diff --git a/src/dsp/alpha_processing_sse2.c b/src/dsp/alpha_processing_sse2.c
index bdf141e5..93608c15 100644
--- a/src/dsp/alpha_processing_sse2.c
+++ b/src/dsp/alpha_processing_sse2.c
@@ -151,8 +151,8 @@ static int ExtractAlpha(const uint8_t* argb, int argb_stride,
 
 // We can't use a 'const int' for the SHUFFLE value, because it has to be an
 // immediate in the _mm_shufflexx_epi16() instruction. We really need a macro.
-// We use: v / 255 = (v + 1 + (v >> 8)) >> 8, where v = alpha * {r,g,b} is
-// a 16bit value.
+// We use: v / 255 = (v * 0x8081) >> 23, where v = alpha * {r,g,b} is a 16bit
+// value.
 #define APPLY_ALPHA(RGBX, SHUFFLE) do {                           \
   const __m128i argb0 = _mm_loadl_epi64((const __m128i*)&(RGBX)); \
   const __m128i argb1 = _mm_unpacklo_epi8(argb0, zero);           \
@@ -161,18 +161,16 @@ static int ExtractAlpha(const uint8_t* argb, int argb_stride,
   const __m128i alpha2 = _mm_shufflehi_epi16(alpha1, SHUFFLE);    \
   /* alpha2 = [ff a0 a0 a0][ff a1 a1 a1] */                       \
   const __m128i A0 = _mm_mullo_epi16(alpha2, argb1);              \
-  const __m128i A1 = _mm_srli_epi16(A0, 8);                       \
-  const __m128i A2 = _mm_add_epi16(A1, A0);                       \
-  const __m128i A3 = _mm_add_epi16(A2, one);                      \
-  const __m128i A4 = _mm_srli_epi16(A3, 8);                       \
-  const __m128i A5 = _mm_packus_epi16(A4, zero);                  \
-  _mm_storel_epi64((__m128i*)&(RGBX), A5);                        \
+  const __m128i A1 = _mm_mulhi_epu16(A0, kMult);                  \
+  const __m128i A2 = _mm_srli_epi16(A1, 7);                       \
+  const __m128i A3 = _mm_packus_epi16(A2, zero);                  \
+  _mm_storel_epi64((__m128i*)&(RGBX), A3);                        \
 } while (0)
 
 static void ApplyAlphaMultiply_SSE2(uint8_t* rgba, int alpha_first,
                                     int w, int h, int stride) {
   const __m128i zero = _mm_setzero_si128();
-  const __m128i one = _mm_set1_epi16(1);
+  const __m128i kMult = _mm_set1_epi16(0x8081u);
   const __m128i kMask = _mm_set_epi16(0, 0xff, 0xff, 0, 0, 0xff, 0xff, 0);
   const int kSpan = 2;
   while (h-- > 0) {
@@ -214,10 +212,11 @@ static void MultARGBRow_SSE2(uint32_t* const ptr, int width, int inverse) {
     const int kSpan = 2;
     const __m128i zero = _mm_setzero_si128();
     const __m128i k128 = _mm_set1_epi16(128);
+    const __m128i kMult = _mm_set1_epi16(0x0101);
     const __m128i kMask = _mm_set_epi16(0, 0xff, 0, 0, 0, 0xff, 0, 0);
     for (x = 0; x + kSpan <= width; x += kSpan) {
       // To compute 'result = (int)(a * x / 255. + .5)', we use:
-      //   tmp = a * v + 128, result = (tmp + (tmp >> 8)) >> 8
+      //   tmp = a * v + 128, result = (tmp * 0x0101u) >> 16
       const __m128i A0 = _mm_loadl_epi64((const __m128i*)&ptr[x]);
       const __m128i A1 = _mm_unpacklo_epi8(A0, zero);
       const __m128i A2 = _mm_or_si128(A1, kMask);
@@ -226,10 +225,8 @@ static void MultARGBRow_SSE2(uint32_t* const ptr, int width, int inverse) {
       // here, A4 = [ff a0 a0 a0][ff a1 a1 a1]
       const __m128i A5 = _mm_mullo_epi16(A4, A1);
       const __m128i A6 = _mm_add_epi16(A5, k128);
-      const __m128i A7 = _mm_srli_epi16(A6, 8);
-      const __m128i A8 = _mm_add_epi16(A7, A6);
-      const __m128i A9 = _mm_srli_epi16(A8, 8);
-      const __m128i A10 = _mm_packus_epi16(A9, zero);
+      const __m128i A7 = _mm_mulhi_epu16(A6, kMult);
+      const __m128i A10 = _mm_packus_epi16(A7, zero);
       _mm_storel_epi64((__m128i*)&ptr[x], A10);
     }
   }
@@ -241,23 +238,19 @@ static void MultRow_SSE2(uint8_t* const ptr, const uint8_t* const alpha,
                          int width, int inverse) {
   int x = 0;
   if (!inverse) {
-    const int kSpan = 8;
     const __m128i zero = _mm_setzero_si128();
-    const __m128i kRound = _mm_set1_epi16(1 << 7);
-    const int w2 = width & ~(kSpan - 1);
-    for (x = 0; x < w2; x += kSpan) {
+    const __m128i k128 = _mm_set1_epi16(128);
+    const __m128i kMult = _mm_set1_epi16(0x0101);
+    for (x = 0; x + 8 <= width; x += 8) {
       const __m128i v0 = _mm_loadl_epi64((__m128i*)&ptr[x]);
+      const __m128i a0 = _mm_loadl_epi64((const __m128i*)&alpha[x]);
       const __m128i v1 = _mm_unpacklo_epi8(v0, zero);
-      const __m128i alpha0 = _mm_loadl_epi64((const __m128i*)&alpha[x]);
-      const __m128i alpha1 = _mm_unpacklo_epi8(alpha0, zero);
-      const __m128i alpha2 = _mm_unpacklo_epi8(alpha0, alpha0);
-      const __m128i v2 = _mm_mulhi_epu16(v1, alpha2);
-      const __m128i v3 = _mm_mullo_epi16(v1, alpha1);
-      const __m128i v4 = _mm_adds_epu16(v2, v3);
-      const __m128i v5 = _mm_adds_epu16(v4, kRound);
-      const __m128i v6 = _mm_srli_epi16(v5, 8);
-      const __m128i v7 = _mm_packus_epi16(v6, zero);
-      _mm_storel_epi64((__m128i*)&ptr[x], v7);
+      const __m128i a1 = _mm_unpacklo_epi8(a0, zero);
+      const __m128i v2 = _mm_mullo_epi16(v1, a1);
+      const __m128i v3 = _mm_add_epi16(v2, k128);
+      const __m128i v4 = _mm_mulhi_epu16(v3, kMult);
+      const __m128i v5 = _mm_packus_epi16(v4, zero);
+      _mm_storel_epi64((__m128i*)&ptr[x], v5);
     }
   }
   width -= x;