diff --git a/libImaging/FilterSIMD_5x5i_4u8.c b/libImaging/FilterSIMD_5x5i_4u8.c
index 5fff7c68a..5c6dced03 100644
--- a/libImaging/FilterSIMD_5x5i_4u8.c
+++ b/libImaging/FilterSIMD_5x5i_4u8.c
@@ -15,6 +15,111 @@ ImagingFilter5x5i_4u8(Imaging imOut, Imaging im, const INT16* kernel,
         INT32* in1 = im->image32[y+1];
         INT32* in2 = im->image32[y+2];
         INT32* out = imOut->image32[y];
+#if defined(__AVX2__)
+
+    #define MM_KERNEL_LOAD(row, x) \
+        pix0##row = _mm256_castsi128_si256(_mm_shuffle_epi8(_mm_loadu_si128((__m128i*) &in2[x]), shuffle)); \
+        pix0##row = _mm256_inserti128_si256(pix0##row, _mm256_castsi256_si128(pix0##row), 1); \
+        pix1##row = _mm256_castsi128_si256(_mm_shuffle_epi8(_mm_loadu_si128((__m128i*) &in1[x]), shuffle)); \
+        pix1##row = _mm256_inserti128_si256(pix1##row, _mm256_castsi256_si128(pix1##row), 1); \
+        pix2##row = _mm256_castsi128_si256(_mm_shuffle_epi8(_mm_loadu_si128((__m128i*) &in0[x]), shuffle)); \
+        pix2##row = _mm256_inserti128_si256(pix2##row, _mm256_castsi256_si128(pix2##row), 1); \
+        pix3##row = _mm256_castsi128_si256(_mm_shuffle_epi8(_mm_loadu_si128((__m128i*) &in_1[x]), shuffle)); \
+        pix3##row = _mm256_inserti128_si256(pix3##row, _mm256_castsi256_si128(pix3##row), 1); \
+        pix4##row = _mm256_castsi128_si256(_mm_shuffle_epi8(_mm_loadu_si128((__m128i*) &in_2[x]), shuffle)); \
+        pix4##row = _mm256_inserti128_si256(pix4##row, _mm256_castsi256_si128(pix4##row), 1);
+
+    #define MM_KERNEL_SUM(ss, row, unpack_epi8, krow, kctl) \
+        ss = _mm256_add_epi32(ss, _mm256_madd_epi16( \
+            unpack_epi8(pix0##row, zero), _mm256_shuffle_epi32(kernel0##krow, kctl))); \
+        ss = _mm256_add_epi32(ss, _mm256_madd_epi16( \
+            unpack_epi8(pix1##row, zero), _mm256_shuffle_epi32(kernel1##krow, kctl))); \
+        ss = _mm256_add_epi32(ss, _mm256_madd_epi16( \
+            unpack_epi8(pix2##row, zero), _mm256_shuffle_epi32(kernel2##krow, kctl))); \
+        ss = _mm256_add_epi32(ss, _mm256_madd_epi16( \
+            unpack_epi8(pix3##row, zero), _mm256_shuffle_epi32(kernel3##krow, kctl))); \
+        ss = _mm256_add_epi32(ss, _mm256_madd_epi16( \
+            unpack_epi8(pix4##row, zero), _mm256_shuffle_epi32(kernel4##krow, kctl)));
+
+        __m128i shuffle = _mm_set_epi8(15,11,14,10,13,9,12,8, 7,3,6,2,5,1,4,0);
+        __m256i kernel00 = _mm256_set_epi16(
+            0, 0, kernel[4], kernel[3], kernel[2], kernel[1], kernel[0], 0,
+            0, 0, 0, kernel[4], kernel[3], kernel[2], kernel[1], kernel[0]);
+        __m256i kernel10 = _mm256_set_epi16(
+            0, 0, kernel[9], kernel[8], kernel[7], kernel[6], kernel[5], 0,
+            0, 0, 0, kernel[9], kernel[8], kernel[7], kernel[6], kernel[5]);
+        __m256i kernel20 = _mm256_set_epi16(
+            0, 0, kernel[14], kernel[13], kernel[12], kernel[11], kernel[10], 0,
+            0, 0, 0, kernel[14], kernel[13], kernel[12], kernel[11], kernel[10]);
+        __m256i kernel30 = _mm256_set_epi16(
+            0, 0, kernel[19], kernel[18], kernel[17], kernel[16], kernel[15], 0,
+            0, 0, 0, kernel[19], kernel[18], kernel[17], kernel[16], kernel[15]);
+        __m256i kernel40 = _mm256_set_epi16(
+            0, 0, kernel[24], kernel[23], kernel[22], kernel[21], kernel[20], 0,
+            0, 0, 0, kernel[24], kernel[23], kernel[22], kernel[21], kernel[20]);
+        __m256i zero = _mm256_setzero_si256();
+        __m256i pix00, pix10, pix20, pix30, pix40;
+        __m256i pix01, pix11, pix21, pix31, pix41;
+
+        out[0] = in0[0];
+        out[1] = in0[1];
+        x = 2;
+
+        MM_KERNEL_LOAD(0, 0);
+
+        for (; x < im->xsize-2-3; x += 4) {
+            __m256i ss0, ss2;
+            __m128i ssout;
+
+            ss0 = _mm256_set1_epi32(offset);
+            ss2 = _mm256_set1_epi32(offset);
+
+            MM_KERNEL_SUM(ss0, 0, _mm256_unpacklo_epi8, 0, 0x00);
+            MM_KERNEL_SUM(ss0, 0, _mm256_unpackhi_epi8, 0, 0x55);
+            MM_KERNEL_SUM(ss2, 0, _mm256_unpackhi_epi8, 0, 0x00);
+            
+            MM_KERNEL_LOAD(1, x+2);
+            MM_KERNEL_SUM(ss0, 1, _mm256_unpacklo_epi8, 0, 0xaa);
+            ss0 = _mm256_srai_epi32(ss0, PRECISION_BITS);
+            MM_KERNEL_SUM(ss2, 1, _mm256_unpacklo_epi8, 0, 0x55);
+            MM_KERNEL_SUM(ss2, 1, _mm256_unpackhi_epi8, 0, 0xaa);
+            ss2 = _mm256_srai_epi32(ss2, PRECISION_BITS);
+
+            pix00 = pix01;
+            pix10 = pix11;
+            pix20 = pix21;
+            pix30 = pix31;
+            pix40 = pix41;
+
+            ss0 = _mm256_packs_epi32(ss0, ss2);
+            ssout = _mm_packus_epi16(
+                _mm256_extracti128_si256(ss0, 0),
+                _mm256_extracti128_si256(ss0, 1));
+            ssout = _mm_shuffle_epi32(ssout, 0xd8);
+            _mm_storeu_si128((__m128i*) &out[x], ssout);
+        }
+        for (; x < im->xsize-2; x++) {
+            __m256i ss0;
+
+            MM_KERNEL_LOAD(0, x-2);
+            MM_KERNEL_LOAD(1, x+2);
+
+            ss0 = _mm256_set1_epi32(offset);
+            MM_KERNEL_SUM(ss0, 0, _mm256_unpacklo_epi8, 0, 0x00);
+            MM_KERNEL_SUM(ss0, 0, _mm256_unpackhi_epi8, 0, 0x55);
+            MM_KERNEL_SUM(ss0, 1, _mm256_unpacklo_epi8, 0, 0xaa);
+            ss0 = _mm256_srai_epi32(ss0, PRECISION_BITS);
+
+            ss0 = _mm256_packs_epi32(ss0, ss0);
+            ss0 = _mm256_packus_epi16(ss0, ss0);
+            out[x] = _mm_cvtsi128_si32(_mm256_castsi256_si128(ss0));
+        }
+        out[x] = in0[x];
+        out[x+1] = in0[x+1];
+    #undef MM_KERNEL_LOAD
+    #undef MM_KERNEL_SUM
+
+#else
 
     #define MM_KERNEL_LOAD(row, x) \
         pix0##row = _mm_shuffle_epi8(_mm_loadu_si128((__m128i*) &in2[x]), shuffle); \
@@ -37,35 +142,25 @@ ImagingFilter5x5i_4u8(Imaging imOut, Imaging im, const INT16* kernel,
 
         __m128i shuffle = _mm_set_epi8(15,11,14,10,13,9,12,8, 7,3,6,2,5,1,4,0);
         __m128i kernel00 = _mm_set_epi16(
-            0, 0, 0, kernel[4],
-            kernel[3], kernel[2], kernel[1], kernel[0]);
+            0, 0, 0, kernel[4], kernel[3], kernel[2], kernel[1], kernel[0]);
         __m128i kernel10 = _mm_set_epi16(
-            0, 0, 0, kernel[9],
-            kernel[8], kernel[7], kernel[6], kernel[5]);
+            0, 0, 0, kernel[9], kernel[8], kernel[7], kernel[6], kernel[5]);
         __m128i kernel20 = _mm_set_epi16(
-            0, 0, 0, kernel[14],
-            kernel[13], kernel[12], kernel[11], kernel[10]);
+            0, 0, 0, kernel[14], kernel[13], kernel[12], kernel[11], kernel[10]);
         __m128i kernel30 = _mm_set_epi16(
-            0, 0, 0, kernel[19],
-            kernel[18], kernel[17], kernel[16], kernel[15]);
+            0, 0, 0, kernel[19], kernel[18], kernel[17], kernel[16], kernel[15]);
         __m128i kernel40 = _mm_set_epi16(
-            0, 0, 0, kernel[24],
-            kernel[23], kernel[22], kernel[21], kernel[20]);
+            0, 0, 0, kernel[24], kernel[23], kernel[22], kernel[21], kernel[20]);
         __m128i kernel01 = _mm_set_epi16(
-            0, 0, kernel[4], kernel[3],
-            kernel[2], kernel[1], kernel[0], 0);
+            0, 0, kernel[4], kernel[3], kernel[2], kernel[1], kernel[0], 0);
         __m128i kernel11 = _mm_set_epi16(
-            0, 0, kernel[9], kernel[8],
-            kernel[7], kernel[6], kernel[5], 0);
+            0, 0, kernel[9], kernel[8], kernel[7], kernel[6], kernel[5], 0);
         __m128i kernel21 = _mm_set_epi16(
-            0, 0, kernel[14], kernel[13],
-            kernel[12], kernel[11], kernel[10], 0);
+            0, 0, kernel[14], kernel[13], kernel[12], kernel[11], kernel[10], 0);
         __m128i kernel31 = _mm_set_epi16(
-            0, 0, kernel[19], kernel[18],
-            kernel[17], kernel[16], kernel[15], 0);
+            0, 0, kernel[19], kernel[18], kernel[17], kernel[16], kernel[15], 0);
         __m128i kernel41 = _mm_set_epi16(
-            0, 0, kernel[24], kernel[23],
-            kernel[22], kernel[21], kernel[20], 0);
+            0, 0, kernel[24], kernel[23], kernel[22], kernel[21], kernel[20], 0);
         __m128i zero = _mm_setzero_si128();
         __m128i pix00, pix10, pix20, pix30, pix40;
         __m128i pix01, pix11, pix21, pix31, pix41;
@@ -76,7 +171,7 @@ ImagingFilter5x5i_4u8(Imaging imOut, Imaging im, const INT16* kernel,
 
         MM_KERNEL_LOAD(0, 0);
 
-        for (; x < im->xsize-2-4; x += 4) {
+        for (; x < im->xsize-2-3; x += 4) {
             __m128i ss0, ss1, ss2, ss3;
 
             ss0 = _mm_set1_epi32(offset);
@@ -132,10 +227,11 @@ ImagingFilter5x5i_4u8(Imaging imOut, Imaging im, const INT16* kernel,
         }
         out[x] = in0[x];
         out[x+1] = in0[x+1];
+    #undef MM_KERNEL_LOAD
+    #undef MM_KERNEL_SUM
+
+#endif
     }
     memcpy(imOut->image32[y], im->image32[y], im->linesize);
     memcpy(imOut->image32[y+1], im->image32[y+1], im->linesize);
-
-#undef MM_KERNEL_LOAD
-#undef MM_KERNEL_SUM
 }