
+
+    add x0, x0, x13
+    add x2, x2, x14
+    add x3, x3, x14
+    add x4, x4, x14
+    add x5, x5, x14
+
+    subs w15, w15, #2
+    bgt get_sad_x4_128_y
+
+    addp v26.4s, v26.4s, v26.4s
+    addp v26.4s, v26.4s, v26.4s
+    addp v27.4s, v27.4s, v27.4s
+    addp v27.4s, v27.4s, v27.4s
+    addp v28.4s, v28.4s, v28.4s
+    addp v28.4s, v28.4s, v28.4s
+    addp v29.4s, v29.4s, v29.4s
+    addp v29.4s, v29.4s, v29.4s
+
+    st1 {v26.s}[0], [x7], #4
+    st1 {v27.s}[0], [x7], #4
+    st1 {v28.s}[0], [x7], #4
+    st1 {v29.s}[0], [x7], #4
+
+ret
+
+
 #endif
 #endif
\ No newline at end of file