diff --git a/source/row_any.cc b/source/row_any.cc
index 26d2ce3b9..0d0a64354 100644
--- a/source/row_any.cc
+++ b/source/row_any.cc
@@ -834,8 +834,12 @@ ANY21C(NV12ToRGB565Row_Any_LASX, NV12ToRGB565Row_LASX, 1, 1, 2, 2, 15)
 #define ANY21CT(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, T, SBPP, BPP, MASK)      \
   void NAMEANY(const T* y_buf, const T* uv_buf, uint8_t* dst_ptr,              \
                const struct YuvConstants* yuvconstants, int width) {           \
-    SIMD_ALIGNED(T vin[16 * 2]);                                               \
+    SIMD_ALIGNED(T vin[32 * 2]);                                               \
+    static_assert(SS(MASK + 1, UVSHIFT) * SBPP * 2 <= sizeof(vin) / 2,         \
+                  "vin buffer too small");                                     \
     SIMD_ALIGNED(uint8_t vout[64]);                                            \
+    static_assert(SS(MASK + 1, DUVSHIFT) * BPP <= sizeof(vout),                \
+                  "vout buffer too small");                                    \
     memset(vin, 0, sizeof(vin)); /* for msan */                                \
     int r = width & MASK;                                                      \
     int n = width & ~MASK;                                                     \
@@ -844,8 +848,8 @@ ANY21C(NV12ToRGB565Row_Any_LASX, NV12ToRGB565Row_LASX, 1, 1, 2, 2, 15)
     }                                                                          \
     ptrdiff_t np = n;                                                          \
     memcpy(vin, y_buf + np, r * SBPP);                                         \
-    memcpy(vin + 16, uv_buf + 2 * (np >> UVSHIFT), SS(r, UVSHIFT) * SBPP * 2); \
-    ANY_SIMD(vin, vin + 16, vout, yuvconstants, MASK + 1);                     \
+    memcpy(vin + 32, uv_buf + 2 * (np >> UVSHIFT), SS(r, UVSHIFT) * SBPP * 2); \
+    ANY_SIMD(vin, vin + 32, vout, yuvconstants, MASK + 1);                     \
     memcpy(dst_ptr + (np >> DUVSHIFT) * BPP, vout, SS(r, DUVSHIFT) * BPP);     \
   }
 
diff --git a/unit_test/convert_argb_test.cc b/unit_test/convert_argb_test.cc
index 8411039d1..c65419823 100644
--- a/unit_test/convert_argb_test.cc
+++ b/unit_test/convert_argb_test.cc
@@ -2736,6 +2736,82 @@ TEST_F(LibYUVConvertTest, TestUYVYToARGB) {
 #endif
 }
 
+// Call P010ToARGBMatrixFilter() with a width that is not a multiple of 16.
+// Verify there is no stack buffer overflow in P410ToARGBRow_Any_AVX2().
+TEST_F(LibYUVConvertTest, P010ToARGBMatrixFilterOverflow) {
+  int width = 1024 + 15;  // 1039
+  int height = 4;
+  int dst_stride_argb = width * 4;
+  align_buffer_page_end_16(src_y, width * height);
+  align_buffer_page_end_16(src_uv, width * ((height + 1) / 2));
+  align_buffer_page_end(dst_argb, dst_stride_argb * height);
+  memset(src_y, 0x02, width * height * sizeof(uint16_t));
+  memset(src_uv, 0x02, width * ((height + 1) / 2) * sizeof(uint16_t));
+  EXPECT_EQ(0, P010ToARGBMatrixFilter(src_y, width, src_uv, width, dst_argb,
+                                      dst_stride_argb, &kYuvI601Constants,
+                                      width, height, kFilterBilinear));
+  free_aligned_buffer_page_end_16(src_y);
+  free_aligned_buffer_page_end_16(src_uv);
+  free_aligned_buffer_page_end(dst_argb);
+}
+
+// Call P210ToARGBMatrixFilter() with a width that is not a multiple of 16.
+// Verify there is no stack buffer overflow in P410ToARGBRow_Any_AVX2().
+TEST_F(LibYUVConvertTest, P210ToARGBMatrixFilterOverflow) {
+  int width = 1024 + 15;  // 1039
+  int height = 4;
+  int dst_stride_argb = width * 4;
+  align_buffer_page_end_16(src_y, width * height);
+  align_buffer_page_end_16(src_uv, width * height);
+  align_buffer_page_end(dst_argb, dst_stride_argb * height);
+  memset(src_y, 0x02, width * height * sizeof(uint16_t));
+  memset(src_uv, 0x02, width * height * sizeof(uint16_t));
+  EXPECT_EQ(0, P210ToARGBMatrixFilter(src_y, width, src_uv, width, dst_argb,
+                                      dst_stride_argb, &kYuvI601Constants,
+                                      width, height, kFilterBilinear));
+  free_aligned_buffer_page_end_16(src_y);
+  free_aligned_buffer_page_end_16(src_uv);
+  free_aligned_buffer_page_end(dst_argb);
+}
+
+// Call P010ToAR30MatrixFilter() with a width that is not a multiple of 16.
+// Verify there is no stack buffer overflow in P410ToAR30Row_Any_AVX2().
+TEST_F(LibYUVConvertTest, P010ToAR30MatrixFilterOverflow) {
+  int width = 1024 + 15;  // 1039
+  int height = 4;
+  int dst_stride_argb = width * 4;
+  align_buffer_page_end_16(src_y, width * height);
+  align_buffer_page_end_16(src_uv, width * ((height + 1) / 2));
+  align_buffer_page_end(dst_argb, dst_stride_argb * height);
+  memset(src_y, 0x02, width * height * sizeof(uint16_t));
+  memset(src_uv, 0x02, width * ((height + 1) / 2) * sizeof(uint16_t));
+  EXPECT_EQ(0, P010ToAR30MatrixFilter(src_y, width, src_uv, width, dst_argb,
+                                      dst_stride_argb, &kYuvI601Constants,
+                                      width, height, kFilterBilinear));
+  free_aligned_buffer_page_end_16(src_y);
+  free_aligned_buffer_page_end_16(src_uv);
+  free_aligned_buffer_page_end(dst_argb);
+}
+
+// Call P210ToAR30MatrixFilter(() with a width that is not a multiple of 16.
+// Verify there is no stack buffer overflow in P410ToAR30Row_Any_AVX2().
+TEST_F(LibYUVConvertTest, P210ToAR30MatrixFilterOverflow) {
+  int width = 1024 + 15;  // 1039
+  int height = 4;
+  int dst_stride_ar30 = width * 4;
+  align_buffer_page_end_16(src_y, width * height);
+  align_buffer_page_end_16(src_uv, width * height);
+  align_buffer_page_end(dst_ar30, dst_stride_ar30 * height);
+  memset(src_y, 0x02, width * height * sizeof(uint16_t));
+  memset(src_uv, 0x02, width * height * sizeof(uint16_t));
+  EXPECT_EQ(0, P210ToAR30MatrixFilter(src_y, width, src_uv, width, dst_ar30,
+                                      dst_stride_ar30, &kYuvI601Constants,
+                                      width, height, kFilterBilinear));
+  free_aligned_buffer_page_end_16(src_y);
+  free_aligned_buffer_page_end_16(src_uv);
+  free_aligned_buffer_page_end(dst_ar30);
+}
+
 #ifdef ENABLE_ROW_TESTS
 TEST_F(LibYUVConvertTest, TestARGBToUVMatrixRow_Opt) {
   const int kMaxWidth = 128;