xref: /third_party/libpng/libpng_optimize.patch

diff --git a/arm/arm_init.c b/arm/arm_init.c
index 3a89998ab..05aa2c0d9 100644
--- a/arm/arm_init.c
+++ b/arm/arm_init.c
@@ -113,13 +113,23 @@ png_init_filter_functions_neon(png_structp pp, unsigned int bpp)
     * initialization function.)
     */
    pp->read_filter[PNG_FILTER_VALUE_UP-1] = png_read_filter_row_up_neon;
-
+#ifdef PNG_MULTY_LINE_ENABLE
+   // OH ISSUE: png optimize
+   pp->read_filter[PNG_FILTER_VALUE_UP_X2-1] = png_read_filter_row_up_x2_neon;
+#endif
    if (bpp == 3)
    {
       pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub3_neon;
       pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg3_neon;
       pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
          png_read_filter_row_paeth3_neon;
+#ifdef PNG_MULTY_LINE_ENABLE
+      // OH ISSUE: png optimize
+      pp->read_filter[PNG_FILTER_VALUE_AVG_X2-1] =
+         png_read_filter_row_avg3_x2_neon;
+      pp->read_filter[PNG_FILTER_VALUE_PAETH_X2-1] =
+         png_read_filter_row_paeth3_x2_neon;
+#endif
    }
 
    else if (bpp == 4)
@@ -128,6 +138,13 @@ png_init_filter_functions_neon(png_structp pp, unsigned int bpp)
       pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg4_neon;
       pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
           png_read_filter_row_paeth4_neon;
+#ifdef PNG_MULTY_LINE_ENABLE
+      // OH ISSUE: png optimize
+      pp->read_filter[PNG_FILTER_VALUE_AVG_X2-1] =
+         png_read_filter_row_avg4_x2_neon;
+      pp->read_filter[PNG_FILTER_VALUE_PAETH_X2-1] =
+         png_read_filter_row_paeth4_x2_neon;
+#endif
    }
 }
 #endif /* PNG_ARM_NEON_OPT > 0 */
diff --git a/arm/filter_neon_intrinsics.c b/arm/filter_neon_intrinsics.c
index 4466d48b2..4ff810a19 100644
--- a/arm/filter_neon_intrinsics.c
+++ b/arm/filter_neon_intrinsics.c
@@ -47,6 +47,7 @@
 
 #if PNG_ARM_NEON_OPT > 0
 
+#ifndef PNG_MULTY_LINE_ENABLE
 void
 png_read_filter_row_up_neon(png_row_infop row_info, png_bytep row,
    png_const_bytep prev_row)
@@ -396,7 +397,1351 @@ png_read_filter_row_paeth4_neon(png_row_infop row_info, png_bytep row,
       vst4_lane_u32(png_ptr(uint32_t,rp), vdest_val, 0);
    }
 }
+#else
+// OH ISSUE: png optimize
+// according to definition: row_info->rowbytes = row_width * row_info->channels,
+// the input rowbytes must be 3 or 4 times the channel size, so:
+// for RGB neon process 12 bytes at once,the tail must be 3,6,9;
+// for RGBA neon process 16 or 8 bytes at once,the tail must be 4;
+// filter operators are internal function, row_info and row ensure non empty outside.
+#define STEP_RGB (12) // 3 channel RGB process 12 bytes at once
+#define TAIL_RGB3 (9) // tail 3 pixels have 9 bytes
+#define TAIL_RGB2 (6) // tail 2 pixels have 6 bytes
+#define TAIL_RGB1 (3) // tail 1 pixel have 3 bytes
+#define STEP_RGBA (16) // GBA neon process 16 bytes at once
+#define STEP_RGBA_HALF (8) // GBA neon process 8 bytes at once
+#define TAIL_RGBA (4) // tail 1 pixel have 4 bytes
+#define IND3 (3) // index 3
+#define IND2 (2) // index 2
+#define OFFSET3 (3) // RGB offset 3 bytes
+#define OFFSET6 (6) // RGB offset 6 bytes
+void png_read_filter_row_up_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   int count = row_info->rowbytes;
+
+   png_debug(1, "in png_read_filter_row_up_neon");
+
+   uint8x16_t qrp, qpp;
+   while (count >= STEP_RGBA) {
+      qrp = vld1q_u8(rp);
+      qpp = vld1q_u8(pp);
+      qrp = vaddq_u8(qrp, qpp);
+      vst1q_u8(rp, qrp);
+      rp += STEP_RGBA;
+      pp += STEP_RGBA;
+      count -= STEP_RGBA;
+   }
+
+   if (count >= STEP_RGBA_HALF) {
+      uint8x8_t qrp1, qpp1;
+      qrp1 = vld1_u8(rp);
+      qpp1 = vld1_u8(pp);
+      qrp1 = vadd_u8(qrp1, qpp1);
+      vst1_u8(rp, qrp1);
+      rp += STEP_RGBA_HALF;
+      pp += STEP_RGBA_HALF;
+      count -= STEP_RGBA_HALF;
+   }
+
+   for (int i = 0; i < count; i++) {
+      *rp = (png_byte)(((int)(*rp) + (int)(*pp++)) & 0xff);
+      rp++;
+   }
+}
+
+void png_read_filter_row_up_x2_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   int count = row_info->rowbytes;
+   png_bytep np = row + row_info->rowbytes + 1;
+
+   png_debug(1, "in png_read_filter_row_up_x2_neon");
+
+   uint8x16_t qrp, qpp, qnp;
+   while (count >= STEP_RGBA) {
+      qrp = vld1q_u8(rp);
+      qpp = vld1q_u8(pp);
+      qnp = vld1q_u8(np);
+      qrp = vaddq_u8(qrp, qpp);
+      qnp = vaddq_u8(qnp, qrp);
+      vst1q_u8(rp, qrp);
+      vst1q_u8(np, qnp);
+      rp += STEP_RGBA;
+      pp += STEP_RGBA;
+      np += STEP_RGBA;
+      count -= STEP_RGBA;
+   }
+
+   if (count >= STEP_RGBA_HALF) {
+      uint8x8_t qrp1, qpp1, qnp1;
+      qrp1 = vld1_u8(rp);
+      qpp1 = vld1_u8(pp);
+      qnp1 = vld1_u8(np);
+      qrp1 = vadd_u8(qrp1, qpp1);
+      qnp1 = vadd_u8(qnp1, qrp1);
+      vst1_u8(rp, qrp1);
+      vst1_u8(np, qnp1);
+      rp += STEP_RGBA_HALF;
+      pp += STEP_RGBA_HALF;
+      np += STEP_RGBA_HALF;
+      count -= STEP_RGBA_HALF;
+   }
+
+   for (int i = 0; i < count; i++) {
+      *rp = (png_byte)(((int)(*rp) + (int)(*pp++)) & 0xff);
+      *np = (png_byte)(((int)(*np) + (int)(*rp++)) & 0xff);
+      np++;
+   }
+}
+
+void png_read_filter_row_sub3_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+
+   uint8x16_t vtmp = vld1q_u8(rp);
+   uint8x8x2_t *vrpt = png_ptr(uint8x8x2_t, &vtmp);
+   uint8x8x2_t vrp = *vrpt;
+
+   uint8x8x4_t vdest;
+   vdest.val[IND3] = vdup_n_u8(0);
+
+   uint8x8_t vtmp1, vtmp2;
+   uint32x2_t *temp_pointer;
+
+   png_debug(1, "in png_read_filter_row_sub3_neon");
+
+   size_t tail_bytes = row_info->rowbytes % STEP_RGB;
+   png_byte last_byte = *rp_stop;
+   png_bytep rp_stop_new = rp_stop - tail_bytes;
+   for (; rp < rp_stop_new;)
+   {
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vdest.val[0] = vadd_u8(vdest.val[IND3], vrp.val[0]);
+      vtmp2 = vext_u8(vrp.val[0], vrp.val[1], OFFSET6);
+      vdest.val[1] = vadd_u8(vdest.val[0], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
+      vdest.val[IND2] = vadd_u8(vdest.val[1], vtmp2);
+      vdest.val[IND3] = vadd_u8(vdest.val[IND2], vtmp1);
+
+      vtmp = vld1q_u8(rp + STEP_RGB);
+      vrpt = png_ptr(uint8x8x2_t, &vtmp);
+      vrp = *vrpt;
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND2]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND3]), 0);
+      rp += OFFSET3;
+   }
+
+   if (tail_bytes == TAIL_RGB1) {
+      vdest.val[0] = vadd_u8(vdest.val[IND3], vrp.val[0]);
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+   } else if (tail_bytes == TAIL_RGB2) {
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vdest.val[0] = vadd_u8(vdest.val[IND3], vrp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[0], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+   } else if (tail_bytes == TAIL_RGB3) {
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vdest.val[0] = vadd_u8(vdest.val[IND3], vrp.val[0]);
+      vtmp2 = vext_u8(vrp.val[0], vrp.val[1], OFFSET6);
+      vdest.val[1] = vadd_u8(vdest.val[0], vtmp1);
+      vdest.val[IND2] = vadd_u8(vdest.val[1], vtmp2);
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND2]), 0);
+   }
+   *rp_stop = last_byte;
+
+   PNG_UNUSED(prev_row)
+}
+
+void png_read_filter_row_sub4_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   int count = row_info->rowbytes;
+
+   uint8x8x4_t vdest;
+   vdest.val[IND3] = vdup_n_u8(0);
+
+   png_debug(1, "in png_read_filter_row_sub4_neon");
+
+   uint32x2x4_t vtmp;
+   uint8x8x4_t *vrpt;
+   uint8x8x4_t vrp;
+   uint32x2x4_t vdest_val;
+   while (count >= STEP_RGBA) {
+      uint32x2x4_t *temp_pointer;
+      vtmp = vld4_u32(png_ptr(uint32_t, rp));
+      vrpt = png_ptr(uint8x8x4_t, &vtmp);
+      vrp = *vrpt;
+
+      vdest.val[0] = vadd_u8(vdest.val[IND3], vrp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[0], vrp.val[1]);
+      vdest.val[IND2] = vadd_u8(vdest.val[1], vrp.val[IND2]);
+      vdest.val[IND3] = vadd_u8(vdest.val[IND2], vrp.val[IND3]);
+
+      vdest_val = png_ldr(uint32x2x4_t, &vdest);
+      vst4_lane_u32(png_ptr(uint32_t, rp), vdest_val, 0);
+
+      rp += STEP_RGBA;
+      count -= STEP_RGBA;
+   }
+
+   if (count >= STEP_RGBA_HALF) {
+      uint32x2x2_t vtmp1 = vld2_u32(png_ptr(uint32_t, rp));
+      uint8x8x2_t *vrpt1 = png_ptr(uint8x8x2_t, &vtmp1);
+      uint8x8x2_t vrp1 = *vrpt1;
+      uint32x2x2_t *temp_pointer;
+      uint32x2x2_t vdest_val1;
+
+      vdest.val[0] = vadd_u8(vdest.val[IND3], vrp1.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[0], vrp1.val[1]);
+      vdest.val[IND3] = vdest.val[1];
+      vdest_val1 = png_ldr(uint32x2x2_t, &vdest);
+      vst2_lane_u32(png_ptr(uint32_t, rp), vdest_val1, 0);
+
+      rp += STEP_RGBA_HALF;
+      count -= STEP_RGBA_HALF;
+   }
+
+   if (count == 0) {
+      return;
+   }
+
+   uint32x2_t vtmp2 = vld1_u32(png_ptr(uint32_t, rp));
+   uint8x8_t *vrpt2 = png_ptr(uint8x8_t, &vtmp2);
+   uint8x8_t vrp2 = *vrpt2;
+   uint32x2_t *temp_pointer;
+   uint32x2_t vdest_val2;
+
+   vdest.val[0] = vadd_u8(vdest.val[IND3], vrp2);
+   vdest_val2 = png_ldr(uint32x2_t, &vdest);
+   vst1_lane_u32(png_ptr(uint32_t, rp), vdest_val2, 0);
+
+   PNG_UNUSED(prev_row)
+}
+
+void png_read_filter_row_avg3_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+
+   uint8x16_t vtmp;
+   uint8x8x2_t *vrpt;
+   uint8x8x2_t vrp;
+   uint8x8x4_t vdest;
+   vdest.val[IND3] = vdup_n_u8(0);
+
+   vtmp = vld1q_u8(rp);
+   vrpt = png_ptr(uint8x8x2_t, &vtmp);
+   vrp = *vrpt;
+
+   png_debug(1, "in png_read_filter_row_avg3_neon");
+
+   uint8x8_t vtmp1, vtmp2, vtmp3;
+   uint8x8x2_t *vppt;
+   uint8x8x2_t vpp;
+   uint32x2_t *temp_pointer;
+
+   size_t tail_bytes = row_info->rowbytes % STEP_RGB;
+   png_byte last_byte = *rp_stop;
+   png_bytep rp_stop_new = rp_stop - tail_bytes;
+   for (; rp < rp_stop_new; pp += STEP_RGB)
+   {
+      vtmp = vld1q_u8(pp);
+      vppt = png_ptr(uint8x8x2_t, &vtmp);
+      vpp = *vppt;
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET3);
+      vtmp3 = vext_u8(vrp.val[0], vrp.val[1], OFFSET6);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vtmp2);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET6);
+      vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
+
+      vtmp = vld1q_u8(rp + STEP_RGB);
+      vrpt = png_ptr(uint8x8x2_t, &vtmp);
+      vrp = *vrpt;
+
+      vdest.val[IND2] = vhadd_u8(vdest.val[1], vtmp2);
+      vdest.val[IND2] = vadd_u8(vdest.val[IND2], vtmp3);
+
+      vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
+
+      vdest.val[IND3] = vhadd_u8(vdest.val[IND2], vtmp2);
+      vdest.val[IND3] = vadd_u8(vdest.val[IND3], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND2]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND3]), 0);
+      rp += OFFSET3;
+   }
+
+   vtmp = vld1q_u8(pp);
+   vppt = png_ptr(uint8x8x2_t, &vtmp);
+   vpp = *vppt;
+
+   if (tail_bytes == TAIL_RGB1) {
+      vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+   } else if (tail_bytes == TAIL_RGB2) {
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET3);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vtmp2);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+   } else if (tail_bytes == TAIL_RGB3) {
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET3);
+      vtmp3 = vext_u8(vrp.val[0], vrp.val[1], OFFSET6);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vtmp2);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET6);
+
+      vdest.val[IND2] = vhadd_u8(vdest.val[1], vtmp2);
+      vdest.val[IND2] = vadd_u8(vdest.val[IND2], vtmp3);
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND2]), 0);
+   }
+   *rp_stop = last_byte;
+}
+
+void png_read_filter_row_avg3_x2_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+   png_bytep np = rp_stop + 1;
+
+   uint8x16_t vtmp;
+   uint8x8x2_t *vrpt;
+   uint8x8x2_t vrp;
+   uint8x8x4_t vdest;
+   vdest.val[IND3] = vdup_n_u8(0);
+
+   vtmp = vld1q_u8(rp);
+   vrpt = png_ptr(uint8x8x2_t, &vtmp);
+   vrp = *vrpt;
+
+   uint8x8x2_t *vnpt;
+   uint8x8x2_t vnp;
+   uint8x8x4_t vdestN;
+   vdestN.val[IND3] = vdup_n_u8(0);
+
+   vtmp = vld1q_u8(np);
+   vnpt = png_ptr(uint8x8x2_t, &vtmp);
+   vnp = *vnpt;
+
+   png_debug(1, "in png_read_filter_row_x2_avg3_neon");
+
+   uint8x8_t vtmp1, vtmp2, vtmp3;
+   uint8x8x2_t *vppt;
+   uint8x8x2_t vpp;
+   uint32x2_t *temp_pointer;
+
+   size_t tail_bytes = row_info->rowbytes % STEP_RGB;
+   png_byte last_byte = *rp_stop;
+   png_byte last_byte_next = *(rp_stop + row_info->rowbytes + 1);
+   png_bytep rp_stop_new = rp_stop - tail_bytes;
+   for (; rp < rp_stop_new; pp += STEP_RGB)
+   {
+      vtmp = vld1q_u8(pp);
+      vppt = png_ptr(uint8x8x2_t, &vtmp);
+      vpp = *vppt;
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET3);
+      vtmp3 = vext_u8(vrp.val[0], vrp.val[1], OFFSET6);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vtmp2);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET6);
+      vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
+
+      vtmp = vld1q_u8(rp + STEP_RGB);
+      vrpt = png_ptr(uint8x8x2_t, &vtmp);
+      vrp = *vrpt;
+
+      vdest.val[IND2] = vhadd_u8(vdest.val[1], vtmp2);
+      vdest.val[IND2] = vadd_u8(vdest.val[IND2], vtmp3);
+
+      vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
+
+      vdest.val[IND3] = vhadd_u8(vdest.val[IND2], vtmp2);
+      vdest.val[IND3] = vadd_u8(vdest.val[IND3], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND2]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND3]), 0);
+      rp += OFFSET3;
+
+      vtmp1 = vext_u8(vnp.val[0], vnp.val[1], OFFSET3);
+      vdestN.val[0] = vhadd_u8(vdestN.val[IND3], vdest.val[0]);
+      vdestN.val[0] = vadd_u8(vdestN.val[0], vnp.val[0]);
+
+      vtmp3 = vext_u8(vnp.val[0], vnp.val[1], OFFSET6);
+      vdestN.val[1] = vhadd_u8(vdestN.val[0], vdest.val[1]);
+      vdestN.val[1] = vadd_u8(vdestN.val[1], vtmp1);
+
+      vtmp1 = vext_u8(vnp.val[1], vnp.val[1], 1);
+
+      vtmp = vld1q_u8(np + STEP_RGB);
+      vnpt = png_ptr(uint8x8x2_t, &vtmp);
+      vnp = *vnpt;
+
+      vdestN.val[IND2] = vhadd_u8(vdestN.val[1], vdest.val[IND2]);
+      vdestN.val[IND2] = vadd_u8(vdestN.val[IND2], vtmp3);
+
+      vdestN.val[IND3] = vhadd_u8(vdestN.val[IND2], vdest.val[IND3]);
+      vdestN.val[IND3] = vadd_u8(vdestN.val[IND3], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[0]), 0);
+      np += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[1]), 0);
+      np += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[IND2]), 0);
+      np += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[IND3]), 0);
+      np += OFFSET3;
+   }
+
+   vtmp = vld1q_u8(pp);
+   vppt = png_ptr(uint8x8x2_t, &vtmp);
+   vpp = *vppt;
+
+   if (tail_bytes == TAIL_RGB1) {
+      vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+
+      vdestN.val[0] = vhadd_u8(vdestN.val[IND3], vdest.val[0]);
+      vdestN.val[0] = vadd_u8(vdestN.val[0], vnp.val[0]);
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[0]), 0);
+   } else if (tail_bytes == TAIL_RGB2) {
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET3);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vtmp2);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+
+      vtmp1 = vext_u8(vnp.val[0], vnp.val[1], OFFSET3);
+      vdestN.val[0] = vhadd_u8(vdestN.val[IND3], vdest.val[0]);
+      vdestN.val[0] = vadd_u8(vdestN.val[0], vnp.val[0]);
+
+      vdestN.val[1] = vhadd_u8(vdestN.val[0], vdest.val[1]);
+      vdestN.val[1] = vadd_u8(vdestN.val[1], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[0]), 0);
+      np += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[1]), 0);
+   } else if (tail_bytes == TAIL_RGB3) {
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET3);
+      vtmp3 = vext_u8(vrp.val[0], vrp.val[1], OFFSET6);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vtmp2);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET6);
+
+      vdest.val[IND2] = vhadd_u8(vdest.val[1], vtmp2);
+      vdest.val[IND2] = vadd_u8(vdest.val[IND2], vtmp3);
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND2]), 0);
+
+      vtmp1 = vext_u8(vnp.val[0], vnp.val[1], OFFSET3);
+      vdestN.val[0] = vhadd_u8(vdestN.val[IND3], vdest.val[0]);
+      vdestN.val[0] = vadd_u8(vdestN.val[0], vnp.val[0]);
+
+      vtmp3 = vext_u8(vnp.val[0], vnp.val[1], OFFSET6);
+      vdestN.val[1] = vhadd_u8(vdestN.val[0], vdest.val[1]);
+      vdestN.val[1] = vadd_u8(vdestN.val[1], vtmp1);
+
+      vdestN.val[IND2] = vhadd_u8(vdestN.val[1], vdest.val[IND2]);
+      vdestN.val[IND2] = vadd_u8(vdestN.val[IND2], vtmp3);
+
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[0]), 0);
+      np += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[1]), 0);
+      np += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[IND2]), 0);
+   }
+   *rp_stop = last_byte;
+   *(rp_stop + row_info->rowbytes + 1) = last_byte_next;
+}
+
+void png_read_filter_row_avg4_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   int count = row_info->rowbytes;
+
+   uint8x8x4_t vdest;
+   vdest.val[IND3] = vdup_n_u8(0);
+
+   png_debug(1, "in png_read_filter_row_avg4_neon");
+
+   uint32x2x4_t vtmp;
+   uint8x8x4_t *vrpt, *vppt;
+   uint8x8x4_t vrp, vpp;
+   uint32x2x4_t vdest_val;
+   while (count >= STEP_RGBA) {
+      uint32x2x4_t *temp_pointer;
+      vtmp = vld4_u32(png_ptr(uint32_t, rp));
+      vrpt = png_ptr(uint8x8x4_t, &vtmp);
+      vrp = *vrpt;
+      vtmp = vld4_u32(png_ptrc(uint32_t, pp));
+      vppt = png_ptr(uint8x8x4_t, &vtmp);
+      vpp = *vppt;
+
+      vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vpp.val[1]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]);
+      vdest.val[IND2] = vhadd_u8(vdest.val[1], vpp.val[IND2]);
+      vdest.val[IND2] = vadd_u8(vdest.val[IND2], vrp.val[IND2]);
+      vdest.val[IND3] = vhadd_u8(vdest.val[IND2], vpp.val[IND3]);
+      vdest.val[IND3] = vadd_u8(vdest.val[IND3], vrp.val[IND3]);
+
+      vdest_val = png_ldr(uint32x2x4_t, &vdest);
+      vst4_lane_u32(png_ptr(uint32_t, rp), vdest_val, 0);
+
+      rp += STEP_RGBA;
+      pp += STEP_RGBA;
+      count -= STEP_RGBA;
+   }
+
+   if (count >= STEP_RGBA_HALF) {
+      uint32x2x2_t vtmp1;
+      uint8x8x2_t *vrpt1, *vppt1;
+      uint8x8x2_t vrp1, vpp1;
+      uint32x2x2_t *temp_pointer;
+      uint32x2x2_t vdest_val1;
+
+      vtmp1 = vld2_u32(png_ptr(uint32_t, rp));
+      vrpt1 = png_ptr(uint8x8x2_t, &vtmp1);
+      vrp1 = *vrpt1;
+      vtmp1 = vld2_u32(png_ptrc(uint32_t, pp));
+      vppt1 = png_ptr(uint8x8x2_t, &vtmp1);
+      vpp1 = *vppt1;
+
+      vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp1.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp1.val[0]);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vpp1.val[1]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vrp1.val[1]);
+      vdest.val[IND3] = vdest.val[1];
+      vdest_val1 = png_ldr(uint32x2x2_t, &vdest);
+      vst2_lane_u32(png_ptr(uint32_t, rp), vdest_val1, 0);
+
+      rp += STEP_RGBA_HALF;
+      pp += STEP_RGBA_HALF;
+      count -= STEP_RGBA_HALF;
+   }
+
+   if (count == 0) {
+      return;
+   }
+
+   uint32x2_t vtmp2;
+   uint8x8_t *vrpt2, *vppt2;
+   uint8x8_t vrp2, vpp2;
+   uint32x2_t *temp_pointer;
+   uint32x2_t vdest_val2;
+
+   vtmp2 = vld1_u32(png_ptr(uint32_t, rp));
+   vrpt2 = png_ptr(uint8x8_t, &vtmp2);
+   vrp2 = *vrpt2;
+   vtmp2 = vld1_u32(png_ptrc(uint32_t, pp));
+   vppt2 = png_ptr(uint8x8_t, &vtmp2);
+   vpp2 = *vppt2;
+
+   vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp2);
+   vdest.val[0] = vadd_u8(vdest.val[0], vrp2);
 
+   vdest_val2 = png_ldr(uint32x2_t, &vdest);
+   vst1_lane_u32(png_ptr(uint32_t, rp), vdest_val2, 0);
+}
+
+void png_read_filter_row_avg4_x2_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   int count = row_info->rowbytes;
+   png_bytep np = row + count + 1;
+
+   uint8x8x4_t vdest;
+   vdest.val[IND3] = vdup_n_u8(0);
+
+   png_debug(1, "in png_read_filter_row_avg4_x2_neon");
+
+   uint32x2x4_t vtmp;
+   uint8x8x4_t *vrpt, *vppt;
+   uint8x8x4_t vrp, vpp;
+   uint32x2x4_t vdest_val;
+
+   uint8x8x4_t *vnpt;
+   uint8x8x4_t vnp;
+   uint8x8x4_t vdestN;
+   vdestN.val[IND3] = vdup_n_u8(0);
+
+   while (count >= STEP_RGBA) {
+      uint32x2x4_t *temp_pointer;
+      vtmp = vld4_u32(png_ptr(uint32_t, rp));
+      vrpt = png_ptr(uint8x8x4_t, &vtmp);
+      vrp = *vrpt;
+      vtmp = vld4_u32(png_ptrc(uint32_t, pp));
+      vppt = png_ptr(uint8x8x4_t, &vtmp);
+      vpp = *vppt;
+      vtmp = vld4_u32(png_ptrc(uint32_t, np));
+      vnpt = png_ptr(uint8x8x4_t, &vtmp);
+      vnp = *vnpt;
+
+      vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vpp.val[1]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]);
+      vdest.val[IND2] = vhadd_u8(vdest.val[1], vpp.val[IND2]);
+      vdest.val[IND2] = vadd_u8(vdest.val[IND2], vrp.val[IND2]);
+      vdest.val[IND3] = vhadd_u8(vdest.val[IND2], vpp.val[IND3]);
+      vdest.val[IND3] = vadd_u8(vdest.val[IND3], vrp.val[IND3]);
+
+      vdest_val = png_ldr(uint32x2x4_t, &vdest);
+      vst4_lane_u32(png_ptr(uint32_t, rp), vdest_val, 0);
+
+      vdestN.val[0] = vhadd_u8(vdestN.val[IND3], vdest.val[0]);
+      vdestN.val[0] = vadd_u8(vdestN.val[0], vnp.val[0]);
+      vdestN.val[1] = vhadd_u8(vdestN.val[0], vdest.val[1]);
+      vdestN.val[1] = vadd_u8(vdestN.val[1], vnp.val[1]);
+      vdestN.val[IND2] = vhadd_u8(vdestN.val[1], vdest.val[IND2]);
+      vdestN.val[IND2] = vadd_u8(vdestN.val[IND2], vnp.val[IND2]);
+      vdestN.val[IND3] = vhadd_u8(vdestN.val[IND2], vdest.val[IND3]);
+      vdestN.val[IND3] = vadd_u8(vdestN.val[IND3], vnp.val[IND3]);
+
+      vdest_val = png_ldr(uint32x2x4_t, &vdestN);
+      vst4_lane_u32(png_ptr(uint32_t, np), vdest_val, 0);
+
+      rp += STEP_RGBA;
+      pp += STEP_RGBA;
+      np += STEP_RGBA;
+      count -= STEP_RGBA;
+   }
+
+   if (count >= STEP_RGBA_HALF) {
+      uint32x2x2_t vtmp1;
+      uint8x8x2_t *vrpt1, *vppt1, *vnpt1;
+      uint8x8x2_t vrp1, vpp1, vnp1;
+      uint32x2x2_t *temp_pointer;
+      uint32x2x2_t vdest_val1;
+
+      vtmp1 = vld2_u32(png_ptr(uint32_t, rp));
+      vrpt1 = png_ptr(uint8x8x2_t, &vtmp1);
+      vrp1 = *vrpt1;
+      vtmp1 = vld2_u32(png_ptrc(uint32_t, pp));
+      vppt1 = png_ptr(uint8x8x2_t, &vtmp1);
+      vpp1 = *vppt1;
+      vtmp1 = vld2_u32(png_ptrc(uint32_t, np));
+      vnpt1 = png_ptr(uint8x8x2_t, &vtmp1);
+      vnp1 = *vnpt1;
+
+      vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp1.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp1.val[0]);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vpp1.val[1]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vrp1.val[1]);
+      vdest.val[IND3] = vdest.val[1];
+      vdest_val1 = png_ldr(uint32x2x2_t, &vdest);
+      vst2_lane_u32(png_ptr(uint32_t, rp), vdest_val1, 0);
+
+      vdestN.val[0] = vhadd_u8(vdestN.val[IND3], vdest.val[0]);
+      vdestN.val[0] = vadd_u8(vdestN.val[0], vnp1.val[0]);
+      vdestN.val[1] = vhadd_u8(vdestN.val[0], vdest.val[1]);
+      vdestN.val[1] = vadd_u8(vdestN.val[1], vnp1.val[1]);
+      vdestN.val[IND3] = vdestN.val[1];
+      vdest_val1 = png_ldr(uint32x2x2_t, &vdestN);
+      vst2_lane_u32(png_ptr(uint32_t, np), vdest_val1, 0);
+
+      rp += STEP_RGBA_HALF;
+      pp += STEP_RGBA_HALF;
+      np += STEP_RGBA_HALF;
+      count -= STEP_RGBA_HALF;
+   }
+
+   if (count == 0) {
+      return;
+   }
+
+   uint32x2_t vtmp2;
+   uint8x8_t *vrpt2, *vppt2, *vnpt2;
+   uint8x8_t vrp2, vpp2, vnp2;
+   uint32x2_t *temp_pointer;
+   uint32x2_t vdest_val2;
+
+   vtmp2 = vld1_u32(png_ptr(uint32_t, rp));
+   vrpt2 = png_ptr(uint8x8_t, &vtmp2);
+   vrp2 = *vrpt2;
+   vtmp2 = vld1_u32(png_ptrc(uint32_t, pp));
+   vppt2 = png_ptr(uint8x8_t, &vtmp2);
+   vpp2 = *vppt2;
+   vtmp2 = vld1_u32(png_ptrc(uint32_t, np));
+   vnpt2 = png_ptr(uint8x8_t, &vtmp2);
+   vnp2 = *vnpt2;
+
+   vdest.val[0] = vhadd_u8(vdest.val[IND3], vpp2);
+   vdest.val[0] = vadd_u8(vdest.val[0], vrp2);
+
+   vdest_val2 = png_ldr(uint32x2_t, &vdest);
+   vst1_lane_u32(png_ptr(uint32_t, rp), vdest_val2, 0);
+
+   vdestN.val[0] = vhadd_u8(vdestN.val[IND3], vdest.val[0]);
+   vdestN.val[0] = vadd_u8(vdestN.val[0], vnp2);
+
+   vdest_val2 = png_ldr(uint32x2_t, &vdestN);
+   vst1_lane_u32(png_ptr(uint32_t, np), vdest_val2, 0);
+}
+
+static uint8x8_t paeth(uint8x8_t a, uint8x8_t b, uint8x8_t c)
+{
+   uint8x8_t d, e;
+   uint16x8_t p1, pa, pb, pc;
+
+   p1 = vaddl_u8(a, b); /* a + b */
+   pc = vaddl_u8(c, c); /* c * 2 */
+   pa = vabdl_u8(b, c); /* pa */
+   pb = vabdl_u8(a, c); /* pb */
+   pc = vabdq_u16(p1, pc); /* pc */
+
+   p1 = vcleq_u16(pa, pb); /* pa <= pb */
+   pa = vcleq_u16(pa, pc); /* pa <= pc */
+   pb = vcleq_u16(pb, pc); /* pb <= pc */
+
+   p1 = vandq_u16(p1, pa); /* pa <= pb && pa <= pc */
+
+   d = vmovn_u16(pb);
+   e = vmovn_u16(p1);
+
+   d = vbsl_u8(d, b, c);
+   e = vbsl_u8(e, a, d);
+
+   return e;
+}
+
+void png_read_filter_row_paeth3_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+
+   uint8x16_t vtmp;
+   uint8x8x2_t *vrpt;
+   uint8x8x2_t vrp;
+   uint8x8_t vlast = vdup_n_u8(0);
+   uint8x8x4_t vdest;
+   vdest.val[IND3] = vdup_n_u8(0);
+
+   vtmp = vld1q_u8(rp);
+   vrpt = png_ptr(uint8x8x2_t, &vtmp);
+   vrp = *vrpt;
+
+   uint8x8x2_t *vppt;
+   uint8x8x2_t vpp;
+   uint8x8_t vtmp1, vtmp2, vtmp3;
+   uint32x2_t *temp_pointer;
+
+   png_debug(1, "in png_read_filter_row_paeth3_neon");
+
+   size_t tail_bytes = row_info->rowbytes % STEP_RGB;
+   png_byte last_byte = *rp_stop;
+   png_bytep rp_stop_new = rp_stop - tail_bytes;
+   for (; rp < rp_stop_new; pp += STEP_RGB)
+   {
+      vtmp = vld1q_u8(pp);
+      vppt = png_ptr(uint8x8x2_t, &vtmp);
+      vpp = *vppt;
+
+      vdest.val[0] = paeth(vdest.val[IND3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET3);
+      vdest.val[1] = paeth(vdest.val[0], vtmp2, vpp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET6);
+      vtmp3 = vext_u8(vpp.val[0], vpp.val[1], OFFSET6);
+      vdest.val[IND2] = paeth(vdest.val[1], vtmp3, vtmp2);
+      vdest.val[IND2] = vadd_u8(vdest.val[IND2], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
+      vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
+
+      vtmp = vld1q_u8(rp + STEP_RGB);
+      vrpt = png_ptr(uint8x8x2_t, &vtmp);
+      vrp = *vrpt;
+
+      vdest.val[IND3] = paeth(vdest.val[IND2], vtmp2, vtmp3);
+      vdest.val[IND3] = vadd_u8(vdest.val[IND3], vtmp1);
+
+      vlast = vtmp2;
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND2]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND3]), 0);
+      rp += OFFSET3;
+   }
+
+   vtmp = vld1q_u8(pp);
+   vppt = png_ptr(uint8x8x2_t, &vtmp);
+   vpp = *vppt;
+
+   if (tail_bytes == TAIL_RGB1) {
+      vdest.val[0] = paeth(vdest.val[IND3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+   } else if (tail_bytes == TAIL_RGB2) {
+      vdest.val[0] = paeth(vdest.val[IND3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET3);
+      vdest.val[1] = paeth(vdest.val[0], vtmp2, vpp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+   } else if (tail_bytes == TAIL_RGB3) {
+      vdest.val[0] = paeth(vdest.val[IND3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET3);
+      vdest.val[1] = paeth(vdest.val[0], vtmp2, vpp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET6);
+      vtmp3 = vext_u8(vpp.val[0], vpp.val[1], OFFSET6);
+      vdest.val[IND2] = paeth(vdest.val[1], vtmp3, vtmp2);
+      vdest.val[IND2] = vadd_u8(vdest.val[IND2], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND2]), 0);
+   }
+   *rp_stop = last_byte;
+}
+
+void png_read_filter_row_paeth3_x2_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+   png_bytep np = rp_stop + 1;
+
+   uint8x16_t vtmp;
+   uint8x8x2_t *vrpt;
+   uint8x8x2_t vrp;
+   uint8x8_t vlast = vdup_n_u8(0);
+   uint8x8x4_t vdest;
+   vdest.val[IND3] = vdup_n_u8(0);
+
+   vtmp = vld1q_u8(rp);
+   vrpt = png_ptr(uint8x8x2_t, &vtmp);
+   vrp = *vrpt;
+
+   uint8x8x2_t *vppt;
+   uint8x8x2_t vpp;
+   uint8x8_t vtmp1, vtmp2, vtmp3;
+   uint32x2_t *temp_pointer;
+
+   uint8x8x2_t *vnpt;
+   uint8x8x2_t vnp;
+   uint8x8_t vlastN = vdup_n_u8(0);
+   uint8x8x4_t vdestN;
+   vdestN.val[IND3] = vdup_n_u8(0);
+
+   vtmp = vld1q_u8(np);
+   vnpt = png_ptr(uint8x8x2_t, &vtmp);
+   vnp = *vnpt;
+
+   png_debug(1, "in png_read_filter_row_paeth3_x2_neon");
+
+   size_t tail_bytes = row_info->rowbytes % STEP_RGB;
+   png_byte last_byte = *rp_stop;
+   png_byte last_byte_next = *(rp_stop + row_info->rowbytes + 1);
+   png_bytep rp_stop_new = rp_stop - tail_bytes;
+
+   for (; rp < rp_stop_new; pp += STEP_RGB)
+   {
+      vtmp = vld1q_u8(pp);
+      vppt = png_ptr(uint8x8x2_t, &vtmp);
+      vpp = *vppt;
+
+      vdest.val[0] = paeth(vdest.val[IND3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET3);
+      vdest.val[1] = paeth(vdest.val[0], vtmp2, vpp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET6);
+      vtmp3 = vext_u8(vpp.val[0], vpp.val[1], OFFSET6);
+      vdest.val[IND2] = paeth(vdest.val[1], vtmp3, vtmp2);
+      vdest.val[IND2] = vadd_u8(vdest.val[IND2], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
+      vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
+
+      vtmp = vld1q_u8(rp + STEP_RGB);
+      vrpt = png_ptr(uint8x8x2_t, &vtmp);
+      vrp = *vrpt;
+
+      vdest.val[IND3] = paeth(vdest.val[IND2], vtmp2, vtmp3);
+      vdest.val[IND3] = vadd_u8(vdest.val[IND3], vtmp1);
+
+      vlast = vtmp2;
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND2]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND3]), 0);
+      rp += OFFSET3;
+
+      vdestN.val[0] = paeth(vdestN.val[IND3], vdest.val[0], vlastN);
+      vdestN.val[0] = vadd_u8(vdestN.val[0], vnp.val[0]);
+
+      vtmp1 = vext_u8(vnp.val[0], vnp.val[1], OFFSET3);
+      vdestN.val[1] = paeth(vdestN.val[0], vdest.val[1], vdest.val[0]);
+      vdestN.val[1] = vadd_u8(vdestN.val[1], vtmp1);
+
+      vtmp1 = vext_u8(vnp.val[0], vnp.val[1], OFFSET6);
+      vdestN.val[IND2] = paeth(vdestN.val[1], vdest.val[IND2], vdest.val[1]);
+      vdestN.val[IND2] = vadd_u8(vdestN.val[IND2], vtmp1);
+
+      vtmp1 = vext_u8(vnp.val[1], vnp.val[1], 1);
+
+      vtmp = vld1q_u8(np + STEP_RGB);
+      vnpt = png_ptr(uint8x8x2_t, &vtmp);
+      vnp = *vnpt;
+
+      vdestN.val[IND3] = paeth(vdestN.val[IND2], vdest.val[IND3], vdest.val[IND2]);
+      vdestN.val[IND3] = vadd_u8(vdestN.val[IND3], vtmp1);
+
+      vlastN = vdest.val[IND3];
+
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[0]), 0);
+      np += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[1]), 0);
+      np += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[IND2]), 0);
+      np += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[IND3]), 0);
+      np += OFFSET3;
+   }
+
+   vtmp = vld1q_u8(pp);
+   vppt = png_ptr(uint8x8x2_t, &vtmp);
+   vpp = *vppt;
+
+   if (tail_bytes == TAIL_RGB1) {
+      vdest.val[0] = paeth(vdest.val[IND3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+
+      vdestN.val[0] = paeth(vdestN.val[IND3], vdest.val[0], vlastN);
+      vdestN.val[0] = vadd_u8(vdestN.val[0], vnp.val[0]);
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[0]), 0);
+   } else if (tail_bytes == TAIL_RGB2) {
+      vdest.val[0] = paeth(vdest.val[IND3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET3);
+      vdest.val[1] = paeth(vdest.val[0], vtmp2, vpp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+
+      vdestN.val[0] = paeth(vdestN.val[IND3], vdest.val[0], vlastN);
+      vdestN.val[0] = vadd_u8(vdestN.val[0], vnp.val[0]);
+
+      vtmp1 = vext_u8(vnp.val[0], vnp.val[1], OFFSET3);
+      vdestN.val[1] = paeth(vdestN.val[0], vdest.val[1], vdest.val[0]);
+      vdestN.val[1] = vadd_u8(vdestN.val[1], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[0]), 0);
+      np += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[1]), 0);
+   } else if (tail_bytes == TAIL_RGB3) {
+      vdest.val[0] = paeth(vdest.val[IND3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET3);
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], OFFSET3);
+      vdest.val[1] = paeth(vdest.val[0], vtmp2, vpp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], OFFSET6);
+      vtmp3 = vext_u8(vpp.val[0], vpp.val[1], OFFSET6);
+      vdest.val[IND2] = paeth(vdest.val[1], vtmp3, vtmp2);
+      vdest.val[IND2] = vadd_u8(vdest.val[IND2], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[0]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[1]), 0);
+      rp += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, rp), png_ldr(uint32x2_t, &vdest.val[IND2]), 0);
+
+      vdestN.val[0] = paeth(vdestN.val[IND3], vdest.val[0], vlastN);
+      vdestN.val[0] = vadd_u8(vdestN.val[0], vnp.val[0]);
+
+      vtmp1 = vext_u8(vnp.val[0], vnp.val[1], OFFSET3);
+      vdestN.val[1] = paeth(vdestN.val[0], vdest.val[1], vdest.val[0]);
+      vdestN.val[1] = vadd_u8(vdestN.val[1], vtmp1);
+
+      vtmp1 = vext_u8(vnp.val[0], vnp.val[1], OFFSET6);
+      vdestN.val[IND2] = paeth(vdestN.val[1], vdest.val[IND2], vdest.val[1]);
+      vdestN.val[IND2] = vadd_u8(vdestN.val[IND2], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[0]), 0);
+      np += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[1]), 0);
+      np += OFFSET3;
+      vst1_lane_u32(png_ptr(uint32_t, np), png_ldr(uint32x2_t, &vdestN.val[IND2]), 0);
+   }
+   *rp_stop = last_byte;
+   *(rp_stop + row_info->rowbytes + 1) = last_byte_next;
+}
+
+void png_read_filter_row_paeth4_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   int count = row_info->rowbytes;
+   png_const_bytep pp = prev_row;
+
+   uint8x8_t vlast = vdup_n_u8(0);
+   uint8x8x4_t vdest;
+   vdest.val[IND3] = vdup_n_u8(0);
+
+   png_debug(1, "in png_read_filter_row_paeth4_neon");
+
+   uint32x2x4_t vtmp;
+   uint8x8x4_t *vrpt, *vppt;
+   uint8x8x4_t vrp, vpp;
+   uint32x2x4_t vdest_val;
+   while (count >= STEP_RGBA) {
+      uint32x2x4_t *temp_pointer;
+      vtmp = vld4_u32(png_ptr(uint32_t, rp));
+      vrpt = png_ptr(uint8x8x4_t, &vtmp);
+      vrp = *vrpt;
+      vtmp = vld4_u32(png_ptrc(uint32_t, pp));
+      vppt = png_ptr(uint8x8x4_t, &vtmp);
+      vpp = *vppt;
+
+      vdest.val[0] = paeth(vdest.val[IND3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+      vdest.val[1] = paeth(vdest.val[0], vpp.val[1], vpp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]);
+      vdest.val[IND2] = paeth(vdest.val[1], vpp.val[IND2], vpp.val[1]);
+      vdest.val[IND2] = vadd_u8(vdest.val[IND2], vrp.val[IND2]);
+      vdest.val[IND3] = paeth(vdest.val[IND2], vpp.val[IND3], vpp.val[IND2]);
+      vdest.val[IND3] = vadd_u8(vdest.val[IND3], vrp.val[IND3]);
+
+      vlast = vpp.val[IND3];
+
+      vdest_val = png_ldr(uint32x2x4_t, &vdest);
+      vst4_lane_u32(png_ptr(uint32_t, rp), vdest_val, 0);
+
+      rp += STEP_RGBA;
+      pp += STEP_RGBA;
+      count -= STEP_RGBA;
+   }
+
+   if (count >= STEP_RGBA_HALF) {
+      uint32x2x2_t vtmp1;
+      uint8x8x2_t *vrpt1, *vppt1;
+      uint8x8x2_t vrp1, vpp1;
+      uint32x2x2_t *temp_pointer;
+      uint32x2x2_t vdest_val1;
+
+      vtmp1 = vld2_u32(png_ptr(uint32_t, rp));
+      vrpt1 = png_ptr(uint8x8x2_t, &vtmp1);
+      vrp1 = *vrpt1;
+      vtmp1 = vld2_u32(png_ptrc(uint32_t, pp));
+      vppt1 = png_ptr(uint8x8x2_t, &vtmp1);
+      vpp1 = *vppt1;
+
+      vdest.val[0] = paeth(vdest.val[IND3], vpp1.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp1.val[0]);
+      vdest.val[1] = paeth(vdest.val[0], vpp1.val[1], vpp1.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vrp1.val[1]);
+      vlast = vpp1.val[1];
+
+      vdest_val1 = png_ldr(uint32x2x2_t, &vdest);
+      vst2_lane_u32(png_ptr(uint32_t, rp), vdest_val1, 0);
+      vdest.val[IND3] = vdest.val[1];
+
+      rp += STEP_RGBA_HALF;
+      pp += STEP_RGBA_HALF;
+      count -= STEP_RGBA_HALF;
+   }
+
+   if (count == 0) {
+      return;
+   }
+
+   uint32x2_t vtmp2;
+   uint8x8_t *vrpt2, *vppt2;
+   uint8x8_t vrp2, vpp2;
+   uint32x2_t *temp_pointer;
+   uint32x2_t vdest_val2;
+
+   vtmp2 = vld1_u32(png_ptr(uint32_t, rp));
+   vrpt2 = png_ptr(uint8x8_t, &vtmp2);
+   vrp2 = *vrpt2;
+   vtmp2 = vld1_u32(png_ptrc(uint32_t, pp));
+   vppt2 = png_ptr(uint8x8_t, &vtmp2);
+   vpp2 = *vppt2;
+
+   vdest.val[0] = paeth(vdest.val[IND3], vpp2, vlast);
+   vdest.val[0] = vadd_u8(vdest.val[0], vrp2);
+
+   vdest_val2 = png_ldr(uint32x2_t, &vdest);
+   vst1_lane_u32(png_ptr(uint32_t, rp), vdest_val2, 0);
+}
+
+void png_read_filter_row_paeth4_x2_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   int count = row_info->rowbytes;
+   png_const_bytep pp = prev_row;
+   png_bytep np = row + row_info->rowbytes + 1;
+
+   uint8x8_t vlast = vdup_n_u8(0);
+   uint8x8x4_t vdest;
+   vdest.val[IND3] = vdup_n_u8(0);
+
+   png_debug(1, "in png_read_filter_row_paeth4_x2_neon");
+
+   uint32x2x4_t vtmp;
+   uint8x8x4_t *vrpt, *vppt;
+   uint8x8x4_t vrp, vpp;
+   uint32x2x4_t vdest_val;
+
+   uint8x8x4_t *vnpt;
+   uint8x8x4_t vnp;
+   uint8x8_t vlastN = vdup_n_u8(0);
+   uint8x8x4_t vdestN;
+   vdestN.val[IND3] = vdup_n_u8(0);
+
+   while (count >= STEP_RGBA) {
+      uint32x2x4_t *temp_pointer;
+      vtmp = vld4_u32(png_ptr(uint32_t, rp));
+      vrpt = png_ptr(uint8x8x4_t, &vtmp);
+      vrp = *vrpt;
+      vtmp = vld4_u32(png_ptrc(uint32_t, pp));
+      vppt = png_ptr(uint8x8x4_t, &vtmp);
+      vpp = *vppt;
+      vtmp = vld4_u32(png_ptrc(uint32_t, np));
+      vnpt = png_ptr(uint8x8x4_t, &vtmp);
+      vnp = *vnpt;
+
+      vdest.val[0] = paeth(vdest.val[IND3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+      vdest.val[1] = paeth(vdest.val[0], vpp.val[1], vpp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]);
+      vdest.val[IND2] = paeth(vdest.val[1], vpp.val[IND2], vpp.val[1]);
+      vdest.val[IND2] = vadd_u8(vdest.val[IND2], vrp.val[IND2]);
+      vdest.val[IND3] = paeth(vdest.val[IND2], vpp.val[IND3], vpp.val[IND2]);
+      vdest.val[IND3] = vadd_u8(vdest.val[IND3], vrp.val[IND3]);
+
+      vlast = vpp.val[IND3];
+
+      vdest_val = png_ldr(uint32x2x4_t, &vdest);
+      vst4_lane_u32(png_ptr(uint32_t, rp), vdest_val, 0);
+
+      vdestN.val[0] = paeth(vdestN.val[IND3], vdest.val[0], vlastN);
+      vdestN.val[0] = vadd_u8(vdestN.val[0], vnp.val[0]);
+      vdestN.val[1] = paeth(vdestN.val[0], vdest.val[1], vdest.val[0]);
+      vdestN.val[1] = vadd_u8(vdestN.val[1], vnp.val[1]);
+      vdestN.val[IND2] = paeth(vdestN.val[1], vdest.val[IND2], vdest.val[1]);
+      vdestN.val[IND2] = vadd_u8(vdestN.val[IND2], vnp.val[IND2]);
+      vdestN.val[IND3] = paeth(vdestN.val[IND2], vdest.val[IND3], vdest.val[IND2]);
+      vdestN.val[IND3] = vadd_u8(vdestN.val[IND3], vnp.val[IND3]);
+
+      vlastN = vdest.val[IND3];
+
+      vdest_val = png_ldr(uint32x2x4_t, &vdestN);
+      vst4_lane_u32(png_ptr(uint32_t, np), vdest_val, 0);
+
+      rp += STEP_RGBA;
+      pp += STEP_RGBA;
+      np += STEP_RGBA;
+      count -= STEP_RGBA;
+   }
+
+   if (count >= STEP_RGBA_HALF) {
+      uint32x2x2_t vtmp1;
+      uint8x8x2_t *vrpt1, *vppt1, *vnpt1;
+      uint8x8x2_t vrp1, vpp1, vnp1;
+      uint32x2x2_t *temp_pointer;
+      uint32x2x2_t vdest_val1;
+
+      vtmp1 = vld2_u32(png_ptr(uint32_t, rp));
+      vrpt1 = png_ptr(uint8x8x2_t, &vtmp1);
+      vrp1 = *vrpt1;
+      vtmp1 = vld2_u32(png_ptrc(uint32_t, pp));
+      vppt1 = png_ptr(uint8x8x2_t, &vtmp1);
+      vpp1 = *vppt1;
+      vtmp1 = vld2_u32(png_ptrc(uint32_t, np));
+      vnpt1 = png_ptr(uint8x8x2_t, &vtmp1);
+      vnp1 = *vnpt1;
+
+      vdest.val[0] = paeth(vdest.val[IND3], vpp1.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp1.val[0]);
+      vdest.val[1] = paeth(vdest.val[0], vpp1.val[1], vpp1.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vrp1.val[1]);
+
+      vlast = vpp1.val[1];
+
+      vdest_val1 = png_ldr(uint32x2x2_t, &vdest);
+      vst2_lane_u32(png_ptr(uint32_t, rp), vdest_val1, 0);
+
+      vdest.val[IND3] = vdest.val[1];
+
+      vdestN.val[0] = paeth(vdestN.val[IND3], vdest.val[0], vlastN);
+      vdestN.val[0] = vadd_u8(vdestN.val[0], vnp1.val[0]);
+      vdestN.val[1] = paeth(vdestN.val[0], vdest.val[1], vdest.val[0]);
+      vdestN.val[1] = vadd_u8(vdestN.val[1], vnp1.val[1]);
+
+      vlastN = vdest.val[1];
+
+      vdest_val1 = png_ldr(uint32x2x2_t, &vdestN);
+      vst2_lane_u32(png_ptr(uint32_t, np), vdest_val1, 0);
+
+      vdestN.val[IND3] = vdestN.val[1];
+
+      rp += STEP_RGBA_HALF;
+      pp += STEP_RGBA_HALF;
+      np += STEP_RGBA_HALF;
+      count -= STEP_RGBA_HALF;
+   }
+
+   if (count == 0) {
+      return;
+   }
+
+   uint32x2_t vtmp2;
+   uint8x8_t *vrpt2, *vppt2, *vnpt2;
+   uint8x8_t vrp2, vpp2, vnp2;
+   uint32x2_t *temp_pointer;
+   uint32x2_t vdest_val2;
+
+   vtmp2 = vld1_u32(png_ptr(uint32_t, rp));
+   vrpt2 = png_ptr(uint8x8_t, &vtmp2);
+   vrp2 = *vrpt2;
+   vtmp2 = vld1_u32(png_ptrc(uint32_t, pp));
+   vppt2 = png_ptr(uint8x8_t, &vtmp2);
+   vpp2 = *vppt2;
+   vtmp2 = vld1_u32(png_ptrc(uint32_t, np));
+   vnpt2 = png_ptr(uint8x8_t, &vtmp2);
+   vnp2 = *vnpt2;
+
+   vdest.val[0] = paeth(vdest.val[IND3], vpp2, vlast);
+   vdest.val[0] = vadd_u8(vdest.val[0], vrp2);
+
+   vdest_val2 = png_ldr(uint32x2_t, &vdest);
+   vst1_lane_u32(png_ptr(uint32_t, rp), vdest_val2, 0);
+
+   vdestN.val[0] = paeth(vdestN.val[IND3], vdest.val[0], vlastN);
+   vdestN.val[0] = vadd_u8(vdestN.val[0], vnp2);
+
+   vdest_val2 = png_ldr(uint32x2_t, &vdestN);
+   vst1_lane_u32(png_ptr(uint32_t, np), vdest_val2, 0);
+}
+#endif /* PNG_MULTY_LINE_ENABLE */
 #endif /* PNG_ARM_NEON_OPT > 0 */
 #endif /* PNG_ARM_NEON_IMPLEMENTATION == 1 (intrinsics) */
 #endif /* READ */
diff --git a/pngpread.c b/pngpread.c
index e283627b7..bb12f61ea 100644
--- a/pngpread.c
+++ b/pngpread.c
@@ -264,9 +264,22 @@ png_push_read_chunk(png_structrp png_ptr, png_inforp info_ptr)
       png_ptr->idat_size = png_ptr->push_length;
       png_ptr->process_mode = PNG_READ_IDAT_MODE;
       png_push_have_info(png_ptr, info_ptr);
-      png_ptr->zstream.avail_out =
-          (uInt) PNG_ROWBYTES(png_ptr->pixel_depth,
-          png_ptr->iwidth) + 1;
+#ifdef PNG_MULTY_LINE_ENABLE
+      // OH ISSUE: png optimize
+      if (png_ptr->interlaced == 0 && png_ptr->bit_depth == 8 &&
+         (png_ptr->transformations & PNG_CHECK) == 0) {
+         int rest = png_ptr->num_rows - png_ptr->row_number;
+         int row_num = rest < PNG_INFLATE_ROWS ? rest : PNG_INFLATE_ROWS;
+         png_ptr->zstream.avail_out = (uInt)(PNG_ROWBYTES(png_ptr->pixel_depth,
+             png_ptr->iwidth) + 1) * row_num;
+      }
+      else
+#endif
+      {
+         png_ptr->zstream.avail_out =
+            (uInt) PNG_ROWBYTES(png_ptr->pixel_depth,
+            png_ptr->iwidth) + 1;
+      }
       png_ptr->zstream.next_out = png_ptr->row_buf;
       return;
    }
@@ -623,6 +636,92 @@ png_push_read_IDAT(png_structrp png_ptr)
    }
 }
 
+#ifdef PNG_MULTY_LINE_ENABLE
+// OH ISSUE: png optimize
+static void png_push_process_row_x2(png_structrp png_ptr,
+   png_row_info row_info_in)
+{
+   png_debug(1, "in png_push_process_row_x2");
+   png_row_info row_info = row_info_in;
+   png_read_filter_row(png_ptr, &row_info, png_ptr->row_buf + 1,
+      png_ptr->prev_row + 1, png_ptr->row_buf[0] + 4);
+
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+   if (png_ptr->transformations != 0)
+      png_do_read_transformations(png_ptr, &row_info);
+#endif
+
+   if (png_ptr->transformed_pixel_depth == 0)
+   {
+      png_ptr->transformed_pixel_depth = row_info.pixel_depth;
+      if (row_info.pixel_depth > png_ptr->maximum_pixel_depth)
+         png_error(png_ptr, "progressive row overflow");
+   }
+
+   png_push_have_row(png_ptr, png_ptr->row_buf + 1);
+   png_read_push_finish_row(png_ptr);
+
+   png_ptr->row_buf = png_ptr->row_buf + png_ptr->rowbytes + 1;
+
+   // do it again
+   if (png_ptr->transformations != 0)
+   {
+      memcpy(png_ptr->prev_row, png_ptr->row_buf, row_info.rowbytes + 1);
+   }
+   else
+   {
+      png_ptr->prev_row = png_ptr->row_buf;
+   }
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+   if (png_ptr->transformations != 0)
+      png_do_read_transformations(png_ptr, &row_info);
+#endif
+
+   png_push_have_row(png_ptr, png_ptr->row_buf + 1);
+   png_read_push_finish_row(png_ptr);
+}
+
+static void png_push_process_multi_rows(png_structrp png_ptr, int row_num)
+{
+   png_debug(1, "in png_push_process_multi_rows");
+   uInt row_bytes =  png_ptr->rowbytes + 1;
+
+   png_row_info row_info;
+   row_info.width = png_ptr->iwidth;
+   row_info.color_type = png_ptr->color_type;
+   row_info.bit_depth = png_ptr->bit_depth;
+   row_info.channels = png_ptr->channels;
+   row_info.pixel_depth = png_ptr->pixel_depth;
+   row_info.rowbytes = png_ptr->rowbytes;
+
+   png_bytep temp_row = png_ptr->row_buf;
+   png_bytep temp_prev_row = png_ptr->prev_row;
+
+   for (int i = 0; i < row_num; i++) {
+      // check if the x2_filter is effective: only supports channels 3 or 4
+      if ((png_ptr->channels == 3 || png_ptr->channels == 4) &&
+          i < row_num -1 && png_ptr->row_buf[0] > PNG_FILTER_VALUE_SUB &&
+          png_ptr->row_buf[0] < PNG_FILTER_VALUE_LAST &&
+          png_ptr->row_buf[0] == png_ptr->row_buf[row_bytes])
+      {
+         png_push_process_row_x2(png_ptr, row_info);
+         png_ptr->row_buf = png_ptr->row_buf + row_bytes;
+         i++;
+         continue;
+      }
+      png_push_process_row(png_ptr);
+      png_ptr->row_buf = png_ptr->row_buf + row_bytes;
+   }
+
+   if (png_ptr->transformations == 0 && png_ptr->interlaced == 0)
+   {
+      png_ptr->prev_row = temp_prev_row;
+      memcpy(png_ptr->prev_row, png_ptr->row_buf - row_bytes, row_bytes);
+   }
+   png_ptr->row_buf = temp_row;
+}
+#endif
+
 void /* PRIVATE */
 png_process_IDAT_data(png_structrp png_ptr, png_bytep buffer,
     size_t buffer_length)
@@ -639,6 +738,17 @@ png_process_IDAT_data(png_structrp png_ptr, png_bytep buffer,
    /* TODO: WARNING: TRUNCATION ERROR: DANGER WILL ROBINSON: */
    png_ptr->zstream.avail_in = (uInt)buffer_length;
 
+#ifdef PNG_MULTY_LINE_ENABLE
+   // OH ISSUE: png optimize
+   int row_num = 1;
+   if (png_ptr->interlaced == 0 && png_ptr->bit_depth == 8 &&
+       (png_ptr->transformations & PNG_CHECK) == 0)
+   {
+      int rest = png_ptr->num_rows - png_ptr->row_number;
+      row_num = rest < PNG_INFLATE_ROWS ? rest : PNG_INFLATE_ROWS;
+   }
+#endif
+
    /* Keep going until the decompressed data is all processed
     * or the stream marked as finished.
     */
@@ -655,9 +765,20 @@ png_process_IDAT_data(png_structrp png_ptr, png_bytep buffer,
       if (!(png_ptr->zstream.avail_out > 0))
       {
          /* TODO: WARNING: TRUNCATION ERROR: DANGER WILL ROBINSON: */
+#ifdef PNG_MULTY_LINE_ENABLE
+         // OH ISSUE: png optimize
+         if (png_ptr->interlaced == 0 && png_ptr->bit_depth == 8 &&
+             (png_ptr->transformations & PNG_CHECK) == 0)
+         {
+            int rest = png_ptr->num_rows - png_ptr->row_number;
+            row_num = rest < PNG_INFLATE_ROWS ? rest : PNG_INFLATE_ROWS;
+         }
+         png_ptr->zstream.avail_out = (uInt)(PNG_ROWBYTES(png_ptr->pixel_depth,
+             png_ptr->iwidth) + 1) * row_num;
+#else
          png_ptr->zstream.avail_out = (uInt)(PNG_ROWBYTES(png_ptr->pixel_depth,
              png_ptr->iwidth) + 1);
-
+#endif
          png_ptr->zstream.next_out = png_ptr->row_buf;
       }
 
@@ -719,7 +840,12 @@ png_process_IDAT_data(png_structrp png_ptr, png_bytep buffer,
 
          /* Do we have a complete row? */
          if (png_ptr->zstream.avail_out == 0)
+#ifdef PNG_MULTY_LINE_ENABLE
+            // OH ISSUE: png optimize
+            png_push_process_multi_rows(png_ptr, row_num);
+#else
             png_push_process_row(png_ptr);
+#endif
       }
 
       /* And check for the end of the stream. */
@@ -738,6 +864,7 @@ png_process_IDAT_data(png_structrp png_ptr, png_bytep buffer,
 void /* PRIVATE */
 png_push_process_row(png_structrp png_ptr)
 {
+   png_debug(1, "in png_push_process_row");
    /* 1.5.6: row_info moved out of png_struct to a local here. */
    png_row_info row_info;
 
@@ -762,8 +889,17 @@ png_push_process_row(png_structrp png_ptr)
     * it may not be in the future, so this was changed just to copy the
     * interlaced row count:
     */
-   memcpy(png_ptr->prev_row, png_ptr->row_buf, row_info.rowbytes + 1);
-
+#ifdef PNG_MULTY_LINE_ENABLE
+   // OH ISSUE: png optimize
+   if (png_ptr->transformations == 0 && png_ptr->interlaced == 0)
+   {
+      png_ptr->prev_row = png_ptr->row_buf;
+   }
+   else
+#endif
+   {
+      memcpy(png_ptr->prev_row, png_ptr->row_buf, row_info.rowbytes + 1);
+   }
 #ifdef PNG_READ_TRANSFORMS_SUPPORTED
    if (png_ptr->transformations != 0)
       png_do_read_transformations(png_ptr, &row_info);
diff --git a/pngpriv.h b/pngpriv.h
index fb521cf00..81300fbd8 100644
--- a/pngpriv.h
+++ b/pngpriv.h
@@ -189,6 +189,19 @@
 #     define PNG_ARM_NEON_IMPLEMENTATION 0
 #endif /* PNG_ARM_NEON_OPT > 0 */
 
+#if defined(PNG_ARM_NEON_IMPLEMENTATION) && defined(PNG_ARM_NEON)
+// OH ISSUE: png optimize
+#  if PNG_ARM_NEON_IMPLEMENTATION == 1
+#    define PNG_MULTY_LINE_ENABLE
+#    define PNG_INFLATE_MAX_SIZE (65536)
+#    define PNG_INFLATE_ROWS (50)
+#    define PNG_CHECK (PNG_EXPAND | PNG_STRIP_ALPHA | PNG_RGB_TO_GRAY | PNG_ENCODE_ALPHA | \
+       PNG_PACKSWAP | PNG_GRAY_TO_RGB | PNG_COMPOSE | PNG_SCALE_16_TO_8 | PNG_16_TO_8 | \
+       PNG_BACKGROUND_EXPAND | PNG_EXPAND_16 | PNG_PACK | PNG_ADD_ALPHA | PNG_EXPAND_tRNS | \
+       PNG_RGB_TO_GRAY_ERR | PNG_RGB_TO_GRAY_WARN | PNG_FILLER | PNG_USER_TRANSFORM)
+#  endif
+#endif
+
 #ifndef PNG_MIPS_MSA_OPT
 #  if defined(__mips_msa) && (__mips_isa_rev >= 5) && defined(PNG_ALIGNED_MEMORY_SUPPORTED)
 #     define PNG_MIPS_MSA_OPT 2
@@ -351,8 +364,14 @@
 #endif
 
 #ifndef PNG_INTERNAL_FUNCTION
+// OH ISSUE: png optimize
+#  ifdef PNG_MULTY_LINE_ENABLE
+#    define PNG_HIDE __attribute__((visibility("hidden")))
+#  else
+#    define PNG_HIDE
+#  endif
 #  define PNG_INTERNAL_FUNCTION(type, name, args, attributes)\
-      PNG_LINKAGE_FUNCTION PNG_FUNCTION(type, name, args, PNG_EMPTY attributes)
+      PNG_LINKAGE_FUNCTION PNG_FUNCTION(type, name, args, PNG_HIDE attributes)
 #endif
 
 #ifndef PNG_INTERNAL_CALLBACK
@@ -1304,6 +1323,19 @@ PNG_INTERNAL_FUNCTION(void,png_read_filter_row_paeth3_neon,(png_row_infop
     row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
 PNG_INTERNAL_FUNCTION(void,png_read_filter_row_paeth4_neon,(png_row_infop
     row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+#ifdef PNG_MULTY_LINE_ENABLE
+// OH ISSUE: png optimize
+PNG_INTERNAL_FUNCTION(void, png_read_filter_row_up_x2_neon, (png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row), PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void, png_read_filter_row_avg3_x2_neon, (png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row), PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void, png_read_filter_row_avg4_x2_neon, (png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row), PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void, png_read_filter_row_paeth3_x2_neon, (png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row), PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void, png_read_filter_row_paeth4_x2_neon, (png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row), PNG_EMPTY);
+#endif
 #endif
 
 #if PNG_MIPS_MSA_OPT > 0
diff --git a/pngread.c b/pngread.c
index 8fa7d9f16..ed5a25307 100644
--- a/pngread.c
+++ b/pngread.c
@@ -54,7 +54,12 @@ png_create_read_struct_2,(png_const_charp user_png_ver, png_voidp error_ptr,
        * required (it will be zero in a write structure.)
        */
 #     ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+#ifdef PNG_MULTY_LINE_ENABLE
+         // OH ISSUE: png optimize
+         png_ptr->IDAT_read_size = PNG_INFLATE_MAX_SIZE;
+#else
          png_ptr->IDAT_read_size = PNG_IDAT_READ_SIZE;
+#endif
 #     endif
 
 #     ifdef PNG_BENIGN_READ_ERRORS_SUPPORTED
@@ -684,6 +689,224 @@ png_read_rows(png_structrp png_ptr, png_bytepp row,
 #endif /* SEQUENTIAL_READ */
 
 #ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+
+#ifdef PNG_MULTY_LINE_ENABLE
+// OH ISSUE: png optimize
+static void png_read_two_rows(png_structrp png_ptr, png_bytepp rows, png_uint_32 i,
+                         png_row_info row_info)
+{
+   png_debug1(1, "in png_read_two_rows %d", png_ptr->row_buf[0]);
+   png_read_filter_row(png_ptr, &row_info, png_ptr->row_buf + 1,
+      png_ptr->prev_row + 1, png_ptr->row_buf[0] + 4);
+
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+   if ((png_ptr->mng_features_permitted & PNG_FLAG_MNG_FILTER_64) != 0 &&
+      (png_ptr->filter_type == PNG_INTRAPIXEL_DIFFERENCING))
+   {
+      /* Intrapixel differencing */
+      png_do_read_intrapixel(&row_info, png_ptr->row_buf + 1);
+   }
+#endif
+
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+   if (png_ptr->transformations
+#       ifdef PNG_CHECK_FOR_INVALID_INDEX_SUPPORTED
+         || png_ptr->num_palette_max >= 0
+#       endif
+      )
+      png_do_read_transformations(png_ptr, &row_info);
+#endif
+
+   /* The transformed pixel depth should match the depth now in row_info. */
+   if (png_ptr->transformed_pixel_depth == 0)
+   {
+      png_ptr->transformed_pixel_depth = row_info.pixel_depth;
+      if (row_info.pixel_depth > png_ptr->maximum_pixel_depth)
+         png_error(png_ptr, "sequential row overflow");
+   }
+
+   else if (png_ptr->transformed_pixel_depth != row_info.pixel_depth)
+      png_error(png_ptr, "internal sequential row size calculation error");
+
+   if (rows[i] != NULL)
+      png_combine_row(png_ptr, rows[i], -1);
+
+   png_read_finish_row(png_ptr);
+
+   if (png_ptr->read_row_fn != NULL)
+      (*(png_ptr->read_row_fn))(png_ptr, png_ptr->row_number, png_ptr->pass);
+
+   png_ptr->row_buf = png_ptr->row_buf + row_info.rowbytes + 1;
+
+   // do again next line
+   memcpy(png_ptr->prev_row, png_ptr->row_buf, row_info.rowbytes + 1);
+
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+   if ((png_ptr->mng_features_permitted & PNG_FLAG_MNG_FILTER_64) != 0 &&
+      (png_ptr->filter_type == PNG_INTRAPIXEL_DIFFERENCING))
+   {
+      /* Intrapixel differencing */
+      png_do_read_intrapixel(&row_info, png_ptr->row_buf + 1);
+   }
+#endif
+
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+   if (png_ptr->transformations
+#       ifdef PNG_CHECK_FOR_INVALID_INDEX_SUPPORTED
+         || png_ptr->num_palette_max >= 0
+#       endif
+      )
+      png_do_read_transformations(png_ptr, &row_info);
+#endif
+
+   /* The transformed pixel depth should match the depth now in row_info. */
+   if (png_ptr->transformed_pixel_depth == 0)
+   {
+      png_ptr->transformed_pixel_depth = row_info.pixel_depth;
+      if (row_info.pixel_depth > png_ptr->maximum_pixel_depth)
+         png_error(png_ptr, "sequential row overflow");
+   }
+
+   else if (png_ptr->transformed_pixel_depth != row_info.pixel_depth)
+      png_error(png_ptr, "internal sequential row size calculation error");
+
+   if (rows[i+1] != NULL)
+      png_combine_row(png_ptr, rows[i+1], -1);
+
+   png_read_finish_row(png_ptr);
+
+   if (png_ptr->read_row_fn != NULL)
+      (*(png_ptr->read_row_fn))(png_ptr, png_ptr->row_number, png_ptr->pass);
+
+   png_ptr->row_buf = png_ptr->row_buf + row_info.rowbytes + 1;
+}
+
+static void png_read_muilty_rows(png_structrp png_ptr, png_bytepp rows,
+   png_uint_32 row_num, png_row_info row_info_in)
+{
+   if (png_ptr == NULL)
+      return;
+
+   png_debug2(1, "in png_read_muilty_rows (row %lu, pass %d)",
+       (unsigned long)png_ptr->row_number, png_ptr->pass);
+
+   if ((png_ptr->mode & PNG_HAVE_IDAT) == 0)
+         png_error(png_ptr, "Invalid attempt to read row data");
+
+   /* Fill the row with IDAT data: */
+   uInt row_bytes =  row_info_in.rowbytes;
+   png_ptr->row_buf[0]=255; /* 255 to force error if no data was found */
+   png_read_IDAT_data(png_ptr, png_ptr->row_buf, (row_bytes + 1) * row_num);
+   png_bytep temp_row = png_ptr->row_buf;
+
+   for (png_uint_32 i = 0; i < row_num; i++) {
+      png_row_info row_info = row_info_in;
+      // check if the x2_filter is effective: only supports channels 3 or 4
+      if ((row_info_in.channels == 3 || row_info_in.channels == 4) &&
+          i < row_num -1 && png_ptr->row_buf[0] > PNG_FILTER_VALUE_SUB &&
+          png_ptr->row_buf[0] < PNG_FILTER_VALUE_LAST &&
+          png_ptr->row_buf[0] == png_ptr->row_buf[row_info_in.rowbytes + 1])
+      {
+         png_read_two_rows(png_ptr, rows, i, row_info);
+         i++;
+         continue;
+      }
+      if (png_ptr->row_buf[0] > PNG_FILTER_VALUE_NONE)
+      {
+         if (png_ptr->row_buf[0] < PNG_FILTER_VALUE_LAST)
+            png_read_filter_row(png_ptr, &row_info, png_ptr->row_buf + 1,
+               png_ptr->prev_row + 1, png_ptr->row_buf[0]);
+         else
+            png_debug1(1, "bad adaptive filter value %d", png_ptr->row_buf[0]);
+      }
+
+      memcpy(png_ptr->prev_row, png_ptr->row_buf, row_info_in.rowbytes + 1);
+
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+      if ((png_ptr->mng_features_permitted & PNG_FLAG_MNG_FILTER_64) != 0 &&
+         (png_ptr->filter_type == PNG_INTRAPIXEL_DIFFERENCING))
+      {
+         /* Intrapixel differencing */
+         png_do_read_intrapixel(&row_info, png_ptr->row_buf + 1);
+      }
+#endif
+
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+      if (png_ptr->transformations
+#        ifdef PNG_CHECK_FOR_INVALID_INDEX_SUPPORTED
+            || png_ptr->num_palette_max >= 0
+#        endif
+         )
+         png_do_read_transformations(png_ptr, &row_info);
+#endif
+
+      /* The transformed pixel depth should match the depth now in row_info. */
+      if (png_ptr->transformed_pixel_depth == 0)
+      {
+         png_ptr->transformed_pixel_depth = row_info.pixel_depth;
+         if (row_info.pixel_depth > png_ptr->maximum_pixel_depth)
+            png_error(png_ptr, "sequential row overflow");
+      }
+
+      else if (png_ptr->transformed_pixel_depth != row_info.pixel_depth)
+         png_error(png_ptr, "internal sequential row size calculation error");
+
+      if (rows[i] != NULL)
+         png_combine_row(png_ptr, rows[i], -1);
+
+      png_read_finish_row(png_ptr);
+
+      if (png_ptr->read_row_fn != NULL)
+         (*(png_ptr->read_row_fn))(png_ptr, png_ptr->row_number, png_ptr->pass);
+
+      png_ptr->row_buf = png_ptr->row_buf + row_bytes + 1;
+   }
+   png_ptr->row_buf = temp_row;
+}
+
+static void png_warn_check(png_structrp png_ptr)
+{
+#ifdef PNG_WARNINGS_SUPPORTED
+   /* Check for transforms that have been set but were defined out */
+#if defined(PNG_WRITE_INVERT_SUPPORTED) && !defined(PNG_READ_INVERT_SUPPORTED)
+   if ((png_ptr->transformations & PNG_INVERT_MONO) != 0)
+      png_warning(png_ptr, "PNG_READ_INVERT_SUPPORTED is not defined");
+#endif
+
+#if defined(PNG_WRITE_FILLER_SUPPORTED) && !defined(PNG_READ_FILLER_SUPPORTED)
+   if ((png_ptr->transformations & PNG_FILLER) != 0)
+      png_warning(png_ptr, "PNG_READ_FILLER_SUPPORTED is not defined");
+#endif
+
+#if defined(PNG_WRITE_PACKSWAP_SUPPORTED) && \
+    !defined(PNG_READ_PACKSWAP_SUPPORTED)
+   if ((png_ptr->transformations & PNG_PACKSWAP) != 0)
+      png_warning(png_ptr, "PNG_READ_PACKSWAP_SUPPORTED is not defined");
+#endif
+
+#if defined(PNG_WRITE_PACK_SUPPORTED) && !defined(PNG_READ_PACK_SUPPORTED)
+   if ((png_ptr->transformations & PNG_PACK) != 0)
+      png_warning(png_ptr, "PNG_READ_PACK_SUPPORTED is not defined");
+#endif
+
+#if defined(PNG_WRITE_SHIFT_SUPPORTED) && !defined(PNG_READ_SHIFT_SUPPORTED)
+   if ((png_ptr->transformations & PNG_SHIFT) != 0)
+      png_warning(png_ptr, "PNG_READ_SHIFT_SUPPORTED is not defined");
+#endif
+
+#if defined(PNG_WRITE_BGR_SUPPORTED) && !defined(PNG_READ_BGR_SUPPORTED)
+   if ((png_ptr->transformations & PNG_BGR) != 0)
+      png_warning(png_ptr, "PNG_READ_BGR_SUPPORTED is not defined");
+#endif
+
+#if defined(PNG_WRITE_SWAP_SUPPORTED) && !defined(PNG_READ_SWAP_SUPPORTED)
+   if ((png_ptr->transformations & PNG_SWAP_BYTES) != 0)
+      png_warning(png_ptr, "PNG_READ_SWAP_SUPPORTED is not defined");
+#endif
+#endif /* WARNINGS */
+}
+#endif // PNG_MULTY_LINE_ENABLE
+
 /* Read the entire image.  If the image has an alpha channel or a tRNS
  * chunk, and you have called png_handle_alpha()[*], you will need to
  * initialize the image to the current image that PNG will be overlaying.
@@ -745,13 +968,45 @@ png_read_image(png_structrp png_ptr, png_bytepp image)
 
    image_height=png_ptr->height;
 
-   for (j = 0; j < pass; j++)
-   {
+#ifdef PNG_MULTY_LINE_ENABLE
+   // OH ISSUE: png optimize
+   if (png_ptr->interlaced == 0 && png_ptr->bit_depth == 8 &&
+       (png_ptr->transformations & PNG_CHECK) == 0) {
+      if ((png_ptr->flags & PNG_FLAG_ROW_INIT) == 0)
+         png_read_start_row(png_ptr);
+
+      png_warn_check(png_ptr);
+      png_row_info row_info;
+      row_info.width = png_ptr->iwidth;
+      row_info.color_type = png_ptr->color_type;
+      row_info.bit_depth = png_ptr->bit_depth;
+      row_info.channels = png_ptr->channels;
+      row_info.pixel_depth = png_ptr->pixel_depth;
+      row_info.rowbytes = png_ptr->rowbytes;
+
       rp = image;
-      for (i = 0; i < image_height; i++)
+      int row_num = PNG_INFLATE_ROWS;
+      for (i = 0; i < image_height; i += PNG_INFLATE_ROWS)
       {
-         png_read_row(png_ptr, *rp, NULL);
-         rp++;
+         if (image_height - i < PNG_INFLATE_ROWS)
+         {
+            row_num = image_height - i;
+         }
+         png_read_muilty_rows(png_ptr, rp, row_num, row_info);
+         rp += row_num;
+      }
+   }
+   else
+#endif
+   {
+      for (j = 0; j < pass; j++)
+      {
+         rp = image;
+         for (i = 0; i < image_height; i++)
+         {
+            png_read_row(png_ptr, *rp, NULL);
+            rp++;
+         }
       }
    }
 }
diff --git a/pngrutil.c b/pngrutil.c
index 9ac8ec11f..f9c65927d 100644
--- a/pngrutil.c
+++ b/pngrutil.c
@@ -4134,7 +4134,12 @@ png_read_filter_row(png_structrp pp, png_row_infop row_info, png_bytep row,
     * PNG_FILTER_OPTIMIZATIONS to a function that overrides the generic
     * implementations.  See png_init_filter_functions above.
     */
+#ifdef PNG_MULTY_LINE_ENABLE
+   // OH ISSUE: png optimize
+   if (filter > PNG_FILTER_VALUE_NONE && filter < PNG_FILTER_VALUE_LAST_X2)
+#else
    if (filter > PNG_FILTER_VALUE_NONE && filter < PNG_FILTER_VALUE_LAST)
+#endif
    {
       if (pp->read_filter[0] == NULL)
          png_init_filter_functions(pp);
@@ -4606,7 +4611,24 @@ defined(PNG_USER_TRANSFORM_PTR_SUPPORTED)
              row_bytes + 48);
 
       else
+      {
+#ifdef PNG_MULTY_LINE_ENABLE
+         // OH ISSUE: png optimize
+         png_uint_32 row_num = 1;
+         if (png_ptr->bit_depth == 8 &&
+             (png_ptr->transformations & PNG_CHECK) == 0)
+         {
+            row_num = png_ptr->height < PNG_INFLATE_ROWS ?
+               png_ptr->height : PNG_INFLATE_ROWS;
+         }
+         png_ptr->big_row_buf = (png_bytep)png_malloc(
+            png_ptr, row_bytes * row_num + 48);
+         if (png_ptr->big_row_buf == NULL)
+            png_error(png_ptr, "png_malloc failed");
+#else
          png_ptr->big_row_buf = (png_bytep)png_malloc(png_ptr, row_bytes + 48);
+#endif
+      }
 
       png_ptr->big_prev_row = (png_bytep)png_malloc(png_ptr, row_bytes + 48);
 
diff --git a/pngstruct.h b/pngstruct.h
index e591d94d5..7c3846475 100644
--- a/pngstruct.h
+++ b/pngstruct.h
@@ -140,6 +140,14 @@ typedef const png_colorspace * PNG_RESTRICT png_const_colorspacerp;
 #define PNG_COLORSPACE_CANCEL(flags)        (0xffff ^ (flags))
 #endif /* COLORSPACE || GAMMA */
 
+#ifdef PNG_MULTY_LINE_ENABLE
+// OH ISSUE: png optimize
+#define PNG_FILTER_VALUE_UP_X2      (6) // PNG_FILTER_VALUE_UP + 4
+#define PNG_FILTER_VALUE_AVG_X2     (7) // PNG_FILTER_VALUE_AVG + 4
+#define PNG_FILTER_VALUE_PAETH_X2   (8) // PNG_FILTER_VALUE_PAETH + 4
+#define PNG_FILTER_VALUE_LAST_X2    (9) // PNG_FILTER_VALUE_LAST + 4
+#endif
+
 struct png_struct_def
 {
 #ifdef PNG_SETJMP_SUPPORTED
@@ -467,8 +475,14 @@ struct png_struct_def
    png_bytep big_prev_row;
 
 /* New member added in libpng-1.5.7 */
+#ifdef PNG_MULTY_LINE_ENABLE
+   // OH ISSUE: png optimize
+   void (*read_filter[PNG_FILTER_VALUE_LAST_X2 - 1])(png_row_infop row_info,
+      png_bytep row, png_const_bytep prev_row);
+#else
    void (*read_filter[PNG_FILTER_VALUE_LAST-1])(png_row_infop row_info,
       png_bytep row, png_const_bytep prev_row);
+#endif
 
 #ifdef PNG_READ_SUPPORTED
 #if defined(PNG_COLORSPACE_SUPPORTED) || defined(PNG_GAMMA_SUPPORTED)
diff --git a/pngtrans.c b/pngtrans.c
index 1100f46eb..9addf3423 100644
--- a/pngtrans.c
+++ b/pngtrans.c
@@ -13,6 +13,19 @@
 
 #include "pngpriv.h"
 
+#ifdef PNG_MULTY_LINE_ENABLE
+#  if defined(_MSC_VER) && !defined(__clang__) && defined(_M_ARM64)
+#    include <arm64_neon.h>
+#  else
+#    include <arm_neon.h>
+#  endif
+#  define STEP_GRAY (16)
+#  define STEP_GA (32)
+#  define STEP_RGB (48)
+#  define STEP_RGBA (64)
+#  define INDEX2 (2)
+#endif
+
 #if defined(PNG_READ_SUPPORTED) || defined(PNG_WRITE_SUPPORTED)
 
 #if defined(PNG_READ_BGR_SUPPORTED) || defined(PNG_WRITE_BGR_SUPPORTED)
@@ -269,13 +282,19 @@ png_do_invert(png_row_infop row_info, png_bytep row)
    if (row_info->color_type == PNG_COLOR_TYPE_GRAY)
    {
       png_bytep rp = row;
-      size_t i;
-      size_t istop = row_info->rowbytes;
-
-      for (i = 0; i < istop; i++)
+      png_bytep rp_stop = row + row_info->rowbytes;
+#ifdef PNG_MULTY_LINE_ENABLE
+      png_bytep rp_stop_neon = rp_stop - STEP_GRAY;
+      for (; rp < rp_stop_neon; rp += STEP_GRAY)
+      {
+         uint8x16_t gray = vld1q_u8(rp);
+         gray = ~gray;
+         vst1q_u8(rp, gray);
+      }
+#endif
+      for (; rp < rp_stop; rp++)
       {
          *rp = (png_byte)(~(*rp));
-         rp++;
       }
    }
 
@@ -283,13 +302,19 @@ png_do_invert(png_row_infop row_info, png_bytep row)
       row_info->bit_depth == 8)
    {
       png_bytep rp = row;
-      size_t i;
-      size_t istop = row_info->rowbytes;
-
-      for (i = 0; i < istop; i += 2)
+      png_bytep rp_stop = row + row_info->rowbytes;
+#ifdef PNG_MULTY_LINE_ENABLE
+      png_bytep rp_stop_neon = rp_stop - STEP_GA;
+      for (; rp < rp_stop_neon; rp += STEP_GA)
+      {
+         uint8x16x2_t gray_alpha = vld2q_u8(rp);
+         gray_alpha.val[0] = ~gray_alpha.val[0];
+         vst2q_u8(rp, gray_alpha);
+      }
+#endif
+      for (; rp < rp_stop; rp += 2)
       {
          *rp = (png_byte)(~(*rp));
-         rp += 2;
       }
    }
 
@@ -298,14 +323,21 @@ png_do_invert(png_row_infop row_info, png_bytep row)
       row_info->bit_depth == 16)
    {
       png_bytep rp = row;
-      size_t i;
-      size_t istop = row_info->rowbytes;
-
-      for (i = 0; i < istop; i += 4)
+      png_bytep rp_stop = row + row_info->rowbytes;
+#ifdef PNG_MULTY_LINE_ENABLE
+      png_bytep rp_stop_neon = rp_stop - STEP_RGBA;
+      for (; rp < rp_stop_neon; rp += STEP_RGBA)
+      {
+         uint8x16x4_t gray_alpha = vld4q_u8(rp);
+         gray_alpha.val[0] = ~gray_alpha.val[0];
+         gray_alpha.val[1] = ~gray_alpha.val[1];
+         vst4q_u8(rp, gray_alpha);
+      }
+#endif
+      for (; rp < rp_stop; rp += 4)
       {
          *rp = (png_byte)(~(*rp));
          *(rp + 1) = (png_byte)(~(*(rp + 1)));
-         rp += 4;
       }
    }
 #endif
@@ -323,10 +355,19 @@ png_do_swap(png_row_infop row_info, png_bytep row)
    if (row_info->bit_depth == 16)
    {
       png_bytep rp = row;
-      png_uint_32 i;
-      png_uint_32 istop= row_info->width * row_info->channels;
-
-      for (i = 0; i < istop; i++, rp += 2)
+      png_bytep rp_stop = row + row_info->rowbytes;
+#ifdef PNG_MULTY_LINE_ENABLE
+      png_bytep rp_stop_neon = rp_stop - STEP_GA;
+      for (; rp < rp_stop_neon; rp += STEP_GA)
+      {
+         uint8x16x2_t gray = vld2q_u8(rp);
+         uint8x16_t tmp = gray.val[0];
+         gray.val[0] = gray.val[1];
+         gray.val[1] = tmp;
+         vst2q_u8(rp, gray);
+      }
+#endif
+      for (; rp < rp_stop; rp += 2)
       {
 #ifdef PNG_BUILTIN_BSWAP16_SUPPORTED
          /* Feature added to libpng-1.6.11 for testing purposes, not
@@ -622,15 +663,24 @@ png_do_bgr(png_row_infop row_info, png_bytep row)
 
    if ((row_info->color_type & PNG_COLOR_MASK_COLOR) != 0)
    {
-      png_uint_32 row_width = row_info->width;
       if (row_info->bit_depth == 8)
       {
          if (row_info->color_type == PNG_COLOR_TYPE_RGB)
          {
-            png_bytep rp;
-            png_uint_32 i;
-
-            for (i = 0, rp = row; i < row_width; i++, rp += 3)
+            png_bytep rp = row;
+            png_bytep rp_stop = row + row_info->rowbytes;
+#ifdef PNG_MULTY_LINE_ENABLE
+            png_bytep rp_stop_neon = rp_stop - STEP_RGB;
+            for (; rp < rp_stop_neon; rp += STEP_RGB)
+            {
+               uint8x16x3_t bgr = vld3q_u8(rp);
+               uint8x16_t tmp = bgr.val[INDEX2];
+               bgr.val[INDEX2] = bgr.val[0];
+               bgr.val[0] = tmp;
+               vst3q_u8(rp, bgr);
+            }
+#endif
+            for (; rp < rp_stop; rp += 3)
             {
                png_byte save = *rp;
                *rp = *(rp + 2);
@@ -640,10 +690,20 @@ png_do_bgr(png_row_infop row_info, png_bytep row)
 
          else if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
          {
-            png_bytep rp;
-            png_uint_32 i;
-
-            for (i = 0, rp = row; i < row_width; i++, rp += 4)
+            png_bytep rp = row;
+            png_bytep rp_stop = row + row_info->rowbytes;
+#ifdef PNG_MULTY_LINE_ENABLE
+            png_bytep rp_stop_neon = rp_stop - STEP_RGBA;
+            for (; rp < rp_stop_neon; rp += STEP_RGBA)
+            {
+               uint8x16x4_t bgra = vld4q_u8(rp);
+               uint8x16_t tmp = bgra.val[INDEX2];
+               bgra.val[INDEX2] = bgra.val[0];
+               bgra.val[0] = tmp;
+               vst4q_u8(rp, bgra);
+            }
+#endif
+            for (; rp < rp_stop; rp += 4)
             {
                png_byte save = *rp;
                *rp = *(rp + 2);
@@ -657,10 +717,20 @@ png_do_bgr(png_row_infop row_info, png_bytep row)
       {
          if (row_info->color_type == PNG_COLOR_TYPE_RGB)
          {
-            png_bytep rp;
-            png_uint_32 i;
-
-            for (i = 0, rp = row; i < row_width; i++, rp += 6)
+            png_bytep rp = row;
+            png_bytep rp_stop = row + row_info->rowbytes;
+#ifdef PNG_MULTY_LINE_ENABLE
+            png_bytep rp_stop_neon = rp_stop - STEP_RGB;
+            for (; rp < rp_stop_neon; rp += STEP_RGB)
+            {
+               uint16x8x3_t bgr = vld3q_u16((unsigned short *)rp);
+               uint16x8_t tmp = bgr.val[INDEX2];
+               bgr.val[INDEX2] = bgr.val[0];
+               bgr.val[0] = tmp;
+               vst3q_u16((unsigned short *)rp, bgr);
+            }
+#endif
+            for (; rp < rp_stop; rp += 6)
             {
                png_byte save = *rp;
                *rp = *(rp + 4);
@@ -673,10 +743,20 @@ png_do_bgr(png_row_infop row_info, png_bytep row)
 
          else if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
          {
-            png_bytep rp;
-            png_uint_32 i;
-
-            for (i = 0, rp = row; i < row_width; i++, rp += 8)
+            png_bytep rp = row;
+            png_bytep rp_stop = row + row_info->rowbytes;
+#ifdef PNG_MULTY_LINE_ENABLE
+            png_bytep rp_stop_neon = rp_stop - STEP_RGBA;
+            for (; rp < rp_stop_neon; rp += STEP_RGBA)
+            {
+               uint16x8x4_t bgra = vld4q_u16((unsigned short *)rp);
+               uint16x8_t tmp = bgra.val[INDEX2];
+               bgra.val[INDEX2] = bgra.val[0];
+               bgra.val[0] = tmp;
+               vst4q_u16((unsigned short *)rp, bgra);
+            }
+#endif
+            for (; rp < rp_stop; rp += 8)
             {
                png_byte save = *rp;
                *rp = *(rp + 4);