diff --git a/CMakeLists.txt b/CMakeLists.txt
index 6a7dde4..6c11f03 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -548,6 +548,27 @@
 endif()
 
 if(WITH_SIMD)
+  if(CPU_TYPE STREQUAL "arm64" OR CPU_TYPE STREQUAL "arm")
+    # GCC doesn't yet have a full or optimal set of Neon intrinsics, so for
+    # performance reasons, when using GCC, we default to using the older GAS
+    # implementation of the Neon SIMD extensions for certain algorithms.
+    if(CMAKE_COMPILER_IS_GNUCC)
+      set(DEFAULT_NEON_INTRINSICS 0)
+    else()
+      set(DEFAULT_NEON_INTRINSICS 1)
+    endif()
+    option(NEON_INTRINSICS
+      "Because GCC doesn't yet have a full or optimal set of Neon intrinsics, for performance reasons, the default when building libjpeg-turbo with GCC is to continue using the older GAS implementation of the Neon SIMD extensions for certain algorithms.  Setting this option forces the full Neon intrinsics implementation to be used with all compilers.  Unsetting this option forces the hybrid GAS/intrinsics implementation to be used with all compilers."
+      ${DEFAULT_NEON_INTRINSICS})
+    boolean_number(NEON_INTRINSICS)
+    if(NEON_INTRINSICS)
+      add_definitions(-DNEON_INTRINSICS)
+      message(STATUS "Use full Neon SIMD intrinsics implementation (NEON_INTRINSICS = ${NEON_INTRINSICS})")
+    else()
+      message(STATUS "Use partial Neon SIMD intrinsics implementation (NEON_INTRINSICS = ${NEON_INTRINSICS})")
+    endif()
+  endif()
+
   add_subdirectory(simd)
 elseif(NOT WITH_12BIT)
   message(STATUS "SIMD extensions: None (WITH_SIMD = ${WITH_SIMD})")
diff --git a/jversion.h b/jversion.h
index 660f518..d1d94f5 100644
--- a/jversion.h
+++ b/jversion.h
@@ -39,7 +39,7 @@
 #define JCOPYRIGHT \
   "Copyright (C) 2009-2020 D. R. Commander\n" \
   "Copyright (C) 2015, 2020 Google, Inc.\n" \
-  "Copyright (C) 2019 Arm Limited\n" \
+  "Copyright (C) 2019-2020 Arm Limited\n" \
   "Copyright (C) 2015-2016, 2018 Matthieu Darbois\n" \
   "Copyright (C) 2011-2016 Siarhei Siamashka\n" \
   "Copyright (C) 2015 Intel Corporation\n" \
diff --git a/simd/CMakeLists.txt b/simd/CMakeLists.txt
index ce304ae..2fd98d5 100644
--- a/simd/CMakeLists.txt
+++ b/simd/CMakeLists.txt
@@ -208,7 +208,7 @@
 
 
 ###############################################################################
-# Arm (GAS)
+# Arm (Intrinsics or GAS)
 ###############################################################################
 
 elseif(CPU_TYPE STREQUAL "arm64" OR CPU_TYPE STREQUAL "arm")
@@ -265,7 +265,15 @@
 
 file(REMOVE ${CMAKE_CURRENT_BINARY_DIR}/gastest.S)
 
-add_library(simd OBJECT ${CPU_TYPE}/jsimd_neon.S ${CPU_TYPE}/jsimd.c)
+if(NEON_INTRINSICS)
+  set(SIMD_SOURCES arm/jccolor-neon.c)
+  if(BITS EQUAL 32)
+    set_source_files_properties(${SIMD_SOURCES} COMPILE_FLAGS -mfpu=neon)
+  endif()
+endif()
+
+add_library(simd OBJECT ${SIMD_SOURCES} arm/aarch${BITS}/jsimd_neon.S
+  arm/aarch${BITS}/jsimd.c)
 
 if(CMAKE_POSITION_INDEPENDENT_CODE OR ENABLE_SHARED)
   set_target_properties(simd PROPERTIES POSITION_INDEPENDENT_CODE 1)
diff --git a/simd/arm/aarch32/jccolext-neon.c b/simd/arm/aarch32/jccolext-neon.c
new file mode 100644
index 0000000..38f90c4
--- /dev/null
+++ b/simd/arm/aarch32/jccolext-neon.c
@@ -0,0 +1,146 @@
+/*
+ * jccolext-neon.c - colorspace conversion (32-bit Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* This file is included by jccolor-neon.c */
+
+
+/* RGB -> YCbCr conversion is defined by the following equations:
+ *    Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+ *    Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + 128
+ *    Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B  + 128
+ *
+ * Avoid floating point arithmetic by using shifted integer constants:
+ *    0.29899597 = 19595 * 2^-16
+ *    0.58700561 = 38470 * 2^-16
+ *    0.11399841 =  7471 * 2^-16
+ *    0.16874695 = 11059 * 2^-16
+ *    0.33125305 = 21709 * 2^-16
+ *    0.50000000 = 32768 * 2^-16
+ *    0.41868592 = 27439 * 2^-16
+ *    0.08131409 =  5329 * 2^-16
+ * These constants are defined in jccolor-neon.c
+ *
+ * We add the fixed-point equivalent of 0.5 to Cb and Cr, which effectively
+ * rounds up or down the result via integer truncation.
+ */
+
+void jsimd_rgb_ycc_convert_neon(JDIMENSION image_width, JSAMPARRAY input_buf,
+                                JSAMPIMAGE output_buf, JDIMENSION output_row,
+                                int num_rows)
+{
+  /* Pointer to RGB(X/A) input data */
+  JSAMPROW inptr;
+  /* Pointers to Y, Cb, and Cr output data */
+  JSAMPROW outptr0, outptr1, outptr2;
+
+  /* Set up conversion constants. */
+#if defined(__clang__)
+  const uint16x4x2_t consts = vld1_u16_x2(jsimd_rgb_ycc_neon_consts);
+#else
+  /* GCC does not currently support the intrinsic vld1_<type>_x2(). */
+  const uint16x4_t consts1 = vld1_u16(jsimd_rgb_ycc_neon_consts);
+  const uint16x4_t consts2 = vld1_u16(jsimd_rgb_ycc_neon_consts + 4);
+  const uint16x4x2_t consts = { { consts1, consts2 } };
+#endif
+  const uint32x4_t scaled_128_5 = vdupq_n_u32((128 << 16) + 32767);
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    output_row++;
+
+    int cols_remaining = image_width;
+    for (; cols_remaining > 0; cols_remaining -= 8) {
+
+      /* To prevent buffer overread by the vector load instructions, the last
+       * (image_width % 8) columns of data are first memcopied to a temporary
+       * buffer large enough to accommodate the vector load.
+       */
+      if (cols_remaining < 8) {
+        ALIGN(16) uint8_t tmp_buf[8 * RGB_PIXELSIZE];
+        memcpy(tmp_buf, inptr, cols_remaining * RGB_PIXELSIZE);
+        inptr = tmp_buf;
+      }
+
+#if RGB_PIXELSIZE == 4
+      uint8x8x4_t input_pixels = vld4_u8(inptr);
+#else
+      uint8x8x3_t input_pixels = vld3_u8(inptr);
+#endif
+      uint16x8_t r = vmovl_u8(input_pixels.val[RGB_RED]);
+      uint16x8_t g = vmovl_u8(input_pixels.val[RGB_GREEN]);
+      uint16x8_t b = vmovl_u8(input_pixels.val[RGB_BLUE]);
+
+      /* Compute Y = 0.29900 * R + 0.58700 * G + 0.11400 * B */
+      uint32x4_t y_low = vmull_lane_u16(vget_low_u16(r), consts.val[0], 0);
+      y_low = vmlal_lane_u16(y_low, vget_low_u16(g), consts.val[0], 1);
+      y_low = vmlal_lane_u16(y_low, vget_low_u16(b), consts.val[0], 2);
+      uint32x4_t y_high = vmull_lane_u16(vget_high_u16(r), consts.val[0], 0);
+      y_high = vmlal_lane_u16(y_high, vget_high_u16(g), consts.val[0], 1);
+      y_high = vmlal_lane_u16(y_high, vget_high_u16(b), consts.val[0], 2);
+
+      /* Compute Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + 128 */
+      uint32x4_t cb_low = scaled_128_5;
+      cb_low = vmlsl_lane_u16(cb_low, vget_low_u16(r), consts.val[0], 3);
+      cb_low = vmlsl_lane_u16(cb_low, vget_low_u16(g), consts.val[1], 0);
+      cb_low = vmlal_lane_u16(cb_low, vget_low_u16(b), consts.val[1], 1);
+      uint32x4_t cb_high = scaled_128_5;
+      cb_high = vmlsl_lane_u16(cb_high, vget_high_u16(r), consts.val[0], 3);
+      cb_high = vmlsl_lane_u16(cb_high, vget_high_u16(g), consts.val[1], 0);
+      cb_high = vmlal_lane_u16(cb_high, vget_high_u16(b), consts.val[1], 1);
+
+      /* Compute Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B  + 128 */
+      uint32x4_t cr_low = scaled_128_5;
+      cr_low = vmlal_lane_u16(cr_low, vget_low_u16(r), consts.val[1], 1);
+      cr_low = vmlsl_lane_u16(cr_low, vget_low_u16(g), consts.val[1], 2);
+      cr_low = vmlsl_lane_u16(cr_low, vget_low_u16(b), consts.val[1], 3);
+      uint32x4_t cr_high = scaled_128_5;
+      cr_high = vmlal_lane_u16(cr_high, vget_high_u16(r), consts.val[1], 1);
+      cr_high = vmlsl_lane_u16(cr_high, vget_high_u16(g), consts.val[1], 2);
+      cr_high = vmlsl_lane_u16(cr_high, vget_high_u16(b), consts.val[1], 3);
+
+      /* Descale Y values (rounding right shift) and narrow to 16-bit. */
+      uint16x8_t y_u16 = vcombine_u16(vrshrn_n_u32(y_low, 16),
+                                      vrshrn_n_u32(y_high, 16));
+      /* Descale Cb values (right shift) and narrow to 16-bit. */
+      uint16x8_t cb_u16 = vcombine_u16(vshrn_n_u32(cb_low, 16),
+                                       vshrn_n_u32(cb_high, 16));
+      /* Descale Cr values (right shift) and narrow to 16-bit. */
+      uint16x8_t cr_u16 = vcombine_u16(vshrn_n_u32(cr_low, 16),
+                                       vshrn_n_u32(cr_high, 16));
+      /* Narrow Y, Cb, and Cr values to 8-bit and store to memory.  Buffer
+       * overwrite is permitted up to the next multiple of ALIGN_SIZE bytes.
+       */
+      vst1_u8(outptr0, vmovn_u16(y_u16));
+      vst1_u8(outptr1, vmovn_u16(cb_u16));
+      vst1_u8(outptr2, vmovn_u16(cr_u16));
+
+      /* Increment pointers. */
+      inptr += (8 * RGB_PIXELSIZE);
+      outptr0 += 8;
+      outptr1 += 8;
+      outptr2 += 8;
+    }
+  }
+}
diff --git a/simd/arm/jsimd.c b/simd/arm/aarch32/jsimd.c
similarity index 98%
rename from simd/arm/jsimd.c
rename to simd/arm/aarch32/jsimd.c
index 709656c..7c73267 100644
--- a/simd/arm/jsimd.c
+++ b/simd/arm/aarch32/jsimd.c
@@ -6,6 +6,7 @@
  * Copyright (C) 2009-2011, 2013-2014, 2016, 2018, D. R. Commander.
  * Copyright (C) 2015-2016, 2018, Matthieu Darbois.
  * Copyright (C) 2019, Google LLC.
+ * Copyright (C) 2020, Arm Limited.
  *
  * Based on the x86 SIMD extension for IJG JPEG library,
  * Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -17,12 +18,12 @@
  */
 
 #define JPEG_INTERNALS
-#include "../../jinclude.h"
-#include "../../jpeglib.h"
+#include "../../../jinclude.h"
+#include "../../../jpeglib.h"
+#include "../../../jsimd.h"
+#include "../../../jdct.h"
+#include "../../../jsimddct.h"
 #include "../../jsimd.h"
-#include "../../jdct.h"
-#include "../../jsimddct.h"
-#include "../jsimd.h"
 
 #include <stdio.h>
 #include <string.h>
diff --git a/simd/arm/jsimd_neon.S b/simd/arm/aarch32/jsimd_neon.S
similarity index 99%
rename from simd/arm/jsimd_neon.S
rename to simd/arm/aarch32/jsimd_neon.S
index 30f9cc6..c85801a 100644
--- a/simd/arm/jsimd_neon.S
+++ b/simd/arm/aarch32/jsimd_neon.S
@@ -1607,6 +1607,8 @@
 .purgem do_store
 
 
+#ifndef NEON_INTRINSICS
+
 /*****************************************************************************/
 
 /*
@@ -1916,6 +1918,8 @@
 .purgem do_load
 .purgem do_store
 
+#endif  /* NEON_INTRINSICS */
+
 
 /*****************************************************************************/
 
diff --git a/simd/arm/aarch64/jccolext-neon.c b/simd/arm/aarch64/jccolext-neon.c
new file mode 100644
index 0000000..756aeda
--- /dev/null
+++ b/simd/arm/aarch64/jccolext-neon.c
@@ -0,0 +1,316 @@
+/*
+ * jccolext-neon.c - colorspace conversion (64-bit Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* This file is included by jccolor-neon.c */
+
+
+/* RGB -> YCbCr conversion is defined by the following equations:
+ *    Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+ *    Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + 128
+ *    Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B  + 128
+ *
+ * Avoid floating point arithmetic by using shifted integer constants:
+ *    0.29899597 = 19595 * 2^-16
+ *    0.58700561 = 38470 * 2^-16
+ *    0.11399841 =  7471 * 2^-16
+ *    0.16874695 = 11059 * 2^-16
+ *    0.33125305 = 21709 * 2^-16
+ *    0.50000000 = 32768 * 2^-16
+ *    0.41868592 = 27439 * 2^-16
+ *    0.08131409 =  5329 * 2^-16
+ * These constants are defined in jccolor-neon.c
+ *
+ * We add the fixed-point equivalent of 0.5 to Cb and Cr, which effectively
+ * rounds up or down the result via integer truncation.
+ */
+
+void jsimd_rgb_ycc_convert_neon(JDIMENSION image_width, JSAMPARRAY input_buf,
+                                JSAMPIMAGE output_buf, JDIMENSION output_row,
+                                int num_rows)
+{
+  /* Pointer to RGB(X/A) input data */
+  JSAMPROW inptr;
+  /* Pointers to Y, Cb, and Cr output data */
+  JSAMPROW outptr0, outptr1, outptr2;
+
+  /* Set up conversion constants. */
+  const uint16x8_t consts = vld1q_u16(jsimd_rgb_ycc_neon_consts);
+  const uint32x4_t scaled_128_5 = vdupq_n_u32((128 << 16) + 32767);
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    output_row++;
+
+    int cols_remaining = image_width;
+    for (; cols_remaining >= 16; cols_remaining -= 16) {
+
+#if RGB_PIXELSIZE == 4
+      uint8x16x4_t input_pixels = vld4q_u8(inptr);
+#else
+      uint8x16x3_t input_pixels = vld3q_u8(inptr);
+#endif
+      uint16x8_t r_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_RED]));
+      uint16x8_t g_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_GREEN]));
+      uint16x8_t b_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_BLUE]));
+      uint16x8_t r_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_RED]));
+      uint16x8_t g_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_GREEN]));
+      uint16x8_t b_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_BLUE]));
+
+      /* Compute Y = 0.29900 * R + 0.58700 * G + 0.11400 * B */
+      uint32x4_t y_ll = vmull_laneq_u16(vget_low_u16(r_l), consts, 0);
+      y_ll = vmlal_laneq_u16(y_ll, vget_low_u16(g_l), consts, 1);
+      y_ll = vmlal_laneq_u16(y_ll, vget_low_u16(b_l), consts, 2);
+      uint32x4_t y_lh = vmull_laneq_u16(vget_high_u16(r_l), consts, 0);
+      y_lh = vmlal_laneq_u16(y_lh, vget_high_u16(g_l), consts, 1);
+      y_lh = vmlal_laneq_u16(y_lh, vget_high_u16(b_l), consts, 2);
+      uint32x4_t y_hl = vmull_laneq_u16(vget_low_u16(r_h), consts, 0);
+      y_hl = vmlal_laneq_u16(y_hl, vget_low_u16(g_h), consts, 1);
+      y_hl = vmlal_laneq_u16(y_hl, vget_low_u16(b_h), consts, 2);
+      uint32x4_t y_hh = vmull_laneq_u16(vget_high_u16(r_h), consts, 0);
+      y_hh = vmlal_laneq_u16(y_hh, vget_high_u16(g_h), consts, 1);
+      y_hh = vmlal_laneq_u16(y_hh, vget_high_u16(b_h), consts, 2);
+
+      /* Compute Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + 128 */
+      uint32x4_t cb_ll = scaled_128_5;
+      cb_ll = vmlsl_laneq_u16(cb_ll, vget_low_u16(r_l), consts, 3);
+      cb_ll = vmlsl_laneq_u16(cb_ll, vget_low_u16(g_l), consts, 4);
+      cb_ll = vmlal_laneq_u16(cb_ll, vget_low_u16(b_l), consts, 5);
+      uint32x4_t cb_lh = scaled_128_5;
+      cb_lh = vmlsl_laneq_u16(cb_lh, vget_high_u16(r_l), consts, 3);
+      cb_lh = vmlsl_laneq_u16(cb_lh, vget_high_u16(g_l), consts, 4);
+      cb_lh = vmlal_laneq_u16(cb_lh, vget_high_u16(b_l), consts, 5);
+      uint32x4_t cb_hl = scaled_128_5;
+      cb_hl = vmlsl_laneq_u16(cb_hl, vget_low_u16(r_h), consts, 3);
+      cb_hl = vmlsl_laneq_u16(cb_hl, vget_low_u16(g_h), consts, 4);
+      cb_hl = vmlal_laneq_u16(cb_hl, vget_low_u16(b_h), consts, 5);
+      uint32x4_t cb_hh = scaled_128_5;
+      cb_hh = vmlsl_laneq_u16(cb_hh, vget_high_u16(r_h), consts, 3);
+      cb_hh = vmlsl_laneq_u16(cb_hh, vget_high_u16(g_h), consts, 4);
+      cb_hh = vmlal_laneq_u16(cb_hh, vget_high_u16(b_h), consts, 5);
+
+      /* Compute Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B  + 128 */
+      uint32x4_t cr_ll = scaled_128_5;
+      cr_ll = vmlal_laneq_u16(cr_ll, vget_low_u16(r_l), consts, 5);
+      cr_ll = vmlsl_laneq_u16(cr_ll, vget_low_u16(g_l), consts, 6);
+      cr_ll = vmlsl_laneq_u16(cr_ll, vget_low_u16(b_l), consts, 7);
+      uint32x4_t cr_lh = scaled_128_5;
+      cr_lh = vmlal_laneq_u16(cr_lh, vget_high_u16(r_l), consts, 5);
+      cr_lh = vmlsl_laneq_u16(cr_lh, vget_high_u16(g_l), consts, 6);
+      cr_lh = vmlsl_laneq_u16(cr_lh, vget_high_u16(b_l), consts, 7);
+      uint32x4_t cr_hl = scaled_128_5;
+      cr_hl = vmlal_laneq_u16(cr_hl, vget_low_u16(r_h), consts, 5);
+      cr_hl = vmlsl_laneq_u16(cr_hl, vget_low_u16(g_h), consts, 6);
+      cr_hl = vmlsl_laneq_u16(cr_hl, vget_low_u16(b_h), consts, 7);
+      uint32x4_t cr_hh = scaled_128_5;
+      cr_hh = vmlal_laneq_u16(cr_hh, vget_high_u16(r_h), consts, 5);
+      cr_hh = vmlsl_laneq_u16(cr_hh, vget_high_u16(g_h), consts, 6);
+      cr_hh = vmlsl_laneq_u16(cr_hh, vget_high_u16(b_h), consts, 7);
+
+      /* Descale Y values (rounding right shift) and narrow to 16-bit. */
+      uint16x8_t y_l = vcombine_u16(vrshrn_n_u32(y_ll, 16),
+                                    vrshrn_n_u32(y_lh, 16));
+      uint16x8_t y_h = vcombine_u16(vrshrn_n_u32(y_hl, 16),
+                                    vrshrn_n_u32(y_hh, 16));
+      /* Descale Cb values (right shift) and narrow to 16-bit. */
+      uint16x8_t cb_l = vcombine_u16(vshrn_n_u32(cb_ll, 16),
+                                     vshrn_n_u32(cb_lh, 16));
+      uint16x8_t cb_h = vcombine_u16(vshrn_n_u32(cb_hl, 16),
+                                     vshrn_n_u32(cb_hh, 16));
+      /* Descale Cr values (right shift) and narrow to 16-bit. */
+      uint16x8_t cr_l = vcombine_u16(vshrn_n_u32(cr_ll, 16),
+                                     vshrn_n_u32(cr_lh, 16));
+      uint16x8_t cr_h = vcombine_u16(vshrn_n_u32(cr_hl, 16),
+                                     vshrn_n_u32(cr_hh, 16));
+      /* Narrow Y, Cb, and Cr values to 8-bit and store to memory.  Buffer
+       * overwrite is permitted up to the next multiple of ALIGN_SIZE bytes.
+       */
+      vst1q_u8(outptr0, vcombine_u8(vmovn_u16(y_l), vmovn_u16(y_h)));
+      vst1q_u8(outptr1, vcombine_u8(vmovn_u16(cb_l), vmovn_u16(cb_h)));
+      vst1q_u8(outptr2, vcombine_u8(vmovn_u16(cr_l), vmovn_u16(cr_h)));
+
+      /* Increment pointers. */
+      inptr += (16 * RGB_PIXELSIZE);
+      outptr0 += 16;
+      outptr1 += 16;
+      outptr2 += 16;
+    }
+
+    if (cols_remaining > 8) {
+      /* To prevent buffer overread by the vector load instructions, the last
+       * (image_width % 16) columns of data are first memcopied to a temporary
+       * buffer large enough to accommodate the vector load.
+       */
+      ALIGN(16) uint8_t tmp_buf[16 * RGB_PIXELSIZE];
+      memcpy(tmp_buf, inptr, cols_remaining * RGB_PIXELSIZE);
+      inptr = tmp_buf;
+
+#if RGB_PIXELSIZE == 4
+      uint8x16x4_t input_pixels = vld4q_u8(inptr);
+#else
+      uint8x16x3_t input_pixels = vld3q_u8(inptr);
+#endif
+      uint16x8_t r_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_RED]));
+      uint16x8_t g_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_GREEN]));
+      uint16x8_t b_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_BLUE]));
+      uint16x8_t r_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_RED]));
+      uint16x8_t g_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_GREEN]));
+      uint16x8_t b_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_BLUE]));
+
+      /* Compute Y = 0.29900 * R + 0.58700 * G + 0.11400 * B */
+      uint32x4_t y_ll = vmull_laneq_u16(vget_low_u16(r_l), consts, 0);
+      y_ll = vmlal_laneq_u16(y_ll, vget_low_u16(g_l), consts, 1);
+      y_ll = vmlal_laneq_u16(y_ll, vget_low_u16(b_l), consts, 2);
+      uint32x4_t y_lh = vmull_laneq_u16(vget_high_u16(r_l), consts, 0);
+      y_lh = vmlal_laneq_u16(y_lh, vget_high_u16(g_l), consts, 1);
+      y_lh = vmlal_laneq_u16(y_lh, vget_high_u16(b_l), consts, 2);
+      uint32x4_t y_hl = vmull_laneq_u16(vget_low_u16(r_h), consts, 0);
+      y_hl = vmlal_laneq_u16(y_hl, vget_low_u16(g_h), consts, 1);
+      y_hl = vmlal_laneq_u16(y_hl, vget_low_u16(b_h), consts, 2);
+      uint32x4_t y_hh = vmull_laneq_u16(vget_high_u16(r_h), consts, 0);
+      y_hh = vmlal_laneq_u16(y_hh, vget_high_u16(g_h), consts, 1);
+      y_hh = vmlal_laneq_u16(y_hh, vget_high_u16(b_h), consts, 2);
+
+      /* Compute Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + 128 */
+      uint32x4_t cb_ll = scaled_128_5;
+      cb_ll = vmlsl_laneq_u16(cb_ll, vget_low_u16(r_l), consts, 3);
+      cb_ll = vmlsl_laneq_u16(cb_ll, vget_low_u16(g_l), consts, 4);
+      cb_ll = vmlal_laneq_u16(cb_ll, vget_low_u16(b_l), consts, 5);
+      uint32x4_t cb_lh = scaled_128_5;
+      cb_lh = vmlsl_laneq_u16(cb_lh, vget_high_u16(r_l), consts, 3);
+      cb_lh = vmlsl_laneq_u16(cb_lh, vget_high_u16(g_l), consts, 4);
+      cb_lh = vmlal_laneq_u16(cb_lh, vget_high_u16(b_l), consts, 5);
+      uint32x4_t cb_hl = scaled_128_5;
+      cb_hl = vmlsl_laneq_u16(cb_hl, vget_low_u16(r_h), consts, 3);
+      cb_hl = vmlsl_laneq_u16(cb_hl, vget_low_u16(g_h), consts, 4);
+      cb_hl = vmlal_laneq_u16(cb_hl, vget_low_u16(b_h), consts, 5);
+      uint32x4_t cb_hh = scaled_128_5;
+      cb_hh = vmlsl_laneq_u16(cb_hh, vget_high_u16(r_h), consts, 3);
+      cb_hh = vmlsl_laneq_u16(cb_hh, vget_high_u16(g_h), consts, 4);
+      cb_hh = vmlal_laneq_u16(cb_hh, vget_high_u16(b_h), consts, 5);
+
+      /* Compute Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B  + 128 */
+      uint32x4_t cr_ll = scaled_128_5;
+      cr_ll = vmlal_laneq_u16(cr_ll, vget_low_u16(r_l), consts, 5);
+      cr_ll = vmlsl_laneq_u16(cr_ll, vget_low_u16(g_l), consts, 6);
+      cr_ll = vmlsl_laneq_u16(cr_ll, vget_low_u16(b_l), consts, 7);
+      uint32x4_t cr_lh = scaled_128_5;
+      cr_lh = vmlal_laneq_u16(cr_lh, vget_high_u16(r_l), consts, 5);
+      cr_lh = vmlsl_laneq_u16(cr_lh, vget_high_u16(g_l), consts, 6);
+      cr_lh = vmlsl_laneq_u16(cr_lh, vget_high_u16(b_l), consts, 7);
+      uint32x4_t cr_hl = scaled_128_5;
+      cr_hl = vmlal_laneq_u16(cr_hl, vget_low_u16(r_h), consts, 5);
+      cr_hl = vmlsl_laneq_u16(cr_hl, vget_low_u16(g_h), consts, 6);
+      cr_hl = vmlsl_laneq_u16(cr_hl, vget_low_u16(b_h), consts, 7);
+      uint32x4_t cr_hh = scaled_128_5;
+      cr_hh = vmlal_laneq_u16(cr_hh, vget_high_u16(r_h), consts, 5);
+      cr_hh = vmlsl_laneq_u16(cr_hh, vget_high_u16(g_h), consts, 6);
+      cr_hh = vmlsl_laneq_u16(cr_hh, vget_high_u16(b_h), consts, 7);
+
+      /* Descale Y values (rounding right shift) and narrow to 16-bit. */
+      uint16x8_t y_l = vcombine_u16(vrshrn_n_u32(y_ll, 16),
+                                    vrshrn_n_u32(y_lh, 16));
+      uint16x8_t y_h = vcombine_u16(vrshrn_n_u32(y_hl, 16),
+                                    vrshrn_n_u32(y_hh, 16));
+      /* Descale Cb values (right shift) and narrow to 16-bit. */
+      uint16x8_t cb_l = vcombine_u16(vshrn_n_u32(cb_ll, 16),
+                                     vshrn_n_u32(cb_lh, 16));
+      uint16x8_t cb_h = vcombine_u16(vshrn_n_u32(cb_hl, 16),
+                                     vshrn_n_u32(cb_hh, 16));
+      /* Descale Cr values (right shift) and narrow to 16-bit. */
+      uint16x8_t cr_l = vcombine_u16(vshrn_n_u32(cr_ll, 16),
+                                     vshrn_n_u32(cr_lh, 16));
+      uint16x8_t cr_h = vcombine_u16(vshrn_n_u32(cr_hl, 16),
+                                     vshrn_n_u32(cr_hh, 16));
+      /* Narrow Y, Cb, and Cr values to 8-bit and store to memory.  Buffer
+       * overwrite is permitted up to the next multiple of ALIGN_SIZE bytes.
+       */
+      vst1q_u8(outptr0, vcombine_u8(vmovn_u16(y_l), vmovn_u16(y_h)));
+      vst1q_u8(outptr1, vcombine_u8(vmovn_u16(cb_l), vmovn_u16(cb_h)));
+      vst1q_u8(outptr2, vcombine_u8(vmovn_u16(cr_l), vmovn_u16(cr_h)));
+
+    } else if (cols_remaining > 0) {
+      /* To prevent buffer overread by the vector load instructions, the last
+       * (image_width % 8) columns of data are first memcopied to a temporary
+       * buffer large enough to accommodate the vector load.
+       */
+      ALIGN(16) uint8_t tmp_buf[8 * RGB_PIXELSIZE];
+      memcpy(tmp_buf, inptr, cols_remaining * RGB_PIXELSIZE);
+      inptr = tmp_buf;
+
+#if RGB_PIXELSIZE == 4
+      uint8x8x4_t input_pixels = vld4_u8(inptr);
+#else
+      uint8x8x3_t input_pixels = vld3_u8(inptr);
+#endif
+      uint16x8_t r = vmovl_u8(input_pixels.val[RGB_RED]);
+      uint16x8_t g = vmovl_u8(input_pixels.val[RGB_GREEN]);
+      uint16x8_t b = vmovl_u8(input_pixels.val[RGB_BLUE]);
+
+      /* Compute Y = 0.29900 * R + 0.58700 * G + 0.11400 * B */
+      uint32x4_t y_l = vmull_laneq_u16(vget_low_u16(r), consts, 0);
+      y_l = vmlal_laneq_u16(y_l, vget_low_u16(g), consts, 1);
+      y_l = vmlal_laneq_u16(y_l, vget_low_u16(b), consts, 2);
+      uint32x4_t y_h = vmull_laneq_u16(vget_high_u16(r), consts, 0);
+      y_h = vmlal_laneq_u16(y_h, vget_high_u16(g), consts, 1);
+      y_h = vmlal_laneq_u16(y_h, vget_high_u16(b), consts, 2);
+
+      /* Compute Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + 128 */
+      uint32x4_t cb_l = scaled_128_5;
+      cb_l = vmlsl_laneq_u16(cb_l, vget_low_u16(r), consts, 3);
+      cb_l = vmlsl_laneq_u16(cb_l, vget_low_u16(g), consts, 4);
+      cb_l = vmlal_laneq_u16(cb_l, vget_low_u16(b), consts, 5);
+      uint32x4_t cb_h = scaled_128_5;
+      cb_h = vmlsl_laneq_u16(cb_h, vget_high_u16(r), consts, 3);
+      cb_h = vmlsl_laneq_u16(cb_h, vget_high_u16(g), consts, 4);
+      cb_h = vmlal_laneq_u16(cb_h, vget_high_u16(b), consts, 5);
+
+      /* Compute Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B  + 128 */
+      uint32x4_t cr_l = scaled_128_5;
+      cr_l = vmlal_laneq_u16(cr_l, vget_low_u16(r), consts, 5);
+      cr_l = vmlsl_laneq_u16(cr_l, vget_low_u16(g), consts, 6);
+      cr_l = vmlsl_laneq_u16(cr_l, vget_low_u16(b), consts, 7);
+      uint32x4_t cr_h = scaled_128_5;
+      cr_h = vmlal_laneq_u16(cr_h, vget_high_u16(r), consts, 5);
+      cr_h = vmlsl_laneq_u16(cr_h, vget_high_u16(g), consts, 6);
+      cr_h = vmlsl_laneq_u16(cr_h, vget_high_u16(b), consts, 7);
+
+      /* Descale Y values (rounding right shift) and narrow to 16-bit. */
+      uint16x8_t y_u16 = vcombine_u16(vrshrn_n_u32(y_l, 16),
+                                      vrshrn_n_u32(y_h, 16));
+      /* Descale Cb values (right shift) and narrow to 16-bit. */
+      uint16x8_t cb_u16 = vcombine_u16(vshrn_n_u32(cb_l, 16),
+                                       vshrn_n_u32(cb_h, 16));
+      /* Descale Cr values (right shift) and narrow to 16-bit. */
+      uint16x8_t cr_u16 = vcombine_u16(vshrn_n_u32(cr_l, 16),
+                                       vshrn_n_u32(cr_h, 16));
+      /* Narrow Y, Cb, and Cr values to 8-bit and store to memory.  Buffer
+       * overwrite is permitted up to the next multiple of ALIGN_SIZE bytes.
+       */
+      vst1_u8(outptr0, vmovn_u16(y_u16));
+      vst1_u8(outptr1, vmovn_u16(cb_u16));
+      vst1_u8(outptr2, vmovn_u16(cr_u16));
+    }
+  }
+}
diff --git a/simd/arm64/jsimd.c b/simd/arm/aarch64/jsimd.c
similarity index 97%
rename from simd/arm64/jsimd.c
rename to simd/arm/aarch64/jsimd.c
index 90c3803..976bf27 100644
--- a/simd/arm64/jsimd.c
+++ b/simd/arm/aarch64/jsimd.c
@@ -3,8 +3,9 @@
  *
  * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
  * Copyright (C) 2011, Nokia Corporation and/or its subsidiary(-ies).
- * Copyright (C) 2009-2011, 2013-2014, 2016, 2018, D. R. Commander.
+ * Copyright (C) 2009-2011, 2013-2014, 2016, 2018, 2020, D. R. Commander.
  * Copyright (C) 2015-2016, 2018, Matthieu Darbois.
+ * Copyright (C) 2020, Arm Limited.
  *
  * Based on the x86 SIMD extension for IJG JPEG library,
  * Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -16,12 +17,12 @@
  */
 
 #define JPEG_INTERNALS
-#include "../../jinclude.h"
-#include "../../jpeglib.h"
+#include "../../../jinclude.h"
+#include "../../../jpeglib.h"
+#include "../../../jsimd.h"
+#include "../../../jdct.h"
+#include "../../../jsimddct.h"
 #include "../../jsimd.h"
-#include "../../jdct.h"
-#include "../../jsimddct.h"
-#include "../jsimd.h"
 #include "jconfigint.h"
 
 #include <stdio.h>
@@ -238,20 +239,28 @@
 
   switch (cinfo->in_color_space) {
   case JCS_EXT_RGB:
+#ifndef NEON_INTRINSICS
     if (simd_features & JSIMD_FASTLD3)
+#endif
       neonfct = jsimd_extrgb_ycc_convert_neon;
+#ifndef NEON_INTRINSICS
     else
       neonfct = jsimd_extrgb_ycc_convert_neon_slowld3;
+#endif
     break;
   case JCS_EXT_RGBX:
   case JCS_EXT_RGBA:
     neonfct = jsimd_extrgbx_ycc_convert_neon;
     break;
   case JCS_EXT_BGR:
+#ifndef NEON_INTRINSICS
     if (simd_features & JSIMD_FASTLD3)
+#endif
       neonfct = jsimd_extbgr_ycc_convert_neon;
+#ifndef NEON_INTRINSICS
     else
       neonfct = jsimd_extbgr_ycc_convert_neon_slowld3;
+#endif
     break;
   case JCS_EXT_BGRX:
   case JCS_EXT_BGRA:
@@ -266,10 +275,14 @@
     neonfct = jsimd_extxrgb_ycc_convert_neon;
     break;
   default:
+#ifndef NEON_INTRINSICS
     if (simd_features & JSIMD_FASTLD3)
+#endif
       neonfct = jsimd_extrgb_ycc_convert_neon;
+#ifndef NEON_INTRINSICS
     else
       neonfct = jsimd_extrgb_ycc_convert_neon_slowld3;
+#endif
     break;
   }
 
diff --git a/simd/arm64/jsimd_neon.S b/simd/arm/aarch64/jsimd_neon.S
similarity index 99%
rename from simd/arm64/jsimd_neon.S
rename to simd/arm/aarch64/jsimd_neon.S
index 933482c..fa47bd1 100644
--- a/simd/arm64/jsimd_neon.S
+++ b/simd/arm/aarch64/jsimd_neon.S
@@ -145,6 +145,8 @@
   .short -128,  -128,   -128,   -128
   .short -128,  -128,   -128,   -128
 
+#ifndef NEON_INTRINSICS
+
 /* Constants for jsimd_*_ycc_neon() */
 
 .balign 16
@@ -154,6 +156,8 @@
   .short 32767, 128, 32767, 128
   .short 32767, 128, 32767, 128
 
+#endif
+
 /* Constants for jsimd_fdct_islow_neon() */
 
 #define F_0_298   2446  /* FIX(0.298631336) */
@@ -1982,6 +1986,8 @@
 .purgem do_store
 
 
+#ifndef NEON_INTRINSICS
+
 /*****************************************************************************/
 
 /*
@@ -2291,6 +2297,8 @@
 .purgem do_load
 .purgem do_store
 
+#endif  /* NEON_INTRINSICS */
+
 
 /*****************************************************************************/
 
diff --git a/simd/arm/align.h b/simd/arm/align.h
new file mode 100644
index 0000000..cff4241
--- /dev/null
+++ b/simd/arm/align.h
@@ -0,0 +1,28 @@
+/*
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* How to obtain memory alignment for structures and variables */
+#if defined(_MSC_VER)
+#define ALIGN(alignment)  __declspec(align(alignment))
+#elif defined(__clang__) || defined(__GNUC__)
+#define ALIGN(alignment)  __attribute__((aligned(alignment)))
+#else
+#error "Unknown compiler"
+#endif
diff --git a/simd/arm/jccolor-neon.c b/simd/arm/jccolor-neon.c
new file mode 100644
index 0000000..1f8d007
--- /dev/null
+++ b/simd/arm/jccolor-neon.c
@@ -0,0 +1,158 @@
+/*
+ * jccolor-neon.c - colorspace conversion (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+
+#include <arm_neon.h>
+
+
+/* RGB -> YCbCr conversion constants */
+
+#define F_0_298  19595
+#define F_0_587  38470
+#define F_0_113  7471
+#define F_0_168  11059
+#define F_0_331  21709
+#define F_0_500  32768
+#define F_0_418  27439
+#define F_0_081  5329
+
+ALIGN(16) static const uint16_t jsimd_rgb_ycc_neon_consts[] = {
+  F_0_298, F_0_587, F_0_113, F_0_168,
+  F_0_331, F_0_500, F_0_418, F_0_081
+};
+
+
+/* Include inline routines for colorspace extensions. */
+
+#if defined(__aarch64__)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+
+#define RGB_RED  EXT_RGB_RED
+#define RGB_GREEN  EXT_RGB_GREEN
+#define RGB_BLUE  EXT_RGB_BLUE
+#define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+#define jsimd_rgb_ycc_convert_neon  jsimd_extrgb_ycc_convert_neon
+#if defined(__aarch64__)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_ycc_convert_neon
+
+#define RGB_RED  EXT_RGBX_RED
+#define RGB_GREEN  EXT_RGBX_GREEN
+#define RGB_BLUE  EXT_RGBX_BLUE
+#define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+#define jsimd_rgb_ycc_convert_neon  jsimd_extrgbx_ycc_convert_neon
+#if defined(__aarch64__)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_ycc_convert_neon
+
+#define RGB_RED  EXT_BGR_RED
+#define RGB_GREEN  EXT_BGR_GREEN
+#define RGB_BLUE  EXT_BGR_BLUE
+#define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+#define jsimd_rgb_ycc_convert_neon  jsimd_extbgr_ycc_convert_neon
+#if defined(__aarch64__)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_ycc_convert_neon
+
+#define RGB_RED  EXT_BGRX_RED
+#define RGB_GREEN  EXT_BGRX_GREEN
+#define RGB_BLUE  EXT_BGRX_BLUE
+#define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+#define jsimd_rgb_ycc_convert_neon  jsimd_extbgrx_ycc_convert_neon
+#if defined(__aarch64__)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_ycc_convert_neon
+
+#define RGB_RED  EXT_XBGR_RED
+#define RGB_GREEN  EXT_XBGR_GREEN
+#define RGB_BLUE  EXT_XBGR_BLUE
+#define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+#define jsimd_rgb_ycc_convert_neon  jsimd_extxbgr_ycc_convert_neon
+#if defined(__aarch64__)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_ycc_convert_neon
+
+#define RGB_RED  EXT_XRGB_RED
+#define RGB_GREEN  EXT_XRGB_GREEN
+#define RGB_BLUE  EXT_XRGB_BLUE
+#define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+#define jsimd_rgb_ycc_convert_neon  jsimd_extxrgb_ycc_convert_neon
+#if defined(__aarch64__)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_ycc_convert_neon
diff --git a/simd/jsimd.h b/simd/jsimd.h
index c7d805b..eb3e283 100644
--- a/simd/jsimd.h
+++ b/simd/jsimd.h
@@ -121,6 +121,8 @@
   (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
    JDIMENSION output_row, int num_rows);
 
+#ifndef NEON_INTRINSICS
+
 EXTERN(void) jsimd_extrgb_ycc_convert_neon_slowld3
   (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
    JDIMENSION output_row, int num_rows);
@@ -128,6 +130,8 @@
   (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
    JDIMENSION output_row, int num_rows);
 
+#endif
+
 EXTERN(void) jsimd_rgb_ycc_convert_dspr2
   (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
    JDIMENSION output_row, int num_rows);