diff --git a/include/libyuv/row.h b/include/libyuv/row.h
index 9ee8af68f..d8fe2137d 100644
--- a/include/libyuv/row.h
+++ b/include/libyuv/row.h
@@ -398,6 +398,7 @@ extern "C" {
 #define HAS_ARGBTOUVJ444ROW_NEON
 #define HAS_ARGBTOUVJROW_NEON
 #define HAS_ARGBTOUVROW_NEON
+// TODO: Fix ARGBTOYROW and test ARGBToI444 tests pass.
 #define HAS_ARGBTOYJROW_NEON
 #define HAS_ARGBTOYROW_NEON
 #define HAS_AYUVTOUVROW_NEON
diff --git a/source/row_neon.cc b/source/row_neon.cc
index d5e36cdef..49d7584dc 100644
--- a/source/row_neon.cc
+++ b/source/row_neon.cc
@@ -1841,18 +1841,22 @@ static void ARGBToUV444MatrixRow_NEON(
       "vdup.u8     d26, d0[2]                    \n"  // UR -0.2969 coefficient
       "vdup.u8     d27, d0[4]                    \n"  // VB -0.1406 coefficient
       "vdup.u8     d28, d0[5]                    \n"  // VG -0.7344 coefficient
+      "vneg.s8     d25, d25                      \n"
+      "vneg.s8     d26, d26                      \n"
+      "vneg.s8     d27, d27                      \n"
+      "vneg.s8     d28, d28                      \n"
       "vmov.u16    q15, #0x8080                  \n"  // 128.5
 
       "1:                                        \n"
       "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 ARGB pixels.
       "subs        %3, %3, #8                    \n"  // 8 processed per loop.
       "vmull.u8    q2, d0, d24                   \n"  // B
-      "vmlal.u8    q2, d1, d25                   \n"  // G
-      "vmlal.u8    q2, d2, d26                   \n"  // R
+      "vmlsl.u8    q2, d1, d25                   \n"  // G
+      "vmlsl.u8    q2, d2, d26                   \n"  // R
 
       "vmull.u8    q3, d2, d24                   \n"  // R
-      "vmlal.u8    q3, d1, d28                   \n"  // G
-      "vmlal.u8    q3, d0, d27                   \n"  // B
+      "vmlsl.u8    q3, d1, d28                   \n"  // G
+      "vmlsl.u8    q3, d0, d27                   \n"  // B
 
       "vaddhn.u16  d0, q2, q15                   \n"  // +128 -> unsigned
       "vaddhn.u16  d1, q3, q15                   \n"  // +128 -> unsigned
diff --git a/source/row_neon64.cc b/source/row_neon64.cc
index 8ef9d4975..c30ef680c 100644
--- a/source/row_neon64.cc
+++ b/source/row_neon64.cc
@@ -2729,19 +2729,23 @@ static void ARGBToUV444MatrixRow_NEON(
       "dup         v26.16b, v0.b[2]              \n"  // UR -0.2969 coefficient
       "dup         v27.16b, v0.b[4]              \n"  // VB -0.1406 coefficient
       "dup         v28.16b, v0.b[5]              \n"  // VG -0.7344 coefficient
+      "neg         v25.16b, v25.16b              \n"
+      "neg         v26.16b, v26.16b              \n"
+      "neg         v27.16b, v27.16b              \n"
+      "neg         v28.16b, v28.16b              \n"
       "movi        v29.16b, #0x80                \n"  // 128.5
 
       "1:                                        \n"
       "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
       "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
       "umull       v4.8h, v0.8b, v24.8b          \n"  // B
-      "umlal       v4.8h, v1.8b, v25.8b          \n"  // G
-      "umlal       v4.8h, v2.8b, v26.8b          \n"  // R
+      "umlsl       v4.8h, v1.8b, v25.8b          \n"  // G
+      "umlsl       v4.8h, v2.8b, v26.8b          \n"  // R
       "prfm        pldl1keep, [%0, 448]          \n"
 
       "umull       v3.8h, v2.8b, v24.8b          \n"  // R
-      "umlal       v3.8h, v1.8b, v28.8b          \n"  // G
-      "umlal       v3.8h, v0.8b, v27.8b          \n"  // B
+      "umlsl       v3.8h, v1.8b, v28.8b          \n"  // G
+      "umlsl       v3.8h, v0.8b, v27.8b          \n"  // B
 
       "addhn       v0.8b, v4.8h, v29.8h          \n"  // +128 -> unsigned
       "addhn       v1.8b, v3.8h, v29.8h          \n"  // +128 -> unsigned