fix 16b overflow in SSE2

the 'accum' variable can be larger than 15b for large
rescale values.

Assert triggered:
 src/dsp/rescaler_sse2.c:249: RescalerExportRowExpand_SSE2: Assertion `v >= 0 && v <= 255' failed.
 src/dsp/rescaler_sse2.c:350: RescalerExportRowShrink_SSE2: Assertion `v >= 0 && v <= 255' failed.

-> fall back to C implementation in this case for now

Change-Id: I7ea1cb72301cafc1459be403f6a6f4e3cbc89bb1

CMakeLists.txt

@@ -10,7 +10,7 @@ option(WEBP_BUILD_GIF2WEBP "Build the gif2webp conversion tool." OFF)
 option(WEBP_BUILD_IMG2WEBP "Build the img2webp animation tool." OFF)
 option(WEBP_BUILD_WEBPINFO "Build the webpinfo command line tool." OFF)
 option(WEBP_BUILD_WEBP_JS "Emscripten build of webp.js." OFF)
-option(WEBP_ENABLE_NEAR_LOSSLESS "Enable near-lossless encoding" ON)
+option(WEBP_NEAR_LOSSLESS "Enable near-lossless encoding" ON)
 option(WEBP_ENABLE_SWAP_16BIT_CSP "Enable byte swap for 16 bit colorspaces." OFF)
 
 if(WEBP_BUILD_WEBP_JS)

examples/cwebp.c

@@ -188,7 +188,7 @@ static void PrintExtraInfoLossless(const WebPPicture* const pic,
     fprintf(stderr, "File:      %s\n", file_name);
     fprintf(stderr, "Dimension: %d x %d\n", pic->width, pic->height);
     fprintf(stderr, "Output:    %d bytes (%.2f bpp)\n", stats->coded_size,
-            1.f * stats->coded_size / pic->width / pic->height);
+            8.f * stats->coded_size / pic->width / pic->height);
     PrintFullLosslessInfo(stats, "ARGB");
   }
 }
@@ -213,7 +213,7 @@ static void PrintExtraInfoLossy(const WebPPicture* const pic, int short_output,
             "           (%.2f bpp)\n",
             stats->coded_size,
             stats->PSNR[0], stats->PSNR[1], stats->PSNR[2], stats->PSNR[3],
-            1.f * stats->coded_size / pic->width / pic->height);
+            8.f * stats->coded_size / pic->width / pic->height);
     if (total > 0) {
       int totals[4] = { 0, 0, 0, 0 };
       fprintf(stderr, "block count:  intra4:     %6d  (%.2f%%)\n"

extras/extras.c

@@ -48,13 +48,14 @@ int WebPImportGray(const uint8_t* gray_data, WebPPicture* pic) {
 
 int WebPImportRGB565(const uint8_t* rgb565, WebPPicture* pic) {
   int x, y;
+  uint32_t* dst;
   if (pic == NULL || rgb565 == NULL) return 0;
   pic->colorspace = WEBP_YUV420;
   pic->use_argb = 1;
   if (!WebPPictureAlloc(pic)) return 0;
+  dst = pic->argb;
   for (y = 0; y < pic->height; ++y) {
     const int width = pic->width;
-    uint32_t* dst = pic->argb + y * pic->argb_stride;
     for (x = 0; x < width; ++x) {
 #ifdef WEBP_SWAP_16BIT_CSP
       const uint32_t rg = rgb565[2 * x + 1];
@@ -70,22 +71,24 @@ int WebPImportRGB565(const uint8_t* rgb565, WebPPicture* pic) {
       r = r | (r >> 5);
       g = g | (g >> 6);
       b = b | (b >> 5);
-      dst[x] = (r << 16) | (g << 8) | b;
+      dst[x] = (0xffu << 24) | (r << 16) | (g << 8) | b;
     }
     rgb565 += 2 * width;
+    dst += pic->argb_stride;
   }
   return 1;
 }
 
 int WebPImportRGB4444(const uint8_t* rgb4444, WebPPicture* pic) {
   int x, y;
+  uint32_t* dst;
   if (pic == NULL || rgb4444 == NULL) return 0;
   pic->colorspace = WEBP_YUV420;
   pic->use_argb = 1;
   if (!WebPPictureAlloc(pic)) return 0;
+  dst = pic->argb;
   for (y = 0; y < pic->height; ++y) {
     const int width = pic->width;
-    uint32_t* dst = pic->argb + y * pic->argb_stride;
     for (x = 0; x < width; ++x) {
 #ifdef WEBP_SWAP_16BIT_CSP
       const uint32_t rg = rgb4444[2 * x + 1];
@@ -106,6 +109,7 @@ int WebPImportRGB4444(const uint8_t* rgb4444, WebPPicture* pic) {
       dst[x] = (a << 24) | (r << 16) | (g << 8) | b;
     }
     rgb4444 += 2 * width;
+    dst += pic->argb_stride;
   }
   return 1;
 }

extras/get_disto.c

@@ -293,7 +293,7 @@ int main(int argc, const char *argv[]) {
   printf("%u %.2f    %.2f %.2f %.2f %.2f [ %.2f bpp ]\n",
          (unsigned int)size1,
          disto[4], disto[0], disto[1], disto[2], disto[3],
-         1.f * size1 / pic1.width / pic1.height);
+         8.f * size1 / pic1.width / pic1.height);
 
   if (output != NULL) {
     uint8_t* data = NULL;

makefile.unix

@@ -112,7 +112,12 @@ endif
 AR = ar
 ARFLAGS = r
 CPPFLAGS = -I. -Isrc/ -Wall
-CFLAGS = -O3 -DNDEBUG $(EXTRA_FLAGS)
+ifeq ($(DEBUG), 1)
+  CFLAGS = -g
+else
+  CFLAGS = -O3 -DNDEBUG
+endif
+CFLAGS += $(EXTRA_FLAGS)
 CC = gcc
 INSTALL = install
 GROFF = /usr/bin/groff

src/dsp/rescaler_sse2.c

@@ -36,7 +36,7 @@ static void LoadTwoPixels_SSE2(const uint8_t* const src, __m128i* out) {
 }
 
 // input: 8 bytes ABCDEFGH -> output: A0B0C0D0E0F0G0H0
-static void LoadHeightPixels_SSE2(const uint8_t* const src, __m128i* out) {
+static void LoadEightPixels_SSE2(const uint8_t* const src, __m128i* out) {
   const __m128i zero = _mm_setzero_si128();
   const __m128i A = _mm_loadl_epi64((const __m128i*)(src));  // ABCDEFGH
   *out = _mm_unpacklo_epi8(A, zero);
@@ -50,13 +50,15 @@ static void RescalerImportRowExpand_SSE2(WebPRescaler* const wrk,
   int accum = x_add;
   __m128i cur_pixels;
 
+  // SSE2 implementation only works with 16b signed arithmetic at max.
+  if (wrk->src_width < 8 || accum >= (1 << 15)) {
+    WebPRescalerImportRowExpand_C(wrk, src);
+    return;
+  }
+
   assert(!WebPRescalerInputDone(wrk));
   assert(wrk->x_expand);
   if (wrk->num_channels == 4) {
-    if (wrk->src_width < 2) {
-      WebPRescalerImportRowExpand_C(wrk, src);
-      return;
-    }
     LoadTwoPixels_SSE2(src, &cur_pixels);
     src += 4;
     while (1) {
@@ -75,11 +77,7 @@ static void RescalerImportRowExpand_SSE2(WebPRescaler* const wrk,
   } else {
     int left;
     const uint8_t* const src_limit = src + wrk->src_width - 8;
-    if (wrk->src_width < 8) {
-      WebPRescalerImportRowExpand_C(wrk, src);
-      return;
-    }
-    LoadHeightPixels_SSE2(src, &cur_pixels);
+    LoadEightPixels_SSE2(src, &cur_pixels);
     src += 7;
     left = 7;
     while (1) {
@@ -94,7 +92,7 @@ static void RescalerImportRowExpand_SSE2(WebPRescaler* const wrk,
         if (--left) {
           cur_pixels = _mm_srli_si128(cur_pixels, 2);
         } else if (src <= src_limit) {
-          LoadHeightPixels_SSE2(src, &cur_pixels);
+          LoadEightPixels_SSE2(src, &cur_pixels);
           src += 7;
           left = 7;
         } else {   // tail