diff --git a/CMakeLists.txt b/CMakeLists.txt
index 8dac532..49c4f90 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -800,29 +800,7 @@
   set(MD5_PPM_3x2_IFAST fd283664b3b49127984af0a7f118fccd)
   set(MD5_JPEG_420_ISLOW_ARI e986fb0a637a8d833d96e8a6d6d84ea1)
   set(MD5_JPEG_444_ISLOW_PROGARI 0a8f1c8f66e113c3cf635df0a475a617)
-  # Since v1.5.1, libjpeg-turbo uses the separate non-fancy upsampling and
-  # YCbCr -> RGB color conversion routines rather than merged upsampling/color
-  # conversion when fancy upsampling is disabled on platforms that have a SIMD
-  # implementation of YCbCr -> RGB color conversion but no SIMD implementation
-  # of merged upsampling/color conversion.  This was intended to improve the
-  # performance of the Arm Neon SIMD extensions, the only SIMD extensions for
-  # which those circumstances currently apply.  The separate non-fancy
-  # upsampling and color conversion routines usually produce bitwise-identical
-  # output to the merged upsampling/color conversion routines, but that is not
-  # the case when skipping scanlines starting at an odd-numbered scanline.  In
-  # libjpeg-turbo 2.0.5 and prior, doing that while using merged h2v2
-  # upsampling caused a segfault, so this test validates the fix for that
-  # segfault.  Unfortunately, however, the test also produces different bitwise
-  # output when using the Neon SIMD extensions, because of the aforementioned
-  # optimization.  The easiest workaround is to use the old test from
-  # libjpeg-turbo 2.0.5 and prior when using the Neon SIMD extensions.  The
-  # aforementioned segfault never would have occurred with the Neon SIMD
-  # extensions anyhow, since merged upsampling is disabled when using them.
-  if((CPU_TYPE STREQUAL "arm64" OR CPU_TYPE STREQUAL "arm") AND WITH_SIMD)
-    set(MD5_PPM_420M_IFAST_ARI 72b59a99bcf1de24c5b27d151bde2437)
-  else()
-    set(MD5_PPM_420M_IFAST_ARI 57251da28a35b46eecb7177d82d10e0e)
-  endif()
+  set(MD5_PPM_420M_IFAST_ARI 57251da28a35b46eecb7177d82d10e0e)
   set(MD5_JPEG_420_ISLOW 9a68f56bc76e466aa7e52f415d0f4a5f)
   set(MD5_PPM_420M_ISLOW_2_1 9f9de8c0612f8d06869b960b05abf9c9)
   set(MD5_PPM_420M_ISLOW_15_8 b6875bc070720b899566cc06459b63b7)
@@ -1223,17 +1201,9 @@
 
   if(WITH_ARITH_DEC)
     # CC: RGB->YCC  SAMP: h2v2 merged  IDCT: ifast  ENT: arith
-    if((CPU_TYPE STREQUAL "arm64" OR CPU_TYPE STREQUAL "arm") AND WITH_SIMD)
-      # Refer to the comment above the definition of MD5_PPM_420M_IFAST_ARI for
-      # an explanation of why this is necessary.
-      add_bittest(djpeg 420m-ifast-ari "-fast;-ppm"
-        testout_420m_ifast_ari.ppm ${TESTIMAGES}/testimgari.jpg
-        ${MD5_PPM_420M_IFAST_ARI})
-    else()
-      add_bittest(djpeg 420m-ifast-ari "-fast;-skip;1,20;-ppm"
-        testout_420m_ifast_ari.ppm ${TESTIMAGES}/testimgari.jpg
-        ${MD5_PPM_420M_IFAST_ARI})
-    endif()
+    add_bittest(djpeg 420m-ifast-ari "-fast;-skip;1,20;-ppm"
+      testout_420m_ifast_ari.ppm ${TESTIMAGES}/testimgari.jpg
+      ${MD5_PPM_420M_IFAST_ARI})
 
     add_bittest(jpegtran 420-islow ""
       testout_420_islow.jpg ${TESTIMAGES}/testimgari.jpg
diff --git a/jdmaster.c b/jdmaster.c
index bc39d1c..cbc8774 100644
--- a/jdmaster.c
+++ b/jdmaster.c
@@ -22,7 +22,6 @@
 #include "jpeglib.h"
 #include "jpegcomp.h"
 #include "jdmaster.h"
-#include "jsimd.h"
 
 
 /*
@@ -70,17 +69,6 @@
       cinfo->comp_info[1]._DCT_scaled_size != cinfo->_min_DCT_scaled_size ||
       cinfo->comp_info[2]._DCT_scaled_size != cinfo->_min_DCT_scaled_size)
     return FALSE;
-#ifdef WITH_SIMD
-  /* If YCbCr-to-RGB color conversion is SIMD-accelerated but merged upsampling
-     isn't, then disabling merged upsampling is likely to be faster when
-     decompressing YCbCr JPEG images. */
-  if (!jsimd_can_h2v2_merged_upsample() && !jsimd_can_h2v1_merged_upsample() &&
-      jsimd_can_ycc_rgb() && cinfo->jpeg_color_space == JCS_YCbCr &&
-      (cinfo->out_color_space == JCS_RGB ||
-       (cinfo->out_color_space >= JCS_EXT_RGB &&
-        cinfo->out_color_space <= JCS_EXT_ARGB)))
-    return FALSE;
-#endif
   /* ??? also need to test for upsample-time rescaling, when & if supported */
   return TRUE;                  /* by golly, it'll work... */
 #else
diff --git a/simd/CMakeLists.txt b/simd/CMakeLists.txt
index f976417..22a887d 100644
--- a/simd/CMakeLists.txt
+++ b/simd/CMakeLists.txt
@@ -266,7 +266,8 @@
 file(REMOVE ${CMAKE_CURRENT_BINARY_DIR}/gastest.S)
 
 set(SIMD_SOURCES arm/jcgray-neon.c arm/jcphuff-neon.c arm/jcsample-neon.c
-  arm/jdsample-neon.c arm/jfdctfst-neon.c arm/jquanti-neon.c)
+  arm/jdmerge-neon.c arm/jdsample-neon.c arm/jfdctfst-neon.c
+  arm/jquanti-neon.c)
 if(NEON_INTRINSICS)
   set(SIMD_SOURCES ${SIMD_SOURCES} arm/jccolor-neon.c arm/jidctint-neon.c)
 endif()
diff --git a/simd/arm/aarch32/jsimd.c b/simd/arm/aarch32/jsimd.c
index e054a45..df13c11 100644
--- a/simd/arm/aarch32/jsimd.c
+++ b/simd/arm/aarch32/jsimd.c
@@ -501,12 +501,34 @@
 GLOBAL(int)
 jsimd_can_h2v2_merged_upsample(void)
 {
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
   return 0;
 }
 
 GLOBAL(int)
 jsimd_can_h2v1_merged_upsample(void)
 {
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
   return 0;
 }
 
@@ -514,12 +536,74 @@
 jsimd_h2v2_merged_upsample(j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
                            JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf)
 {
+  void (*neonfct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+
+  switch (cinfo->out_color_space) {
+    case JCS_EXT_RGB:
+      neonfct = jsimd_h2v2_extrgb_merged_upsample_neon;
+      break;
+    case JCS_EXT_RGBX:
+    case JCS_EXT_RGBA:
+      neonfct = jsimd_h2v2_extrgbx_merged_upsample_neon;
+      break;
+    case JCS_EXT_BGR:
+      neonfct = jsimd_h2v2_extbgr_merged_upsample_neon;
+      break;
+    case JCS_EXT_BGRX:
+    case JCS_EXT_BGRA:
+      neonfct = jsimd_h2v2_extbgrx_merged_upsample_neon;
+      break;
+    case JCS_EXT_XBGR:
+    case JCS_EXT_ABGR:
+      neonfct = jsimd_h2v2_extxbgr_merged_upsample_neon;
+      break;
+    case JCS_EXT_XRGB:
+    case JCS_EXT_ARGB:
+      neonfct = jsimd_h2v2_extxrgb_merged_upsample_neon;
+      break;
+    default:
+      neonfct = jsimd_h2v2_extrgb_merged_upsample_neon;
+      break;
+  }
+
+  neonfct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
 }
 
 GLOBAL(void)
 jsimd_h2v1_merged_upsample(j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
                            JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf)
 {
+  void (*neonfct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+
+  switch (cinfo->out_color_space) {
+    case JCS_EXT_RGB:
+      neonfct = jsimd_h2v1_extrgb_merged_upsample_neon;
+      break;
+    case JCS_EXT_RGBX:
+    case JCS_EXT_RGBA:
+      neonfct = jsimd_h2v1_extrgbx_merged_upsample_neon;
+      break;
+    case JCS_EXT_BGR:
+      neonfct = jsimd_h2v1_extbgr_merged_upsample_neon;
+      break;
+    case JCS_EXT_BGRX:
+    case JCS_EXT_BGRA:
+      neonfct = jsimd_h2v1_extbgrx_merged_upsample_neon;
+      break;
+    case JCS_EXT_XBGR:
+    case JCS_EXT_ABGR:
+      neonfct = jsimd_h2v1_extxbgr_merged_upsample_neon;
+      break;
+    case JCS_EXT_XRGB:
+    case JCS_EXT_ARGB:
+      neonfct = jsimd_h2v1_extxrgb_merged_upsample_neon;
+      break;
+    default:
+      neonfct = jsimd_h2v1_extrgb_merged_upsample_neon;
+      break;
+  }
+
+  neonfct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
 }
 
 GLOBAL(int)
diff --git a/simd/arm/aarch64/jsimd.c b/simd/arm/aarch64/jsimd.c
index a507c2c..74cd6cc 100644
--- a/simd/arm/aarch64/jsimd.c
+++ b/simd/arm/aarch64/jsimd.c
@@ -569,12 +569,34 @@
 GLOBAL(int)
 jsimd_can_h2v2_merged_upsample(void)
 {
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
   return 0;
 }
 
 GLOBAL(int)
 jsimd_can_h2v1_merged_upsample(void)
 {
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
   return 0;
 }
 
@@ -582,12 +604,74 @@
 jsimd_h2v2_merged_upsample(j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
                            JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf)
 {
+  void (*neonfct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+
+  switch (cinfo->out_color_space) {
+    case JCS_EXT_RGB:
+      neonfct = jsimd_h2v2_extrgb_merged_upsample_neon;
+      break;
+    case JCS_EXT_RGBX:
+    case JCS_EXT_RGBA:
+      neonfct = jsimd_h2v2_extrgbx_merged_upsample_neon;
+      break;
+    case JCS_EXT_BGR:
+      neonfct = jsimd_h2v2_extbgr_merged_upsample_neon;
+      break;
+    case JCS_EXT_BGRX:
+    case JCS_EXT_BGRA:
+      neonfct = jsimd_h2v2_extbgrx_merged_upsample_neon;
+      break;
+    case JCS_EXT_XBGR:
+    case JCS_EXT_ABGR:
+      neonfct = jsimd_h2v2_extxbgr_merged_upsample_neon;
+      break;
+    case JCS_EXT_XRGB:
+    case JCS_EXT_ARGB:
+      neonfct = jsimd_h2v2_extxrgb_merged_upsample_neon;
+      break;
+    default:
+      neonfct = jsimd_h2v2_extrgb_merged_upsample_neon;
+      break;
+  }
+
+  neonfct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
 }
 
 GLOBAL(void)
 jsimd_h2v1_merged_upsample(j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
                            JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf)
 {
+  void (*neonfct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+
+  switch (cinfo->out_color_space) {
+    case JCS_EXT_RGB:
+      neonfct = jsimd_h2v1_extrgb_merged_upsample_neon;
+      break;
+    case JCS_EXT_RGBX:
+    case JCS_EXT_RGBA:
+      neonfct = jsimd_h2v1_extrgbx_merged_upsample_neon;
+      break;
+    case JCS_EXT_BGR:
+      neonfct = jsimd_h2v1_extbgr_merged_upsample_neon;
+      break;
+    case JCS_EXT_BGRX:
+    case JCS_EXT_BGRA:
+      neonfct = jsimd_h2v1_extbgrx_merged_upsample_neon;
+      break;
+    case JCS_EXT_XBGR:
+    case JCS_EXT_ABGR:
+      neonfct = jsimd_h2v1_extxbgr_merged_upsample_neon;
+      break;
+    case JCS_EXT_XRGB:
+    case JCS_EXT_ARGB:
+      neonfct = jsimd_h2v1_extxrgb_merged_upsample_neon;
+      break;
+    default:
+      neonfct = jsimd_h2v1_extrgb_merged_upsample_neon;
+      break;
+  }
+
+  neonfct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
 }
 
 GLOBAL(int)
diff --git a/simd/arm/jdmerge-neon.c b/simd/arm/jdmerge-neon.c
new file mode 100644
index 0000000..18fb9d8
--- /dev/null
+++ b/simd/arm/jdmerge-neon.c
@@ -0,0 +1,144 @@
+/*
+ * jdmerge-neon.c - merged upsampling/color 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>
+
+
+/* YCbCr -> RGB conversion constants */
+
+#define F_0_344  11277  /* 0.3441467 = 11277 * 2^-15 */
+#define F_0_714  23401  /* 0.7141418 = 23401 * 2^-15 */
+#define F_1_402  22971  /* 1.4020386 = 22971 * 2^-14 */
+#define F_1_772  29033  /* 1.7720337 = 29033 * 2^-14 */
+
+ALIGN(16) static const int16_t jsimd_ycc_rgb_convert_neon_consts[] = {
+  -F_0_344, F_0_714, F_1_402, F_1_772
+};
+
+
+/* Include inline routines for colorspace extensions. */
+
+#include "jdmrgext-neon.c"
+#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_h2v1_merged_upsample_neon  jsimd_h2v1_extrgb_merged_upsample_neon
+#define jsimd_h2v2_merged_upsample_neon  jsimd_h2v2_extrgb_merged_upsample_neon
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_h2v1_merged_upsample_neon
+#undef jsimd_h2v2_merged_upsample_neon
+
+#define RGB_RED  EXT_RGBX_RED
+#define RGB_GREEN  EXT_RGBX_GREEN
+#define RGB_BLUE  EXT_RGBX_BLUE
+#define RGB_ALPHA  3
+#define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+#define jsimd_h2v1_merged_upsample_neon  jsimd_h2v1_extrgbx_merged_upsample_neon
+#define jsimd_h2v2_merged_upsample_neon  jsimd_h2v2_extrgbx_merged_upsample_neon
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef jsimd_h2v1_merged_upsample_neon
+#undef jsimd_h2v2_merged_upsample_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_h2v1_merged_upsample_neon  jsimd_h2v1_extbgr_merged_upsample_neon
+#define jsimd_h2v2_merged_upsample_neon  jsimd_h2v2_extbgr_merged_upsample_neon
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_h2v1_merged_upsample_neon
+#undef jsimd_h2v2_merged_upsample_neon
+
+#define RGB_RED  EXT_BGRX_RED
+#define RGB_GREEN  EXT_BGRX_GREEN
+#define RGB_BLUE  EXT_BGRX_BLUE
+#define RGB_ALPHA  3
+#define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+#define jsimd_h2v1_merged_upsample_neon  jsimd_h2v1_extbgrx_merged_upsample_neon
+#define jsimd_h2v2_merged_upsample_neon  jsimd_h2v2_extbgrx_merged_upsample_neon
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef jsimd_h2v1_merged_upsample_neon
+#undef jsimd_h2v2_merged_upsample_neon
+
+#define RGB_RED  EXT_XBGR_RED
+#define RGB_GREEN  EXT_XBGR_GREEN
+#define RGB_BLUE  EXT_XBGR_BLUE
+#define RGB_ALPHA  0
+#define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+#define jsimd_h2v1_merged_upsample_neon  jsimd_h2v1_extxbgr_merged_upsample_neon
+#define jsimd_h2v2_merged_upsample_neon  jsimd_h2v2_extxbgr_merged_upsample_neon
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef jsimd_h2v1_merged_upsample_neon
+#undef jsimd_h2v2_merged_upsample_neon
+
+#define RGB_RED  EXT_XRGB_RED
+#define RGB_GREEN  EXT_XRGB_GREEN
+#define RGB_BLUE  EXT_XRGB_BLUE
+#define RGB_ALPHA  0
+#define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+#define jsimd_h2v1_merged_upsample_neon  jsimd_h2v1_extxrgb_merged_upsample_neon
+#define jsimd_h2v2_merged_upsample_neon  jsimd_h2v2_extxrgb_merged_upsample_neon
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef jsimd_h2v1_merged_upsample_neon
diff --git a/simd/arm/jdmrgext-neon.c b/simd/arm/jdmrgext-neon.c
new file mode 100644
index 0000000..fa2ec05
--- /dev/null
+++ b/simd/arm/jdmrgext-neon.c
@@ -0,0 +1,667 @@
+/*
+ * jdmrgext-neon.c - merged upsampling/color conversion (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander.  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 jdmerge-neon.c. */
+
+
+/* These routines combine simple (non-fancy, i.e. non-smooth) h2v1 or h2v2
+ * chroma upsampling and YCbCr -> RGB color conversion into a single function.
+ *
+ * As with the standalone functions, YCbCr -> RGB conversion is defined by the
+ * following equations:
+ *    R = Y                        + 1.40200 * (Cr - 128)
+ *    G = Y - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128)
+ *    B = Y + 1.77200 * (Cb - 128)
+ *
+ * Scaled integer constants are used to avoid floating-point arithmetic:
+ *    0.3441467 = 11277 * 2^-15
+ *    0.7141418 = 23401 * 2^-15
+ *    1.4020386 = 22971 * 2^-14
+ *    1.7720337 = 29033 * 2^-14
+ * These constants are defined in jdmerge-neon.c.
+ *
+ * To ensure correct results, rounding is used when descaling.
+ */
+
+/* Notes on safe memory access for merged upsampling/YCbCr -> RGB conversion
+ * routines:
+ *
+ * Input memory buffers can be safely overread up to the next multiple of
+ * ALIGN_SIZE bytes, since they are always allocated by alloc_sarray() in
+ * jmemmgr.c.
+ *
+ * The output buffer cannot safely be written beyond output_width, since
+ * output_buf points to a possibly unpadded row in the decompressed image
+ * buffer allocated by the calling program.
+ */
+
+/* Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+ */
+
+void jsimd_h2v1_merged_upsample_neon(JDIMENSION output_width,
+                                     JSAMPIMAGE input_buf,
+                                     JDIMENSION in_row_group_ctr,
+                                     JSAMPARRAY output_buf)
+{
+  JSAMPROW outptr;
+  /* Pointers to Y, Cb, and Cr data */
+  JSAMPROW inptr0, inptr1, inptr2;
+
+  const int16x4_t consts = vld1_s16(jsimd_ycc_rgb_convert_neon_consts);
+  const int16x8_t neg_128 = vdupq_n_s16(-128);
+
+  inptr0 = input_buf[0][in_row_group_ctr];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr = output_buf[0];
+
+  int cols_remaining = output_width;
+  for (; cols_remaining >= 16; cols_remaining -= 16) {
+    /* De-interleave Y component values into two separate vectors, one
+     * containing the component values with even-numbered indices and one
+     * containing the component values with odd-numbered indices.
+     */
+    uint8x8x2_t y = vld2_u8(inptr0);
+    uint8x8_t cb = vld1_u8(inptr1);
+    uint8x8_t cr = vld1_u8(inptr2);
+    /* Subtract 128 from Cb and Cr. */
+    int16x8_t cr_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+    int16x8_t cb_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+    /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+    int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+    int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+    g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+    g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+    /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+    int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+                                     vrshrn_n_s32(g_sub_y_h, 15));
+    /* Compute R-Y: 1.40200 * (Cr - 128) */
+    int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
+    /* Compute B-Y: 1.77200 * (Cb - 128) */
+    int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
+    /* Add the chroma-derived values (G-Y, R-Y, and B-Y) to both the "even" and
+     * "odd" Y component values.  This effectively upsamples the chroma
+     * components horizontally.
+     */
+    int16x8_t g_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y.val[0]));
+    int16x8_t r_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y.val[0]));
+    int16x8_t b_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y.val[0]));
+    int16x8_t g_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y.val[1]));
+    int16x8_t r_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y.val[1]));
+    int16x8_t b_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y.val[1]));
+    /* Convert each component to unsigned and narrow, clamping to [0-255].
+     * Re-interleave the "even" and "odd" component values.
+     */
+    uint8x8x2_t r = vzip_u8(vqmovun_s16(r_even), vqmovun_s16(r_odd));
+    uint8x8x2_t g = vzip_u8(vqmovun_s16(g_even), vqmovun_s16(g_odd));
+    uint8x8x2_t b = vzip_u8(vqmovun_s16(b_even), vqmovun_s16(b_odd));
+
+#ifdef RGB_ALPHA
+    uint8x16x4_t rgba;
+    rgba.val[RGB_RED] = vcombine_u8(r.val[0], r.val[1]);
+    rgba.val[RGB_GREEN] = vcombine_u8(g.val[0], g.val[1]);
+    rgba.val[RGB_BLUE] = vcombine_u8(b.val[0], b.val[1]);
+    /* Set alpha channel to opaque (0xFF). */
+    rgba.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
+    /* Store RGBA pixel data to memory. */
+    vst4q_u8(outptr, rgba);
+#else
+    uint8x16x3_t rgb;
+    rgb.val[RGB_RED] = vcombine_u8(r.val[0], r.val[1]);
+    rgb.val[RGB_GREEN] = vcombine_u8(g.val[0], g.val[1]);
+    rgb.val[RGB_BLUE] = vcombine_u8(b.val[0], b.val[1]);
+    /* Store RGB pixel data to memory. */
+    vst3q_u8(outptr, rgb);
+#endif
+
+    /* Increment pointers. */
+    inptr0 += 16;
+    inptr1 += 8;
+    inptr2 += 8;
+    outptr += (RGB_PIXELSIZE * 16);
+  }
+
+  if (cols_remaining > 0) {
+    /* De-interleave Y component values into two separate vectors, one
+     * containing the component values with even-numbered indices and one
+     * containing the component values with odd-numbered indices.
+     */
+    uint8x8x2_t y = vld2_u8(inptr0);
+    uint8x8_t cb = vld1_u8(inptr1);
+    uint8x8_t cr = vld1_u8(inptr2);
+    /* Subtract 128 from Cb and Cr. */
+    int16x8_t cr_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+    int16x8_t cb_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+    /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+    int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+    int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+    g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+    g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+    /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+    int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+                                     vrshrn_n_s32(g_sub_y_h, 15));
+    /* Compute R-Y: 1.40200 * (Cr - 128) */
+    int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
+    /* Compute B-Y: 1.77200 * (Cb - 128) */
+    int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
+    /* Add the chroma-derived values (G-Y, R-Y, and B-Y) to both the "even" and
+     * "odd" Y component values.  This effectively upsamples the chroma
+     * components horizontally.
+     */
+    int16x8_t g_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y.val[0]));
+    int16x8_t r_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y.val[0]));
+    int16x8_t b_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y.val[0]));
+    int16x8_t g_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y.val[1]));
+    int16x8_t r_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y.val[1]));
+    int16x8_t b_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y.val[1]));
+    /* Convert each component to unsigned and narrow, clamping to [0-255].
+     * Re-interleave the "even" and "odd" component values.
+     */
+    uint8x8x2_t r = vzip_u8(vqmovun_s16(r_even), vqmovun_s16(r_odd));
+    uint8x8x2_t g = vzip_u8(vqmovun_s16(g_even), vqmovun_s16(g_odd));
+    uint8x8x2_t b = vzip_u8(vqmovun_s16(b_even), vqmovun_s16(b_odd));
+
+#ifdef RGB_ALPHA
+    uint8x8x4_t rgba_h;
+    rgba_h.val[RGB_RED] = r.val[1];
+    rgba_h.val[RGB_GREEN] = g.val[1];
+    rgba_h.val[RGB_BLUE] = b.val[1];
+    /* Set alpha channel to opaque (0xFF). */
+    rgba_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+    uint8x8x4_t rgba_l;
+    rgba_l.val[RGB_RED] = r.val[0];
+    rgba_l.val[RGB_GREEN] = g.val[0];
+    rgba_l.val[RGB_BLUE] = b.val[0];
+    /* Set alpha channel to opaque (0xFF). */
+    rgba_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+    /* Store RGBA pixel data to memory. */
+    switch (cols_remaining) {
+    case 15:
+      vst4_lane_u8(outptr + 14 * RGB_PIXELSIZE, rgba_h, 6);
+    case 14:
+      vst4_lane_u8(outptr + 13 * RGB_PIXELSIZE, rgba_h, 5);
+    case 13:
+      vst4_lane_u8(outptr + 12 * RGB_PIXELSIZE, rgba_h, 4);
+    case 12:
+      vst4_lane_u8(outptr + 11 * RGB_PIXELSIZE, rgba_h, 3);
+    case 11:
+      vst4_lane_u8(outptr + 10 * RGB_PIXELSIZE, rgba_h, 2);
+    case 10:
+      vst4_lane_u8(outptr + 9 * RGB_PIXELSIZE, rgba_h, 1);
+    case 9:
+      vst4_lane_u8(outptr + 8 * RGB_PIXELSIZE, rgba_h, 0);
+    case 8:
+      vst4_u8(outptr, rgba_l);
+      break;
+    case 7:
+      vst4_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgba_l, 6);
+    case 6:
+      vst4_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgba_l, 5);
+    case 5:
+      vst4_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgba_l, 4);
+    case 4:
+      vst4_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgba_l, 3);
+    case 3:
+      vst4_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgba_l, 2);
+    case 2:
+      vst4_lane_u8(outptr + RGB_PIXELSIZE, rgba_l, 1);
+    case 1:
+      vst4_lane_u8(outptr, rgba_l, 0);
+    default:
+      break;
+    }
+#else
+    uint8x8x3_t rgb_h;
+    rgb_h.val[RGB_RED] = r.val[1];
+    rgb_h.val[RGB_GREEN] = g.val[1];
+    rgb_h.val[RGB_BLUE] = b.val[1];
+    uint8x8x3_t rgb_l;
+    rgb_l.val[RGB_RED] = r.val[0];
+    rgb_l.val[RGB_GREEN] = g.val[0];
+    rgb_l.val[RGB_BLUE] = b.val[0];
+    /* Store RGB pixel data to memory. */
+    switch (cols_remaining) {
+    case 15:
+      vst3_lane_u8(outptr + 14 * RGB_PIXELSIZE, rgb_h, 6);
+    case 14:
+      vst3_lane_u8(outptr + 13 * RGB_PIXELSIZE, rgb_h, 5);
+    case 13:
+      vst3_lane_u8(outptr + 12 * RGB_PIXELSIZE, rgb_h, 4);
+    case 12:
+      vst3_lane_u8(outptr + 11 * RGB_PIXELSIZE, rgb_h, 3);
+    case 11:
+      vst3_lane_u8(outptr + 10 * RGB_PIXELSIZE, rgb_h, 2);
+    case 10:
+      vst3_lane_u8(outptr + 9 * RGB_PIXELSIZE, rgb_h, 1);
+    case 9:
+      vst3_lane_u8(outptr + 8 * RGB_PIXELSIZE, rgb_h, 0);
+    case 8:
+      vst3_u8(outptr, rgb_l);
+      break;
+    case 7:
+      vst3_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgb_l, 6);
+    case 6:
+      vst3_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgb_l, 5);
+    case 5:
+      vst3_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgb_l, 4);
+    case 4:
+      vst3_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgb_l, 3);
+    case 3:
+      vst3_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgb_l, 2);
+    case 2:
+      vst3_lane_u8(outptr + RGB_PIXELSIZE, rgb_l, 1);
+    case 1:
+      vst3_lane_u8(outptr, rgb_l, 0);
+    default:
+      break;
+    }
+#endif
+  }
+}
+
+
+/* Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+ *
+ * See comments above for details regarding color conversion and safe memory
+ * access.
+ */
+
+void jsimd_h2v2_merged_upsample_neon(JDIMENSION output_width,
+                                     JSAMPIMAGE input_buf,
+                                     JDIMENSION in_row_group_ctr,
+                                     JSAMPARRAY output_buf)
+{
+  JSAMPROW outptr0, outptr1;
+  /* Pointers to Y (both rows), Cb, and Cr data */
+  JSAMPROW inptr0_0, inptr0_1, inptr1, inptr2;
+
+  const int16x4_t consts = vld1_s16(jsimd_ycc_rgb_convert_neon_consts);
+  const int16x8_t neg_128 = vdupq_n_s16(-128);
+
+  inptr0_0 = input_buf[0][in_row_group_ctr * 2];
+  inptr0_1 = input_buf[0][in_row_group_ctr * 2 + 1];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr0 = output_buf[0];
+  outptr1 = output_buf[1];
+
+  int cols_remaining = output_width;
+  for (; cols_remaining >= 16; cols_remaining -= 16) {
+    /* For each row, de-interleave Y component values into two separate
+     * vectors, one containing the component values with even-numbered indices
+     * and one containing the component values with odd-numbered indices.
+     */
+    uint8x8x2_t y0 = vld2_u8(inptr0_0);
+    uint8x8x2_t y1 = vld2_u8(inptr0_1);
+    uint8x8_t cb = vld1_u8(inptr1);
+    uint8x8_t cr = vld1_u8(inptr2);
+    /* Subtract 128 from Cb and Cr. */
+    int16x8_t cr_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+    int16x8_t cb_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+    /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+    int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+    int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+    g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+    g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+    /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+    int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+                                     vrshrn_n_s32(g_sub_y_h, 15));
+    /* Compute R-Y: 1.40200 * (Cr - 128) */
+    int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
+    /* Compute B-Y: 1.77200 * (Cb - 128) */
+    int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
+    /* For each row, add the chroma-derived values (G-Y, R-Y, and B-Y) to both
+     * the "even" and "odd" Y component values.  This effectively upsamples the
+     * chroma components both horizontally and vertically.
+     */
+    int16x8_t g0_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y0.val[0]));
+    int16x8_t r0_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y0.val[0]));
+    int16x8_t b0_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y0.val[0]));
+    int16x8_t g0_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y0.val[1]));
+    int16x8_t r0_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y0.val[1]));
+    int16x8_t b0_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y0.val[1]));
+    int16x8_t g1_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y1.val[0]));
+    int16x8_t r1_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y1.val[0]));
+    int16x8_t b1_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y1.val[0]));
+    int16x8_t g1_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y1.val[1]));
+    int16x8_t r1_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y1.val[1]));
+    int16x8_t b1_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y1.val[1]));
+    /* Convert each component to unsigned and narrow, clamping to [0-255].
+     * Re-interleave the "even" and "odd" component values.
+     */
+    uint8x8x2_t r0 = vzip_u8(vqmovun_s16(r0_even), vqmovun_s16(r0_odd));
+    uint8x8x2_t r1 = vzip_u8(vqmovun_s16(r1_even), vqmovun_s16(r1_odd));
+    uint8x8x2_t g0 = vzip_u8(vqmovun_s16(g0_even), vqmovun_s16(g0_odd));
+    uint8x8x2_t g1 = vzip_u8(vqmovun_s16(g1_even), vqmovun_s16(g1_odd));
+    uint8x8x2_t b0 = vzip_u8(vqmovun_s16(b0_even), vqmovun_s16(b0_odd));
+    uint8x8x2_t b1 = vzip_u8(vqmovun_s16(b1_even), vqmovun_s16(b1_odd));
+
+#ifdef RGB_ALPHA
+    uint8x16x4_t rgba0, rgba1;
+    rgba0.val[RGB_RED] = vcombine_u8(r0.val[0], r0.val[1]);
+    rgba1.val[RGB_RED] = vcombine_u8(r1.val[0], r1.val[1]);
+    rgba0.val[RGB_GREEN] = vcombine_u8(g0.val[0], g0.val[1]);
+    rgba1.val[RGB_GREEN] = vcombine_u8(g1.val[0], g1.val[1]);
+    rgba0.val[RGB_BLUE] = vcombine_u8(b0.val[0], b0.val[1]);
+    rgba1.val[RGB_BLUE] = vcombine_u8(b1.val[0], b1.val[1]);
+    /* Set alpha channel to opaque (0xFF). */
+    rgba0.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
+    rgba1.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
+    /* Store RGBA pixel data to memory. */
+    vst4q_u8(outptr0, rgba0);
+    vst4q_u8(outptr1, rgba1);
+#else
+    uint8x16x3_t rgb0, rgb1;
+    rgb0.val[RGB_RED] = vcombine_u8(r0.val[0], r0.val[1]);
+    rgb1.val[RGB_RED] = vcombine_u8(r1.val[0], r1.val[1]);
+    rgb0.val[RGB_GREEN] = vcombine_u8(g0.val[0], g0.val[1]);
+    rgb1.val[RGB_GREEN] = vcombine_u8(g1.val[0], g1.val[1]);
+    rgb0.val[RGB_BLUE] = vcombine_u8(b0.val[0], b0.val[1]);
+    rgb1.val[RGB_BLUE] = vcombine_u8(b1.val[0], b1.val[1]);
+    /* Store RGB pixel data to memory. */
+    vst3q_u8(outptr0, rgb0);
+    vst3q_u8(outptr1, rgb1);
+#endif
+
+    /* Increment pointers. */
+    inptr0_0 += 16;
+    inptr0_1 += 16;
+    inptr1 += 8;
+    inptr2 += 8;
+    outptr0 += (RGB_PIXELSIZE * 16);
+    outptr1 += (RGB_PIXELSIZE * 16);
+  }
+
+  if (cols_remaining > 0) {
+    /* For each row, de-interleave Y component values into two separate
+     * vectors, one containing the component values with even-numbered indices
+     * and one containing the component values with odd-numbered indices.
+     */
+    uint8x8x2_t y0 = vld2_u8(inptr0_0);
+    uint8x8x2_t y1 = vld2_u8(inptr0_1);
+    uint8x8_t cb = vld1_u8(inptr1);
+    uint8x8_t cr = vld1_u8(inptr2);
+    /* Subtract 128 from Cb and Cr. */
+    int16x8_t cr_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+    int16x8_t cb_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+    /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+    int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+    int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+    g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+    g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+    /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+    int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+                                     vrshrn_n_s32(g_sub_y_h, 15));
+    /* Compute R-Y: 1.40200 * (Cr - 128) */
+    int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
+    /* Compute B-Y: 1.77200 * (Cb - 128) */
+    int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
+    /* For each row, add the chroma-derived values (G-Y, R-Y, and B-Y) to both
+     * the "even" and "odd" Y component values.  This effectively upsamples the
+     * chroma components both horizontally and vertically.
+     */
+    int16x8_t g0_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y0.val[0]));
+    int16x8_t r0_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y0.val[0]));
+    int16x8_t b0_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y0.val[0]));
+    int16x8_t g0_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y0.val[1]));
+    int16x8_t r0_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y0.val[1]));
+    int16x8_t b0_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y0.val[1]));
+    int16x8_t g1_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y1.val[0]));
+    int16x8_t r1_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y1.val[0]));
+    int16x8_t b1_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y1.val[0]));
+    int16x8_t g1_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y1.val[1]));
+    int16x8_t r1_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y1.val[1]));
+    int16x8_t b1_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y1.val[1]));
+    /* Convert each component to unsigned and narrow, clamping to [0-255].
+     * Re-interleave the "even" and "odd" component values.
+     */
+    uint8x8x2_t r0 = vzip_u8(vqmovun_s16(r0_even), vqmovun_s16(r0_odd));
+    uint8x8x2_t r1 = vzip_u8(vqmovun_s16(r1_even), vqmovun_s16(r1_odd));
+    uint8x8x2_t g0 = vzip_u8(vqmovun_s16(g0_even), vqmovun_s16(g0_odd));
+    uint8x8x2_t g1 = vzip_u8(vqmovun_s16(g1_even), vqmovun_s16(g1_odd));
+    uint8x8x2_t b0 = vzip_u8(vqmovun_s16(b0_even), vqmovun_s16(b0_odd));
+    uint8x8x2_t b1 = vzip_u8(vqmovun_s16(b1_even), vqmovun_s16(b1_odd));
+
+#ifdef RGB_ALPHA
+    uint8x8x4_t rgba0_h, rgba1_h;
+    rgba0_h.val[RGB_RED] = r0.val[1];
+    rgba1_h.val[RGB_RED] = r1.val[1];
+    rgba0_h.val[RGB_GREEN] = g0.val[1];
+    rgba1_h.val[RGB_GREEN] = g1.val[1];
+    rgba0_h.val[RGB_BLUE] = b0.val[1];
+    rgba1_h.val[RGB_BLUE] = b1.val[1];
+    /* Set alpha channel to opaque (0xFF). */
+    rgba0_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+    rgba1_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+
+    uint8x8x4_t rgba0_l, rgba1_l;
+    rgba0_l.val[RGB_RED] = r0.val[0];
+    rgba1_l.val[RGB_RED] = r1.val[0];
+    rgba0_l.val[RGB_GREEN] = g0.val[0];
+    rgba1_l.val[RGB_GREEN] = g1.val[0];
+    rgba0_l.val[RGB_BLUE] = b0.val[0];
+    rgba1_l.val[RGB_BLUE] = b1.val[0];
+    /* Set alpha channel to opaque (0xFF). */
+    rgba0_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+    rgba1_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+    /* Store RGBA pixel data to memory. */
+    switch (cols_remaining) {
+    case 15:
+      vst4_lane_u8(outptr0 + 14 * RGB_PIXELSIZE, rgba0_h, 6);
+      vst4_lane_u8(outptr1 + 14 * RGB_PIXELSIZE, rgba1_h, 6);
+    case 14:
+      vst4_lane_u8(outptr0 + 13 * RGB_PIXELSIZE, rgba0_h, 5);
+      vst4_lane_u8(outptr1 + 13 * RGB_PIXELSIZE, rgba1_h, 5);
+    case 13:
+      vst4_lane_u8(outptr0 + 12 * RGB_PIXELSIZE, rgba0_h, 4);
+      vst4_lane_u8(outptr1 + 12 * RGB_PIXELSIZE, rgba1_h, 4);
+    case 12:
+      vst4_lane_u8(outptr0 + 11 * RGB_PIXELSIZE, rgba0_h, 3);
+      vst4_lane_u8(outptr1 + 11 * RGB_PIXELSIZE, rgba1_h, 3);
+    case 11:
+      vst4_lane_u8(outptr0 + 10 * RGB_PIXELSIZE, rgba0_h, 2);
+      vst4_lane_u8(outptr1 + 10 * RGB_PIXELSIZE, rgba1_h, 2);
+    case 10:
+      vst4_lane_u8(outptr0 + 9 * RGB_PIXELSIZE, rgba0_h, 1);
+      vst4_lane_u8(outptr1 + 9 * RGB_PIXELSIZE, rgba1_h, 1);
+    case 9:
+      vst4_lane_u8(outptr0 + 8 * RGB_PIXELSIZE, rgba0_h, 0);
+      vst4_lane_u8(outptr1 + 8 * RGB_PIXELSIZE, rgba1_h, 0);
+    case 8:
+      vst4_u8(outptr0, rgba0_l);
+      vst4_u8(outptr1, rgba1_l);
+      break;
+    case 7:
+      vst4_lane_u8(outptr0 + 6 * RGB_PIXELSIZE, rgba0_l, 6);
+      vst4_lane_u8(outptr1 + 6 * RGB_PIXELSIZE, rgba1_l, 6);
+    case 6:
+      vst4_lane_u8(outptr0 + 5 * RGB_PIXELSIZE, rgba0_l, 5);
+      vst4_lane_u8(outptr1 + 5 * RGB_PIXELSIZE, rgba1_l, 5);
+    case 5:
+      vst4_lane_u8(outptr0 + 4 * RGB_PIXELSIZE, rgba0_l, 4);
+      vst4_lane_u8(outptr1 + 4 * RGB_PIXELSIZE, rgba1_l, 4);
+    case 4:
+      vst4_lane_u8(outptr0 + 3 * RGB_PIXELSIZE, rgba0_l, 3);
+      vst4_lane_u8(outptr1 + 3 * RGB_PIXELSIZE, rgba1_l, 3);
+    case 3:
+      vst4_lane_u8(outptr0 + 2 * RGB_PIXELSIZE, rgba0_l, 2);
+      vst4_lane_u8(outptr1 + 2 * RGB_PIXELSIZE, rgba1_l, 2);
+    case 2:
+      vst4_lane_u8(outptr0 + 1 * RGB_PIXELSIZE, rgba0_l, 1);
+      vst4_lane_u8(outptr1 + 1 * RGB_PIXELSIZE, rgba1_l, 1);
+    case 1:
+      vst4_lane_u8(outptr0, rgba0_l, 0);
+      vst4_lane_u8(outptr1, rgba1_l, 0);
+    default:
+      break;
+    }
+#else
+    uint8x8x3_t rgb0_h, rgb1_h;
+    rgb0_h.val[RGB_RED] = r0.val[1];
+    rgb1_h.val[RGB_RED] = r1.val[1];
+    rgb0_h.val[RGB_GREEN] = g0.val[1];
+    rgb1_h.val[RGB_GREEN] = g1.val[1];
+    rgb0_h.val[RGB_BLUE] = b0.val[1];
+    rgb1_h.val[RGB_BLUE] = b1.val[1];
+
+    uint8x8x3_t rgb0_l, rgb1_l;
+    rgb0_l.val[RGB_RED] = r0.val[0];
+    rgb1_l.val[RGB_RED] = r1.val[0];
+    rgb0_l.val[RGB_GREEN] = g0.val[0];
+    rgb1_l.val[RGB_GREEN] = g1.val[0];
+    rgb0_l.val[RGB_BLUE] = b0.val[0];
+    rgb1_l.val[RGB_BLUE] = b1.val[0];
+    /* Store RGB pixel data to memory. */
+    switch (cols_remaining) {
+    case 15:
+      vst3_lane_u8(outptr0 + 14 * RGB_PIXELSIZE, rgb0_h, 6);
+      vst3_lane_u8(outptr1 + 14 * RGB_PIXELSIZE, rgb1_h, 6);
+    case 14:
+      vst3_lane_u8(outptr0 + 13 * RGB_PIXELSIZE, rgb0_h, 5);
+      vst3_lane_u8(outptr1 + 13 * RGB_PIXELSIZE, rgb1_h, 5);
+    case 13:
+      vst3_lane_u8(outptr0 + 12 * RGB_PIXELSIZE, rgb0_h, 4);
+      vst3_lane_u8(outptr1 + 12 * RGB_PIXELSIZE, rgb1_h, 4);
+    case 12:
+      vst3_lane_u8(outptr0 + 11 * RGB_PIXELSIZE, rgb0_h, 3);
+      vst3_lane_u8(outptr1 + 11 * RGB_PIXELSIZE, rgb1_h, 3);
+    case 11:
+      vst3_lane_u8(outptr0 + 10 * RGB_PIXELSIZE, rgb0_h, 2);
+      vst3_lane_u8(outptr1 + 10 * RGB_PIXELSIZE, rgb1_h, 2);
+    case 10:
+      vst3_lane_u8(outptr0 + 9 * RGB_PIXELSIZE, rgb0_h, 1);
+      vst3_lane_u8(outptr1 + 9 * RGB_PIXELSIZE, rgb1_h, 1);
+    case 9:
+      vst3_lane_u8(outptr0 + 8 * RGB_PIXELSIZE, rgb0_h, 0);
+      vst3_lane_u8(outptr1 + 8 * RGB_PIXELSIZE, rgb1_h, 0);
+    case 8:
+      vst3_u8(outptr0, rgb0_l);
+      vst3_u8(outptr1, rgb1_l);
+      break;
+    case 7:
+      vst3_lane_u8(outptr0 + 6 * RGB_PIXELSIZE, rgb0_l, 6);
+      vst3_lane_u8(outptr1 + 6 * RGB_PIXELSIZE, rgb1_l, 6);
+    case 6:
+      vst3_lane_u8(outptr0 + 5 * RGB_PIXELSIZE, rgb0_l, 5);
+      vst3_lane_u8(outptr1 + 5 * RGB_PIXELSIZE, rgb1_l, 5);
+    case 5:
+      vst3_lane_u8(outptr0 + 4 * RGB_PIXELSIZE, rgb0_l, 4);
+      vst3_lane_u8(outptr1 + 4 * RGB_PIXELSIZE, rgb1_l, 4);
+    case 4:
+      vst3_lane_u8(outptr0 + 3 * RGB_PIXELSIZE, rgb0_l, 3);
+      vst3_lane_u8(outptr1 + 3 * RGB_PIXELSIZE, rgb1_l, 3);
+    case 3:
+      vst3_lane_u8(outptr0 + 2 * RGB_PIXELSIZE, rgb0_l, 2);
+      vst3_lane_u8(outptr1 + 2 * RGB_PIXELSIZE, rgb1_l, 2);
+    case 2:
+      vst3_lane_u8(outptr0 + 1 * RGB_PIXELSIZE, rgb0_l, 1);
+      vst3_lane_u8(outptr1 + 1 * RGB_PIXELSIZE, rgb1_l, 1);
+    case 1:
+      vst3_lane_u8(outptr0, rgb0_l, 0);
+      vst3_lane_u8(outptr1, rgb1_l, 0);
+    default:
+      break;
+    }
+#endif
+  }
+}
diff --git a/simd/jsimd.h b/simd/jsimd.h
index 053ea3c..ccc9a10 100644
--- a/simd/jsimd.h
+++ b/simd/jsimd.h
@@ -824,6 +824,50 @@
   (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
    JSAMPARRAY output_buf);
 
+EXTERN(void) jsimd_h2v1_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v1_extrgb_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v1_extrgbx_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v1_extbgr_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v1_extbgrx_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v1_extxbgr_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v1_extxrgb_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+
+EXTERN(void) jsimd_h2v2_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v2_extrgb_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v2_extrgbx_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v2_extbgr_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v2_extbgrx_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v2_extxbgr_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v2_extxrgb_merged_upsample_neon
+  (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+   JSAMPARRAY output_buf);
+
 EXTERN(void) jsimd_h2v1_merged_upsample_dspr2
   (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
    JSAMPARRAY output_buf, JSAMPLE *range);