diff --git a/src/dec/vp8l.c b/src/dec/vp8l.c
index 01c3274f..de3711a0 100644
--- a/src/dec/vp8l.c
+++ b/src/dec/vp8l.c
@@ -1492,9 +1492,8 @@ static void ExtractAlphaRows(VP8LDecoder* const dec, int last_row) {
     const int cache_pixs = width * num_rows_to_process;
     uint8_t* const dst = output + width * cur_row;
     const uint32_t* const src = dec->argb_cache_;
-    int i;
     ApplyInverseTransforms(dec, num_rows_to_process, in);
-    for (i = 0; i < cache_pixs; ++i) dst[i] = (src[i] >> 8) & 0xff;
+    WebPExtractGreen(src, dst, cache_pixs);
     AlphaApplyFilter(alph_dec,
                      cur_row, cur_row + num_rows_to_process, dst, width);
     num_rows -= num_rows_to_process;
@@ -1562,6 +1561,8 @@ int VP8LDecodeAlphaImageStream(ALPHDecoder* const alph_dec, int last_row) {
     return 1;  // done
   }
 
+  if (!alph_dec->use_8b_decode_) WebPInitAlphaProcessing();
+
   // Decode (with special row processing).
   return alph_dec->use_8b_decode_ ?
       DecodeAlphaData(dec, (uint8_t*)dec->pixels_, dec->width_, dec->height_,
diff --git a/src/dsp/alpha_processing.c b/src/dsp/alpha_processing.c
index fd564d5d..4b60e092 100644
--- a/src/dsp/alpha_processing.c
+++ b/src/dsp/alpha_processing.c
@@ -284,9 +284,9 @@ static void ApplyAlphaMultiply_16b(uint8_t* rgba4444,
 #endif
 }
 
-static int DispatchAlpha(const uint8_t* alpha, int alpha_stride,
-                         int width, int height,
-                         uint8_t* dst, int dst_stride) {
+static int DispatchAlpha_C(const uint8_t* alpha, int alpha_stride,
+                           int width, int height,
+                           uint8_t* dst, int dst_stride) {
   uint32_t alpha_mask = 0xff;
   int i, j;
 
@@ -303,9 +303,9 @@ static int DispatchAlpha(const uint8_t* alpha, int alpha_stride,
   return (alpha_mask != 0xff);
 }
 
-static void DispatchAlphaToGreen(const uint8_t* alpha, int alpha_stride,
-                                 int width, int height,
-                                 uint32_t* dst, int dst_stride) {
+static void DispatchAlphaToGreen_C(const uint8_t* alpha, int alpha_stride,
+                                   int width, int height,
+                                   uint32_t* dst, int dst_stride) {
   int i, j;
   for (j = 0; j < height; ++j) {
     for (i = 0; i < width; ++i) {
@@ -316,9 +316,9 @@ static void DispatchAlphaToGreen(const uint8_t* alpha, int alpha_stride,
   }
 }
 
-static int ExtractAlpha(const uint8_t* argb, int argb_stride,
-                        int width, int height,
-                        uint8_t* alpha, int alpha_stride) {
+static int ExtractAlpha_C(const uint8_t* argb, int argb_stride,
+                          int width, int height,
+                          uint8_t* alpha, int alpha_stride) {
   uint8_t alpha_mask = 0xff;
   int i, j;
 
@@ -334,11 +334,17 @@ static int ExtractAlpha(const uint8_t* argb, int argb_stride,
   return (alpha_mask == 0xff);
 }
 
+static void ExtractGreen_C(const uint32_t* argb, uint8_t* alpha, int size) {
+  int i;
+  for (i = 0; i < size; ++i) alpha[i] = argb[i] >> 8;
+}
+
 void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int);
 void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int);
 int (*WebPDispatchAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
 void (*WebPDispatchAlphaToGreen)(const uint8_t*, int, int, int, uint32_t*, int);
 int (*WebPExtractAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
+void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size);
 
 //------------------------------------------------------------------------------
 // Init function
@@ -358,9 +364,11 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessing(void) {
   WebPMultRow = WebPMultRowC;
   WebPApplyAlphaMultiply = ApplyAlphaMultiply;
   WebPApplyAlphaMultiply4444 = ApplyAlphaMultiply_16b;
-  WebPDispatchAlpha = DispatchAlpha;
-  WebPDispatchAlphaToGreen = DispatchAlphaToGreen;
-  WebPExtractAlpha = ExtractAlpha;
+
+  WebPDispatchAlpha = DispatchAlpha_C;
+  WebPDispatchAlphaToGreen = DispatchAlphaToGreen_C;
+  WebPExtractAlpha = ExtractAlpha_C;
+  WebPExtractGreen = ExtractGreen_C;
 
   // If defined, use CPUInfo() to overwrite some pointers with faster versions.
   if (VP8GetCPUInfo != NULL) {
diff --git a/src/dsp/alpha_processing_neon.c b/src/dsp/alpha_processing_neon.c
index ed845729..606a401c 100644
--- a/src/dsp/alpha_processing_neon.c
+++ b/src/dsp/alpha_processing_neon.c
@@ -161,6 +161,17 @@ static int ExtractAlpha_NEON(const uint8_t* argb, int argb_stride,
   return (alpha_mask == 0xffffffffu);
 }
 
+static void ExtractGreen_NEON(const uint32_t* argb,
+                              uint8_t* alpha, int size) {
+  int i;
+  for (i = 0; i + 16 <= size; i += 16) {
+    const uint8x16x4_t rgbX = vld4q_u8((const uint8_t*)(argb + i));
+    const uint8x16_t greens = rgbX.val[1];
+    vst1q_u8(alpha + i, greens);
+  }
+  for (; i < size; ++i) alpha[i] = (argb[i] >> 8) & 0xff;
+}
+
 //------------------------------------------------------------------------------
 
 extern void WebPInitAlphaProcessingNEON(void);
@@ -170,6 +181,7 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingNEON(void) {
   WebPDispatchAlpha = DispatchAlpha_NEON;
   WebPDispatchAlphaToGreen = DispatchAlphaToGreen_NEON;
   WebPExtractAlpha = ExtractAlpha_NEON;
+  WebPExtractGreen = ExtractGreen_NEON;
 }
 
 #else  // !WEBP_USE_NEON
diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index d6d50e77..7ce58856 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -509,6 +509,10 @@ extern int (*WebPExtractAlpha)(const uint8_t* argb, int argb_stride,
                                int width, int height,
                                uint8_t* alpha, int alpha_stride);
 
+// Extract the green values from 32b values in argb[] and pack them into alpha[]
+// (this is the opposite of WebPDispatchAlphaToGreen).
+extern void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size);
+
 // Pre-Multiply operation transforms x into x * A / 255  (where x=Y,R,G or B).
 // Un-Multiply operation transforms x into x * 255 / A.